A Dynamic Prognosis Model of Patients with Chronic Heart Failure: A Prospective Cohort Study Using Follow-Up Data and...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Traditional chronic heart failure (CHF) mortality prediction models use static data, limiting accuracy.
• Longitudinal follow-up data capture disease dynamics for improved risk prediction.
• Dynamic models offer potential for developing tailored interventions in CHF patients.

Purpose of the Study:

• To develop and validate a dynamic prediction model for 3-year all-cause mortality in CHF patients.
• To compare the performance of different machine learning models, including GRU, LSTM, MLP, and LR.
• To identify key predictors of mortality using interpretability techniques.

Main Methods:

• Enrolled 1,333 CHF patients, collecting patient-reported outcome (PRO) measures, lifestyle, and medication data.
• Developed and compared GRU, LSTM, MLP, and LR models using sequential data for 3-year mortality risk.
• Assessed model performance using AUC, accuracy, TNR, TPR, Brier score, and F1-score; utilized TimeSHAP for interpretability.

Main Results:

• The GRU model demonstrated superior predictive accuracy, with performance improving over time.
• Peak performance for the GRU model was achieved by 24 months (AUC: 0.765).
• Key predictors identified included CHF-PRO measures (physical condition, appetite, sleep, independence, anxiety), age, and NYHA class.

Conclusions:

• Longitudinal PRO data combined with GRU modeling provide a robust tool for predicting CHF mortality risk.
• A web-based decision support system was developed for practical risk score calculation.
• This dynamic approach supports evidence-based, tailored interventions for CHF management.