Risk of Mortality Prediction Involving Time-Varying Covariates for Patients with Heart Failure Using Deep Learning.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Prognostics

Background:

• Heart failure (HF) presents significant challenges to healthcare systems.
• Accurate mortality prediction is crucial for effective HF management.
• Existing prognostic models may not fully leverage temporal patient data.

Purpose of the Study:

• To develop and validate a novel deep learning prognostic model for predicting all-cause mortality in HF patients.
• To compare the performance of the proposed model against classical deep learning and traditional statistical models.
• To assess the utility of time-varying covariates in HF mortality prediction.

Main Methods:

• Development and validation of a recurrent neural network-based model (RNNSurv) incorporating time-varying covariates.
• Enrollment of 730 HF patients from Toho University Ohashi Medical Center (April 2016 - March 2020).
• Comparative analysis with a deep feed-forward neural network (DeepSurv) and a multivariate Cox proportional hazard model.

Main Results:

• RNNSurv demonstrated superior prediction performance compared to DeepSurv and Cox models (C-index: 0.839 vs. 0.755 vs. 0.762).
• RNNSurv showed better calibration and risk stratification capabilities, particularly for high-mortality risk patients.
• The model effectively utilized temporal information from time-varying covariates for improved prediction.

Conclusions:

• The RNNSurv model offers enhanced prediction accuracy for HF mortality by incorporating temporal data.
• This deep learning approach can assist in clinical decision-making for HF patient management.
• Mortality risk factors are risk-level specific, advocating for individualized clinical strategies.