Using Deep Learning to Identify High-Risk Patients with Heart Failure with Reduced Ejection Fraction.

Area of Science:

• Cardiology
• Biomedical Informatics
• Artificial Intelligence

Background:

• Deep learning (DL) is not yet well-established for identifying high-risk heart failure (HF) patients.
• Predictive models are crucial for managing patients with heart failure with reduced ejection fraction (HFrEF).

Purpose of the Study:

• To utilize DL models for predicting hospitalizations, worsening HF events, and 30/90-day readmissions in HFrEF patients.
• To compare DL model performance against traditional machine learning algorithms.

Main Methods:

• Analysis of adult HFrEF patient data from IBM MarketScan databases (2015-2017).
• Implementation of a sequential model architecture using bi-directional long short-term memory (Bi-LSTM) layers.
• Comparison with logistic regression, random forest, and XGBoost models, assessing performance via AUC, precision, and recall.

Main Results:

• DL models achieved high AUCs (up to 0.977) for predicting HF hospitalizations and worsening HF events without a buffer window.
• AUCs for 30- and 90-day readmissions were 0.597 and 0.614, respectively; 0.861 for 90-day readmission in 18-64 age group.
• DL models consistently outperformed traditional machine learning models across all assessed outcomes.

Conclusions:

• A DL approach using Bi-LSTM is a feasible and effective tool for predicting HF-related outcomes.
• This technology can aid in identifying high-risk HFrEF patients for targeted interventions.
• Further development is needed to address data limitations and enhance predictive accuracy.