Explainable Risk Prediction on Streaming Data: a LSTMs application to SARS-CoV-2 Patients

The evolution of SARS-CoV-2 disease patients differs greatly depending on their demographic and clinical history. The sudden deterioration of some patients requires prompt early detection to reduce mortality. We demonstrate that a Long-Short Term Memory deep learning approach can predict effectively the risk of SARS-CoV-2 patients developing critical illness based on historical clinical data. In this study, we developed this model using a cohort of 2301 patients diagnosed with SARS-CoV-2 using RT-qPCR from February-May 2020. The patients are from several medical centers of Community of Madrid (Hospitales de Madrid del grupo HM Hospitales). The developed model achieves a f1-score of 92.29% and accuracy of 95.94%. In addition, we include model explainability by analyzing robustness to perturbations, feature importance using Shapley's additive exPlanning technique, and transparency metrics and visualizations to improve the reliability of the AI system and achieve early adoption by clinicians. To provide a clinical tool to the end user we also created an online application for early assessment of patients' risk of serious illness. This proposed rapid response methodology enables efficient allocation of healthcare resources and addresses some pandemic-related challenges.