TinyBioNet: A lightweight time-frequency network for biomedical signal classification on edge devices

This repository provides a pipeline for training a quantized deep learning model on physiological signals (PPG, ACC, ECG, EDA, EMG) using STFT preprocessing and sliding window segmentation. The model is designed to support multiple datasets, cross-validation, weight quantization, and TFLite INT8 deployment.

---

Features

• Multiple datasets supported:
  • AffectiveROAD (low, medium, high)
  • PPG_ACC (rest, squat, step)
  • WEASAD (baseline, stress, amusement, meditation)
• Flexible input selection:
  • PPG, ACC, ECG, EDA, EMG, or combined signals
• Sliding window segmentation with overlap
  • Automatically handles averaging or majority vote for target labels
• STFT preprocessing to transform time-domain signals into time-frequency representations
• Weight quantization with custom bitwidth (default 8-bit)
• Fine-tuning quantized models
• Full INT8 TFLite conversion with representative dataset
• Stratified K-Fold Cross-Validation for robust evaluation
• Metrics:
  • Accuracy
  • Weighted F1-score
  • Detailed classification report