key: cord-1013574-yw81pkrq authors: Huang, Chiou-Jye; Chen, Yung-Hsiang; Ma, Yuxuan; Kuo, Ping-Huan title: Multiple-Input Deep Convolutional Neural Network Model for COVID-19 Forecasting in China date: 2020-03-27 journal: nan DOI: 10.1101/2020.03.23.20041608 sha: a5c5aa8a715139536850864b516ca3b872b1e519 doc_id: 1013574 cord_uid: yw81pkrq COVID-19 is spreading all across the globe. Up until March 23, 2020, the confirmed cases in 173 countries and regions of the globe had surpassed 354,000, and more than 15,300 deaths had resulted. The confirmed cases outside of China had also reached over 81,000, with over 3,200 deaths. In this study, a Convolutional Neural Network (CNN) was proposed to analyze and predict the number of confirmed cases. Several cities with the most confirmed cases in China were the focus of this study, and a COVID-19 forecasting model, based on the CNN deep neural network method, was proposed. To compare the overall efficacies of different algorithms, the indicators of mean absolute error and root mean square error were applied in the experiment of this study. The experiment results indicated that compared with other deep learning methods, the CNN model proposed in this study has the greatest prediction efficacy. The feasibility and practicality of the model in predicting the cumulative number of COVID-19 confirmed cases were also verified in this study. CoV-2), and it has, at the time of writing, spread in over 173 countries, causing more than 15,000 deaths. The 31 epidemic is still spreading, and the WHO has classified the global risk level of the disease as a "pandemic." 32 Currently, the earliest known case symptoms appeared on December 1, 2019, and the first case sought treatment 33 on December 8. On December 26, 2019, Wuhan City Respiratory and Intensive care doctor Zhang Jixian first 34 discovered this pneumonia with unknown cause and suspected that it is an infectious disease. Subsequently, the 35 disease broke out in Wuhan City. On January 20, 2020, Chinese academic Zhong Nanshan publicly announced 36 that the novel coronavirus pneumonia "definitely transmits between people." On January 23, 2020, the Wuhan 37 City Government announced the adoption of lockdown and quarantine measures in the infected areas, which 38 was the first case of lockdown in a major city (with a population of 11 million) in recent public health history. Since January 13, the disease continued to spread to Thailand, Japan, and South Korea. On January 30, three 40 countries outside China were verified as having interpersonal propagation, and the pandemic was thus 41 designated as an international public health emergency event by the WHO. According to reports on the epidemic, "the propagation speed is faster, and the virus propagation power 43 has increased." Infected people can infect others with the virus while exhibiting no symptoms; and the latency 44 from infection to the presentation of symptoms is as long as 14 days. These characteristics also increase the 45 difficulty of controlling the epidemic. As the epidemic continues, there is also a problem of the global 46 undersupply of surgical masks. Currently, no vaccine and remedy for the novel coronavirus have been 47 discovered. WHO assistant director-general Bruce Aylward stated that Remdesivir is currently the only drug 48 . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 March 27, . . https://doi.org/10.1101 March 27, /2020 doi: medRxiv preprint 2 of 16 that is "considered to probably have real efficacy." The China-Japan Friendship Hospital in China and the 49 American National Institute of Allergy and Infectious Diseases are starting to conduct clinical trials for the drug. At present, no sufficient knowledge regarding the disease is known, and key factors, such as virus source, virus 51 birthplace, morbid mechanism, virus pathogenicity, and propagation power are still uncertain. The WHO stated 52 on March 3 that with the novel coronavirus epidemic, the world is in an "unknown state." 53 Up until March 23, 2020, 345,297 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The structure of this study proceeds as follows. The first section introduces the entire study. The second 135 section is a literature review. The following third section details this study's proposed deep neural network 136 algorithm. The fourth section presents and discusses the experiment results in detail. Finally, the fifth section 137 concludes the entire study and underscores the contribution of this study and the predicted trend of the COVID-138 19 epidemic. . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The structure principle of the fully connected layer can be expressed by Equation (2): In (2) is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The difference between the RNN and LSTM is that for LSTM, a cell state is added to store long-term states. In 193 the neural unit model structure in Fig. 4 (a) , the internal structure of LSTM can be divided into the input gate, 194 forget gate, and output gate. The principle of the LSTM input gate is expressed in the following formulae. . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The forget gate of the LSTM uses a sigmoid layer and a dot product to allow information to pass through 203 selectively. Equation (7) allows the LSTM to decide whether to forget the related information of the previous 204 cell, at a certain probability, in which is the weight matrix, and is the offset term. 206 The output gate of LSTM decides which states are required to be maintained by the input ℎ −1 and 207 according to Equations (8) and (9) The deep LSTM structure proposed in this study is illustrated in Fig. 4 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The principle of the GRU reset gate is described in Equation (11). The reset information is reserved 230 after ℎ −1 and are separately combined with the weight matrix. = ( ⋅ [ℎ −1 , ])(11) The input information of the current moment is maintained with the obtained at the weight 233 probability. The following Formula (12) is used to calculate the information of the current moment h . The structure of the proposed deep GRU is shown in Fig. 5 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 . . https://doi.org/10.1101 /2020 Algorithm 1 Convolutional Neural Network Input, target, and hyperparameter setting: Load the training set and testing set. Reshape the input into the corresponding three-dimensional matrix. 1. The first dimension is the sample index. 2. The second dimension is the time step. 3. The third dimension is the characteristic. Reshape the output into a two-dimensional matrix. 1. The first dimension is the sample target index. The established CNN model is used to predict the testing set. The real value is compared with the predicted value, and the prediction error is calculated. Data on confirmed cases of COVID-19 from January 23, 2020 to March 2, 2020, and from January 258 23, 2020 to March 2, 2020, were obtained from Surging News Network (a media outlet) [9] and WHO [31], respectively. In this experiment, the two evaluation indexes of the mean absolute error (MAE) and root 260 mean square error (RMSE) were used; their formulae are presented in Equations (14) and (15) Tables I, II, III, is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 . . https://doi.org/10.1101 /2020 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 . . https://doi.org/10.1101 /2020 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 March 27, . . https://doi.org/10.1101 March 27, /2020 is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint . CC-BY-NC-ND 4.0 International license It is made available under a perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint The copyright holder for this this version posted March 27, 2020 . . https://doi.org/10.1101 /2020 Risk Assessment of Novel Coronavirus 375 COVID-19 Outbreaks Outside China Optimization Method for Forecasting Confirmed Cases of COVID-19 in China Real-Time Estimation of the Risk of Death from Novel Coronavirus 380 (COVID-19) Infection: Inference Using Exported Cases Epidemic Spread of the 2019 Novel Coronavirus Driven by Spring Festival Transportation in China: A 383 Population-Based Study Clinical 385 characteristics and imaging manifestations of the 2019 novel coronavirus disease (COVID-19):A multi-386 center study in Wenzhou city Li Jin and Momiao Xiong Artificial Intelligence Forecasting of Covid-19 in Host and infectivity 390 prediction of Wuhan 2019 novel coronavirus using deep learning algorithm CRISPR-based 392 surveillance for COVID-19 using genomically-comprehensive machine learning design Time-varying transmission dynamics of Novel Coronavirus Pneumonia in China Breaking down of healthcare system: Mathematical modelling 399 for controlling the novel coronavirus (2019-nCoV) outbreak in Wuhan Characteristics of COVID-19 infection in 408 Clinical characteristics and intrauterine vertical transmission potential of 411 COVID-19 infection in nine pregnant women: a retrospective review of medical records. The Lancet Bearing Capacity Using Artificial Neural Network and Random Forest Mechanism Using Artificial Neural Network Based Customer Profiles in Smart Grid The Influence of the Activation Function in a Convolution Neural 419 CNN Method Based on DRNet and RoI Align for Building Detection in Remote Sensing Images Towards a Smart Smoking Cessation App: A Public Environment Emotion Prediction Model Using LSTM 426 Network Semi-Supervised Bidirectional Long Short-Term Memory and 428 Conditional Random Fields Model for Named-Entity Recognition Using Embeddings from Language 429 Hybrid Deep Learning Predictor for Smart Agriculture Sensing Based on Empirical Mode Decomposition and Gated Recurrent Unit Group 432 An Optimal Feature Parameter Set Based on Gated Recurrent Unit is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in (which was not certified by peer review) preprintThe copyright holder for this this version posted March 27, 2020 . . https://doi.org/10.1101 /2020