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

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Public Health

Background:

• Static risk scores are insufficient for rapidly evolving patient conditions like COVID-19.
• Heterogeneous disease trajectories in COVID-19 necessitate dynamic prognostic tools.
• Timely interventions are crucial, as highlighted by COVID-19 mortality from various complications.

Purpose of the Study:

• To propose a dynamic prognostic risk assessment framework using longitudinal hospitalization data for COVID-19.
• To develop an interpretable model for initial prognosis screening and continuous mortality risk updates.
• To provide clinicians with early, explainable warnings to minimize cognitive load and support interventions.

Main Methods:

• Retrospective analysis of 382 COVID-19 patient electronic medical records.
• Gradient boosting decision trees (Light Gradient Boosting Machine) for initial risk prediction.
• Random survival forests (RSF) for dynamic daily mortality risk assessment using longitudinal data.
• SurvSHAP(t) for time-dependent explanation of risk factors.

Main Results:

• Initial prediction model showed good agreement with severity outcomes (AUC up to 0.970 for death).
• Dynamic RSF model achieved high accuracy (C-index 0.941, mean AUC 0.936).
• The dynamic assessment identified high-risk patients 1-2 weeks before adverse outcomes, with evolving key predictors (CRP, SpO2, platelets, β-D-glucan).

Conclusions:

• Integrating static and dynamic predictions enables early identification of high-risk COVID-19 patients.
• The framework supports timely interventions and efficient resource allocation through phase-specific predictors.
• Prospective, multicenter validation is needed to confirm generalizability and clinical impact.