PPG-Based Arrhythmia Detection with Deep Learning

This repository combines signal processing (MATLAB) + deep learning (Python RNN/LSTM) to detect atrial fibrillation (AF) from PPG signals. It also includes a Streamlit web interface for easy prediction and visualization.

---

Background: PPG for Arrhythmia Detection

Photoplethysmography (PPG) measures pulsatile blood volume changes using light, offering a low-cost, wearable-friendly window into cardiac rhythm. While ECG is the gold standard for arrhythmia diagnosis, PPG captures the mechanical consequence of each heartbeat at the periphery. This makes it attractive for continuous, unobtrusive screening, especially on the wrist or finger.

Among arrhythmias, atrial fibrillation (AF) is the most suitable target for PPG: it produces irregular pulse intervals and subtle morphology changes (damped or inconsistent dicrotic notch) that PPG can express as pulse rate variability (PRV) and beat-to-beat shape variability. Other rhythm disorders (e.g., PVCs, tachy-/brady-arrhythmias) can also perturb the peripheral pulse, but are harder to separate from motion artifacts and vasomotor effects.

Key advantages:

• Scalability & comfort: works with commodity LEDs/photodiodes; ideal for wearables.
• Rich signal content: waveform morphology encodes vascular tone and timing (systolic upstroke, dicrotic notch, reflection).
• ML-friendly: irregularity and morphology variability are learnable patterns in both time series and spectro-temporal views.

Key challenges:

• Motion artifacts & perfusion changes (especially at the wrist).
• Sensor/skin variability (skin tone, contact pressure, temperature).
• Labeling: definitive rhythm labels usually require synchronized ECG.

---

What “arrhythmia” looks like in PPG

• AF: highly variable inter-pulse intervals (IPI); morphology consistency drops.
• PVCs/ectopy: alternating strong/weak pulses; may mimic motion artifacts.
• Tachy-/Brady-arrhythmia: sustained rate shifts with altered upstroke time.

---

Project Overview

This project aims to:

1. Preprocess PPG signals in MATLAB to remove noise and detect beats.
2. Segment signals into fixed-length windows.
3. Train a deep learning RNN/LSTM in Python to classify AF vs Normal Rhythm.
4. Provide a Streamlit web interface for prediction and visualization.

---

Dataset

• MIMIC PERform AF Dataset

  • Contains ~20-minute PPG + ECG recordings from 35 subjects (19 AF, 16 Normal) sampled at 125 Hz.
  • Labels: AF vs Non-AF.

---

Workflow

1. MATLAB Signal Processing

• Band-pass filter (0.5–8 Hz) to isolate cardiac components.
• Remove baseline drift and normalize amplitude.
• Detect systolic peaks → compute inter-beat intervals (IBI).
• Segment into overlapping windows (e.g., 5 sec with 50% overlap).
• Export windows and labels for deep learning.

2. Deep Learning (Python)

• Load preprocessed windows (.mat or .csv).
• Build an LSTM/RNN to classify AF vs Normal.
• Apply optional augmentation (noise, scaling, time warping).
• Train, validate, and evaluate model.
• Save trained model as .h5 for later use.

3. Streamlit Web Interface

• Accepts MATLAB .mat files or CSV files.
• Preprocess, segment, and classify uploaded PPG signals.
• Shows AF probability per window and plots example windows.
• Allows downloading window-wise AF predictions as CSV.

---

Repository Structure

.
├── matlab/
│   ├── preprocess_ppg.m        # filter, normalize, beat detection
│   ├── detect_beats.m          # systolic peak detection
│   └── export_windows.m        # segment and save windows
├── code/
│   ├── af_rnn.ipynb            # full Jupyter notebook: LSTM training & evaluation
│   ├── example_generate.py     # generate an example csv file 
│   └── ppg_app.py              # Streamlit web interface for predictions
├── data/
│   └── ppg_windows.mat         # preprocessed windows (example)
├── models/
│   └── ppg_af_lstm.h5          # saved trained model
└── README.md

---

Usage

1. Training & Evaluation (Jupyter Notebook)

1. Open af_rnn.ipynb.
2. Load preprocessed windows from data/ppg_windows.mat.
3. Train LSTM/RNN and evaluate performance.
4. Save trained model to models/ppg_af_lstm.h5.

2. Streamlit Interface

cd code
streamlit run app.py

• Upload .mat or .csv PPG signals.
• View AF probability, example window plot.
• Download predictions as ppg_af_predictions.csv.

---

Requirements

• Python 3.9+
• TensorFlow / Keras
• NumPy, Pandas, SciPy, Matplotlib
• Streamlit

pip install tensorflow numpy pandas scipy matplotlib streamlit

---