R-Peak Detection

Below is an overview of our ECG R-peak detection and classifier training pipeline.

Instruction

1. Download & unzip the PTB-XL dataset (v1.0.3) from PhysioNet.
2. Copy your rpeak_classifier.py script into the root of the unzipped folder:(ptb-xl-a-large-publicly-available-electrocardiography-dataset-1.0.3)
3. cd into that folder:

cd ptb-xl-a-large-publicly-available-electrocardiography-dataset-1.0.3
pip install wfdb numpy scipy pandas torch scikit-learn matplotlib(or any other related module)
python rpeak_classifier.py

4. follow instruction type in the sample rate, press enter to run

Data Preprocessing

1. Lead II Extraction

   • Read raw .dat/.hea files with wfdb.rdsamp(...).
   • Select channel 1 (Lead II).

2. Resampling

   • Prompt for a target_fs (default 500 Hz).
   • Use scipy.signal.resample() to map any original sampling rate to target_fs.

3. Filtering

   • Apply a zero-phase Butterworth band-pass (0.5 – 40 Hz) via

     nyq = 0.5 * fs
     b, a = butter(order=3, [0.5/nyq, 40/nyq], btype='bandpass')
     filtered = filtfilt(b, a, signal)

4. Pan–Tompkins QRS Detection

   • Band-pass 5 – 15 Hz
   • Differentiate → Square → Moving-window integrate (~150 ms)
   • scipy.signal.find_peaks() with a min distance of 250 ms and adaptive threshold

5. Peak Refinement

   • For each candidate, search ±10 samples in the raw signal to locate the true local maximum.

6. Window Extraction

   • Positive windows: 128 samples centered on each refined R-peak.
   • Negative windows: sample outside ±64 samples around true peaks (neg_ratio=5).

---

Model Architecture

Input: 1 × 128 window
 └─ Conv1d(1→16, kernel=5, padding=2)
     └─ ReLU
     └─ MaxPool1d(2)
 └─ Conv1d(16→32, kernel=5, padding=2)
     └─ ReLU
     └─ MaxPool1d(2)
 └─ Flatten → Linear((128/4)*32 → 64)
     └─ ReLU
     └─ Linear(64 → 1)
     └─ Sigmoid
Output: P(R-peak at center)

Hyperparameter Summary

• Sampling Rate (target_fs): 100 – 500 Hz (you choose at runtime; defaults to 500 Hz)
• Window Size: 128 samples (~1.28 s at 100 Hz)
• Negative-to-Positive Ratio (neg_ratio): 5 negative windows for every true R-peak window
• Batch Size: 64 windows per training step
• Learning Rate: 1 × 10⁻³ (Adam optimizer)
• Epochs & Early Stopping: up to 30 epochs with patience=5 on validation F₁
• Detection Threshold: 0.85 probability for classifying a window as “R-peak”
• Minimum Peak Distance: 0.3 s (distance=int(0.3 × fs) samples) between successive detections

Training logs

the training result is saved in classifier_rpeak_testset.csv

SCP codes Detection

Below is an overview of our ECG scp codes detection classifier training pipeline.(Due to limitation of hardware, only trained for 5 labels here)

Instruction

1. Download & unzip the PTB-XL dataset (v1.0.3) from PhysioNet.
2. Copy your rpeak_classifier.py script into the root of the unzipped folder:(ptb-xl-a-large-publicly-available-electrocardiography-dataset-1.0.3)
3. cd into that folder:

cd ptb-xl-a-large-publicly-available-electrocardiography-dataset-1.0.3
pip install wfdb numpy scipy pandas torch scikit-learn matplotlib(or any other related module)
python scp_code_classifier.py

Data Processing

1. Metadata & Label Extraction

• Load ptbxl_database.csv and parse each record’s scp_codes dictionary.
• Filter to the top-5 diagnostic classes (NORM, IMI, ASMI, LVH, NDT), discarding any ECGs with none of these labels.
• Convert each record’s label set into a fixed-length multi-hot vector for training.

2. Train / Validation / Test Split

• Folds 1–8 → Training set
• Fold 9 → Validation set
• Fold 10 → Test set

3. ECG Signal Preprocessing

• Load raw Lead II waveforms from the .dat/.hea files (originally at 500 Hz).
• Resample to the user-specified target rate (e.g. 100, 250, or 500 Hz).
• Band-pass filter (0.5–40 Hz, zero-phase Butterworth) to remove baseline wander and noise.

4. Snippet Windowing

• Window length: 10 seconds (i.e. target_fs × 10 samples).
• Pad shorter signals with zeros; truncate longer signals to ensure uniform length.

5. Dataset & DataLoader

• Wrap each filtered snippet and its multi-hot label into a PyTorch Dataset.
• Use DataLoader to batch, shuffle, and feed data into the model during training, validation, and testing.

Hyperparameters

• Sampling Rate (target_fs): User‐specified (100, 250, or 500 Hz; default 500 Hz)
• Window Length: 10 s of ECG (window_size = target_fs × 10)
• Batch Size: 32
• Optimizer: Adam
• Learning Rate: 1 × 10⁻³
• Max Epochs: 30
• Early Stopping Patience: 7 epochs (stop if validation F₁-micro does not improve)
• Convolutional Blocks: 5 layers, each with
  • Kernel size = 7, padding = 3
  • Filters = [32, 64, 128, 128, 256]
  • BatchNorm → ReLU → MaxPool1d(2)
• Dropout: 0.5 (applied before and after the 256-unit dense layer)
• Classifier Head:
  • Dense(Flattened → 256) → ReLU → Dropout
  • Dense(256 → 5) → Sigmoid
• Inference Threshold: 0.5 (probability cutoff for each SCP code)
• Metrics:
  • AUROC & AUPRC (macro)
  • F₁-micro & F₁-macro
  • Accuracy@3
  • Brier score

Training logs

the training result is saved in scp_code_predictions.csv