12SL is the algorithm used by GE Healthcare Physicians in determining Myocardial Infarction (MI), or heart attack in a more common term from patients' electrocardiogram (ECG). We trained a machine learning model to improve 12SL misclassification by identifying both false positives (missed MI detection) and false negatives (incorrect MI detection). We also further classify MI cases as either acute or non-acute. This enables better evaluation of MI in borderline cases and helps physicians make quicker and more accurate clinical decisions, especially in emergency situations.
-
Using Google Colab.
- Open Jupyter Notebook file.
- Click on the last icon on the left navigation bar, upload all the files in the folder "content" under Google Drive team folder.
- Select GPU (EDIT -> NOTEBOOK SETTINGS -> T4 GPU)
- Run all the cells in the Jupyter Notebook.
We first read in the ECG data provided by our mentors along with header files from the Physionet Challenge. The data preprocessing step involves merging multiple tables into a single consolidated file, standardizing column names, handling missing values, and then splitting the dataset into false positive and false negative groups for separate model training.
Our model is designed as a two-pass ensemble system using three different Light Gradient Boosting Machine (LightGBM) trained upon thousands of ECG records.
In the first pass, two LightGBM classifiers were trained on false positives and false negatives from the 12SL algorithm to classify cases as MI or non-MI (NMI). Records predicted as NMI are re-evaluated using the false negative model to reduce the risk of missing true MIs. These three models work together to correct 12SL misclassifications and improve initial categorization.
In the second pass, a third LightGBM model further classifies MI cases as acute or non-acute, enabling more precise identification of urgent conditions.
After running the model, we generate a final prediction report based on the sample ECG file. The report contains:
- Predicted MI Status: Indicates whether the model classifies the ECG as Positive (MI) or Negative (No MI).
- Predicted Severity: For Positive cases, further classification into Acute MI or Non-Acute MI.
- Model Interpretation: A SHAP graph showing feature importance rankings, and a decision tree illustrating how the models make their classification decisions.
Our model achieved strong model accuracy in improving MI detection. The first model achieved an F1 score of 0.95. The second model achieved an F1 score of 0.65.
Our model achieves high false positive accuracy, strong feature isolation to eliminate distracting features, and high model interpretability through decision trees.
The third model achieved a relatively low F1 score of 0.33 due to low fidelity in some ECG signals and requires extensive parameter tuning.
To enhance the performance of the acute detection model, we can try incorporating additional features and exploring alternative machine learning models, such as deep learning models that better capture temporal ECG patterns.