Explainable machine learning framework for dynamic monitoring of disease prognostic risk

Patients’ conditions continue to change after the diagnosis, with each patient showing a different time course. Here, we propose a dynamic prognostic risk assessment framework based on longitudinal data during hospitalization, using coronavirus disease (COVID-19) as an example. We extracted electronic medical records of 382 COVID-19 cases treated at Tokyo Shinagawa Hospital between 27 January and 30 September 2020. Gradient boosting decision trees were used to predict the maximum clinical deterioration, including deaths, from the data at initial diagnosis. Random survival forests were then used to calculate a 7-day cumulative hazard function to dynamically assess the risk of mortality of patients on each day during hospitalization. SurvSHAP(t) was applied to provide a time-dependent explanation of the contribution of each variable to the prediction. The prediction at initial diagnosis agreed well with the actual severity (area under the receiver operating characteristic curves = 0.717–0.970), but some cases showed discrepancies between actual and predicted prognosis. The dynamic mortality risk assessment during hospitalization could discriminate between deceased and surviving patients 1–2 weeks before the outcome. Early in hospitalization, C-reactive protein (CRP) was an important risk factor for mortality, while in the middle period peripheral oxygen saturation (SpO ₂ ) increased its importance and platelets and β-D-glucan were the main risk factors immediately before death. Dynamic risk assessment considering heterogeneous risk factors and time-to-event is useful for the early detection of patients who deteriorate rapidly after hospitalization. This framework provides healthcare professionals with the explainable real-time guidance for clinical decision-making during hospitalization.