Cardiovascular disease prediction using variational recurrent autoencoders with uncertainty estimation

Area of Science:

• Cardiology
• Machine Learning
• Data Science

Background:

• Cardiovascular diseases (CVDs) are a major global health concern, necessitating accurate and reliable classification methods.
• Current methods often lack calibrated probability estimates, hindering clinical decision-making.
• Benchmark studies are crucial for evaluating and advancing CVD classification techniques.

Purpose of the Study:

• To propose a deep learning framework for accurate CVD classification from static tabular data.
• To incorporate uncertainty estimation for reliable probability predictions.
• To establish a robust benchmark for CVD classification using a Variational Recurrent Autoencoder (VRAE).

Main Methods:

• Developed a VRAE model integrating Gated Recurrent Unit (GRU) encoding, variational latent sampling, and Monte Carlo (MC) dropout.
• Utilized synthetic noise-augmented pseudo-sequences to represent static tabular data, simulating measurement perturbation and missingness.
• Evaluated the model on two public datasets: Heart Failure Prediction and Cardiovascular Disease datasets.

Main Results:

• Achieved high accuracy (95.8% and 96.1%) and F1-scores (95.7% and 96.0%) on both datasets.
• Outperformed traditional classifiers and other deep learning models, including LSTM and GRU-Attention.
• Demonstrated robustness to synthetic noise and provided superior calibration with the lowest Brier scores (0.061 and 0.059).

Conclusions:

• The proposed VRAE model offers uncertainty-aware representation learning for benchmark-level static tabular CVD classification.
• The framework provides accurate predictions and reliable uncertainty quantification.
• Further validation on real-world longitudinal data is needed for clinical deployment and prospective risk prediction.