Data Analytics and Learning for Cardiovascular Disease Detection based on ECG Signals

Overview

During my research internship at North Carolina State University, under the supervision of Associate Professor Zhishan Guo, I developed a one-class classification model to detect cardiovascular abnormalities using ECG signals. The project aimed to address limitations in traditional methods that focused on specific disease categories by creating a model capable of identifying a broader range of abnormalities, including those not present in the training dataset. The PhysioNet dataset was used to train and validate the models.

Results

• Designed and trained a one-class autoencoder model that demonstrated robust performance in distinguishing normal and abnormal ECG signals.
• Integrated convolutional layers, LSTM modules, and self-attention mechanisms to enhance feature extraction and anomaly detection.
• Successfully reduced reconstruction error through refined preprocessing, improving model sensitivity to diverse abnormalities.
• Presented findings at NCSU’s summer research symposium and contributed to a submitted paper for ICCPS25.

Poster Presentation PDF

Technical Details

• Model Development:
  • Developed a one-class autoencoder pipeline leveraging dimensionality reduction and reconstruction errors for abnormality detection.
  • Integrated convolutional layers, LSTM modules, and self-attention mechanisms to improve feature extraction.
  • Benchmarked model designs to optimize accuracy and efficiency for real-time applications.
• Preprocessing:
  • Addressed noise and segmentation inaccuracies in ECG signals using advanced denoising methods.
  • Collaborated with medical experts to design signal templates for precise feature extraction.
• Implementation:
  • Implemented the full pipeline for model training, testing, and visualization.
  • Extended work to adaptive model selection for embedded systems, ensuring low latency and high reliability.

Challenges and Solutions

1. Challenge: Variability in real-world datasets, including noise and segmentation errors.
   Solution: Tested multiple denoising methods and developed specialized signal templates in collaboration with medical experts.

2. Challenge: Limited labeled data for abnormal signals.
   Solution: Exclusively trained on normal signals to utilize reconstruction errors for anomaly detection.

Reflection and Insights

This project deepened my understanding of biomedical signal analysis, especially in feature extraction and pattern recognition. The collaboration with medical experts highlighted the importance of interdisciplinary research in addressing practical challenges. Building on this experience, I am inspired to explore the application of signal-processing models in neuroscience, particularly for analyzing brain neural signals like EEG and fMRI.

Team and Role

This research was conducted under the guidance of Professor Zhishan Guo at NCSU. I led the development of the autoencoder model and preprocessing pipeline, contributed to adaptive model selection strategies, and collaborated with team members on testing, result visualization, and paper writing.