Table 2.

Summary of COVID-19 outbreak (stage 2) related literature

Reference	Dataset	COVID19 Data	Time interval	AI/ML method	Performance	Relevance	Shortcoming
Randhawa et al. (2020)	National Center for Biotechnology Information (NCBI) database	text. Genome of 29,903 base pairs. 28 sequences of COVID-19 virus and the bat Betacoronavirus RaTG13	before Jan 23, 2020	Machine Learning with Digital Signal Processing (ML-DSP) approach, which uses 6 supervised learning approaches like Linear SVM and KNN, augmented by a decision tree approach to the machine learning component	Average ACC: 90.5 – 96.2	i) Confirm the taxonomy of the COVID-19, and possible bat origin; ii) alignment-free methodology adopted to rapidly analyze large datasets.	ML-DSP is a black-box method that does not offer a (biological) explanation for its output and is not able to assign a taxon that it has not been trained on.
Fong et al. (2020a)	Archive of Chinese health authorities	time series. 14 instances of suspected cases	Jan 21 – Feb 3, 2020	polynomial neural network with corrective feedback (PNN+cf)	RMSE: 136.55 RMSE Lin Regressor: 520.16 RMSE ARIMA: 1016.27	Data augmentation to the existing little data and fine-tuning the parameters of an individual forecasting model	Predicted result is very sensitive to the parameters used. Understand why algorithms incur very low or very high errors (for panel selection).
Fong et al. (2020c)	Chinese Center for Disease Control and Prevention	time series. 58010 of recent confirmed cases	Jan 25 – Feb 25, 2020	Broyden–Fletcher–Goldfarb–Shanno optimized polynomial neural network (BFGS-PNN), i.e. PNN enhanced with parameter optimization function.	RMSE: 62077.26 RMSE LinReg: 127693.55	i) BFGS algorithm to optimize the parameters and network structure size using alliteratively hill-climbing technique. ii) Estimate the direct cost that is needed as an urgent part of national budget planning to control the COVID-19 epidemic	Compare and contrast the differences of other techniques and refine input for accuracy.
Wang et al. (2020)	dataset of the radiology department of Huazhong University	images. Chest CT scans. 540 patients including 313 COVID-19 patients	Dec 13, 2019 – Feb 6, 2020	Weakly supervised deep learning framework (DeCoVNet): UNet (pre) & three stages 3D Deep Network & UNet	(ROC) AUC: 0.959	COVID-19 classification and lesion localization using 3D CT volumes	i) UNet model for lung segmentation did not utilize temporal information and it was trained using imperfect ground-truth masks ii) cross-center validations (more hospitals); iii) CT data of (CAP) not collected; iv) explainability
Kang et al. (2020)	Chinese CDC	images. Chest CT images. 2,522 patients including 1,495 COVID-19 patients	Jan 9 – Feb 14, 2020	FCNN: Structured Latent Multi-View Representation Learning	ACC: 95.5% Sens: 96.6%, Spec: 93.2%	Classify COVID-19 vs CAP. Use of multi-view representation learning with multiple features, such as texture, surface, volume histogram, and intensity.	i) Diagnosis with more classes instead of only two types of disease (i.e. COVID-19 and CAP). ii) Clinical characteristics of patients can be useful for diagnosis.
Xu et al. (2020)	2 China Hospitals	images. CT samples. 618 CT samples of which 219 came from 110 patients with COVID-19	Jan 19 – Feb 14, 2020	Residual network (ResNet)-18 by concatenating the location-attention mechanism in the full-connection layer to improve the overall accuracy	binary ACC : 86.7%	Multi-center case study. Location-attention classification model	Only compared the CT manifestation of COVID-19 with that of IAVP. To combine the patient’s contact history, travel history, first symptoms, and laboratory examination.
Mei et al. (2020)	18 medical centers in China	images. CT scan. 905 patients including 419 COVID-19 patients	Jan 17 – Mar 3, 2020	SVM, random forest and MLP classifiers	(ROC) AUC: 0.92 Sens: 84.3 Spec: 82.8	Compared performance to one fellowship-trained thoracic radiologist with 10 years of experience and one thoracic radiology fellow.	Explore various approaches, including 3D deep-learning models and develop the interpretability of CNN models. To validate the robustness of the models, is important to test the AI system in multiple centers.
Liang et al. (2020)	NHC of the People’s Republic of China	time series. baseline clinical features. 1,590 patients	before Jan, 2020	three-layer feedforward neural network & LASSO Cox model	C-index: 0.894 (ROC) AUC: 0.911	Deep Learning Survival Cox model had superior discriminating power compared with the classical Cox model, because it unravels the nonlinear relationships between complex clinical covariates and their hazards.	To extended deep learning model to integrate time-dependent covariates such as vital signs and high-dimensional features such a CT or X-ray images.
Wang et al. (2021a)	5 hospitals; most from hospitals in Wuhan, others from hospitals in Beijing.	multimedia CT images. 850 COVID-19 patients vs 541 non COVID-19	20 Feb 2020	i) classificaiton: ResNet-50, Inception networks, DPN-92, and Attention ResNet-50; ii) segmentation: fully convolutional networks (FCN-8s), U-Net, V-Net and 3D U-Net++.	best AUC: 0.991, with 3D Unet++ & ResNet-50	Experience in building and deploying an AI system	i) Does not perform well when there were multiple types of lesions, or with significant metal or motion artifacts, ii) The system is too dependent on fully annotated CT images.

The timings and data of the research papers analyzed are reported in the table. It can be observed that in this stage of the pandemic most of the works focus on the prediction of infection diffusion and early SARS-CoV-2 induced pneumonia diagnosis. Moreover, it is worth noticing that, since the first outbreaks occurred in China, a large number of the datasets used came from this country