Personalized survival probabilities for SARS-CoV-2 positive patients by explainable machine learning

Area of Science:

• Computational epidemiology
• Clinical informatics
• Biostatistics

Background:

• Accurate risk assessment for SARS-CoV-2 patients is crucial for precision medicine.
• Machine learning offers potential for developing predictive models from complex health data.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting 12-week mortality in SARS-CoV-2 positive patients.
• To identify key risk factors contributing to mortality using electronic health record (EHR) data.

Main Methods:

• A discrete-time survival model was trained on EHR data from 33,938 SARS-CoV-2 cases.
• 2723 variables including demographics, diagnoses, medications, lab results, and vital signs were analyzed.
• Model performance was evaluated using weighted concordance index and area under the precision-recall curve.

Main Results:

• The model achieved a weighted concordance index of 0.95 and an AUC for precision-recall of 0.71.
• Key mortality risk factors identified include age, sex, medication count, prior hospitalizations, and lymphocyte counts.
• The model revealed temporal dynamics of 22 selected risk factors.

Conclusions:

• An explainable survival model using EHR data can predict personalized mortality risk for SARS-CoV-2 patients.
• This approach has the potential to be integrated into routine care for enhanced patient management.
• Further validation is recommended for clinical implementation of personalized survival probability reporting.