AI/ML models for predicting cardio-vascular diseaeses

Abstract

This project presents a comprehensive framework to address class imbalance in the PTB-XL electrocardiogram (ECG) dataset by integrating diffusion models with state-of-the-art deep learning architectures, including 1D-CNN-LSTM and Gap-5 ST-CNN-LSTM. The diffusion model employs Gaussian noise addition with multiple noise schedules, such as cosine, to generate synthetic ECG samples that augment minority classes. This approach ensures the temporal and physio- logical integrity of the signals while balancing the dataset effectively. The diffusion model is powered by a 1D U-Net architecture with spatio-temporal capabilities, incorporating residual blocks, attention mechanisms, and temporal convolutions. The forward diffusion process progressively adds Gaussian noise, while the reverse process reconstructs clean signals. This facilitates the generation of realistic ECG samples that mimic the original signal distribution and enhance model robustness. For classification, we leverage a hybrid pipeline combining Swin Transformers and 1D-CNN- LSTM. The 1D-CNN extracts fine-grained temporal features, while the LSTM captures sequential dependencies, making it well-suited for time-series ECG data. The Gap-5 ST-CNN-LSTM, a state-of-the-art architecture, further enhances the framework with its optimized spatio-temporal convolutional layers and LSTM units, offering superior feature representation and temporal understanding. Evaluation metrics such as F1 score, AUC, and class-wise recall demonstrate significant improvements in both minority class representation and overall classification accuracy. By combining diffusion-based augmentation with advanced deep learning architectures, this framework delivers a robust solution for multi-label, multi-class ECG diagnostics on imbalanced biomedical datasets.