Cardiovascular disease prediction using variational recurrent autoencoders with uncertainty estimation

Area of Science:

• Cardiovascular research
• Machine learning in healthcare
• Biomedical informatics

Background:

• Cardiovascular diseases (CVDs) are a major global cause of mortality.
• Accurate classification and reliable probability estimates are crucial for CVD benchmark studies.
• Existing methods may lack robustness in handling noisy or incomplete clinical data.

Purpose of the Study:

• To propose a deep learning framework for cardiovascular disease classification using static tabular clinical records.
• To develop a model that provides calibrated probability estimates and uncertainty quantification.
• To evaluate the model's performance and robustness against various baselines.

Main Methods:

• A Variational Recurrent Autoencoder (VRAE) framework was developed, integrating Gated Recurrent Unit (GRU) encoding, variational latent sampling, and Monte Carlo (MC) dropout.
• Static tabular datasets were transformed into synthetic pseudo-sequences to simulate noise and missingness for denoising latent representation learning.
• The model was evaluated on the Heart Failure Prediction and Cardiovascular Disease datasets.

Main Results:

• The VRAE model achieved high accuracy (95.8% and 96.1%) and F1-scores (95.7% and 96.0%) on both datasets.
• It outperformed traditional classifiers and other deep learning models, including LSTM and GRU-Attention.
• The model demonstrated robustness under synthetic noise and achieved the lowest Brier scores, indicating improved calibration.

Conclusions:

• Uncertainty-aware representation learning using the VRAE framework is effective for benchmark-level static tabular CVD classification.
• The model offers reliable performance and improved calibration compared to existing methods.
• Further validation on real-world longitudinal electronic health record cohorts is recommended for clinical deployment.