This repository contains machine learning models developed as part of Team 9's work for the Cardio Challenge under the auspices of the U.S. Department of Energy at Lawrence Livermore National Laboratory (LLNL). The primary focus is on detecting cardiac abnormalities using ECG data and predicting activation times in the myocardium. The repository includes various binary and multi-class classification models as well as advanced techniques for feature extraction and activation map reconstruction.

• Atharva Gupta
• Anuvetha Govindarajan
• Brady Snyder
• Davinia Muthalaly
• Noah Gallego

• Problem Solved: Identifying normal versus abnormal heartbeats based on 10-lead ECG data.
• Solution: Implemented a logistic regression model with 5-fold cross-validation, achieving an accuracy of 83% on test data. This solution provides a baseline model for quick diagnosis of irregular heartbeats.

• Problem Solved: Enhancing the detection of normal and abnormal heartbeats.
• Solution: A feed-forward neural network with ReLU activation functions in hidden layers and sigmoid activation in the output layer was implemented. This model improved accuracy over logistic regression.

• Problem Solved: Classifying ECG sequences into one of five arrhythmia categories.
• Solution: Built multiple models, including:
  • K-Nearest Neighbors (KNN): Achieved an F1 score of 0.97 for normal cases but lower for other classes due to data imbalance.
  • Decision Trees: With a max depth of 10, produced an average F1 score of 0.72 for the majority class.
  • Convolutional Neural Networks (CNN): This model achieved 99% accuracy across all classes, solving the imbalance problem through data rebalancing techniques (down-sampling and up-sampling).

• Problem Solved: Predicting the activation time across 75 distinct regions of the myocardium.
• Solution: A hybrid network architecture was developed, combining convolutional and pooling layers with fully connected layers. The model achieved a mean-squared error (MSE) of 104, with predictions within 5.93 milliseconds of actual activation times.

• Problem Solved: Reducing noise in ECG signal data for improved feature extraction and model accuracy.
• Solution: Applied a Fourier Transform to activation time sequences, resulting in a de-noised dataset that improved CNN performance, bringing the MSE down to 104 from previous models.

• Problem Solved: Lack of interpretability in neural networks, particularly CNNs.
• Solution: Integrated Grad-CAM to visualize which features influence the CNN's decision-making process, enhancing the model's transparency and aiding medical professionals in understanding the basis for predictions.

• Data Imbalance: Addressed through rebalancing techniques.
• Accuracy: Best model (CNN) achieved 99% accuracy.
• Activation Time Prediction: Achieved predictions within 5.93ms of actual times.