Deep learning for predicting rehospitalization in acute heart failure: Model foundation and external validation

Area of Science:

• Cardiology
• Artificial Intelligence
• Health Informatics

Background:

• Heart failure (HF) rehospitalization poses a significant challenge in patient management.
• Existing risk prediction models for HF rehospitalization often lack the integration of deep learning and real-world data.
• Accurate prediction of HF rehospitalization is crucial for timely interventions and improved patient outcomes.

Purpose of the Study:

• To develop and validate a deep learning-based prediction model for heart failure (HF) rehospitalization.
• To predict HF rehospitalization at 30, 90, and 365 days post-discharge from acute HF (AHF).
• To identify key prognostic features contributing to HF rehospitalization using real-world data.

Main Methods:

• Utilized real-world data from patients admitted for acute heart failure (AHF) between January 2014 and January 2019.
• Employed deep learning algorithms, including hyperbolic tangent activation layers and recurrent layers with gated recurrent units.
• Assessed prediction performance using Area Under the Curve (AUC), precision, recall, specificity, and F1 score, with Shapley value analysis for feature importance.

Main Results:

• Identified 22 prognostic features (6 time-independent, 16 time-dependent) significantly associated with HF rehospitalization.
• Achieved moderate discrimination for predicting rehospitalization within 30, 90, and 365 days (AUC: 0.63, 0.74, and 0.76, respectively).
• Found that features during follow-up contributed more significantly to HF rehospitalization prediction than those from earlier time points.

Conclusions:

• The developed deep learning model provides valid predictions for heart failure rehospitalization up to one year post-discharge.
• The model can guide the implementation of targeted interventions and care strategies for HF patients.
• Close patient monitoring and regular blood tests are vital for assessing rehospitalization risk.