Application of Machine Learning and Deep Learning in diagnosis Covid19

Predicting the progression of COVID-19 in hospitalised patients was an important challenge during the pandemic, particularly when healthcare resources were limited. This study investigates the use of machine learning (ML) and deep learning (DL) models to predict three clinical outcomes: death, admission to the intensive care unit (ICU), and clinical status (improvement or worsening). The study uses two patient datasets: an Algerian cohort of 80 patients from the Internal Medicine Department of the University Hospital Centre of Constantine, collected between July and August 2020, and a French cohort of 99 patients collected between March and May 2020. Five ML classifiers were evaluated: logistic regression (LR), k-nearest neighbours (KNN), classification and regression trees (CART), Naive Bayes (NB), and support vector machines (SVM). Their performance was examined in combination with three feature-selection methods: analysis of variance (ANOVA), mutual information (MI), and recursive feature elimination (RFE). The results were also compared with multilayer perceptron (MLP) models using one to five hidden layers. Model evaluation included 30 repetitions of five-fold cross-validation on the training data, followed by evaluation on a separate hold-out test set. The results showed that selecting 20 features generally provided better cross-validated accuracy than selecting 30 features or using the full feature set. The best-performing classifier-feature-selection combination varied across the six dataset-outcome tasks. LR combined with RFE was selected for three tasks, LR combined with ANOVA for one task, and CART combined with ANOVA for two tasks. The selected ML models achieved test-set accuracies ranging from 0.47 to 0.95 and higher accuracy than the average MLP results in five of the six tasks. These findings suggest that ML methods can be useful for exploring clinical outcomes in hospitalised COVID-19 patients. However, the small sample sizes and limitations of the available data mean that the results should be interpreted with caution. Further evaluation using larger and independent patient datasets is needed before these models can be considered for clinical use.