← Back

Cardiovascular Disease Detection

Deep Learning for ECG-Based Diagnosis

Aug 2023 — May 2024 Completed

Overview

This project focused on optimizing a convolutional neural network (CNN) to detect the presence of six different cardiovascular diseases from raw 12-lead electrocardiogram (ECG) signals. The goal was to improve diagnostic automation and support clinical decision-making in cardiology.

Clinical Significance

Cardiovascular diseases remain the leading cause of death globally. Early and accurate detection through ECG analysis can significantly improve patient outcomes. However, manual ECG interpretation is time-consuming and subject to inter-observer variability. Automated detection systems can:

  • Provide rapid preliminary screening in high-volume settings
  • Assist healthcare workers in resource-limited environments
  • Reduce diagnostic delays in emergency situations
  • Serve as a second opinion for complex cases

Technical Approach

Data Processing

  • Processed and standardized a dataset of approximately 2.3 million ECG traces
  • Normalized scaling variations across 12-lead recordings from heterogeneous ECG machines
  • Implemented robust preprocessing to handle artifacts and noise in clinical recordings
  • Developed data augmentation strategies specific to physiological signals

Model Architecture

  • Optimized a 1D convolutional neural network architecture for temporal signal processing
  • Implemented multi-task learning for simultaneous detection of six cardiovascular conditions
  • Explored attention mechanisms to identify diagnostically relevant ECG segments

Feature Analysis

  • Investigated feature importance across precordial leads (V1-V6)
  • Analyzed contribution of limb leads (I, II, III, aVR, aVL, aVF)
  • Mitigated multicollinearity in input signals to determine optimal channel configurations
  • Performed ablation studies to understand model behavior

Results

0.89 Final F1-Score
+7.2% Improvement from Baseline
2.3M ECG Traces Processed

The optimized model achieved an F1-score of 0.89, improving from the baseline of 0.83. This 7.2% improvement in diagnostic performance demonstrates enhanced model generalization and clinical applicability.

Technologies Used

Python PyTorch NumPy Pandas SciPy Scikit-learn Matplotlib Signal Processing Wfdb