|
98.44% Test Accuracy |
98.44% F1 Score |
~44K Parameters |
<10ms Inference Time |
This project is a practical deliverable for the Production Internship Course, aiming to explore ECG-signal-based biometric identification technology.
As an emerging biometric modality, ECG offers unique advantages over traditional methods:
| Feature | Advantage |
|---|---|
| 🔐 Liveness Detection | ECG signals can only be captured from living subjects, naturally providing anti-spoofing capability. |
| 🧬 Uniqueness | Each person has unique cardiac electrophysiological characteristics. |
| 🛡️ Hard to Forge | Compared with fingerprints or faces, ECG is extremely difficult to replicate. |
| ⏱️ Continuous Verification | Supports continuous identity monitoring and verification. |
- Data Collection: Collected ECG data from 6 subjects using an ADS1292R acquisition device.
- Sampling Specs: 250Hz sampling rate, 24-bit ADC precision, ~5 minutes of data per person.
- Core Technologies: Signal preprocessing + Pan-Tompkins R-peak detection + 1D-CNN deep learning classification.
- Final Result: Test accuracy reached 98.44%, meeting practical application requirements.
|
|
||||||||||||||||||||
|
|
Visualization of signal preprocessing, R-peak detection, and heartbeat segmentation
Real-time identity recognition with confidence display
Confusion matrix, accuracy, F1 score, and other evaluation metrics
DeepSeek AI-assisted report generation
| 🎯 Metric | 📈 Value | 🎯 Metric | 📈 Value |
|---|---|---|---|
| Test Accuracy | 98.44% | Validation Accuracy | 99.22% |
| Precision | 98.45% | Sensitivity/Recall | 98.44% |
| F1 Score | 98.44% | Specificity | 99.68% |
| A | B | C | D | E | F | Recall | |
|---|---|---|---|---|---|---|---|
| A | 72 | 0 | 1 | 0 | 0 | 0 | 98.6% |
| B | 0 | 104 | 0 | 0 | 0 | 1 | 99.0% |
| C | 2 | 0 | 103 | 0 | 0 | 0 | 98.1% |
| D | 0 | 0 | 1 | 78 | 0 | 1 | 97.5% |
| E | 0 | 0 | 1 | 0 | 72 | 0 | 98.6% |
| F | 0 | 1 | 0 | 0 | 0 | 75 | 98.7% |
Total test samples: 512 | Correct classifications: 504 | Misclassifications: 8
|
|
# 1️⃣ Clone repository
git clone https://github.com/Marways7/ECG_IdentificationX.git
cd ECG_IdentificationX
# 2️⃣ Install dependencies
pip install -r requirements.txt
# 3️⃣ Configure API key
echo "DEEPSEEK_API_KEY=your_api_key_here" > .env
echo "DEEPSEEK_BASE_URL=https://api.deepseek.com/v1" >> .env
# 4️⃣ Prepare ECG data (see details below)
# Put your ECG files into `ECG_data/`
# 5️⃣ Train model
python train.py --epochs 50 --model_type lightweight
# 6️⃣ Launch web interface
streamlit run app.py📁 How to prepare ECG data?
⚠️ Privacy Notice: ECG data is personally identifiable biometric data. To protect privacy, this repository does not include raw ECG recordings.
Place your ECG files in ECG_data/ with the following format:
| Field | Description |
|---|---|
timestamp |
Timestamp |
Channel 1 |
ECG signal (24-bit ADC raw value) |
Channel 2 |
Respiration signal (optional) |
Example file structure:
ECG_data/
├── A.csv # Data for subject A
├── B.csv # Data for subject B
└── ... # More subjects
CSV file format:
timestamp,Channel 1,Channel 2,Channel 3
1767759832,128363,11421861,652
1767759832,128774,11421871,652
...📌 How to get a DeepSeek API Key?
- Visit DeepSeek Platform
- Sign up / log in
- Create an API key
- Copy it into
.env
🌐 After startup, visit http://localhost:8501
flowchart LR
subgraph A[" 📥 Acquisition "]
A1["ADS1292R<br/>250Hz"]
end
subgraph B[" 🔧 Preprocessing "]
B1["Filtering"] --> B2["Denoising"]
end
subgraph C[" 📊 Features "]
C1["R-peak Detection"]
C2["HRV/MFCC"]
end
subgraph D[" 🧠 Model "]
D1["1D-CNN"]
end
subgraph E[" 📤 Output "]
E1["Identity A-F"]
end
A1 --> B1
B2 --> C1
C1 --> C2
C2 --> D1
D1 --> E1
style A fill:#e3f2fd
style B fill:#fff3e0
style C fill:#e8f5e9
style D fill:#f3e5f5
style E fill:#ffebee
📦 ECG_IdentificationX
┣ 📂 ECG_data/ # Raw data (A-F, total 2558 beats)
┣ 📂 src/
┃ ┣ 📂 preprocessing/ # Signal preprocessing
┃ ┣ 📂 feature_extraction/ # Feature extraction
┃ ┣ 📂 models/ # Deep learning models
┃ ┗ 📂 utils/ # Utility functions
┣ 📂 docs/ # Technical docs
┣ 📄 app.py # Web interface
┣ 📄 train.py # Training script
┗ 📄 .env # API config
flowchart LR
A["Raw ADC"] --> B["DC removal"] --> C["50Hz notch"] --> D["Band-pass filtering"] --> E["Wavelet denoising"] --> F["Clean ECG"]
style A fill:#ffcdd2
style F fill:#c8e6c9
flowchart LR
A["ECG"] --> B["Band-pass 5-15Hz"] --> C["Differentiation"] --> D["Squaring"] --> E["Integration"] --> F["Thresholding"] --> G["R peaks"]
style A fill:#e3f2fd
style G fill:#c8e6c9
| ⏱️ Time Domain | 📊 Frequency Domain | 🔀 Nonlinear |
|---|---|---|
| SDNN | VLF | SD1/SD2 |
| RMSSD | LF | ApEn |
| pNN50 | HF | SampEn |
| pNN20 | LF/HF | DFA |
Input[175] → Conv(32,k7) → Conv(64,k5) → Conv(128,k3) → GAP → FC(64) → Output[6]
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| Chip | ADS1292R | Sampling rate | 250Hz |
| ADC | 24-bit | Gain | 6x |
| Reference voltage | 2.42V | Total beats | 2,558 |
ADC conversion formula:
| Category | Technology |
|---|---|
| 🐍 Language | Python 3.10+ |
| 🧠 Deep Learning | PyTorch 2.0+ |
| 📊 Signal Processing | SciPy, NumPy, PyWavelets |
| 🤖 AI | DeepSeek API |
| 🌐 Web | Streamlit, Plotly |
| 📄 Document | 📝 Content |
|---|---|
| 📘 System Overview | Architecture, data specs |
| 📗 Signal Preprocessing | Filtering, denoising |
| 📙 R-peak Detection | Pan-Tompkins |
| 📕 Feature Extraction | HRV, MFCC |
| 📓 Deep Learning | CNN architecture |
| 📔 Training & Evaluation | Training strategy |
| 📒 Performance Summary | Result analysis |
This project is licensed under the MIT License