Explainable Machine Learning Framework for Dynamic Monitoring of Disease Prognostic Risk: Retrospective Cohort Study

Background: Patients' clinical status often evolves rapidly after an initial diagnosis, with each patient exhibiting a distinct disease trajectory. As a result, static risk scores fall short in supporting timely interventions-an issue highlighted by COVID-19, where deaths have stemmed from heterogeneous pathways such as pneumonia, multiorgan failure, or exacerbation of preexisting conditions. Objective: This study aims to propose a dynamic prognostic risk assessment framework based on longitudinal data collected during hospitalization, using COVID-19 as an example. Our aim was to develop and validate an interpretable framework that (1) screens prognosis at admission and (2) dynamically updates mortality risk throughout hospitalization, thereby providing clinicians with early, explainable warnings while minimizing additional cognitive load. Methods: In this retrospective study, we extracted electronic medical records of 382 COVID-19 cases treated at Tokyo Shinagawa Hospital between January 27 and September 30, 2020. At admission, gradient boosting decision trees (Light Gradient Boosting Machine) were used to predict the maximum clinical deterioration, including death, based on data available at initial diagnosis. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). For in-hospital monitoring, random survival forests (RSF) were trained on a longitudinal dataset that combined static demographic characteristics with serially measured vital signs and laboratory results. The model dynamically assessed daily mortality risk by calculating a 7-day cumulative hazard function, with risk scores recalculated each day during hospitalization. RSF accuracy was evaluated in an independent one-third test set using the concordance index (C-index), an integrated Brier score (1-50 days), and mean time-dependent AUC. SurvSHAP(t), an extension of Shapley Additive Explanations, was applied to provide time-dependent explanations of each variable's contribution to the prediction. Results: The prediction at initial diagnosis showed good agreement with the actual severity outcomes (AUC of 0.717 for predicting hospitalization/severity ≥2; 0.878 for severity ≥3; 0.951 for severity ≥4; 0.952 for severity ≥5; and 0.970 for death/severity=6), although some cases exhibited discrepancies between the predicted and actual prognoses. The dynamic mortality risk assessment during hospitalization using the RSF achieved a test-set C-index of 0.941, an integrated Brier score of 0.315, and a mean time-dependent AUC of 0.936. This dynamic assessment was able to distinguish between dead and surviving patients as early as 1-2 weeks before the outcome. Early in hospitalization, C-reactive protein was an important risk factor for mortality; during the middle period, peripheral oxygen saturation (SpO2) gained importance; and immediately before death, platelets and β-D-glucan were the primary risk factors. Conclusions: Integrating static admission triage with daily, explainable RSF predictions enables early identification of patients with COVID-19 at high risk of deterioration. By surfacing phase-specific, actionable predictors, the framework supports timely interventions and more efficient resource allocation. Prospective, multicenter studies are warranted to validate its generalizability and clinical impact.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.