Technical Detail
Training and Evaluation Methodology
This page summarizes the implemented training pipeline, model architecture strategy, and deployment setup used for the ECG arrhythmia classification system.
Data and Label Construction
- Dataset source: PTB-XL public ECG corpus with 12-lead waveforms and metadata.
- SCP code aggregation transforms raw statements into five diagnostic superclasses.
- Signals are loaded at configured sampling rates and standardized per training workflow.
- Metadata features include age, sex, height, weight, infarction staging, and pacemaker status.
Modeling Strategy
- Model01: metadata-only dense network for baseline comparisons.
- Model02: dual-input architecture (metadata branch + Conv1D ECG branch).
- Model03: augmentation-enhanced variant using windowing and additive noise.
- All models are trained as multi-label classifiers with sigmoid outputs.
Pipeline Components
01
Preprocessing
Metadata engineering, target creation, standardized scaling, and split preparation.
02
Training
TensorFlow/Keras training with early stopping and model checkpoint tracking.
03
Evaluation
Held-out set performance measurement using shared post-processing logic.
04
Deployment
Streamlit inference interface with embedded web presentation for review sessions.
Reproducible Training Commands
Representative commands from the project workflow:
# prepare and train all models
python train.py --data-path /path/to/ptbxl --model all
# train the best-performing architecture
python train.py --data-path /path/to/ptbxl --model model02
# evaluate trained checkpoints
python evaluate.py --model all
Use Boundary
This platform is presented as a computational cardiology research demonstrator.
It does not replace clinical interpretation, formal diagnosis, or physician-led decision workflows.