Development and validation of an interpretable machine learning scoring tool for estimating time to emergency...

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Predictive Analytics

Background:

• Emergency readmissions create significant burdens for patients and healthcare systems.
• Accurate risk stratification is crucial for effective transitional care interventions.
• Existing risk prediction models may lack accuracy or interpretability for temporal risk assessment.

Purpose of the Study:

• To develop and validate an interpretable machine learning risk scoring system for predicting short- and intermediate-term emergency readmission risks.
• To provide data for temporal risk stratification and inform clinical decision-making.
• To enhance transitional care interventions aimed at reducing readmissions.

Main Methods:

• Retrospective analysis of emergency admission episodes from a tertiary hospital in Singapore (2009-2016).
• Development of the Score for Emergency ReAdmission Prediction (SERAP) using an interpretable machine learning system for time-to-event outcomes.
• SERAP incorporates six variables: prior emergency admissions, age, malignancy history, renal disease history, serum creatinine, and serum albumin.

Main Results:

• The study included 293,589 admission episodes; 27.3% resulted in readmission within 90 days.
• SERAP demonstrated strong predictive performance in the testing cohort with an integrated AUC of 0.737.
• SERAP outperformed established scores like LACE and HOSPITAL for 30-day readmission prediction and can predict readmission at various time points within 90 days.

Conclusions:

• The SERAP tool offers superior performance in risk prediction compared to existing scores.
• SERAP provides accurate, interpretable information for temporal risk stratification and clinical decision-making.
• Further external validation is recommended to assess SERAP's real-world performance in diverse settings.