Prediction of clinical outcome in patients with COVID-19 is important for timely intervention, appropriate ICU triage, and best utilization of hospital resources. We propose a deep learning framework to predict the patient outcomes by utilizing longitudinal clinical data from a tertiary care hospital in India. After quality control (QC) preprocessing and filtering, 126,716 diagnostic records with 13 key laboratory and vital sign variables were retained for analysis. We propose and test several neural architectures for the classification of ICU vs ward patients, namely GRU, LSTM, BiLSTM, CNN-LSTM, and Transformer models. A two-stage experimental setup was employed. We began by doing an exploratory train-test analysis for hyperparameter tuning. Next, we performed a rigorous comparative analysis based on the five-fold cross-validation results. We evaluated the performance according to such criteria as accuracy, sensitivity, specificity, recall, precision, F1-score, ROC curve and AUC. Of the models we tested, the Transformer performs best overall, with 91.9% accuracy, 85.4% sensitivity, 75.6% specificity, and 83.1% recall. It clearly demonstrates a good specificity in the view of high-risk patients, but with a minimum false negative rate. The CNN-LSTM had a comparably good performance and the GRU was an attractive low-cost alternative with reasonable predictive power and computational burden. The generalizability of our models was validated externally, where the Transformer also yielded the best performance on unseen clinical data. These results highlight the promise of deep learning, in particular Transformer-based models, as robust instruments for COVID-19 prognosis and ICU admission management.
K et al. (Sun,) studied this question.