Heartbeat Classification using CNN on UWB Radar Data

Overview

This project uses a pre-trained Convolutional Neural Network (CNN) to classify heartbeats detected from Ultra-Wideband (UWB) radar signals. The UWB data is processed to extract heartbeat-related signals, which are then normalized and fed into the CNN for classification.

Features

• Loads a trained CNN model for heartbeat classification.
• Processes raw UWB signals by applying a bandpass filter (1-3 Hz).
• Segments UWB data into ECG-sized windows (187 samples per segment).
• Normalizes the data using a pre-trained scaler.
• Classifies each heartbeat segment into five categories:
  • Normal (N)
  • Supraventricular (S)
  • Ventricular (V)
  • Fusion (F)
  • Unknown (Q)
• Displays classification results and visualizes detected heartbeats.

Definitions

1. CNN (Convolutional Neural Network)

A type of deep learning model that is particularly effective for analyzing time-series and spatial data, such as ECG waveforms. It extracts important patterns and features from input signals.

2. UWB (Ultra-Wideband) Radar

A wireless communication and sensing technology that uses short pulses over a wide frequency range to detect objects, movements, and vital signs such as heartbeats.

3. ECG (Electrocardiogram)

A recording of the electrical activity of the heart, commonly used for diagnosing cardiac conditions. The CNN model in this project was trained on ECG data to recognize heartbeat patterns.

4. Bandpass Filtering

A signal processing technique that allows signals within a specific frequency range (1-3 Hz in this case) to pass through while removing noise and irrelevant frequencies.

5. Normalization

A data preprocessing step where values are scaled to a common range (e.g., between -1 and 1) so that the model processes them effectively and consistently.

6. Segmentation

Dividing a continuous signal into fixed-size windows (187 samples in this case) to extract individual heartbeats for classification.

7. Scaler

A pre-trained model that was used to standardize the ECG data during training. The same scaler is used to normalize new UWB signal inputs before feeding them into the CNN.

8. Classification

The process of assigning labels to input data based on learned patterns. The CNN predicts the heartbeat type (N, S, V, F, or Q) for each segment.

Model Performance and Training Summary

1. Model Performance During Training

• Final Training Accuracy: 99.17% (Epoch 20)
• Final Training Loss: 0.0243
• Validation Accuracy: Fluctuated slightly but remained around 99.05%-99.17%.
• Best Model Saved at Epoch 20: Model improved in the last epoch, leading to the best accuracy being saved at 99.17%.

📌 What This Means:

• A 99.17% accuracy suggests that the model rarely makes mistakes on training data.
• The low loss (0.0243) indicates high confidence in predictions.
• Stable performance across epochs, avoiding overfitting.

2. Model Performance on Test Data

• Final Test Accuracy: 98.47%
• Final Test Loss: 0.0975

📌 What This Means:

• A 98.47% accuracy on unseen test data shows strong generalization.
• The slightly higher loss (0.0975) compared to training loss (0.0243) is expected in real-world applications.
• The small accuracy drop from 99.17% (training) → 98.47% (test) suggests minimal overfitting.

3. Key Observations

✅ Consistently High Accuracy: The model maintained 99% training accuracy and 98.47% test accuracy, showing strong learning capabilities.
✅ Minimal Overfitting: The small gap between training and test accuracy suggests the model didn't just memorize the training data but learned meaningful patterns.
✅ Efficient Training Progression: Improved over 20 epochs, showing good learning behavior.
✅ Best Model Saved Automatically: The best-performing model was saved (best_ecg_cnn_model.h5) based on validation accuracy.

Installation

Requirements

Ensure you have the following dependencies installed:

pip install numpy pandas tensorflow scipy joblib matplotlib

Explanation of the Pipeline

1. Load the trained CNN model: Uses a previously trained model on ECG data.
2. Read the UWB signal: Loads the raw UWB radar signal from a CSV file.
3. Filter the signal: Applies a bandpass filter (1-3 Hz) to extract heartbeat components.
4. Segment into ECG-sized windows: Splits the signal into segments of 187 samples.
5. Normalize the data: Uses a pre-trained scaler to match the CNN’s expected input format.
6. Classify the heartbeats: Feeds segments into the CNN for classification.
7. Display results: Prints predictions and visualizes heartbeat waveforms.

Example Output

Loading trained CNN model...
Loading raw UWB dataset...
Filtering UWB signal for heartbeat frequencies (1-3 Hz)...
Segmenting UWB signal into ECG-sized windows...
Classifying UWB segments using CNN...

Heartbeat Classification Results:
Segment 1: Classified as Normal (N)
Segment 2: Classified as Ventricular (V)
...
Total detected heartbeats classified: 5

Visualization

Raw UWB Data:

The script plots detected heartbeats with their classifications:

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

License

This project is licensed under the MIT License.

Credits

The UWB dataset was provided by Moro, G.; Di Luca, F.; Dardari, D.; Frisoni, G. Human Being Detection from UWB NLOS Signals: Accuracy and Generality of Advanced Machine Learning Models. Sensors 2022, 22, 1656. https://doi.org/10.3390/s22041656.

The CNN was trained with the pubilcally available ECG dataset (https://www.kaggle.com/datasets/shayanfazeli/heartbeat)