Cardio-XAI

An AI that doesn't just detect heart attacks; it shows you where they are.

Comment

💡 Inspiration

Cardiovascular disease remains the leading cause of death globally. While the 12-lead ECG is the gold standard for diagnosis, interpreting these signals is notoriously difficult—even for residents. A subtle ST-elevation or a T-wave inversion can easily be missed in a busy ER.

We were inspired to build Cardio-AI not just to "classify numbers," but to solve the "Black Box" problem in medical AI. We didn't want a system that just says "Probability: 90%." We wanted a system that could act as a true digital colleague, one that points to the exact heartbeat and says, "Look here. This T-wave morphology is suspicious."

🩺 What it does

Cardio-AI is an interpretable Deep Learning system that detects 5 major cardiovascular pathologies (including Myocardial Infarction and Ischemia) from raw 12-lead ECG signals.

It goes beyond simple classification by generating a Clinical Dashboard that:

  1. Stratifies Risk: Uses calibrated thresholds to flag patients as "Low," "Moderate," or "High" risk.
  2. Visualizes Evidence: Uses Grad-CAM (Gradient-weighted Class Activation Mapping) to project heatmaps onto the ECG, highlighting the specific waveforms (like the ST-segment) that triggered the diagnosis.

⚙️ How we built it

We adopted a "First Principles" approach to Geometric Deep Learning.

  • Architecture: We built a custom 1D-ResNet backbone optimized for time-series data. We enhanced it with Squeeze-and-Excitation (SE) Blocks, which allow the network to dynamically "attend" to specific leads (e.g., prioritizing Lead II over Lead aVR) depending on the signal context.
  • Data Pipeline: We used the PTB-XL dataset (21,000+ records). We implemented a rigorous preprocessing pipeline including 100Hz downsampling and Z-score normalization.
  • Math & Optimization: To handle the massive class imbalance (healthy patients far outnumber sick ones), we replaced standard Cross-Entropy Loss with Focal Loss:

This mathematically down-weights "easy" examples and forces the model to focus its gradient energy on hard, rare cases like Conduction Disturbances.

🚧 Challenges we ran into

Our biggest hurdle was the "Clean Data Bias."

Midway through the project, we discovered a critical failure mode: The model assigned a higher risk score (28%) to a healthy patient with a loose electrode artifact (Patient 2400) than to a patient with a confirmed heart attack (Patient 63, 26%). The model had learned that "Signal Noise = Disease."

The Fix: We couldn't just tune hyperparameters. We engineered a Robust Training Pipeline with Online Noise Injection.

  1. We mathematically generated random Gaussian noise and amplitude scaling artifacts on the fly during training.
  2. We forced the model to classify these noisy signals as "Normal."
  3. We implemented Low-LR Fine-Tuning (learning rate ) to settle the weights.

The Result: The model learned to distinguish artifacts from pathology. In the final version, the Heart Attack patient's score jumped to 49%, while the Noisy Healthy patient correctly stayed at 27%.

🏆 Accomplishments that we're proud of

  • SOTA Performance: Achieving a Macro-AUC of 0.920 on the test set, with Ischemia detection reaching 0.93.
  • The "Safety Net": We successfully calibrated the model so that no confirmed heart attack in our validation set was missed (Sensitivity prioritization).
  • Interpretability: Seeing the Grad-CAM heatmap correctly light up the T-Wave on a patient with ischemia was a definitive "Aha!" moment, proving the model learned the actual cardiology, not just statistical shortcuts.

🧠 What we learned

  • Data-Centric AI is Key: The improvements didn't come from making the neural network deeper; they came from fixing how the model perceived noise via augmentation.
  • Calibration Matters: A raw probability score is useless to a doctor. Mapping those probabilities to clinically relevant risk tiers ("High" vs "Moderate") is the bridge between code and care.

🚀 What's next for Cardio-AI

  • Real-Time Deployment: Porting the PyTorch model to ONNX for edge deployment on portable ECG devices.
  • Text Generation: Integrating an LLM to generate natural language reports (e.g., "Sinus rhythm with marked ST-elevation in Lead II") based on the model's feature vectors.

Built With

Share this project:

Updates