Predicting readmission rates in critically ill heart failure patients during a 90-day vulnerable phase using...

Area of Science:

• Cardiology
• Medical Informatics
• Predictive Analytics

Background:

• Critically ill Heart Failure (HF) patients face high readmission rates during the vulnerable post-discharge period.
• Existing predictive models often lack comprehensive data on comorbidities and medication histories, limiting their accuracy.

Purpose of the Study:

• To develop interpretable predictive models for Heart Failure readmission risk.
• To identify key comorbidities and medications associated with HF readmission.

Main Methods:

• Utilized critically ill HF patient data from MIMIC-IV and MIMIC-III databases for training and validation.
• Constructed and evaluated four predictive models, including Neural Multitasking Logistic Regression (NMTLR).
• Employed Shapley Additive Explanations (SHAP) for feature importance interpretation and model simplification.

Main Results:

• The NMTLR model achieved superior predictive performance, with AUCs of 0.752 (internal) and 0.785 (external).
• SHAP analysis revealed comorbidities and medications, including psychoanaleptics, as significant predictors.
• A compact NMTLR model also demonstrated strong predictive capabilities.

Conclusions:

• Both full and compact NMTLR models effectively predict Heart Failure readmission.
• Psychoanaleptics emerged as a potential novel predictor for HF readmission, requiring further clinical investigation.