PII: 0898-1221(87)90228-8 Comput. Math. Applic. Vol. 14, No. 9-12. pp. 793-802, 1987 009%4943/87 $3.00+0.00 Printed in Great Britain Pergamon Journals Ltd E X P E R T S Y S T E M S F O R C A N C E R C H E M O T H E R A P Y S. GAGLIO Institute of Electrical Engineering, University of Palermo, Italy M. GENOVESI, C. RUGGmRO a n d G . SPrNELLI Department of Communications, Computer and System Science, Genoa University, Italy C. NICOLINI Chair of Biophysics, School of Medicine, Genoa University, Italy G . BONADONNA a n d P. VALAGUSSA National Cancer Institute, Milan, Italy Abstract--In the present paper we describe BREASTCAN and NEWCHEM, two expert systems for the characterization of optimal adjuvant cancer therapies. The purpose of BREASTCAN is to support physicians in the postoperative breast cancer therapy, on the basis of currently used therapy protocols. It was developed in Prolog and positively validated, referring to the chemotherapies used by oncologists for some patients in the National Cancer Institute in Milan. NEWCHEM is a system oriented to the development of new cancer therapies, based on pharmaco-cell kinetic modeling and the newest molecular knowledge about neoplastic process. The system is being built and at first it will be validated by experiments on mice. Our aim with NEWCHEM is to extend our knowledge base and our rules to incorporate also all the most advanced knowledge at the molecular and cellular level, both theoretical and experimental, to make readily accessible to the health-community a system which will be really as expert as the present state of the art allows. 1. I N T R O D U C T I O N T h e use o f A I (artificial intelligence) techniques, with its logical-deductive m e c h a n i s m s , t h a t r e p r o - duce the h u m a n r e a s o n i n g w i t h o u t the restriction o f h u m a n capabilities in m e m o r y a n d association, is p a r t i c u l a r l y suitable f o r t h o s e medical p r o b l e m s , in which y o u m u s t r e g a r d a g r e a t n u m b e r o f p r o g n o s e s a n d their effects a n d interactions. This is the case o f c a n c e r c h e m o t h e r a p y , used f o r 10 y e a r s with e n c o u r a g i n g results. I n the p r e s e n t p a p e r we describe B R E A S T C A N a n d N E W C H E M , t w o e x p e r t systems f o r the c h a r a c t e r i z a t i o n o f o p t i m a l a d j u v a n t c a n c e r therapies. T h e y are b o t h i m p l e m e n t e d o n a V A X 11/750 a t the D e p a r t m e n t o f C o m m u n i c a t i o n s , C o m p u t e r a n d System Science o f the U n i v e r s i t y o f G e n o a , with the c o - o p e r a t i o n o f the C h a i r o f Biophysics o f School o f Medicine o f the G e n o a U n i v e r s i t y a n d the N a t i o n a l C a n c e r I n s t i t u t e in Milan. T h e p u r p o s e o f B R E A S T C A N is t o s u p p o r t physicians in the p o s t o p e r a t i v e b r e a s t c a n c e r t h e r a p y , o n the basis o f c u r r e n t l y used t h e r a p y p r o t o c o l s . I t was d e v e l o p e d in P r o l o g a n d positively validated, referring t o the c h e m o t h e r a p i e s used b y oncologists f o r s o m e p a t i e n t s in the N a t i o n a l C a n c e r I n s t i t u t e in Milan. I n B R E A S T C A N the k n o w l e d g e is represented using production rules, t h a t are s t r u c t u r e s w h o s e f o r m is I F . . . T H E N , and frames, c o m p l e x d a t a structures f o r m o d e l i n g s t e r e o t y p e d situations. E v e r y f r a m e describes a c e r t a i n s i t u a t i o n o r hypothesis. N E W C H E M is a s y s t e m o r i e n t e d t o the d e v e l o p m e n t o f new c a n c e r therapies, b a s e d o n p h a r m a c o - c e l l kinetic m o d e l i n g a n d the newest m o l e c u l a r k n o w l e d g e a b o u t neoplastic process. T h e s y s t e m is being built a n d a t first it will b e v a l i d a t e d b y e x p e r i m e n t s o n mice. I t is c o n s t i t u t e d b y a set o f frames, a deduction rule system, a n agenda-based c o n t r o l system, a n hierarchical planning system, a d a t a b a s e a b o u t drugs, a qualitative reasoning system, m a t h e m a t i c a l m o d e l s f o r s i m u l a t i o n a n d o p t i m a l c o n t r o l o f p h a m a c o - e n z y m e ( M i c h a e l i s - M e n t e n ) a n d p h a r m a c o - c e l l processes. 2. B R E A S T C A N : A N E X P E R T S Y S T E M F O R P O S T O P E R A T I V E B R E A S T C A N C E R T H E R A P Y C h e m o t h e r a p y as a d j u v a n t t r e a t m e n t in p a t i e n t s with o p e r a b l e b r e a s t c a n c e r a n d histologically positive axillary n o d e s h a s b e e n used f o r a b o u t 10 years, achieving results t h a t s u p p o r t the use o f a d j u v a n t c h e m o t h e r a p y in clinical practice [1-3], 793 794 s. GAGLIO et al. Numerous clinical trials have been designed and activated, employing adjuvant chemotherapy alone or combined with endocrine or immunotherapy. Both study design and results o f adjuvant protocols are influenced by a great number o f prognostic variables (e.g. subgroups o f patients with different size o f primary tumor (T) and extent o f axillary node (N) involvement, presence or absence o f hormone receptors, quantities o f drugs to be administered and intervals between each drug and treatment cycles, combinations o f drugs, optimal treatment duration) whose significance and possible interactions need to be assessed, as well as by a great variety o f clinical problems often originating from chemotherapy during the postoperative course o f breast cancer (e.g, treatment discontinuation because o f toxicity, continuous low drug dosage, different forms o f salvage regimen, occurrence o f general medical problems not related to adjuvant therapy). It is therefore felt that AI techniques [4] could make valuable contributions in this field. In particular it would be worth while to set up an expert system to assist physicians in the selection and performance o f optimal treatment--maximizing therapeutic effectiveness and limiting toxicity. The use o f the expert system should facilitate an optimal decision making at each treatment cycle by taking into account all foreseeable medical variables. An expert system for oncology protocol management, focused on the treatment o f Hodgkin's Disease and the non-Hodgkin's lymphomas, has been recently set up [5]. The present paper describes an expert system for adjuvant breast cancer chemotherapy--based on few widely accepted clinical protocols [2]--which has been implemented to assist clinicians treating patients by standard chemotherapy. 2.1. Outline o f B R E A S T C A N B R E A S T C A N is written in Prolog [6]--a programming language that has been used in many areas o f AI research--that has been chosen because it allows to quickly write clear, concise and readable programs, and provides an efficient built-in deduction system. B R E A S T C A N is written employing production rules--that is general statements about objects and their relationships--and frames--that is data structures that model stereotype situations. Production rules encode single pieces o f expert knowledge and are written as Prolog clauses. They specify particular solutions to a given problem according to given conditions. As an example, the following B R E A S T C A N rule rule(4,3,ctx,2,Time),300,Atruth,10):-- f (look__for (platelets,2,Time, Platelets,Truth 1 )), Platelets < 1 0 0 0 0 0 , f (look__for (leukocytes,2,Time, Leukocytes,Truth2)), Leukocytes < 3800, f (Iook__for(n__delays,2,Time, Ndelays, 10) ), Ndelays = 2, min(Truthl ,Truth2,Atruth). establishes that the suggested dose o f Ctx (cyclophosphamide) is 300 mg/m 2 if platelets count is less than 100,000, leukocytes count is less than 3800, and two consecutive delays have already occurred in drug administration. A rule is characterized by a truth value which is combined with truth values resulting from the satisfaction o f antecedent conditions to provide a measure o f evidence o f the conclusion. Production rules have been grouped, in BREASTCAN, into frames. Each frame corresponds to a given situation or hypothesis. One frame (named "initial") contain the rules establishing whether the patient is eligible for chemotherapy Or not; another frame (named "registration") manages the description o f the patient pathology prior to chemotherapy, to be used for adjuvent therapy choice and to be recorded for subsequent statistical studies that the user may want to carry out, and recommends one treatment protocol, each treatment protocol specifying one chemotherapy cycle. The user can then select one treatment protocol (accepting the system recommendation or not). At every stage the user can obtain explanations from the system, that can describe the pathway followed to reach any stage o f consultation. Frames are organized in a tree-like structure as shown in Fig. 1. Expert systems for cancer c h e m o t h e r a p y 795 + . . . . . . . . . . . . I N I T I A L [ FRAME DESCRIPTOR + . . . . . . . . . . . . REGISTRATION [ SUBFRAME POINTERS + . . . . . . . . . . . . FLOW-SHEET I S T E P S J + . . . . . T I l I R U L E S + . . . . . . . . . . . . PRIMARY TREATMENT - - + . . . . . T2 l I DEFAULT RULES I + ..... Tn I TRIGGERS + . . . . . . . . . . . . FOLLOW-UP I COMPLETION ACTIONS l + ..... FTI I TREATMENT I U T I L I T I E S + . . . . . . . . . . . . UPON . . . . . . . . . . + . . . . . FT2 FAILURE I ITEM DESCRIPTORS + . . . . . FTn Fig. 1 Fig. 2 The treatment protocols are subframes of one frame (named "treatment") that takes into account the stage o f the treatment, the possible presence of pathologies developed during it, and suggests (activating one of the subframes) the drugs to be administered and their quantities, and the date of the next drug administration. 2.2. Control structure The control system activates and executes frames. A frame can be in one of the following states: --inactive; --active (that is ready for execution); - - c u r r e n t (in execution); --completed. The control system provides a number of procedures not available directly in Prolog, designed to assist consultation. Frames are selected for instantiation and execution according to an "applicability value", that evalutes the agreement between the specific situation considered and the situation to which the frame relates; this is often accomplished comparing, for each frame, the postsurgical conditions characterizing the frame with the patient conditions. 2.3. Frame structure and execution Frames consist of a sequence of steps, rules and facts. Steps consist of groups of statements that are executed sequentially, and steps are executed sequentially. Some steps are conditioned, that is their execution depends on a condition to be tested: if the condition is not true, the step is not executed and the control passes to the following step. The steps of a frame encode the procedural knowledge ("what to do") of the stereotype situation modelled by the frame. The rules of a frame encode the declarative knowledge (what is true and "how to do" things within the given situation). The facts are pairs item--value that encode what is known about particular situations (measures and information inferred by the system. Each value is provided with a descriptor containing the way the value has been determined (asked, deduced, default) and certainty factor, which expresses the extent of certainty associated with the value. Certainty factors are propagated and modified in the course of rule application as described in Buchanan and Shortliffe [7]. The structure o f a frame is shown in Fig. 2. The frame descriptor contains some general information about the frame, as its name, the number of steps, its possibility, and the condition for activation. Subframe pointers indicate the subframes of the frame. Triggers are forward rules that perform particular actions when critical values of the parameters are added to the set of items. Completion actions are performed when the frame enters the completed status, for instance the activation of another frame. 796 S. GAGLIO et al. Utilities are auxiliary procedures used within the steps. Item descriptors encode the constraints to be satisfied by the values of the items, the questions to be asked o f the user concerning the items, and take care of data integrity (for instance, menopause age cannot be greater than the current age of the patient). When, during the execution of a frame, the value of an item is needed, the system first consults its data base (the pairs item--value in the frames); if it does not find it, it tries to apply the rules of the frame to deduce it; finally, if it fails, it asks the user the value; when the answer is not even provided by the user, the system uses default rules and values. 2.4. Interfacing with the user BREASTCAN offers the possibility to obtain indications about the line of reasoning that leads backwards from the current answer by the system. This capability is useful in many respects: realizing how the system comes to a conclusion enables users to make the most o f the consultative advice and also helps users in difficult, "non-standard" situations, in which they may wish to violate a rule; moreover, it makes it easier for the user to make changes in the program. When starting consultation about a patient for the first time, the user is asked questions that enable the system to decide whether chemotherapy is appropriate for that patient. If the answer is no the system gives indications about the reasons of this conclusion and some advice about that particular case. At the end of this stage the user can choose whether he wishes to carry on or not (irrespective of the conclusions of the system about the eligibility of the patient). The following stage, relating to the description o f the patient pathology prior to chemotherapy, consists of six parts; at the end of each of them the data entered by the user in that section are displayed and the user is enabled to change values that he may have entered by mistake. The system now suggests one or more treatments and the user can choose which treatment he wishes to perform (following the suggestions by the system or not). 2,5. Assessment and further developments BREASTCAN has been proven capable to closely reproduce the actual decision taken over the period of 6 months by medical oncologists at the National Cancer Institute of Milan, Italy, in the treatment of a few patients, taken as test-cases in a preliminary retrospective confirmation of the system validity. Our next step is then to extend this study to a large number o f patients and to incorporate into the decision-making process all other possible strategies suggested by other alternative treatment protocols from worldwide clinical trials on the breast cancer. 3. N E W C H E M : AN EXPERT SYSTEM FOR THE T R E A T M E N T OF D I S S E M I N A T E D C A N C E R Since early 1974 a comprehensive program was developed at the Biophysics Division o f Temple University in Philadelphia (U.S.A.), aiming to develop a rational basis for the chemotherapy of cancer [8]. This approach resulted in a widely interdisciplinary effort which, by a constant feedback between experimentation on animals and theoretical simulations, has been able to predict and explain multiple drug-action and interaction at molecular [9] and cellular [10] level for a variety o f normal and cancer tissues. Through the aid o f optimal control theory [11] such a complex modeling can furthermore suggest the optimal treatment, with the sequence of rest-periods and drug administration at the proper timing and dosage. Recent preliminary results have indeed been quite encouraging on the treatment of lung metastates from melanoma B-16 in mice [12]. There appears to be however a basic premise for the successful utilization of the sophisticated pharmaco-enzyme and pharmaco-cell kinetic modeling, namely the detailed molecular knowledge of the drug-tissue metabolism and o f the physico-chemical properties o f both normal and cancer cell in the various functional state (cycling, non-cycling and in varying stage o f differentiation). Fortunately, modern biophysical techniques [13] permit such a characterization at single cell level with high accuracy and frequently on real-time. The transfer of such a complex and multifold approach to routing medical practice appears however nearly prohibitive, for its highly analytical Expert systems for cancer chemotherapy 797 ÷ . . . . . . . . . . . . . ÷ I AGENDA I ÷ . . . . . . . . . ÷ ÷ . . . . . . . . . . . . . ÷ + . . . . . . . . . . . ÷ I l I I I I I RULES I I TASK - I I I PLANNER I I I i T i i + . . . . . . . . . + I TASK - 2 I + . . . . . . . . . . . * I i . . . I I i . . . . . . . . . . . . . . . I I I . . . . . . . . . v I v v ÷ . . . . . . . . . . . . ÷ ÷ . . . . . . . . . . . . . . . . . ÷ + . . . . . . . . . . . . . ÷ I I < - - I I t QUALITATIVE I I DEDUCTION t TASK I I I • ->I I l<--I REASONING l I SYSTEM ) PROCESSOR I I I I I--> I - I SYSTEM I ÷ . . . . . . . . . . . . + . . . . . . . . . . . . . . . . . . I + . . . . . . . . . . . . . + ÷ . . . . . . . ÷ I I v ÷ . . . . . . . . . . + + . . . . . . . . . . . ÷ + . . . . . . . . . . . + I I I I I I I TASK I I DRUGS l I USER I I I I i I I I RULES I I DATA BASE I I INTERFACE I I I I I I I ÷ . . . . . . . . . . + + . . . . . . . . . . . + + . . . . . . . . . . . + + . . . . . . . . . . + + . . . . . . . . . . . . . . . . . + I I I I +-->1 FRAMES i < . . . . + 1 MATHEMATICAL I I I I I + . . . . . . . . . . + I MODELS I I I ÷ . . . . . . . . . . . . . . . . . ÷ Fig. 3 nature and for the large number of parameters involved (constantly expanding due to the progress in cell and molecular biology). We have then recently considered to use AI to bridge such a gap, by developing an ad hoc expert system named N E W C H E M to guide our present animal experimentation but with the aim to improve in the near future human cancer treatment. 3. I. The structure o f N E W C H E M The structure of N E W C H E M is shown in the diagram of Fig. 3. It is based on many paradigms of AI, since it puts together into one integrated system, a set of components, which are: a production rule-based system, a frame stucture, an agenda-based control, a planner, a data base about drugs, a qualitative reasoning system, a blackboard system. All those components cooperate to produce a plan (i.e. a sequence of actions) to be carried out by the user to achieve a given goal. We adopt such a composite structure, because we believe that intelligent behaviour and accurate answers can be only obtained using specific techniques for the different kinds of knowledge and strategies to be embodied by the system. This is in agreement with current trends in expert system research. For instance, more recent systems like C E N T A U R [14], ONCOCIN [5] and ABEL [15] use multiple representations of knowledge. Furthermore, N E W C H E M uses mathematical models for pharmaco-cell kinetics [12], pharmaco-enzyme kinetics [9], and optimal control theory [11], written in F O R T R A N , which are queried when the system needs precise answers concerning simulation and optimization of parameters at various levels of the decision-making process. The system is written in Franz Lisp and uses the PEARL package [15] for efficient storage and retrieval, and the FLAVORS package for object oriented programming. The interaction between N E W C H E M and the experimenter is as follows: (a) The system asks the experimenter to provide some data describing the status of a subject to be treated. (b) The system produces an optimal treatment plan together with a set of justifica- tions for its choices and the expected results. 798 s. GAGLIO et aL (c) The experimenter can propose changes in the plan and ask about possible consequences. (d) A new treatment plan is proposed whenever the results are different from the expected ones. In the following a short description of each component of the system is given. 3.2. Frames Frames in N E W C H E M are associated with different points of view of the problem. They collect specific data and knowledge about inferences and actions concerning the specific domain. Data are related to different aspects of the case being examined. Frames are associated respectively with primary tumor, metastates, different tissues and organs (like bone marrow and small intestine). Different frames are also associated with discriptions respectively at the cellular level and the molecular level. For each aspect, a frame acts as a blackboard in which all data are collected together with an indication about how they have been obtained and a measure o f uncertainty. Consistency checking among data is also performed within the frame. Frames also contain action blocks, which are sequences of tasks to be performed to fill the data entries and for different operations of the whole system. Tasks may refer, for instance, to the activation of other frames, to queries to the declarative knowledge base and to the mathematical model, and to planning activity. A particular action block is the activation block, to be executed at the activation o f the frame. The way a task is executed is based on a set of task rules, that for each task select the appropriate operations, depending on the data known so far (the known states o f the subject). As an example, in the frame "treatment", a rule for the task "select a treatment plan" is the following: (Task-rule (Task "select a treatment plan") (Context treatment) (Condition (isknown general__strategy)) (Conclude~action '(progn (enter__date) (enter__new__data) (deduce general__strategy) (ask__user) (deduce strategy) (ask__user) (deduce treatment) (ask__user) (find plan) (ask_.user)))) Some data within a frame can be inferred from other data by means o f backward production rules. For instance, the rule (Rule (Ruleno 45) (Context treatment) (Condition (and ( > =subject performance~tatus 50) (for ?organ in (kidney, heart, liver, lungs) (same ?organ sufficiency satisfactory)) (same__strategy kill__tumor)) ) (ConcludLvalue_treatment_type drug__combination)) gives treatment__type the value drug__combination if the performance status of the subject is > = 50, the sufficiency o f kidney, heart, liver and lungs is satisfactory and the strategy is to kill the tumor. Expert systems for cancer chemotherapy 799 W i t h an item (or d a t a entry) in a frame, o n e can associate a " t r i g g e r " , which p r o m o t e s a p p r o p r i a t e actions when a critical value for the item is obtained. Items can have multiple values, d e t e r m i n e d by m ean s o f different rules. Different values have, however, associated different scores representing their uncertainty. At a given time a task can be active or inactive. W h e n a fram e enters the active status, the set o f tasks, m a k i n g its activation block, is loaded in the ag en d a o f the system, waiting to be executed. G r o u p s o f frames are organized in a tree-like fashion w h en ev er some o f t h em can be considered m o r e specific points o f view. In such a case a hierarchical relation " s u b f r a m e " holds between pairs o f frames. 3.3. The agenda and the task processor T h e c o n t r o l o f the system uses an agenda o f waiting tasks a n d a task processor. T h e tasks in the agenda belong to active frames. F o r each o f them the following d a t a are maintained: • T h e n a m e o f the task. • T h e frame to which it belongs. • A priority value. • A justification. At the end o f the execution o f a task, a new task is selected acco rd i n g to its priority. T h e task processor is a p r o b l e m solver that behaves in the following ways: • F o r each task, it searches the associated task rules (within the fram e the tasks belong to) to find a set o f possible actions. Each action is a L I S P co d e to which it is assigned a score, expressing h o w m u c h it is a p p r o p r i a t e fo r the c u r r e n t situation. • It selects the action with the highest score • It executes the selected action. Actions m a y require to insert new tasks in the agenda. F o r each o f t h e m a p ri o ri t y value is c o m p u t e d and the c u r r e n t task is inserted in the list o f its justifications. 3.4. The deduction system T h e d e d u c t i o n system is invoked whenever in the system a value fo r an item is required. T h e d e d u c t i o n system first explores the frame to which the item belongs t o see if a value is already k n o w n . I f this is n o t the case, it looks f o r rules to deduce it. Rules are first searched fo r in the c u r r e n t f r a m e (the f r a m e o f the c u r r e n t task), then in the frames which are ancestors o f the c u r r e n t one in the subframe hierarchy, and, finally, in the fram e to which the item belongs. I f n o rule is f o u n d , the value is asked o f the user or a default value, if specified, is used. T h e d e d u c t i o n system provides for an item, which is an attribute associated with an object, a list o f P E A R L structures o f the following kind: (Item (Object lisp) (Attribute lisp) (Context lisp) (Time lisp) (Value lisp) (Uncertainty integer) ) E a c h structure expresses a value for the item at a given time with an associated u n c e r t a i n t y value. 3.5. Planner T h e task o f the p l a n n e r is to p r o d u c e a plan fo r the t r e a t m e n t o f the subject, i.e. a partially o r d e r e d set o f actions to be p e r f o r m e d by the experimenter. T h e o rd eri n g o f the actions refers to a t e m p o r a l sequence, in the sense that some actions m u st be p e r f o r m e d b efo re others. T h e p l a n n e r is i n v o k e d by the task executor an d receives as i n p u t an initial p l an t o refine. Refinement is p e r f o r m e d by substituting a given action in the set with a n o t h e r partially o r d e r e d 800 S. OAOLIO e t al. ;Task_plan RS : ;IF the type of tumor is metastatized, ; its progress istanzla reduced and ; the state of the treatment istanzia in progress ;THEN change only one drug (ci Task_plan R3 (Name First_strategy) (Number 3) (Context General_strategy) (Type father) (Tax_father nil) (Tax_child nil) (Condition (prog (Date Listglob) (setq Date (daymonthyear)) (setq Listglob (list 'cell_descr 'cell_descr 0)) (return (fu_and (same 'type 'tumor 'metastatized Date Listglob) (same 'progress 'tumor 'reduced Date Listglob) (same 'status 'treatment 'in_progress Date Listglob))))) (Action (conclude 'First_strategy 'change_only_drug 9.2)) (Description (prog () (printterpr " I find that the TYPE of TUMOR is METASTATIZED (CERTAINTY FACTOR : ~f) ,, (same 'type 'tumor 'metastatized (daymonthyear) (list 'cell_descr 'cell_descr 0)) 5 I) (printterpr " the PROGRESS of TUMOR is REDUCED (CERTAINTY FACTOR : %f) ,, (same 'progress 'tumor 'reduced (daymonthyear) (list 'cell_descr 'cell_descr e)) fi 2) (printterpr " and the STATUS of TREATMENT is IN PROGRESS (CERTAINTY FACTOR : ~f) . (same 'status 'treatment 'in_progress (daymonthyear) (list 'cell_descr 'cell_descr 0)) 5 2) (cprinttab " THEN I recommend to CHANGE ONLY ONE DRUG " nil 5)))) (insertdb R3) Fig. 4 set o f more specific actions. This is accomplished using refinement rules, which are context- dependent production rules, that replace a given action in their left part with the set o f actions specified in their fight part. Consistency tests are also performed on the overall plan to resolve interactions and conflicts. This planning is similar to the hierarchical planning performed by N O A H [17]. The actions specified in a plan refer to different kinds o f intervention on the subject, mainly surgery, radiology and administration o f drugs. Some constraints, describing the particular modalities o f the intervention (for instance, the properties o f the drugs to be administered) are associated with these actions. These constraints are used to select, at the end o f the planning activity, the most appropriate intervention through a query to the declarative knowledge base--for instance, the specific drug needed. This selection is made only when the plan is completely refined, in order to avoid an early choice that can compromise the whole plan. This is what is called a least commitment strategy. Constraints can also be propagated within the plan as in M O L G E N [18]. The plan is associated as a value with the item plan in the frame treatment. An example o f a rule o f the Planner is shown in the Fig. 4. 3.6. The data base about drugs The data base about drugs o f the system contains data concerning all drugs that are to be used during the treatment considered by the system and their properties. Drugs are represented as structured types o f LISP objects, stored in an internal form as blocks o f memory, regarding logical groupings o f betorogeneous data as slots and slot fillers. One main structure defines the way in which all drugs are characterized, while each drug is represented as an individual instance o f this structure. The main structure can be regarded as a reference model o f a drug in which the greatest number o f attributes and properties are available, Expert systems for cancer chemotherapy 801 and each individual drug is represented filling up all known attributes o f the main structure and specifying properties to the greatest possible detail. This knowledge representation is implemented in P E A R L an AI language allowing to create hierarchically defined slot filler representations and to handle them efficiently, by easy inserting and fetching. P E A R L is implemented in LISP, so the knowledge base can be directly addressed within a LISP system as the one described in the previous sections. Specific rules contained in the system directly address attributes o f the drugs represented in P E A R L in the declarative data base. The main structure adopted is implemented in P E A R L as follows: (drug (name symbol) (type symbol) (main___attribute symbol) (class symbol) (toxicity symbol) (action struct)) According to this main structure, each specific drug considered by the system is represented as follows: (drug (name cyclophosphamide) (type alchilant) (main__attribute polifunctional) (action (type inhibits) (object mitosis))) 3. 7. The qualitative reasoning system The qualitative reasoning system uses qualitative process theory, as developed by Forbus [19], to reason about processes at a cellular level. This component is used within the whole system to perform qualitative simulations in order to "roughly" predict the consequence o f specific actions. In qualitative process theory physical parameters are characterized by a quantity space in which only intervals o f values are considered. Usually, what interests is if a quantity is positive, negative or null, and if it is increasing, decreasing or stationary. Furthermore, physical situations and processes are described symbolically, in a language based on predicate calculus. F o r a given process, the objects involved, the conditions in which the process takes place, the relations holding among the parameters, and the causes o f change (influences) are specified. The qualitative reasoning sytem is implemented in an object oriented style o f programming which is provided by the "flavors package" o f Franz Lisp. As an example the inhibition o f D N A replication by a substance can be described as follows: (make_instance 'process ':individuals '(DNA cell ?S) ': preconditions '(and (substance ?S) (in hibitor__D N/k~synthesis ?S)) ': influences '(I- (Am (synthesis DNA)) (Am (concentration ?S)))) The set o f possible sequences o f processes that follow a given action on a cell or on a cell population (for instance, an increase in concentration o f a substance) is the result o f a qualitative simulation. This result is useful during the planning process in order t o take care o f the possible consequences o f the actions that have been selected. 802 S. GAGLIO et al. 3.8. T h e m a t h e m a t i c a l m o d e l T h e m o d e l i n g c o n c e r n s v a r i o u s levels o f d e c i s i o n - m a k i n g a n d a t t e m p t s t o p r o v i d e a n i n t e g r a t e d a n d realistic a p p r o a c h t o o p t i m a l c a n c e r t r e a t m e n t . E a c h a s p e c t o f this m o d e l i n g h a s b e e n extensively d e a l t in p r e v i o u s p u b l i c a t i o n [5, 8, 9 - 1 1 , 18], a n d is b a s e d o n n u m e r o u s b i o p h y s i c a l p a r a m e t e r s , w h i c h a r e r e v i e w e d in details in a f o r t h c o m i n g b o o k [13]. E a c h m o d e l a c t s s y n e r g i s t i c a l l y w i t h i n a p r o p e r m u l t i p l e time scale, n a m e l y : - - a t t h e level o f s - m i n - h , a m o d e l ( D N A M E T ) f o r D N A e n z y m a t i c s y n t h e s i s d e t e r m i n e s m u l t i p l e D N A a n t i m e t a b o l i t e s i n t e r a c t i o n t o i n d u c e e i t h e r differential s y n c h r o n y o r differential killing b e t w e e n n o r m a l a n d c a n c e r tissues; a t t h e level o f h - d a y s , a cell kinetics m o d e l d e t e r m i n e s d r u g - t i s s u e s t o s u g g e s t ( S I V F I T ) o p t i m a l s h o r t t e r m s d r u g p r o t o c o l s ( t y p e , d o s a g e a n d timing); - - a t t h e level o f w e e k s - m o n t h s , a n o p t i m a l c o n t r o l - b a s e d m o d e l ( S I V F I T 2) p r e d i c t s l o n g - t e r m t i m i n g o f t r e a t m e n t , w h e r e rest p e r i o d s a n d c h a n g e s in t h e t y p e o f d r u g s a r e d e t e r m i n e d t a k i n g i n t o c o n s i d e r a t i o n l o n g - t e r m effects like d r u g resistance, d u e t o gene a m p l i f i c a t i o n a n d / o r cell m u t a t i o n . 4. C O N C L U S I O N T w o e x p e r t s y s t e m s in O n c o l o g y h a v e b e e n d e s c r i b e d . T h e first o f t h e m , B R E A S T C A N , w h i c h f o l l o w s a t r a d i t i o n a l a p p r o a c h , is b a s e d o n l y o n t h e t r a d i t i o n a l i n p u t d e r i v e d f r o m clinical e x p e r i e n c e w i t h e m p i r i c a l trials, w h i c h a r e c a r r i e d o u t o n r a n d o m i z e d p a t i e n t s w i t h the p r e v a i l i n g c h e m o t h e r a p e u t i c p r o t o c o l s . O u r a i m w i t h t h e s e c o n d o n e , N E W C H E M is i n s t e a d t o e x t e n d o u r k n o w l e d g e b a s e a n d o u r rules t o i n c o r p o r a t e a l s o all t h e m o s t a d v a n c e d k n o w l e d g e a t t h e m o l e c u l a r a n d c e l l u l a r level, b o t h t h e o r e t i c a l a n d e x p e r i m e n t a l , t o m a k e r e a d i l y accessible t o the h e a l t h - c o m m u n i t y a s y s t e m w h i c h will be really as e x p e r t as the p r e s e n t state o f t h e a r t allows. R E F E R E N C E S 1. G. Bonadonna and P. Valagussa, Chemotherapy of breast cancer: current views and results. Int. J. Radiat. Oncol. Biol. Phys. 9, 279-297 (1983). 2. S. E. Jones and S. E. Salmon, Adjuvant systemic therapy o f cancer IV. Grune & Stratton, New York (1984). 3. G. Bonadonna and P. Valagussa, Adjuvant systemic therapy for resectable breast cancer. J. Clin. Oncol. (in press). 4. J. C. Kunz, E. H. Shortliffe, B. G. Buchanan and E. A. Feigenbaum, Comparison of techniques of computer-assisted decision-making in medicine. Modeling and Analysis in Biomedicine, pp. 335-367. World Scientific, Singapore (1984). 5. E. H. Shortliffe, A. C. Scott, M. B. Bischoff, A. B. Campbell, W. Van Melle and C. D. Jacobs, ONCOCIN: An expert system for oncology protocol management. Proc. 7th Int. Joint Conf. on Artificial Intelligence, Vancouver, B.C., pp. 876-881 (1981). 6. W. F. Clocksin and C. S. Mellish, Programming in Prolog. Springer, New York (1981). 7. B. G. Buchanan and E. H. Shortliffe (Eds), Rule-Based Expert Systems: The MYC1N Experiments o f the Stanford Heuristic Programming Project. Addison-Wesley, Reading, Mass (1984). 8. C. Nicolini, The principles and methods of cell synchronization in cancer chemotherapy. Biophys. Biochem. Acta 458, 243-282 (1976). 9. C. Nicolini, A. Belmont, M. Grattarola, C. Moore and E. Milgram, Pharmaco-enzyme kinetics simulations of experimental interactions among multiple antineopolastic drugs. Biochem. Pharm. 28, 2891-2908 (1979). 10. S. Zietz, C. Desaive, S. Lessin and C. Nicolini, Modeling to determine dose dependence of drug and cell kinetic parameters. Comput. Biomed. Res. 13, 297-305 (1980). 11. S. Zietz and C. Nicolini, Mathematical approaches to optimization of cancer chemotherapy. Bull. Math. Biol. 41, 305-324 (1979). 12. C. Nicolini, S. Zietz, M. Tansy and S. Lessin, Modeling and analysis in cancer treatment. Modeling and Analysis in Biomedicine, (Ed. C. Nicolini) pp. 291-333. World Scientific, Singapore. 13. C. Nicolini, Biophysics and Cancer. Plenum, New York (1985). 14. J. S. Aikins, Prototipical knowledge for expert systems. Artificial Intelligence 20, 163-210 (1983). 15. R. S. Patil, P. Szolovits and W. B. Schwartz, Modeling knowledge of the patient in acid-base and electrolyte disorders. Artificial Intelligence in Medicine (Ed. P. Szolovits) pp. 191-226. Westview Press, Boulder, Co. (1982). 16. M. Deering, J. Faletti and R. Wilensky, Using the PEARL AI package. Technical Report Department of EECS, University of California, Berkeley (1982). 17. E. D. Sacerdoti, NOAH: Planning in a hierarchy o f abstraction spaces. Artificial Intelligence 5, 115-135 (1974). 18. M. Stefik, Planning with constraints. Artificial Intelligence 16, 111-140 (1981). 19. K. D. Forbus, Qualitative process theory. Artificial Intelligence 24, 85-168 (1984).