Document Type

Restricted Campus Only

Publication Date

4-28-2026

Abstract

Cardiac diseases are one of the leading causes of death across the world, and early detection heavily impacts the patient’s outcomes. This project focuses on developing an AI-based system capable of accurately classifying heart conditions from electrocardiogram (ECG) signals. The team designed a system that processes ECG data and identifies nine distinct cardiac conditions using different machine learning techniques to address these issues. The system is designed to be reliable, scalable, and accessible, to people in environments with limited access to medical professionals.

The system architecture consists of ECG data extraction from WFDB records, preprocessing and standardization, dataset organization and labeling, and integration with machine learning models. The development throughout the first semester showed that 10-second ECG segments provide enough information for classification. Nevertheless, we identified several limitations, including inconsistent file formatting, missing data, and limited validation tools. Furthermore we altered the system to make significant improvements throughout the second semester. The teams implemented interactive signal visualization tools, R-peak-based segmentation for higher-quality data, improved error handling, and standardized preprocessing techniques to support both 1D and 2D neural network models.

Multiple machine learning approaches were evaluated, among them, convolutional neural networks (CNNs) outperformed all other methods with accuracies above project criterion of 75%. The implementation of 1D CNNs using raw ECG data was the cornerstone of the prototype’s classification system.

Validation and testing was conducted using separate data from training to ensure unbiased evaluation. While training data used oversampling and data augmentation to expand the data the models can learn from. The final system demonstrated strong classification performance with a simple user experience. These results show that the deep learning approaches that implement raw ECG signals are more effective than feature-engineered methods for this application.

Overall, the project successfully delivered a functional AI-based ECG classification system that exceeds the performance requirements. Our final system, an ensemble of two 1D-CNNs, has been packaged as a black box with 10 second ECG inputs and a simple .CSV file output.

Comments

Mehran Aminian, Team Advisor

Download

Share

COinS