Deep learning-based methods for predicting COVID-19: A critical review

Abstract

The COVID-19 pandemic, caused by SARS-CoV-2 and first identified in Wuhan, China, disrupted human life worldwide. Numerous modeling studies have aimed to understand its dynamics. This systematic review analyzes and summarizes deep learning-based techniques used to predict COVID-19 cases. A research approach is outlined to perform a systematic literature review, focusing on formulating research questions, defining search criteria, and extracting relevant data. In total, 102 papers were considered after meeting the inclusion and exclusion criteria. Regarding scaling methods, two were reported, with normalization being the most widely used (39.22%). Eight hyperparameter selection processes were reported, with grid search (6.87%), Bayesian optimization (2.94%), and manual search (2.94%) being the most commonly applied. Regarding model validation strategies, split validation (91.2%) was the most commonly used. For COVID-19 prediction, findings revealed that long short-term memory networks were the most extensively used (48.31%). Other frequently used models included hybrid models, gated recurrent units, convolutional neural networks, recurrent neural networks, compartmental models enhanced with deep learning, multi-layer perceptrons, and deep neural networks. Regarding what was predicted, confirmed cases were the most frequently predicted (29.95%), followed by deaths, new cases, recoveries, epidemiological parameters, active cases, intensive care unit admissions, and hospitalizations. However, several papers failed to report valuable and anticipated information derived from deep learning models. Scaling methods were not reported in 57 articles (55.88%), and hyperparameter selection processes were only noted in 19.61% of the papers. Additionally, the software used was reported in only 29.41% of the studies. To ensure the reliability of COVID-19 predictions, researchers must explicitly document the conditions under which deep learning models perform optimally and identify biases that could affect predictions.