Enhancing heart failure treatment decisions: interpretable machine learning models for advanced therapy eligibility...

Area of Science:

• Cardiology
• Biomedical Informatics
• Machine Learning

Background:

• Timely referral for advanced heart failure therapies is crucial for patient outcomes.
• Current decision-making is complex, time-consuming, and requires specialized expertise.
• Electronic Health Records (EHRs) contain valuable sequential data for prediction.

Purpose of the Study:

• To develop and evaluate logistic tensor regression (LTR) models for predicting heart failure patients requiring advanced therapies.
• To analyze model performance at both population and individual levels using EHR data.
• To identify key clinical features and comorbidities associated with advanced therapy referral.

Main Methods:

• Proposed two LTR models (standard and personalized) utilizing irregularly spaced sequential EHR data.
• Collected clinical features from previous visits for predictions at the start of subsequent visits.
• Conducted patient-wise ten-fold cross-validation and compared with other machine learning models.

Main Results:

• Standard LTR achieved an average F1 score of 0.708, AUC of 0.903, and AUPRC of 0.836.
• Personalized LTR achieved an F1 score of 0.670, AUC of 0.869, and AUPRC of 0.839.
• The LTR models outperformed other machine learning methods and improved their performance through weight transfer.

Conclusions:

• The proposed LTR models effectively predict the need for advanced heart failure therapies using EHR data.
• Identified significant clinical features like chronic kidney disease and hypotension, validated by experts.
• The models offer insights into risk factors for heart failure progression at population and individual levels.