key: cord-0035452-mnbhbxaq
authors: Belciug, Smaranda; Gorunescu, Florin
title: Era of Intelligent Systems in Healthcare
date: 2019-03-21
journal: Intelligent Decision Support Systems—A Journey to Smarter Healthcare
DOI: 10.1007/978-3-030-14354-1_1
sha: 5d573ffa6e09b83d1e8fafa4d349747c246118fd
doc_id: 35452
cord_uid: mnbhbxaq

The aim of this chapter is to prepare the reader for the outstanding trip that she/he embarked when starting reading this book. At first, we shall try to look for answers to some of the most important questions regarding the connection between intelligent systems and healthcare. What are intelligent systems? How can they be used in healthcare? Have they got benefits and prospects? Let us highlight some of the decisive factors for a successful deployment of intelligent systems in healthcare, including intelligent clinical support and intelligent patient management.

To have a more comprehensive understanding of this concept, let us to go back to the year 1950, and see what A.M. Turing said about the 'thinking machines'. In Turing [1] , Turing had to answer the following question "Can machines think?" Analyzing different aspects related to this 'simple' question, he concluded that the problem should be reformulated as follows: "The new form of the problem can be described in terms of a game which we call the 'imitation game'. It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either "X is A and Y is B" or "X is B and Y is A". The interrogator is allowed to put questions to A and B", with the answers, ideally, typewritten; the question is related to sex "C: Will X please tell me the length of his or her hair?" In this context, Turing brought up the issue: "We now ask the question, 'What will happen when a machine takes the part of A in this game?' Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, 'Can machines think?'. Summarizing the above, if we assume that we have a person, a machine, and an interrogator, then the object of the 'imitation game' for the interrogator is to determine which of the other two is the person, and which is the machine. However, the story does not end here, because "The question we put … will not be quite definite until we have specified what we mean by the word 'machine.'" Turing concludes that "We are the more ready to do so in view of the fact that the present interest in 'thinking machines' has been aroused by a particular kind of machine, usually called an 'electronic computer' or 'digital computer'. Following this suggestion we only permit digital computers to take part in our game." Starting from this presentation, we come to the phrase "The Turing test", as a way of answering the question whether machines can think or not. An interesting fact, according to the Stanford Encyclopedia of Philosophy [2] , is the idea that the Turing's test has deeper roots in history. Thus, it is suggested that the Turing test is foreshadowed in R. Descartes' "Discourse on the Method of Rightly Conducting one's Reason and Seeking Truth in the Sciences", or, shortly "Discourse on the Method (1637)". Finally, let us mention that there are currently opinions that say the classical Turing test is out of date and that we need to find other ways to measure the machines' "intelligence" [3] [4] [5] [6] . In this context, different approaches have emerged to assess the Turing test capacity to detect the presence of consciousness. We can mention the Chinese room argument [Internet Encyclopedia of Philosophy], which is an experiment of John Searle [7] designed to show that the Turing test is insufficient to detect the presence of consciousness, even if the 'room' can behave or function as a conscious mind would ( Fig. 1.1) .

One might think that ISs are one step away from the Turing test. If we consider that a computer is deemed to have Artificial Intelligence (AI) capabilities, if it can mimic human responses under specific conditions, everything is clear, but (AI) is more than a question of imitation, it is a matter of understanding. There are many ways to describe the concept of AI. According to Encyclopedia Britannica [Encyclopaedia Britannica], AI means "the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings", having thus the capabilities of "developing systems endowed with the intellectual processes characteristic of humans, such as the ability to reason, discover meaning, generalize, or learn from past experience." It is noteworthy that Turing has provided the earliest significant work regarding AI, the first manifesto, all modern computers being considered as "Turing machines". In an unpublished 1948 report, entitled "Intelligent Machinery", he introduced many of the central concepts of AI. More details about A. Turing and his substantial contribution to AI, including the report mentioned above, are to be found in Copeland [8] .

After this brief history about the beginnings of AI, let us return to our main story regarding ISs. Consistent with the definition of ISs presented at the start, let us also point out that ISs can be found in different forms: from AI models processing huge datasets to AI models controlling robots. The ISs field represents an interdisciplinary research domain bringing together ideas from AI, machine learning (ML), and a range of fields such as psychology, linguistics and brain sciences, connected by many interdisciplinary relationships ( Fig. 1.2) . Nowadays, a wide variety of ISs have been developed, such as:

• expert systems, • fuzzy systems, • artificial neural networks, • evolutionary computation (genetic/evolutionary algorithms, genetic programming, evolutionary strategies), • support vector machines, • particle swarm optimization, • ant colony systems, • memetic algorithms, • ant colony optimization, • clustering, • Bayesian (learning) model, • deep learning, • hybrid models. (neuro-genetic, neuro-fuzzy, fuzzy-genetic, etc.)

The general areas of applications of modern ISs include the following topics:

• Artificial Intelligence More details related to ISs in Padhy [9] , Shin and Xu [10] , Hopgood [11] , Grosan and Abraham [12] , Wilamowski and David Irwin [13] , Pap [14] , Kryszkiewicz et al. [15] , Martínez de Pisón et al. [16] , Bi et al. [17] .

With regard to the practical applications of ISs, there is a huge collection of real situations where they show their potential to effectively solve the problems that arise. In what follows, we will briefly address some well-known real-world applications of ISs, so that the reader can get an idea of the potential of these "smart tools" and their application area.

• Fuzzy logic control [18] [19] [20] [21] [22] ; • Business Intelligence/Management Intelligent Systems [23] [24] [25] [26] ;

• Intelligent Bioinformatics Systems [27] [28] [29] [30] [31] ;

• Intelligent Healthcare Systems [32] [33] [34] [35] [36] ;

• Intelligent Game [37] [38] [39] [40] [41] ;

• Intelligent Multimedia [42] [43] [44] [45] [46] ; • User Interfaces and Human Computer Interaction [47] [48] [49] [50] ;

• Knowledge-based Software Engineering/ Knowledge Engineering/ Management [51] [52] [53] [54] [55] ;

• Speech recognition [56] [57] [58] [59] [60] ; • Brain-Machine Interface Systems [61] [62] [63] [64] [65] ;

• Intelligent Robotic Systems [66] [67] [68] [69] [70] ;

• Intelligent Transportation Systems [71] [72] [73] ;

• Medical Imaging [74] [75] [76] [77] [78] ;

• Psychology [79] [80] [81] [82] 160] ;

• Military applications [83] [84] [85] [86] 161] ;

• Engineering problems [87] [88] [89] [90] 162] ;

• Smart cities [91] [92] [93] [94] [95] ;

• Internet of Things (IoT) [96] [97] [98] [99] [100] .

Besides presenting some basic theoretical and practical concepts of the most known ISs applications, we believe that is far more interesting for the reader to see their impact on everyday life. Basically, we will talk now about the well-known intelligent devices. An intelligent/smart device is any type of equipment, instrument, or machine that has its own computing capability. Moreover, these devices are, generally, connected to other devices via wireless networks. Most of them are now parts of the so-called "smart home". In this context, the IoT describes how the Internet will link intelligent devices, to allow these endpoints to generate and share data. Examples of intelligent devices are smartphones, phablets, tablets, smart watches, wearable computers, smart home devices (e.g., smart thermostat, smart lighting bulbs, smart security cameras, smart switch, smart speakers, smart AC control, etc.), intelligent vehicle technologies, etc. (Figs. 1.3, 1.4, 1.5) .

Besides these everyday in use smart devices, let us not forget to mention the famous virtual personal assistants: Siri, Google Assistant, and Cortana-the intelligent digital personal assistants on various platforms (e.g., iOS, Android, and Windows Mobile)that help us find useful information when vocally asked ( Fig. 1.6 ).

It is interesting to go through a brief history of AI/IS to see the impressive rhythm of development of this important field of Computer Science in just 60 years from its beginning [101] :

• One considers that AI started at "Dartmouth Artificial Intelligence" conference, held at Dartmouth College, Hanover, New Hampshire, USA, in 1956 (https:// www.livinginternet.com/i/ii_ai.htm-accessed November 24, 2017) . In this regard, let us mention the "Dartmouth AI Project Proposal", by J. McCarthy (Dartmouth College), M. L. Minsky (Harvard University), N. Rochester (IBM Corporation), and C.E. Shannon (Bell Telephone Laboratories)- August 31, 1955 , states that "We propose that a 2 month, 10 man study of artificial (http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html accessed November 24, 2017). • A milestone for the use of AI in real-world applications was the idea of creating a 'perfect machine translator', funded by the Defense Advanced Research Projects Agency (DARPA). Unfortunately, two reports (Automatic Language Processing Advisory Committee (ALPAC) report-1966, and Lighthill report-1973) "concluded negatively about the possibility of creating a machine that could learn or be considered intelligent". • Another landmark in AI history was the occurrence of the "expert systems" in the eighties, giving a 'breather' to the field by UK and Japan funding. It is noteworthy that expert systems first emerged in the early 1950s when the Rand-Carnegie team of Newell, Shaw, and Simon developed a General Problem Solver that dealt with theorems proof, geometric problems and chess playing. • Another revival of the AI field began in 1993 with the MIT "Cog" project, aiming to build a humanoid robot, and with the Dynamic Analysis and Replanning Tool (DART), aiming to optimize and schedule transportation of supplies or personnel and solve other logistical problems, used by U.S. military. • We now reach February 10, 1996, when "Deep Blue", the chess-playing super computer developed by IBM, won its first game against a world champion, Garry Kasparov. Overall, Kasparov defeated Deep Blue by a score of 4-2. However, in 1997, after a heavily upgrade, Deep Blue defeated Kasparov in a rematch, the first defeat of a reigning world chess champion by a computer under tournament conditions (score 3.5/2.5). It is noteworthy that C.E. Shannon was the first to think about developing a chess-playing program. Then, until 2012, the AI development has materialized mainly through academic  research, with no significant practical impact. The milestone was December 4,  2012, when a convolutional neural network was presented at the Neural  Information Processing Systems-NIPS 2012 Conference (Advances in Neural  Information Processing Systems 25) . Since then, the AI development trend has been steadily increasing.

The field of AI went through phases of rapid progress and failures, followed by a cooling in investment and interest often referred to as "AI winters" (inspired by the "nuclear winter"). The first frost occurred in the 1970s, as progress slowed and government funding dried up, and another one in the 1980s because of the failure of the expected commercial impact.

It is interesting to mention an important fact regarding the place and the moment of the occurrence of AI, beyond the official chronology. Thus, according to H. Bruderer (https://cacm.acm.org/blogs/blog-cacm/222486-the-birthplace-ofartificial-intelligence/fulltext-accessed November 24, 2017), the birthplace of AI would have been Paris, on the occasion of the Colloques internationaux du Centre national de la recherche scientifique, Paris, Janvier [8] [9] [10] [11] [12] [13] 1951 AI is a research field so vast that we cannot keep track of everything. Among the leading global publishers of AI journals (books) that serve and support the research community, we can quote Springer, Elsevier, Wiley, IEEE, ACM, etc. Because of the substantial significance related to the AI field, both from theoretical and Since in the past and, mainly, in the recent years, we have been experiencing an exponential growth of the AI/IS development, every year many prestigious international conferences, covering the most diverse aspects of the field, take place in the most diverse locations worldwide. As examples of such recent events, we mention the following ones: The impetuous development that AI has known, led to the emergence of an intensely debated issue today by well-known scientists as well as ordinary people. The question is: What scares us when we see the unprecedented development of AI? Most people worry about the occurrence of the "singularity point", that is the milestone in history when AI will surpass human intelligence and its consequences. Below we list the opinions of some famous people in this area. To conclude, the AI's existential risk resides in the possibility that substantial progress in AI could someday result in human extinction, or some other unrecoverable global catastrophe (see also Wikipedia: "Existential risk from artificial general intelligence") (https://en.wikipedia.org/wiki/Existential_risk_from_ artificial_general_intelligence#Reactions-accessed November 24, 2017)

Across all the research fields, from engineering, meteorology, and business to healthcare, sociology, and multimedia and so forth, data are being collected and accumulated at a very fast pace. Under these circumstances, there is an urgent need for developing advanced ISs to assist humans in extracting useful information/ knowledge from the huge volumes of digital data in order to make real-time accurate decisions. As it is well known, healthcare deals with thorough processes of the diagnosis, treatment and prevention of disease, injury, physical, and mental impairments. At the same time, it also covers the hospital/patient management. As it rapidly evolved in most countries, the healthcare industry has become the generator of a massive amount of data, including electronic medical records, administrative reports, and other useful knowledge. Figures 1.7 and 1.8 provide a synthetic picture of the way ISs assist and, especially help optimizing the healthcare process, both in terms of computerized/automated medical diagnosis and intelligent patient management.

The healthcare domain includes many industries and companies that are involved in products and services related to health. Among the most important branches of healthcare, one can mention:

• Pharmaceutical industry, which is the part of the healthcare sector that deals with medicines. It comprises different fields pertaining to the discovery, development, production, and marketing of pharmaceutical drugs. The pharmaceutical industry worldwide revenue was grosso modo estimated over one trillion U.S. dollars in 2014 (Berkrot, B., Reuters, #Health News, April 20, 2010). After this short overview of the current 'landscape' in this field, we now focus on the "revolution" brought by AI in healthcare. Let us first take a brief look on the history of using AI tools in medicine.

Let us recall the DENDRAL project (coming from "DENDRitic ALgorithm"), also known as the "grandfather of experts systems", started at Stanford University, California, in 1965 by Edward Feigenbaum, Bruce G. Buchanan, Joshua Lederberg, and Carl Djerassi, along with a team of research associates and students [104] , that was initially intended as a chemical-analysis expert system. MYCIN, developed by Edward Shortliffe (1972) as a doctoral dissertation for Stanford Medical School [105] , was a clinical expert system, derived from DENDRAL, and designed for selection of antibiotics for patients with serious blood infections, hence the name "mycin"-suffix of many antibiotics. Mycin consisted of three components (sub-systems): (a) consultation system, (b) explanation system, and (c) rule acquisition system. It is noteworthy that MYCIN was never publicly used in clinical practice.

Other To see how AI can be an important player in the healthcare domain, we need to get a first look inside the physician's activity.

Let us first briefly analyze the interaction patient-physician. This complex interaction includes empathy, information management, application of expertise in a given context, negotiation with multiple stakeholders, and unpredictable physical response in special situations, such as surgery or post-op, some actions taking place in real time (for instance, patient on the operating table or patient in the ICU (Intensive Care Unit)). In this context, we are wondering: "Are these important real-life aspects viable AI-applications or not?" Obviously, these are not AI-applicable functions. Thus, according to a study regarding the way a physician spends his/her time at work (U.S. ambulatory care in four specialties in four states) [106] , it resulted that physicians spent 27.0% of their total time on direct clinical face time with patients and 49.2% of their time on EHR (Electronic Health Record), and desk work during the office day. Under these circumstances, physicians have less time to directly practice medicine, do necessary research, master new medical/ AI technology, and improve their skills in order to become better doctors. To conclude, apart from the measures that need to be taken to increase the doctor-patient interaction time against the clerical work, it will be important for physicians and patients to understand and engage the evolution of automation in medicine in order to optimize the patient care. Physicians must be open to the rapidly advancing AI technology, and, more importantly, they should embrace this opportunity rather than fear it. In this sense, there are many concerns of the medical staff on this subject, due to recent news in the media. Thus, in the near future, robots and AI technologies in healthcare could lead to a doctorless hospital [107] . Although hospitals have been slow in adopting AI technologies and robotics, these new approaches took their place gradually, fundamentally changing the landscape. Thus, doctors in near future will have to have new skills vs. today's doctors (e.g., computer skills, robotics), to successfully compete with the new technologies. Let us hope that we will not turn the current hospitals into robot-factory hospitals. It is noteworthy to mention in this sense, that financial pressures will inevitably force the recognition of the fact that the medical robots, powered by the AI technology, will be the only way to help doctors provide a significantly higher medical service than the current one. If driverless cars are going to reduce traffic accidents and congestion, but under human supervision, then we may hope that the intelligent systems grafted in the hospitals of the future will one day save more lives and reduce the cost of healthcare, without removing doctors from the medical activities and decisions ( Fig. 1.9 ).

To conclude this brief glance over a possible scenario of future hospitals' functioning, let us mention the opinion of Naveen Jain, an entrepreneur and philanthropist driven to solve the world's biggest challenges through innovation, founder/CEO of some successful companies, such as Moon Express, BlueDot, Intelius, Talent Wise, Viome, InfoSpace. In a CNCBC TV interview (published by Kharpal, A, Fri, 1 Dec 2017, 1:13 AM ET) from the Slush technology conference in Helsinki, Finland, 2017 (http://www.slush.org/), where he spoke about the contribution of AI in healthcare, comparing it with a "tsunami", because of its considerable role in processing huge amount of data, impossible to be done by humans. The latest AI implications and achievements in healthcare come to justify this visionary opinion. Let us recall, in this regard, only that a Chinese AI-powered robot (Xiaoyi, meaning "little doctor") became the world's first machine to pass successfully a medical exam on Nov. 6, 2017 (posted by Mara P. on November 24, 2017, © 2016 TechThe Lead-http://techthelead.com/robot-chinas-medical-exam/).

We have seen from the above facts how important AI's involvement in healthcare can become. There are current discussions about the "Fourth Industrial Revolution" [108] , involving new cutting-edge technologies (AI technologies, intelligent robots, driverless cars, nanotechnologies, quantum computing, IoT, etc.), which will affect various domains such as healthcare, economy, and industry, etc. The basic idea is that enormous databases have been collected in recent years, and for the year 2020 being foreseen, for instance, an amount of data equaling about 44 zettabytes (44 trillion gigabytes), according to an IDC Digital Universe study (IDC Analyze the Future), Digital Universe "Big Data, Bigger Digital Shadows, and Biggest Growth in the Far East" (sponsored by EMC 2 , DellEMC), © 2017 KDnuggets (https://www.kdnuggets.com/2012/12/idc-digital-universe-2020.html)

As mentioned above, we dispose nowadays of a huge amount of data, from the medical field in our case, which have to be processed somehow to discover useful, often hidden, information in order to make an optimal decision in real time. Let us recall, in this respect, the well-known UCI Machine Learning Repository, a collection of databases used by the ML community for the empirical analysis of ML algorithms (https://archive.ics.uci.edu/ml/datasets.html). We also recall the Kaggle datasets (https://www.kaggle.com/datasets), maintaining important datasets as a service to the ML community, the platform for researchers to share their data Academic Torrents (http://academictorrents.com/) or the dataset repository from GitHub (https://github.com/awesomedata/awesome-public-datasets/blob/master/ README.rst-https://github.com/awesomedata/awesome-public-datasets/blob/ master/README.rst). Therefore, there are many public or private databases that can be used for different research purposes, e.g., diagnoses, medical procedures, medication, demographics, cost/charge of healthcare services, etc. There are also the efficacious tools provided by AI to process this data. We are, metaphorically speaking, in the situation of Ali Baba in the cave of the forty thieves, searching for valuable information helping to raise the healthcare quality.

In order to better understand the "treasure" of information/knowledge contained in these healthcare databases, let us restrict ourselves only to those concerning medical diagnosis. The huge size datasets containing both the symptoms and the correct diagnosis of a particular illness represent the collection of data from hundreds or thousands of physicians, data gathered over many years. When making a particular diagnosis, the physician processes only the data recorded in his/her memory, data gathered throughout his/her whole career up to that point. The effectiveness and speed of diagnosis-making differ significantly from a resident with a few years of practice under the supervision of an attending physician to an experienced doctor with many years of medical practice, in which he/she has seen numerous cases. How can AI support the decision-making process in this context? Very simple: by using the powerful data analysis and decision-making mechanisms (the classification technology, in short) applied to the "experience" of hundreds or thousands of doctors, through the corresponding databases, in a blink of an eye. Metaphorically speaking, this computerized data processing is the synergy of information provided by all doctors who have collected data over time and not just a second medical opinion.

Finally, yet importantly, let us talk about the well-known "curse of dimensionality", Bellman [109] Nowadays, we dispose of very sophisticated and advanced medical equipment, making thus possible the collection of a huge amount of features, more or less important in the medical diagnosis process or patient management. We may have to process datasets with a relatively small number of attributes (e.g., the thyroid disease dataset with 21 attributes from UCI Machine Learning Repository-http://archive.ics.uci.edu/ml/datasets/thyroid+disease). On the other hand, we have to deal with high dimension datasets containing medical data obtained through DNA microarray analysis (e.g., 24,481 attributes, as in the case of the breast cancer Kent Ridge dataset from Machine Learning Data Set Repository (http://mldata.org/). Thus, we have to tackle datasets with a huge number of variables (attributes), therefore becoming "suffocated" by the richness of information. Who gets to decide objectively and effectively which features are important and which are redundant, even in the case of 21 attributes? The answer is very simple: not the human, but the "machine" using its "artificial intelligence" to perform the so-called feature selection and feature extraction techniques.

We have tried above to outline the role of ISs in the medical field, thus answering the question "Why ISs in healthcare?" We will now present some of the most important ways in which ISs are involved in this domain, answering the question "How ISs in healthcare?" Thus, according to the IBM response to the White House Office of Science and Technology Policy's Request for information (Summer 2016), IBM [110], systems can advance precision medicine by ingesting patients' electronic medical history and relevant medical literature, performing cohort analysis, identifying micro-segments of similar patients, evaluating standard-of-care practices and available treatment options, ranking by relevance, risk and preference, and ultimately recommending the most effective treatments for their patients."

When we think about how ISs are involved in the healthcare decision-making, two approaches that work in tandem come to mind. Firstly, the natural language processing technology deals with the information extraction from unstructured data (e.g., clinical notes, medical journals, etc.) in order to improve the machine-readable structured medical data. Secondly, using both structured data contained in medical datasets, or unstructured data 'translated' by NLP, cutting-edge ML/DM algorithms attempt to extract valuable information/knowledge in order to support efficiently the human decision-making. Let us say a few words about each approach.

• The Natural Language Processing (NLP) Technology. Starting with the Turing's test to assess the machines' intelligence", the interaction between human and computers through the natural languages has remained a very important and topical issue even today. In the context of healthcare supported by ISs/AI, we are interested in how to apply NLP to unstructured data. By unstructured data, or unstructured information, we broadly understand information that either does not have a pre-defined data model and/or does not fit well into relational tables. We can mention as unstructured data: health records, documents, images, audio/ video files, analog data, sentiments about a given topic, etc. In the healthcare field, it is very important to have direct information about data or, indirectly, data about data, that is, metadata. For instance, grammar (syntax) is metadata. Both structured languages (e.g., Java), and unstructured languages (e.g., the English language) have grammars, describing them (data about data) by clarifying, for example, the relations between words in a sentence. On the other hand, we may be interested by the meaning/interpretation of words or groups of words within a certain medical narrative context, in order to better understand some symptoms declared by the patient (semantics). For more details regarding natural language understanding of unstructured data, see Trim [111] . There are different NLP techniques for extracting information/knowledge from unstructured data, depending on their type (e.g., texts, images, audio, sentiments, etc.). In the healthcare case, there are both structured and unstructured data, clinical data, for instance, being either in structured or unstructured form. There are clinical data presented in some templates (e.g., tables), i.e., structured data, which can be processed directly by different ML techniques. There are also unstructured clinical data in the form of free text narratives (e.g., observations and/or thoughts regarding the patient, obtained during the doctor-patient dialogue) that can only be addressed with the NLP techniques [92] . The traditional approach to support clinical decision-making by NLP in the case of free text narratives is a two-fold process. Firstly, there is text processing enabling the discovery of disease-relevant keywords in the unstructured data. Secondly, the validated keywords through a sensitivity analysis focused on their effects on the classification of the normal and abnormal cases are used to supplement the available structured data. A well-known example of a computer system, which has been directly applied in NLP for healthcare, is the famous "IBM Watson" (https:// www.ibm.com/watson/). IBM announced in 2013 the application of IBM Watson at Memorial Sloan Kettering Cancer Center, New York City. Watson Oncology is a cognitive computing system designed to support the oncology community of physicians as they consider treatment options with their patients. Concretely, it interprets cancer patients' clinical information and identifies individualized, evidence-based treatment options (https://www.mskcc.org/about/ innovative-collaborations/watson-oncology). Let us also mention, in this context, the natural language application-programming interface (API) from Google-Google Cloud Natural Language (https://cloud.google.com/natural-language/). Of the most popular NLP applications we mention just the following ones:

• Medicine, supporting physicians to extract and summarize information of any symptoms, drug dosage, and response data. Consequently, one can identify possible side effects of any medicine while highlighting or flagging significant items in data.

• Machine translation, with the focus on keeping the meaning of sentences intact along with grammar and tenses. • Text categorization, used for assignation of different documents to predefined categories or indices. • E-mail spam filtering, using a set of protocols to determine which of the incoming messages are spams and which are not. • Information extraction, dealing with the extraction of entities, such as names, places, events, dates, times and prices, increasing thus both accuracy and efficiency of a directed search. • Summarization of information, used to summarize huge amount of the data while keeping the meaning intact.

• Classical and Modern ML/DM Algorithms. ML is a very important topic of Computer Science that study theories, algorithms, and related real-life applications that give computers the ability to learn like humans, in other words, acting without being explicitly programmed. While DM can be seen as the science of exploring big data (large or even huge datasets) for extracting implicit, previously unknown and potentially useful information, ML represents the underlying technology to accomplish this task. Depending on the learning paradigm, ML may be broadly classified into learning with a teacher and learning without a teacher. Briefly, learning under the supervision of a 'teacher', or supervised learning, seen as a 'past experience' of the model, means the process of establishing a correspondence between an input and an output, using a training dataset. The purpose of supervised learning is to predict the output value for any new input after completion of the training process, done under the supervision of a 'teacher'. Metaphorically speaking, this is the case of a student who learns from his/her teacher through a question-answer process. The (supervised) classification process represents a classical example of the supervised learning. Unlike supervised learning, in learning without a teacher there is no 'teacher' to monitor the learning process. There are two categories of such a learning paradigm. Firstly, reinforcement learning which aims to connect situations to actions by maximizing a reward (reinforcement) signal. The learning of an input-output mapping is performed by repeated interactions with the environment, in order to maximize the performance, and a "supervisor" does not conduct the student's learning process. Secondly, self-organized learning or unsupervised learning which operates with no external 'teacher' (or referee) to monitor the learning process. In the unsupervised learning, the model is adapted to observations, being distinguished by the fact that there is no a priori output (the learner is fed with only questions without answers, therefore, no teacher). Classical examples of unsupervised learning are the clustering process, or outlier detection. In this context, we cannot fail to mention the statistical learning. Briefly, statistical learning (SL) refers to various tools to model and understand complex datasets by using both statistics techniques and ML algorithms. As in the ML case, there is both supervised and unsupervised SL. While supervised SL builds a statistical model for predicting/estimating an output based on one or more inputs, unsupervised SL deals with inputs without supervising output. More about DM, ML, and SL can be found, for instance, in Gorunescu [112] , James et al. [113] , Sugiyama [114] . Below we will provide more details about the mechanisms of DM/ML/SL, along with some healthcare problems they can efficiently solve. • Classification/Decision-Making. First, let us explain from a theoretical point of view the notion of classification/decision-making, also known as pattern recognition, class prediction or discriminant analysis. Thus, (supervised) classification represents a specific regression analysis from the statistical modeling point of view. Based on one or more independent variables (called predictor variables), this particular type of regression has to predict a categorical dependent variable (called grouping variable). From the AI/ML point of view, the predictor variables represent the object's features, while the grouping variable represents the class label. To conclude, given a set of features, one has to guess the class label, this 'guess' being called a decision, and the process itself is called decisionmaking. In this context, the process of automatic diagnosis, supported by ISs, represents practically a classification issue. Taking into account the signs and symptoms of a patient, he/she must be classified into one of the possible classes corresponding to the different types of possible diseases which can cause these signs/symptoms. The process of classification is based on four basic components:

• Class, which is the dependent (categorical) variable of the model, representing the 'label' (diagnosis type) put on the object (patient) after its classification. An example of such a class (label) is "type 2 diabetes". • Predictors, which are the independent variables of the model, representing the characteristics/attributes (results of the medical tests) of the objects (patients) that have to be classified (diagnosed). For example, the medical test needed to diagnose type 2 diabetes is the glycated hemoglobin (A1C) test, or its alternative test, consisting of random blood sugar test, fasting blood sugar test, and oral glucose tolerance test (for specific details, see, for instance, (https://www.mayoclinic.org/). • Training dataset, which is the set of data containing values for both the class and the corresponding attributes. It is used in the "learning phase" for 'training' the model to recognize the appropriate class, based on available predictors. Examples of such sets are the publicly available datasets provided on Internet for people who work with data. One of the best known and used by the researchers is UCI Machine Learning Repository (https://archive.ics. uci.edu/ml/datasets.html), which contains a rich collection of medical data on different types of cancer, heart diseases, dermatology, hepatitis, HIV, molecular biology, Parkinson, diabetes, and so on. • Testing dataset, which contains new data, formed by attributes only, without the corresponding class. Based on its previous learning experience, obtained through the training process, the classification model will choose the most appropriate class, and the classification accuracy (model performance) can be thus evaluated in real conditions.

Summarizing, a (supervised) classification process is characterized by:

• Input, which is formed by the training dataset containing objects with attributes;

• Output, which is the assigned label for each object, based on the its attributes;

• Classifier, which is fed with input and used to predict the class of a new object (output) with unknown label.

We have illustrated in Fig. 1 .10 the design stages of building a classification model.

Among the most popular classification models (algorithms) today, we could mention:

• Classification/Decision Tree. Random Forest. The classification/decision tree (CT/DT) paradigm can be summarized by the following idea. Having a data space, we want to divide it into distinct classes by using a classification rule. Such a rule has the simplified form: "If the value of the attribute A 1 is lower than a 1 , then if the value of the attribute A 2 is higher than a 2 , …, then the object with attributes A 1 , A 2 , … belongs to class C i. ". The decision boundaries are a 1 , a 2 , …, and the values of the attributes are compared with these thresholds. Once the whole data space is well divided into data sub-spaces as homogeneous as possible, the classification rule is built upon the optimal values of the decision boundaries obtained during the tree induction, (the "tree growth" through the training (learning) process), and it will be used to classify new objects. The tree induction is characterized by:

• Each internal node of the tree (i.e., non-terminal node) expresses the testing based on a certain attribute; • Each 'branch' of the tree is the test's result; • The 'leaf' nodes (i.e., terminal nodes) represent the (decision) classes.

Let us just say a few words about how to split the tree's branches (i.e., split criteria), considered as measures of node 'impurity'. Among the most known Starting from a (classification) tree, why not use an entire forest? This is, actually, the basic idea of considering the random forest (RF) model. The idea behind this paradigm is to build many trees using random objects sampled from a dataset. Seen as an extension of CART, RF builds a multiple CART model with different sample and different initial variables. A random vector is generated for the kth tree, independently from the random vectors corresponding to the (k -1) past trees, and the remaining data are used for class prediction. In this way, we can consider RF like bootstrapping algorithm with CART model. Among the main advantages of RF let us mention the ability of managing a large number of variables (features), performing a sensitivity analysis (i.e., ranking variables according to the impact on the class variable), providing information about the relation between the variables and the classification, handling missing values, no overfitting, detecting interactions between variables, etc. Instead of growing a single tree, RF grows many trees. To classify a new (still unclassified) object, corresponding to an input vector, we assign the vector to each of the trees in the forest. Then, just like in a voting system, each tree "votes" for a certain class and RF chooses the classification having the most votes over all the trees in the forest [115] , (https://www.stat.berkeley.edu/*breiman/RandomForests/).

• Bayesian Classifier. Let us first mention the roots of this paradigm of great importance in the modern theory of decisions. Recall that the English Presbyterian minister and mathematician Thomas Bayes discovered the following famous result [116] , particularly important through its applications. Simply, Bayes' theorem describes the probability of an event, based on prior knowledge of conditions that might be related to the event. Formally, the theorem can be described by the conditional probability formula:

where H stands for a hypothesis and E represents the corresponding evidence. In the Bayesian terminology, PðHjEÞ is called posterior probability, P H ð Þ is called prior probability, P E ð Þ is the evidence, and PðEjHÞ is called likelihood. Thus, the Bayes' formula may be written as:

The Bayesian classifier is based on the minimization of the expected risk when making a decision regarding the choice of a certain class. When assigning an object to a class, one makes a decision, thus we often speak either of Bayesian classification or of Bayesian decision. The Bayesian classification/decision rule can be summarized as follows. Let D k be the decision rule regarding the 'natural' state A k , and let PðerrorjxÞ ¼ 1 À PðA k jxÞ be the error related to A k , given measurement x. The goal of the Bayesian decision is to minimize PðerrorjxÞ, by making the optimal choice of D k . In most cases, when there is no real danger of a wrong decision, all the states/attributes A k are considered to be independent, hence we speak about the naïve Bayes decision/classification.

• Artificial Neural Networks. Artificial neural networks (ANNs), or, simply, neural networks (NNs) may be regarded as non-programmed (non-algorithmic) adaptive information processing systems. NNs learn from examples and behave like the human brain. While the network through the learning process acquires knowledge, the (synaptic) weights, quantifying the intensities of the inter-neuron connections, are used to store the gained knowledge. Once the learning/training process is completed and the synaptic weights are set up, the network is used to classify new objects without known class/label. NN is composed of a large number of interconnected processing elements (artificial neurons), working in parallel to solve a specific problem. Basically, NN consists of the input units fed with information from the environment, the computational units within the network (hidden neurons), controlling the actions in the network, and the output units, which synthesize the network's response. Perhaps the most known and used type of NN is the multi-layer perceptron (MLP), consisting of multiple layers of computational units, usually interconnected in a feed-forward way. Each neuron in one layer is directly connected to the neurons of the subsequent layer, and the computational units use a certain activation function, usually a hyperbolic tangent function. It is noteworthy that, according to the universal approximation theorem for NNs, just a one hidden layer MLP (3-layer MLP) can approximate arbitrarily closely every continuous function, mapping intervals of real numbers to some output interval of real numbers. The way NN is used for classification is relatively simple. In the training phase, the correct class for each record is known (supervised training), and the output nodes can assign correct values or not. Then, one compares the network's computed values for the output nodes to the target values, after that it calculates an error term for each node. The errors are then propagated back through the network. The errors are used to adjust the synaptic weights so that, during the next iteration, the output values will be closer to the target values. The iterative learning process consists in the presentation of the training samples to the network one at a time, and the weights associated with the input values being adjusted each time, this process being often repeated a certain number of epochs. The idea behind this learning paradigm is that the network is trained by adjusting the synaptic weights in order to predict the correct class label of input samples. Once the NN architecture has been designed (structure complexity and parameters), the network is ready to be trained. To start this process, the initial weights are chosen randomly and the training process begins. Once the learning phase has been completed and the synaptic weights have been adjusted, the network is ready to classify new (unlabeled) cases. Advantages of NNs include the high tolerance to noisy data, as well as the ability to classify patterns on which they have not been trained (unsupervised NNs). For more details, see Gorunescu [112] , Haykin [117] . • Deep Learning. NNs, developed in the 1950s, attempted to simulate the way the brain worked in greatly simplified form due to the computational constraints at that moment. It has been demonstrated that "shallow" NNs, i.e., feedforward NNs with a single hidden layer of finite size have the capacity to approximate continuous functions, result subsequently generalized to feed-forward multi-layer architectures (see universal approximation theorem). As it is well known, the basic idea underlying the continuing development of NNs was to build an algorithm attempting to mimic the activity in layers of neurons in the neocortex, a part of the cerebral cortex where the thinking occurs (higher-order brain functions, such as sensory perception, cognition, spatial reasoning and language, etc.). However, NNs have led to many disappointments as compared to other breakthroughs. Rapid development of modern computing tools, and improvements in mathematical formulas, enable computer scientists to build up NNs with a large number of layers than ever before, which is infeasible for classical ones. Under these circumstances, deep learning (DL) is a modern extension of the classical NNs paradigm. Simplifying, one can view DL as a NN with many layers, in an attempt to mimic better the human neocortex. The other reason for the recent popularity of DL is the increase of the data volume and complexity (big data). Among the most popular DL applications we can mention automatic speech recognition, image recognition, natural language processing, image restoration, healthcare, etc. Medical records such as doctors' reports, test results and medical images are a gold mine of health information for DL. A clear majority of DL techniques is used in imaging analysis, which makes sense given that images are naturally complex and high volume. Medical images such as MRIs, CT scans, and X-rays are among the most important tools doctors use in diagnosing diseases ranging from spine injuries to heart disease and cancer. Analyzing medical images can often be a difficult and time-consuming process requiring thus predominant use of DL. In the medical applications, the commonly used DL algorithms include convolution neural network (CNN), recurrent neural network, deep belief network, and deep neural network. CNN has been developed as a much better alternative to the classical ML algorithms when handling high dimensional data, i.e., data with a large number of traits. In the classical approach for high-dimensional image analysis, the solution is to perform dimension reduction. Firstly, preselect a subset of pixels as features, and, secondly, perform the ML algorithms on the resulting lower dimensional features. However, the usual feature selection procedures may lose information from the images. Alternatively, the inputs for the CNN are the properly normalized pixel values on the images. CNN then transfers the pixel values in the image through weighting in the convolution layers and, alternatively, sampling in the sub-sampling layers. The final output is a recursive function of the weighted input values. The weights are trained to minimize the average error between the outcomes and the predictions [118] . The implementation of a CNN has been included in popular software packages such as Caffe from Berkeley AI Research, CNTK from Microsoft, TensorFlow from Google, Keras. https://en. wikipedia.org/wiki/Comparison_of_deep_learning_software. • Support Vector Machines. Suppose we dispose of a set of objects, and we want to separate them into two distinct classes. Imagine a pasture with many sheep grazing, some white, some black, mixed together. The shepherd who looks after them wants, at some point, to divide them by color. If, ideally, the two types of sheep are well separated by an imaginary straight line (i.e., the white ones placed on one side, the black ones placed on the other side), then the (linear) separation issue for the shepherd is completely solved. Unfortunately, in the real-world case, they are well mixed, and the shepherd's problem becomes much more complicated. The support vector machines (SVMs) bring an efficient solution to this problem, i.e., the linear separation of non-linear mixed objects [112, 117, 119, 120] . In order to understand this solution, the shepherd should know what it means to 'teleport' an object in another space, as in the "Star Trek" movies. The idea behind the SVM paradigm is to consider a separation hyperplane (in a higher dimensional space than the one containing the original objects), such that the margin of separation between different types of objects (e.g., white sheet and black sheet) is maximized. To summarize, if the original objects are linearly separable, then there exist a (linear) hyperplane, which separates them. If, as in real-world situations, the original objects are not linearly separable, then the kernel SVM is used based on a separation hyperplane in a high-dimension (kernel space). Thus, SVM employs the kernel trick, i.e., a clever solution to a non-linear separation problem by mapping the original non-linearly separable points into a higher-dimensional space, where a linear classifier exists. For the shepherd's problem, the sheep are 'teleported' into a high-dimensional space where they can be (linearly) separated. In brief, a non-linear classification in the original space is equivalent to a linear classification in the new space, so the main problem remains the linear separation of objects. Under these circumstances, assume the objects are linearly separable. Consequently, there is a linear hyperplane, which separates the two types of objects (e.g., white sheep and black sheep). There could be infinitely many possible such hyperplanes in the absence of additional constraints. For a given hyperplane, the separation between it and the closest object is called margin of separation. The goal of SVM is to find the particular hyperplane, which maximize the margin of separation, the so-called optimal hyperplane (or optimal decision boundary). Let us finally see where the name of SVM comes from. Thus, the "support vectors" are those objects that lie closest to the optimal hyperplane. It can be said that they are 'on the border', so they are difficult to classify, but, on the other hand, they play an essential role for SVM. • k-Nearest Neighbors. Imagine that we are on the edge of a pond and we see more birds on the water. How do we figure out which ones are ducks? A handy solution: using the inductive reasoning of the 'duck test': "If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck". Returning to the k-nearest neighbors (k-NN) algorithm for pattern recognition, the classification is performed by labeling a new object taking into account the resemblance with the k closest neighboring objects. Given a training dataset and a new object to be classified, a certain similarity "distance" between the new object and the training objects is first computed, and the nearest (most similar) k objects are then chosen. The new object is classified by a majority vote of its neighbors, with the object being assigned to the class most common amongst its k nearest neighbors according to the similarity. Unlike other classifiers, there is no training phase for k-NN. For this reason, k-NN is considered a lazy learner because it does not learn a discriminative function from the training data, but just "memorizes" the training dataset instead. In addition, k-NN is a non-parametric classifier because it does not make any assumptions on the underlying data distribution. We only need three elements to use it: (a) a set of stored records (training dataset), (b) a similarity 'measure' to compute the resemblance between objects, and (c) the value of k, representing the necessary number of (neighboring) objects for comparison, based on which we will achieve the classification of a new object. Next, computing the 'similarity' between all the training records and the new object, we identify the k nearest objects (most similar k neighbors), and assign the label that is most frequent among the k training records nearest to that object ("majority voting"). k-NN algorithm is among the simplest of all ML algorithms, being insensitive to outliers, versatile in applications (both classification and regression), having (relatively) high accuracy, and non-parametric feature (no assumptions about data). On the other hand, it is computationally expensive, requires high memory, is sensitive to irrelevant features and the scale of the data, and the prediction speed might be slow for large k. k-NN is particularly well suited for multi-modal classes as well as applications in which an object can have many class labels. • Genetic Algorithms for Classification/Feature Extraction. The genetic algorithms (GAs) are based on the modern theory of evolution, with roots both in the Darwin's natural selection principle presented in the famous "Origin of species" [121] , and in the Mendel's genetics regarding the discreet nature of the hereditary factors transferred from parents to children ("Versuche uber Pflanzenhybride"-"Research about plant hybrids/Experiments in plant hybridization" (1865)). GAs are considered as the most popular case of evolutionary algorithms (EAs). They represent a metaheuristic optimization algorithm, based on a population of potential solutions and using specific mechanisms inspired by the biological evolution, such as: chromosomes, reproduction, mutation, recombination, selection, and survival of the fittest. GAs consist of a population of chromosomes, and multiple operators: selection according to fitness, crossover to produce new offspring, and random mutation of new offspring. A GA algorithm has the following steps.

Step 1: the data are encoded in a vector form and the recombination and mutation rates are picked;

Step 2: the population, consisting of a certain number n of chromosomes, is chosen;

Step 3: the fitness function is computed for each chromosome;

Step 4: the iteration takes place through selection, crossover and mutation until n chromosomes have been generated;

Step 5: the current population is replaced by the new one;

Step 6: a termination criterion is used to stop the evolutionary process. When GAs are used for a classification task, they usually attempt to find either (decision) boundaries between classes (dividing hyperplanes), or sets of classification rules, or feature extraction/selection, etc. • Rule-Based Classification. A rule-based classifier uses a set of IF-THEN rules to classify objects in different classes. The process of rule-based classification uses a training dataset of labeled objects from which classification rules are extracted in order to build the classifier. The set of rules are then used in a given order to classify new (unlabeled) objects. The IF part of the rule is called rule antecedent or precondition. The THEN part of the rule is called rule consequent. The antecedent part consists of one or more attribute tests that are logically connected by AND. The consequent part consists of class prediction. Mathematically speaking, a rule is an implication of the form X ! Y, where X represents the rule antecedent (a conjunction of attributes values), and Y represents the rule consequent, representing the class label. Rule-based classifiers have some advantages as compared to other classification algorithms: high expressiveness and performance, easy to interpret, easy to generate, and high classification speed. • Logistic Regression. Logistic regression (LR) is a statistical classification model, which aims to determine a certain outcome given by a dichotomous variable with two possible states, based on a dataset containing one or more independent predictor/explanatory variables. To understand the logistic regression, let us briefly recall the regressive model. The basic model is the simple linear regression (SLR). SLR is represented mathematically by a linear equation connecting the response (dependent) variable Y with the predictor (independent) variable X Y ¼ b Á X þ a ð Þ . It is interesting to mention in this context that the regression paradigm has its roots in the work of the famous geneticist Sir Francis Galton regarding "regression towards the mean" ("Regression towards mediocrity in hereditary stature", 1886). According to this theory, the characteristics in the offspring regress towards a 'mediocre' point (the mean, actually), opposed to the assumption that extreme characteristics in parents are passed on completely to their offspring. Instead of analyzing bivariate data as in the case of SLR, in most real-life issues the multivariate data are being analyzed. Thus, the multiple linear regression (MLR) model is represented mathematically by a linear equation connecting the response (dependent) variable Y with the predictor (independent) variables X 1 ;

ð X k þ aÞ: However, there are many research areas, including healthcare, economics, physics, meteorology, astronomy, biology, etc., in which the dependent variable Y is no longer a continuous variable, but a binary (categorical) one, which can be simply encoded as Y ¼ 0 and Y ¼ 1 (logistic regression). In order to transform the value of the dependent variable into a binary one we shall use the following formula: logit p ð Þ ¼ ln p

where p is the proportion of objects that have a certain characteristic After reviewing some well-known classification algorithms, it is natural to ask ourselves the question "which ones are the best?" As is to be expected, the answer to such a question depends on the problem at hand, as well as the data we are working with. A subjective choice would be, irrespective of the (subjective) ranking: Naïve Bayes Classifier, Support Vector Machine, Artificial Neural Networks/Deep Learning, Logistic Regression, Random Forests, Decision Trees, k-Nearest Neighbors. However, the problem remains open, due to the fact that even at this time such algorithms are being developed throughout the world.

Besides the classification models, there are other ML algorithms that need to be briefly discussed:

• Survival Regression Models. In different real-world circumstances, such as clinical investigations, there are many situations where several known quantities potentially affect a certain outcome (e.g., patient prognosis). Statistical models are frequently used to analyze survival with respect to several factors simultaneously, providing at the same time the effect size for each factor. In many medical, biological or engineering researches a common issue is to determine whether or not certain continuous (independent) variables are correlated with the survival or failure times. In this context, the first idea that comes to mind is to apply a multiple regression model, but we will end up with two major issues. On the one hand, the dependent variable of interest (i.e., survival/failure time) is most likely not normally distributed, violating thus an assumption for ordinary least squares multiple regression. On the other hand, some observations will be incomplete, the so-called censored data. One of the most known and used survival regression models is the popular Cox's proportional hazard model (Cox PH model). It is the most general survival regression model because it is not based on any assumptions concerning the nature or shape of the underlying survival distribution. No assumptions are made about the nature or shape of the hazard function, the model assuming that the underlying hazard rate, rather than survival time, is a function of the independent variables (covariates). The standard Cox PH model assumes time independent variables, i.e., variables not changing over time (e.g., sex). A more general approach requires an extension to time dependent variables. One of the advantages of the Cox model is its ability to encompass covariates that change over time. The extension of the Cox model to the case of covariates that change over time is known as the Cox's Proportional hazard model with time-dependent covariates. • Cluster Analysis/Clustering. Cluster analysis or clustering, also called segmentation or taxonomy analysis, is an exploratory data analysis that aims to identify structures within the data. By clustering we mean the technique of dividing a set of data into several groups, called clusters, based on certain predetermined similarities. It is noteworthy that the classification process is different from the clustering process. When we classify a certain object, we assign a certain label to that object, while in clustering we group a set of objects in a certain number of clusters, such that objects in the same cluster are more similar to each other than to those in other clusters. Given a measure of similarity and a set of objects, each of them characterized by a set of attributes, the clustering issue is how to divide them into groups (clusters) such that:

• Objects belonging to a cluster are more similar to one another;

• Objects in different clusters are less similar to one another.

From a ML point of view, the clustering process is a form of unsupervised learning. We can mention three types of cluster structure: (a) a single cluster against the rest of the data, (b) the segmentation/partition of the data in a certain number of clusters, and (c) a (nested) hierarchy of clusters. Classically, there are two major approaches to the clustering process:

• Partitional/flat clustering, which represents a division of a set of objects into non-overlapping subsets (clusters), such that each object is in exactly one subset (hard clustering). Let us also mention the soft clustering in which an object has fractional membership in several clusters; • Hierarchical clustering, which represents a set of nested clusters that are organized as a tree.

Remark. Although partitional clustering is efficient and conceptually simple, it has a number of drawbacks, such as: an unstructured set of clusters; the requirement of a prespecified number of clusters as input; and the techniques are nondeterministic.

On the other hand, the hierarchical clustering produces a (nested) hierarchy that is a more informative structure than the flat clustering, does not require a prespecified the number of clusters, and most techniques used are deterministic. Beyond the classic approach above, we will mention some typical cluster models (https://en.wikipedia.org/wiki/Cluster_analysis), such as connectivity models, centroid models, distribution models, density models, subspace models, graph-based models, neural model, etc.

The clustering process involves three main steps:

1. Defining a similarity measure; 2. Defining evaluation criteria measures for the clusters building process; 3. Building an algorithm to construct clusters based on the chosen evaluation criterion.

The clustering paradigm is based on the similarity between objects. A similarity measure indicates how similar two objects are. The choice of a specific measure essentially depends on the problem at hand. In other words, the resemblance depends on the point of view of the intended purpose (e.g., segmentation according to gender, age, disease type, symptoms, etc.). Let us remark that the choice of a measure of similarity must be always in accordance with the type of available data (e.g., numerical, categorical, rank, fuzzy, etc.).

When we segment a certain dataset into clusters, it is naturally to consider the concept of a good clustering. Obviously, there is no universal definition for what is a good clustering. Nevertheless, several evaluation criteria have been developed in literature. We will mention below some important criteria in this direction:

• Internal validation, which evaluates the clustering result using only quantities and features inherited from the dataset, without respect to external information (e.g., sum of squared error (SSE), other minimum variance criteria, scatter criteria, etc.). • External validation, which consists in comparing the obtained clustering with other segmentation approaches (e.g., mutual information based measure, precision-recall measure, etc.). • Relative validation, which compares two different clustering models or clusters (e.g., statistical testing, using SSE, etc.).

As mentioned above, there are two major classes of clustering models: partitioning methods and hierarchical methods. For the methods in the first category, we mention error-minimization algorithms (e.g., k-means algorithm) and graph-theoretic clustering. Among the hierarchical methods, we mention agglomerative hierarchical clustering (bottom-up), divisive hierarchical clustering (top-down), and conceptual.

For more information regarding the clustering models, see Gorunescu [112] , Rokach and Maimon 122], Mirkin [123] .

• Feature Selection/Feature Extraction. In Bellman [124] , has first introduced the term of "curse of dimensionality", referring as a phenomenon which often occurs when applying ML algorithms to high-dimensional data. In this situation, "the number of samples needed to estimate an arbitrary function with a given level of accuracy grows exponentially with respect to the number of input variables (i.e., dimensionality) of the function" [125] . In the ML practice, there is a maximum number of features in the database, above which the algorithm's performance will decrease rather than increase. To avoid such a situation, one can use the dimensionality reduction technique, obtaining a reduced amount of time and memory required by data processing, better visualization, elimination of irrelevant features and possible noise reduction. As techniques for dimensionality reduction, we can mention typical multivariate exploratory techniques such as factor analysis, principal components analysis, multidimensional scaling, cluster analysis, canonical correlation, etc. There are usually two different approaches for dimensionality reduction: (a) feature selection (FS), and (b) feature extraction (FE). FS uses different methods for selecting a subset of the existing features, the most significant ones for the proposed purpose (e.g., classification, regression). In this way, the selected subset of variables from the input can efficiently describe the input data. As a result, the effects from noise or irrelevant variables are reduced, and this approach provides prediction at least as good, but with a much lower computing effort. So far along, we may think that IS in healthcare is a real miracle. But can we actually trust its results? Is it possible to measure the machine's performance and if so, can we take comfort in the fact that the measurements are accurate? Fortunately, since 1800 we have Statistics to rely on. Data scientists use statistical methods for comparing classifiers. This is not an easy task to tackle. As stated before, there is no single classifier that works best on all given real-world problems. No classifier offers a 'short-cut'. We can say that this phenomenon is related to David Wolpert and William G. Macready theorem 'No-free-lunch'. Each classification algorithm plays the role of a 'restaurant', providing a certain 'dish' at a certain 'price', thus its up to us to determine the 'smart-deal' [126] .

Because the issue at hand is the healthcare system, a very important part is represented by the associated cost regarding correct and incorrect classification. Depending on the case the cost varies. For a better understanding of the process, let us imagine the following situation: let's suppose that we have a classifier that establishes whether a tumor is benign or malignant in a certain type of aggressive cancer. The cost of misclassifying a malignant tumor as benign is greater than misclassifying a benign tumor as malignant? Both situations are serious, whether a person believes he/she is healthy and does not start the treatment (surgery, chemotherapy, radiotherapy, etc.) or the person is healthy and starts an unnecessary treatment [127] .

The classification evaluation counts the correct and incorrect predicted objects. These values are being tabulated in a confusion matrix-see Table 1 .3. In addition, a cost matrix can be created in order to maximize benefit or minimize cost-see Table 1 .4. Depending on the case, sometimes a more accurate classification is desired for certain classes rather than others.

Each cost is assigned according to the problem. The overall cost and accuracy are computed using the following formulas:

Example 1.1 Let us consider the example mentioned above: a case of an aggressive form of cancer. We have two classifiers that C 1 and C 2. Below we have three tables. The first one, Table 1 .5, presents the cost matrix established a priori in this sensitive issue. Tables 1.6 and 1.7 present the confusion matrix for the two classifiers.

Keeping the formulas in mind, we will obtain the following numbers measuring the performance:

Now comes the real problem. We have both the cost and the accuracy of each classifier. These pairs of parameters that determine the measure of performance must be taken into consideration when choosing the best performing model. This remains an open question, since we must find a balance between cost and accuracy.

Besides the cost and the accuracy, two other statistical concepts must be added to the checklist for finding the 'smart-deal' of IS: sensitivity and specificity. It must be noted that these concepts are used for binary classification. Both concepts are equally important, so a classification must be sensitive and specific. Sensitivity measures the proportion of 'true positives' that are correctly classified, whereas specificity the proportion of 'true negatives' correctly classified. 

In order to find the probability that the model will give the correct diagnosis we need to compute two other concepts: positive predictive value (PPV) and negative predictive value (NPV). Basically, PPV computes the proportion of 'positive' cases that are classified correct, and NPV computes the proportion of 'negative' cases that are classified correct.

If one interprets these parameters in our example, 100% sensitivity means that all cancer patients are recognized as having the disease, whereas 100% specificity means all healthy people will be recognized as being healthy. Unfortunately, 100% sensitivity and 100% specificity is just a beautiful dream, impossible to be reached.

For even a more thorough analysis we can use other four indicators: In practice, the medical personnel prefers to use high false positive rate, rather than false negative rate.

Another interesting way of assessing the prediction results is the Receiver Operating Characteristic Curve (ROC). The history of the ROC curves is fascinating. For this, we must time travel back on December 7, 1941. We are located on the island of Oahu, Hawaii. We are during the World War II, in Pearl Harbor. The radar picked up some unusual behavior, described as a flock of birds. The famous reply concerning this situation was: "Don't worry about it". The primitive radar failed to differentiate between enemy aircrafts and birds. 2,403 Americans died that day. 1,178 were wounded. Electrical and radar engineers started developing ROC curves. Since then, they have been used in various applications in medical research [127] [128] [129] [130] [131] [132] .

For a better understanding we shall present in Fig. 1.11 below two types of ROC curves, discrete and continuous depending on the output of the classifiers. A ROC curve is a two-dimensional graph, which plots on the X-axis the FP rate, and on the Y-axis the TP rate. This way we can find the trade-off between the benefits and costs. Now that we see how a ROC curve looks like, we need to interpret it. In order to obtain a good interpretation, we can use the following rule: one point in space is better than another if it is situated to the northwest of the square. Practically, if the TP rate is high and FP rate is low, or both at the same time, the prediction is better. In fact, using the line of no-discrimination, one can say that if the points are above the line, the classification is good, otherwise is poor.

For some, the ROC curve is not as suggestive as a number, so the solution is using the Area under the ROC curve (AUC). Obviously, a picture is worth a thousand words, whereas a number is a number. Using the AUC diminishes the information regarding the pattern of trade-offs of the classifier.

The translation guide of the AUC is: But what is the philosophy behind the AUC? If one is familiar to the Mann-Whitney U test or the Wilcoxon test of ranks, the interpretation of the AUC is equivalent. AUC is the probability that the classifier will assign a higher score to a positive example rather than to the negative one. More information about AUC can be found in Fawcett [133] , Hanley and McNeil [134] .

If a data scientist wants to validate its model, unfortunately the ROC curve and AUC are not sufficient. Let us suppose that again we have two classifiers C 1 and C 2 that we want to compare. If both classifiers are tested on the same database, there shouldn't be any issues, but what if the first one, C 1 has an accuracy of 72% on a sample size of 4500 items, and the second C 2 has an accuracy of 88% on a sample size of 50. Can you actually say the second classifier performs better?

Fortunately, the z-test, also known as the "difference between two proportions" test will provide the correct answer. The test computes the p-value (which determines how much was the hazard involved in the result). If p > 0.05 there is no statistical significant difference between the two models.

At this point, we answered the questions why and how IS in healthcare. Our journey continues to the next level: what are the actual benefits of intelligent healthcare?

Intelligent healthcare… The name resonates with a science fiction movie or book. We shall once again state that is our strong belief that there will be no doctorless hospitals. Even if healthcare without AI can no longer exist, the human sixth sense shall never be replaced. Nevertheless, intelligent healthcare is happening today.

It should be noted, that unfortunately 100% accuracy in diagnosis or patient management is just a dream. People fail, machines fail. Statistics is just numbers that give you an insight on what may or may not happen. Data scientists 'play' with numbers, but unfortunately, those numbers are people, people's characteristics. No one knows exactly what will happen if you pinpoint a certain person, but intelligent healthcare and statistics can give you and the doctors a probability. If that person is part of that probability or not it is not up to us to know or foresee.

In healthcare there are three major concerns: a disease, a disorder and a syndrome. The definition of a disease was published in the British Medical Journal [135] in 1900: "resulting from a pathophysiological response to external or internal factors". A disease has signs and symptoms (e.g. cancer, cardiovascular disease). The definition of a disorder is "abnormal physical or mental condition" [136] (e.g. arrhythmia-abnormal heartbeat). A syndrome is a collection of symptoms that suggests a certain disease (e.g. Down's syndrome, auto-immune syndrome, acute respiratory distress syndrome). If for a disease or a disorder there is treatment, for a syndrome there is no such thing. That is why it is very difficult for the medical personnel to identify and treat a certain syndrome. Here the trial-error approach is used.

The success of using AI in medicine is going up. Billions of dollars are pushed in the system. According to a recent report of Frost and Sullivan (https://ww2.frost.com/ news/press-releases/600-m-6-billion-artificial-intelligence-systems-poised-dramaticmarket-expansion-healthcare-accessed May 4, 2018) the market for AI will reach 6 billion dollars in 2021. AI goes towards personalized medicine gathering information in order to tailor personalized treatments and to monitor patient's response to it.

Next we will present the benefits of using intelligent healthcare in diagnosing and treating diseases and disorders.

The most frightening C word is Cancer. It is everywhere. Let's admit it, each one of us has one or even more loved ones who are/were diagnosed with this disease. Some are still with us struggling each day, praying and hoping for new therapies to be discovered, others lost the battle. The statistic is bad. According to Cancer Research UK (http://www.cancerresearchuk.org/health-professional/cancer-statistics/ worldwide-cancer#heading-Zero-accessed May 4, 2018) in 2012, 14.1 million new cases of cancer occurred worldwide. In 2008 an estimated 169.3 million years of health life were lost globally due to cancer. It is an estimate that by 2030 there will be 23.6 million new cases each year.

In the last episode of Dr. House aired on May 21, 2012, Dr. House tells his best friend, an oncologist who had just been diagnosed with cancer that: "Cancer is boring". Sadly, it isn't true. In Pleasance [137] , the authors state that in fact cancer is actually hundreds or thousands of rare diseases, and that each tumor is to some extent unique. Nothing is boring about that. In some cancers (e.g. esophageal) the chemotherapy protocol has not change in years. If the genomic and outcome data is recorded each unique tumor might have its protocol.

But what is cancer? Cancer happens when a cell mutates and starts to reproduce but has no longer control of its growth. Even in the same type of cancer, cells can grow rapidly and that makes the cancer aggressive or can grow slowly. So, even if it is of the same type, cancer is an individualized disease, so it must be addressed accordingly McDermott et al. [138] .

So now that we have established an example of a devastating disease, let's take a look on how would the healthcare system benefit from the use of AI.

Due to smoking and polluted air, the rates of lung cancer in China have gone up. Only 15% of people that are diagnosed survive. Fortunately, due to the fact that Chinese citizens are required to get regular lung screenings, the disease is caught early. On May 16, 2017, Forbes magazine presented Infervision Inc., a company that focuses on intelligent healthcare (https://www.forbes.com/sites/jenniferhicks/2017/05/16/seehow-artificial-intelligence-can-improve-medical-diagnosis-and-healthcare/#129562306 223-accessed May 5, 2018.). Infervision uses Deep Learning to learn from images produced by XRays, MRIs, CTs and even pathology, in order to identify abnormal tissues in the lungs. The technology is being used in Shanghai Changzheng Hospital.

At the Mayo clinic, the genomic information of brain tumors is discovered without a biopsy, by the use of Deep learning. At Stanford, an ANN is trained to recognize skin cancer.

"Deep-patient" is an unsupervised representation that predicts from EHR, the future of patients from 78 diseases from schizophrenia, severe diabetes to various cancers [139] .

Even if cancer is scary, it is the second leading cause of death worldwide. According to a report of the American Heart Association-http://www.heart.org/ HEARTORG/-accessed May 5, 2018-that compiled data from more than 190 countries, heart diseases took 17.3 millions lives each year. By 2030 the number is expected to be 23.6 million. Comparing cancer and heart disease deaths in expected 2030, we can observe a tie.

About 47% of sudden cardiac deaths occur outside a hospital. Heart problems are currently diagnosed by monitoring the timing of the heartbeats in scans, but cardiologists are not always accurate. During this procedure one in five patients suffers a heart attack or undergoes an unnecessary surgery. Out of 60000 scans performed each year, 12000 are misdiagnosed. The estimated cost is $812 millions.

An AI system that was developed at the John Radcliffe Hospital discovers in scans information that the human eye cannot see. The system has been tested in six cardiology units in clinical trials and the results are expected to be published soon.

Deep learning can find patterns in heterogeneous syndromes and image recognition. Intelligent healthcare can classify new genotypes and phenotypes for: heart failure, Takotsubo cardiomyopathy, hypertrophic cardiomyopathy, hypertension, and coronary artery disease. Just like in cancer, the therapy can be tailored accordingly.

Data Scientists from GOOGLE and Verily have developed a system that assesses a person's risk of heart disease using AI. The deep learning algorithm analysis scans of the back of a patient's eye and gathers information related to the individual's age, blood pressure and even if the person is a smoker or not Poplin et al. [140] . 300 000 patients were used in this study. The fundus of the eye is full of blood vessels that reflect the overall health. Analyzing the fundus important predictors of cardiovascular diseases are found.

In an article that was published by the Journal of the American Medical Informatics Association [141] , the researchers modeled a recurrent neural network in order to learn temporal relations among events in EHR. Thus, the algorithm anticipates early stages of heart failure, leading to a tailored prevention plan. In a recent study, which was published in January 2018 [142] , a smart phone equipped with an AI algorithm measures arterial stiffness. Arterial stiffness appears when the arteries become rigid, increasing blood and pulse pressure. This process causes the tension to travel to the peripheral vasculature, causing organ damage (kidney, brain, etc.). The only current alternatives to this smart phone app are MRI and tonometry.

The 7th leading cause of death in the USA is diabetes according to the American Diabetes Association-http://www.diabetes.org/diabetes-basics/statistics/-accessed May 16, 2018. The cost of caring for diabetes patients is up to $245 billion per year [143] . In 2015, 84.1 million Americans age 18 and older had prediabetes.

Unfortunately this disease is turning into an epidemic. According to Neocleous et al. [144] 28.5% of diabetics have diabetic retinopathy that often leads to blindness. A deep neural network was developed for detection of diabetic retinopathy in 2016, Pomprapa [145] .

Diagnosing diseases and disorders is a difficult task. An even more difficult issue is diagnosing syndromes. Syndromes have no protocols when it comes to diagnostic or treatment. Doctors use the trial error approach in most cases, often leading to misjudge.

But, here comes AI to the rescue (at least in some syndromes…)! AliveCorhttps://www.alivecor.com/-accessed May 14, 2018-together with Mayo Clinic started in 2016 to test whether the long QT syndrome can be identified in EKG results. Long QT syndrome causes almost 4000 sudden deaths per year in children and young adults in the US. According to Knaus et al. [146] the syndrome appears in 1 in 2000 live births. AliveCor sells portable EKG sensors that can be attached to smartphones and an EKG embedded in an Apple Watch band. The system developed by AliveCor uses a deep neural network that learns from the EKG results from over 1000 patients in order to find relevant features of the long QT, leading to an early diagnosis that might prevent sudden deaths.

AI is improving Down syndrome diagnosis also. The results published in Ultrasound in Obstetrics and Gynecology in 2017 [147] , by a research team from Netherlands, Cyprus and the UK, use an ANN to diagnose the syndrome. Current non-invasive procedures compute an estimate of the maternal age, blood tests and ultrasound to detect the presence or absence of fetal nasal bone. Still, the estimates have a considerable false-positive rate. Another non-invasive test is the cell-free DNA, but the costs are very high. The same trained ANN can diagnose Turner syndrome also. An automatic ventilation system using AI was developed for acute respiratory distress syndrome (ARDS). ARDS is hard to diagnose, because it has to be differentiated from pneumonia or congestive heart failure. ARDS has no cure and can leave the survivors with diminished functional capacity, mental illness and low quality of life; medical personnel just provide life support for the patient until the lungs start functioning again. The aim of the study [148] is to develop an automatic control system for mechanical ventilation using AI. Currently this approach can be applied in animal trials, no approval for humans has been granted yet.

AI is used to diagnose rare diseases, such as the Mabry syndrome [149] . The Mabry syndrome is a rare condition that is triggered just by the mutation of a single gene, and it causes mental retardation. The scientists use AI to find in the photographs of 91 patients specific cell surface changes, such as: a narrow tent-shaped upper lip, broad bridge of the nose and wide-set eyes, etc. These features may be obvious or not, thus complicating the differential diagnosis.

If these technologies will be produced at a reasonable cost so that every primary care office could afford them, the detection could be made early and it could save lives.

Besides diagnosing diseases, the benefits of intelligent healthcare are found in:

• Managing Medical Records. As stated earlier in this chapter, only 27% of the workday, the doctor has face time with the patient. An intelligent management of medical records improves that ratio. The origin of medicine comes from the Latin "Ars medicina"-"the art of healing". Beautiful said, right? It is an art after all. Let's sail a little throughout history:

• In ancient Egypt the first known by name physician was Imhotep, 2600 BC. He was able to describe and treat 200 diseases; • In 500 BC Alcmaeon of Croton was able to differentiate arteries from veins; • The first aspirin was prescribed in 400 BC by Hippocrates; Medicine has come a long way from bloodletting for every disease to personalized medicine. To refresh one's memory let us mention that for headaches the cure was pressing a hangman's rope to your head, or for gout the treatment was applying a superb mixture of worms, pig's marrow and herbs, boiled together with the carcass of a red-haired dog.

We hope we have convinced you of the grand benefits that AI brings to healthcare. Let us continue our journey to see what is next for healthcare, what prospects do intelligent systems bring?

1.4 What Are the Prospects of Using Intelligent Systems in Healthcare?

Healthcare… The final frontier… These are the voyages of the Intelligent Systems. Their continuing mission: To explore strange new diseases… To seek out new treatments; new drugs… to boldly go where no one has gone before! This is an adaptation of Captain's Jean-Luc Picard quote from the Star Trek series The Next Generation (1987) that fits perfectly in the context regarding the prospects of Intelligent Systems in healthcare! A hard task is to identify which idea is pure speculation or not. A well-informed opinion can be found in the report delivered in December 2017 by JASON [151] . JASON is a scientific group, which has been advising the US government since 1960. The group's concerns are of sensitive nature. The report on McNamara Line electronic barrier strategy employed in 1966-1968 during the Vietnam War brought the group's notoriety. Even if the group is military focused, JASON now redirects its interest in AI in healthcare.

There are two major actors in this play. First, there is the academia, which is interested in developing new ML techniques. The other one is the private sector, which has expressed an enormous interest in AI in health applications. Currently there are 106 listed startups from 15 different countries (US, UK and Israel, etc.). The question that intrigues everybody: Is this just a step in the long road of intelligent systems applied in healthcare or their end is just around the corner?

The quality and cost of healthcare all over the world have enabled the scientists and the industry to open their minds and seek new solutions. JASON's report states that this time around, AI applied in healthcare will be a blast because the confluence of three forces:

• The current medical system is unsatisfactory • The ubiquity of smart devices in our lives • The success of home services provided through Amazon and others.

One of the prospects of intelligent systems in healthcare is the development of clinical applications. For this idea to be put into practice one major concern regards the availability of relevant data [152] . Finding quality data is a real challenge. Medical datasets have privacy issues and are expensive to collect because of longitudinal studies (repeated observations) and clinical trials. Once researchers collect the data, they tend to keep it to themselves. Other issues concern the lack interoperability of EHR systems.

After the data is collected, it must be labeled. The labeling is done by independent professional assessments of each case, whether we are talking about images or other types of data.

A very interesting fact pointed out by JASON, that we totally support and agree with, is that the academia's focus should be concentrated also on creating rigorous testing and validation approaches, not only on researching for new methods. The problems in implementation must be identified and ameliorated [153, 154] , in order to build up confidence. Modifying existing protocols for diagnosis and treatment can be achieved only if the medical community trusts the intelligent systems. For this, the intelligent systems:

• Should address an important clinical need;

• Should perform at least as well as existing approaches;

• Should be statistically tested and validated in order to be trusted; • Should provide improvements in patient outcomes, life expectancy, quality of life; • Should reduce costs. mHealth (mobile or digital health) apps are becoming nested by smart technologies [156] . The apps provide answers to a full spectrum of health issues. Some questions have risen regarding the fact whether the medical community should integrate mHealth into healthcare? Some of the mHealth apps have been included in clinical trials. The American Medical Association (AMA) has recently embraced a set of principles to promote effective and safe mHealth apps. AMA encourages the medical personnel and users to use the apps, associated devices, sensors and trackers.

According to a survey released by AMA in 2016, 31% of doctors see the potential of digital tools in medicine and half of them believe in the improvement brought by those (https://www.ama-assn.org/sites/default/files/mediabrowser/ specialty%20group/washington/ama-digital-health-report923.pdf).

The AMA report states the fact that academia and the industry must fit the IDSS in the existing medical systems and practices by including data privacy assurance, linkage between EHR and billing, reimbursements.

The use of mHealth apps is two-fold. One the one hand the users can monitor their health, and, the on the other hand, large dataset can be created and then used for training the AI applications. For example, using the Asthma MD app, the user can upload anonymously data into a Google database for research purposes. The data is analyzed in order to find correlations between asthma and environmental factors, triggers, and climate change. Another example is the mHealth app for managing Parkinson disease [155] . The app was available for free in the US through the Apple App Store. 48 k people downloaded the app, and 78 of the 25% eligible individuals accepted to share their personal data. The numbers show that people are open to building medical databases for research purposes.

More and more research proposal grants on IDSS applied in healthcare are written in order for the academia to receive sponsoring from government agencies. Still, it is our strong belief that in some countries there is still not enough funding in this domain.

Assessment through smartphone apps cannot be done in cases like measuring metrics of health (minerals, vitamins, hemoglobin, cholesterol, etc.), viruses, bacteria, and cancer or hearth disease biomarkers. Hopefully, one day, maybe these will be evaluated through small blood samples, and then the dream of monitoring them through mHealth will become true.

All the ideas presented above sound too good to be true. Sadly, like everything else in this world, this internet/smart device healthcare has its Achilles' heel. Because of the enormous sums of money that are invested in this technology, there are and will be many predatory companies that scam people. So the question goes down to this: how does someone separate the wheat from the chaff? How can we know which app to trust?

There are many sites that offer treatments for certain diseases at an unreasonably expensive price. For example, we shall present a case mentioned in the JASON report. It regards a mutation of the Methylenetetrahydrofolate reductase (MTHFR) gene. Doing a simple Google search on this MTHFR gene mutation, one discovers some sites that state that the "conventional" medical doctors ignore this gene as source of your health problems. Some naturopaths have obviously the solution to one's problem. Firstly, you will read all the health issues that this mutation produces: anxiety, fatigue, brain fog, dysplasia and of course cancer, heart attacks, stroke, Alzheimer, miscarriage, autism, etc. A genetic test is then required in order to see whether the gene is mutated or not. The cost of a test plus four-month treatment plan is $3000. A "small" price for finding out what is wrong with you, and for enabling you to take control of your life and health. The mutation will be solved through a "natural remedy plan".

To avoid these scams, you should read carefully the sites you visit. For example, if the site is out of date for several years, that should indicate a problem. Another warning sign is the expert's CV. Check his/hers ISI Web of science, Scopus and Google scholar profile. If his/her citations come only from himself/herself, you should start having doubts. Check out where he/she published her articles. Are the journals trustworthy, or are they predatory?

This example can be easily extrapolated to IDSS applications. Imagine all the dermatological diseases, including skin cancer. Online skin cancer services exist: https://www.directderm.com/-accessed May 21, 2018, https://www.firstderm. com/-accessed May 21, 2018, https://www.skinvision.com/-accessed May 21, 2018. Skinvision is a new company that enables users to send a picture through the app and receive feedback on it. There is little information about the algorithms behind the analysis, and nothing on how was this application validated. In fine print we can read that the website acknowledges that "our solution is not a diagnostic device". Do not get fooled by these websites. There are websites that can be trusted such as: Mayo Clinic-https://www.mayoclinic.org/-accessed May 21, 2018, WebMD -https://www.webmd.com/-accessed May 21, 2018.

Another prospect of intelligent systems in healthcare should be the encouragement of "crowdsourcing" movement. AI competitions stimulated the creation of large corpuses of data, used afterwards in healthcare applications. Some AI competitions are: ImageNet http://www.image-net.org/challenges/LSVRC/-accessed May 21, 2018, Kaggle, The Booz Allen Hamilton Data Science Bolw-http://www. datasciencebowl.com/totheclinic/-accessed May 21, 2018, Zooinverse-https:// www.zooniverse.org/-accessed May 21, 2018.

We appeal again to the famous Dr. House. If you have watched the series, you know that whenever a difficult case was presented to him, Dr. House asked his team to go and search the person's house, neighborhood, work place, etc. This is personalized medicine. Everything we do, our education, our neighborhood, our economic status, our diet, our social context defines and determines our health. If we corroborate all of these with genetic information we can determine certain diseases and elaborate treatment plans.

The Human Genome Project was an international research plan with the goal of determining and mapping all the genes in the human genome. In 1984, the US government started this plan, but the project was launched in 1990 and completed in 2001. In 2003, the human genome sequence, 3 billion DNA bases, was determined. The ultimate goal of the project was to identify correlation between certain gene variations and human diseases. The scientists were looking for the answers to questions such as: Why do some people have cancer or Alzheimer? If you could have your human genome analyzed in order to foresee which disease you will suffer from during your lifetime, would you want to know? Surprisingly enough, it has been found that human disease is rarely linked to specific genetic mutations [157, 158] . This led to the idea that personalized medicine is a combo box of genetics, behavior, environment, family history, life and treatment experiences.

The Project All of Us-https://allofus.nih.gov/-accessed May 21, 2018-brings together National Institutes of Health (NIH) and Precision Medicine Initiative (PMI). The goal of the project is to collect a massive amount of individual health data from genetics to social behavior. The data gathered will contain:

• Basic information on medical history and lifestyle (habits and overall health); • Physical measurements (blood pressure, pulse, height, weight, hip/waist circumference); • Biosamples of blood and urine; • Optional DNA testing;

• EHR (health care visits, diagnosis, procedures, treatments, vital signs, laboratory tests).

All this data will be collected over 10 years. Future goals include EHR and data collected from wireless sensor technologies (mobile/wearable devices) and geospatial and environmental data. All the data will be available to participants, researchers and the public. On the bright side, this means that the academia and the industry can have a large dataset that they can use to improve their IDSSs, but unfortunately the privacy protection of the participants is not guaranteed. PMI has recognized this from the start of the project. The All of Us project is trying to develop privacy, trust principles, data security policy principles and framework.

For the database to be complete, the environmental data must be monitored and recorded (e.g. pathogen exposure). Let us stop for a minute and ask the following question: when you were in your doctor's office or in the E.R., were you ask in what environment you live in? Our guess is that you weren't. Not because the doctor missed or forgot this question, but because this question is not on the standard protocol questionnaire. Environmental exposure can trigger a disease such as cancer or autism. The things that need to be measured are: chemical components of air pollution, allergens, noise, UV intensity, lead, asbestos, radiation, and human pathogens. All this data must be captured and added to the training set, in order for the IDSS to perform better.

Another matter that needs to be addressed is Robotic surgery, O'Sullivan et al. [163] . We do not know whether you are a fan of the series Grey's Anatomy or The Resident, or any other new medical drama show, but for sure you have heard about the da Vinci robot from Intuitive Surgical. The Food and Drug Administration (FDA) cleared da Vinci for surgery in 2000.

The da Vinci robot is a combo box that contains a console and four robotic arms. The surgeon is able to move the robotic arms through the console. The arms can hold objects, and act as scalpels, scissors, bovies or graspers. The da Vinci has better motion than the human hand, and it can also reduce the tremor and refine the surgeon's hand movements. The da Vinci system allows the surgeon to operate while sitting, opposed to the conventional laparoscopy. A procedure using da Vinci is minimally invasive, it has less pain associated with it, less blood loss and thus less need for blood transfusions.

Not all the robotic surgeries involve the use of AI, but some of them use computer vision to identify distances, specific body parts, etc. In the future using AI the robotic surgery might determine the margins of a tumor, separating normal tissue from malignant tissue, etc.

If it can be autonomous driving, why not even autonomous robot surgery? The Smart Tissue Autonomous Robot (STAR) performed a surgery on a pig's small intestines using its own vision tools and artificial intelligence. The mind-blowing fact is that STAR outperformed the human surgeons that were given the same task; the robot's stitches were more consistent and resistant to leaks, Shademan et al. [159] . STAR performed the surgery both in ex vivo tissue in the lab and on in vivo tissue on an anesthetized pig. In 40% of the trials, human intervention was needed.

Current medical systems all over the world need to be improved. Medical care is expensive and these costs must be reduced. AI plays and will play an important role in healthcare, whether it will regard early diagnosis and tailored treatment, protection from insurance fraud or reduction of costs.

One question arises though. When you start driving a car on a trip and you use a navigation app, the app tells you that the driver holds the whole responsibility. So, who holds the responsibility when it comes to IDSS in healthcare? If a machine makes a fatal or near fatal error who pays for it? The hospital that bought the system? The manufacturer? The doctors that did not use it properly?

No matter the advances in the field, the old saying stands still: Medicine is not mathematics. The sixth sense of the human doctor shall never be replaced by a machine, no matter how that machine was trained, which algorithm it has been implemented with, and its statistical validation results. We once again strongly state that intelligent systems in healthcare will provide at best a recommendation.

Computing machinery and intelligence

The Turing Test. The Stanford Encyclopedia of Philosophy

Forget the Turing test -there are better ways of judging AI

The truth about the turing test

Brainy machines need an updated IQ Test, experts say

Judge Weighs in on Chatbot's turing test performance

Minds, Brains, and Programs

The Essential Turing: Seminal Writings in Computing, Logic, Philosophy, Artificial Intelligence, and Artificial Life: Plus The Secrets of Enigma

Artificial Intelligence and Intelligent Systems, 1st edn

Intelligent Systems: Modeling, Optimization, and Control

Intelligent Systems for Engineers and Scientists

Intelligent Systems: A Modern Approach (Intelligent Systems Reference Library Series

Intelligent Systems

Intelligent Systems: Models and Applications. Topics in Intelligent Engineering and Informatics Series

Foundations of Intelligent Systems

Hybrid Artificial Intelligent Systems

Intelligent Systems and Applications

Advanced Fuzzy Logic Technologies in Industrial Applications

Foundations of Fuzzy Logic and Soft Computing

Analysis and Synthesis of Fuzzy Control Systems: A Model-Based Approach

Intelligent Control: A Hybrid Approach Based on Fuzzy Logic Neural Networks and Genetic Algorithms

Introduction to Type-2 Fuzzy Logic Control: Theory and Applications

Business Intelligence for Telecommunications. Auerbach (Taylor & Francis)

Management Intelligent Systems: First International Symposium

Business Intelligence

Business Intelligence Guidebook: From Data Integration to Analytics

Intelligent Bioinformatics: The Application of Artificial Intelligence Techniques to Bioinformatics Problems

Bioinformatics: Problem Solving Paradigms

Computational Intelligence Methods for Bioinformatics and Biostatistics

Intelligent Computing in Bioinformatics

Computational Intelligence Methods for Bioinformatics and Biostatistics

Intelligent Paradigms for Healthcare Enterprises: Systems Thinking

Advanced Computational Intelligence Paradigms in Healthcare -3

Intelligent Patient Management

Advanced Computational Intelligence Paradigms in Healthcare 5

Engage! Transforming Healthcare Through Digital Patient Engagement

Programming Game AI by Example

Advanced Intelligent Paradigms in Computer Games

Artificial Intelligence for Games

Introduction Game AI

Knowledge-Free and Learning-Based Methods in Intelligent Game Playing

Intelligent Multimedia Computing Science: Business Interfaces, Wireless Computing, Databases and Data Mining

Pervasive Computing: Innovations in Intelligent Multimedia and Applications

Intelligent Multimedia Communication: Techniques and Applications

Intelligent Multimedia Databases and Information Retrieval: Advancing Applications and Technologies

Intelligent Multimedia Surveillance: Current Trends and Research

Human-Computer Interaction Handbook

Human-Computer Interaction. Novel User Experiences, Part III

Human-Computer Interaction. User Interface Design, Development and Multimodality, Part I

Human Interface and the Management of Information: Information, Knowledge and Interaction Design, Part I

An Introduction to Knowledge Engineering

Knowledge-Based Software Engineering

Knowledge-Based Software Engineering

Knowledge Management: Value Creation Through Organizational Learning

Advances in Knowledge Management: Celebrating Twenty Years of Research and Practice

Speech in Mobile and Pervasive Environments

Towards Adaptive Spoken Dialog Systems

Computational Paralinguistics: Emotion, Affect and Personality in Speech and Language Processing

Automatic Speech Recognition: A Deep Learning Approach

Natural Language Processing and Computational Linguistics 1: Speech, Morphology and Syntax

Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems (Computational Neuroscience Series

Brain-Computer Interfaces: Principles and Practice

Brain-Computer Interfacing: An Introduction, 1st edn

Towards Practical Brain-Computer Interfaces: Bridging the Gap from Research to Real-World Applications

Brain-Computer Interfaces 1: Methods and Perspectives (Cognitive Science)

Autonomous Robots

Design and Control of Intelligent Robotic Systems

Introduction to Autonomous Robots

The Horizons of Evolutionary Robotics

Introduction to Robotics Mechanics and Control

Intelligent Infrastructure: Neural Networks, Wavelets, and Chaos Theory for Intelligent Transportation Systems and Smart Structures

Modelling Public Transport Passenger Flows in the Era of Intelligent Transport Systems

Intelligent Transportation Systems: 802.11-based Vehicular Communications

Computational Intelligence in Medical Imaging: Techniques and Applications

3D Computer Vision: Efficient Methods and Applications

Computational Vision and Medical Image Processing

Advanced Topics in Computer Vision

Computer Analysis of Images and Patterns

A History of Cognitive Science

Computing the Mind

Artificial Psychology: The Quest for What It Means to Be Human, 1st edn

The Computational Theory of Mind. The Stanford Encyclopedia of Philosophy

Autonomous Military Robotics

Drone Wars: Transforming Conflict, Law, and Policy

Military Robots: Mapping the Moral Landscape (Military and Defense Ethics)

Popular Mechanics Robots: A New Age of Bionics

Intelligent Computational Optimization in Engineering: Techniques & Applications

Energy Pricing Models: Recent Advances, Methods, and Tools

Intelligent Techniques in Engineering Management: Theory and Applications

Electricity Distribution: Intelligent Solutions for Electricity Transmission and Distribution Networks

The New Science of Cities

Smart Cities: Big Data, Civic Hackers, and the Quest for a New Utopia

Intelligent Cities: Enabling Tools and Technology

The Age of Intelligent Cities: Smart Environments and Innovation-for-all Strategies

Smart Cities: Foundations, Principles, and Applications

Designing the Internet of Things

From Machine-to-Machine to the Internet of Things: Introduction to a New Age of Intelligence

The Internet of Things

Building the Internet of Things: Implement New Business Models, Disrupt Competitors, Transform Your Industry

Cognitive Internet of Things: Collaboration to Optimize Action

A Brief History of Artificial Intelligence

Distribution profile and efficiency of the European pharmaceutical full-line wholesaling sector

Bureau of the Census, Total Revenue for Medical and Diagnostic Laboratories, Establishments Subject to Federal Income Tax [REV6215TAXABL144QNSA], retrieved from FRED

DENDRAL and Meta-DENDRAL roots of knowledge systems and expert system applications

Computer-Based Medical Consultations: MYCIN

Allocation of physician time in ambulatory practice: a time and motion study in 4 specialties

Robots in health care could lead to a doctorless hospital). The Conversation-Copyright © 2010-2017. The Conversation Trust (UK) Limited

The Fourth Revolution

IBM: Preparing for the future of artificial intelligence

Natural Language Understanding of Unstructured Data. IBM developerWorks®

Data Mining: Concepts Models and Techniques

An Introduction to Statistical Learning (With Application in R)

Statistical Machine Learning

Machine Learning Approaches to Bioinformatics

An essay towards solving a problem in the doctrine of chances

Neural Networks. A Comprehensive Foundation

Artificial intelligence in healthcare: past, present and future

Support vector machines and evolutionary algorithms doe classification. Single or Together?

Support Vector Machines and Perceptrons. Learning, Optimization, Classification, and Application to Social Networks

On the origin of species (1859)

Data Mining and Knowledge Discovery Handbook

Clustering: A Data Recovery Approach

Adaptive Control Processes

Curse of dimensionality

No free lunch theorems for optimization

The Statistical Evaluation of Medical Tests For Classification and Prediction

Receiver-operating characteristic (ROC) -plots: a fundamental evaluation tool in clinical medicine

The effect of two priors on Bayesian estimation of "Proper" binormal ROC curves from common and degenerate datasets

Sensitivity, specificity, receiver-operating characteristic (ROC) curves and likelihood ratios for electronic fetal heart rate monitoring using new evaluation techniques

A non-inferiority test of areas under two parametric ROC curves

The use of receiver operating characteristic curves in biomedical informatics

An introduction to ROC analysis

The meaning and use of the area under a receiver operating characteristic (ROC) curve

What is a disease?

Genomics and the continuum of cancer care

Deep patient: an unsupervised representation to predict the future of patients from electronic health records

Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning

Predicting changes in hypertension control using electronic health records from a chronic disease management program

Artificial intelligence estimation of carotid-demoral pulse wave velocity using carotid waveform

Prevalence of the congenital long-QT syndrome

Two-stage approache for risk estimation of fetal trisomy 21 and other aneuploidies using computational intelligence systems

Artificial intelligence for closed-loop ventilation therapy with hemodynamic control using the open lung concept

Characterization of glycosylphosphatidylinositol biosynthesis defects by clinical features flow cytometry, and automated image analysis

Digital medicine's march on chronic disease

Automated analysis of retinal images for detection of referable diabetic retinopathy

Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs

Diagnostic errors in the intensive care unit: a systematic review of autopsy studies

Artificial Intelligence for Health and Health Care

Opportunities and obstacles for deep learning in biology and medicine

Opportunities and obstacles for deep learning in biology and medicine

The parable of Google flu: traps in Big data analysis

Formative evaluation of participant experience with mobile eConsent in the app-mediated Parkinson mPower Study: a mixed methods study

Mobile-health: a review of current state in 2015

How many genes underlie the occurrence of common complex diseases in the population?

Genome-wide association studies for complex traits: consensus, uncertainty and challenges

Supervised autonomous robotic soft tissue surgery

Artificial intelligence and cognition

Artificial Intelligence and the Future of Defense: Strategic Implications for Small-and Medium-Sized Force Providers. The Hague Centre for Strategic Studies

Operations Management in Automotive Industries: From Industrial Strategies to Production Resources Management, Through the Industrialization Process and Supply Chain to Pursue Value Creation

Legal, regulatory and ethical frameworks or standards for AI and autonomous robotic surgery. The Int