This project focuses on the categorization of ECG heartbeat signals into five distinct classes using deep learning models: Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) networks. Both models were evaluated on the MIT-BIH Arrhythmia dataset, consisting of labeled heartbeat signals.
-
Data Loading:
- The datasets,
mitbih_train.csvandmitbih_test.csv, were loaded into Pandas DataFrames.
- The datasets,
-
Data Inspection:
- Checked for null values, duplicates, and class distributions.
-
Feature-Label Separation:
- Extracted features and labels for training and testing.
-
Data Scaling:
- Min-max scaling was applied to normalize the input values to the range [0, 1].
-
Data Splitting:
- Test data was further split into validation and test subsets.
-
Reshaping:
- Reshaped data to be compatible with recurrent layers by adding a third dimension.
-
GRU:
- Two GRU layers with dropout regularization.
- Dense output layer with softmax activation for multi-class classification.
-
LSTM:
- Two LSTM layers with dropout regularization.
- Dense output layer with softmax activation for multi-class classification.
| Metric | GRU Model | LSTM Model |
|---|---|---|
| Test Loss | 0.0555 | 0.1297 |
| Test Accuracy | 98.45% | 96.67% |
| Class | Description | GRU Precision | GRU Recall | GRU F1-Score | LSTM Precision | LSTM Recall | LSTM F1-Score |
|---|---|---|---|---|---|---|---|
| 0 | Normal | 99% | 99% | 99% | 97% | 99% | 98% |
| 1 | Atrial Premature | 88% | 78% | 82% | 85% | 54% | 66% |
| 2 | Premature Ventricular Contraction | 96% | 95% | 95% | 89% | 91% | 90% |
| 3 | Fusion of Ventricular and Normal | 80% | 84% | 82% | 69% | 51% | 59% |
| 4 | Fusion of Paced and Normal | 99% | 99% | 99% | 99% | 93% | 96% |
-
Iterative Development:
- Models were iteratively trained and evaluated to improve performance.
-
Early Stopping:
- Implemented to prevent overfitting by monitoring validation loss.
-
Learning Rate Adjustment:
- Used
ReduceLROnPlateauto dynamically adjust the learning rate for faster convergence.
- Used
-
Modular Preprocessing Pipeline:
- Designed a scalable preprocessing pipeline to adapt easily for future datasets.
-
Visualization:
- Confusion matrices and performance graphs provided insights into model performance.
- The GRU model achieved a higher accuracy (98.45%) compared to the LSTM model (96.67%).
- GRU demonstrated superior precision, recall, and F1-scores across most classes, particularly in rare heartbeat categories.
- While LSTM required fewer epochs, its test loss and accuracy suggest GRU’s architecture was more effective for this dataset.
- This project highlights the efficiency of preprocessing and model selection in achieving robust classification results for medical time-series data.