Wearable Multimodal Sleep Stage Classification via Dynamic Spectral–Temporal Feature Fusion

Overview

Sleep quality is closely associated with overall health. Disorders such as sleep apnea, insomnia, and abnormal REM activity are linked to cardiovascular and stress-related diseases.
Traditional sleep-stage evaluation relies on polysomnography (PSG), which is not suitable for long-term home monitoring due to its invasiveness and complexity.

This project proposes a wearable-based sleep stage classification algorithm using ECG, PPG, and IMU signals.
The algorithm extracts dynamic spectral–temporal features to classify four sleep stages:

• Deep Sleep (NREM3)
• Light Sleep (NREM1–2)
• REM Sleep
• Wake

---

Key Features

1. Multimodal Biosignal Fusion

ECG

• R-peak detection and HR estimation
• HRV analysis: SDNN, RMSSD, LF/HF ratios
• Autonomic nervous system activity indicators

PPG

• Pulse morphology (rise time, reflection index)
• Respiratory-synchronized amplitude modulation (RSA)
• Low/High frequency spectral ratios

IMU

• Sleep posture detection (left/right/supine)
• Micro-motion analysis related to respiration
• Activity Index within 30-second windows

---

2. Dynamic Time–Frequency Feature Extraction

• STFT and Wavelet transforms applied to ECG/PPG
• Power spectral density estimation
• Frequency band energy ratio analysis
• IMU-based separation of low-frequency breathing vs. high-frequency motions
• Final feature vector combining HRV + PTT + RSA + motion metrics

---

3. Rule-Based or Lightweight ML Classification

Sleep Stage Classification Rules:

• Deep Sleep (NREM3): high HRV, minimal motion, stable PPG
• Light Sleep (NREM1–2): moderate HR, small movement
• REM: high HR variability, irregular breathing, micro-movements
• Wake: high HR, strong movement amplitude

Supports:

• Heuristic rule-based logic
• Lightweight classifiers (KNN, LDA, etc.)

---

4. Real-Time & Offline Analysis

• Real-time streaming evaluation every 30 seconds
• Offline processing of long-term data
• Export to CSV or MAT
• Hypnogram visualization

---

System Workflow

1. Wearable Device Setup

• ECG electrodes
• Wrist PPG sensor
• IMU-based motion sensor
• 6–8 hours of overnight recording

2. Data Acquisition & Preprocessing

• Sampling rates:
  • ECG: 128 Hz
  • PPG: 64 Hz
  • IMU: 32 Hz
• Noise removal using moving average & band-pass filtering

3. Dynamic Window Analysis

• 30–60 second windows
• Extract features: HR, HRV, PTT, RSA, posture, movement
• STFT / Wavelet for spectral analysis

4. Sleep Stage Classification

• Feature fusion: HRV + PPG + IMU metrics
• Rule-based or ML-based classification
• Output labels: Deep / Light / REM / Wake

5. Visualization & Storage

• Hypnogram generation
• Export abnormal breathing event logs
• Store raw, filtered, and processed features

---

Applications

• Clinical research: PSG alternative/auxiliary indicators
• Consumer wearable devices
• Sleep apnea / hypopnea detection
• Mobile healthcare apps (fatigue & stress index estimation)

---

Expected Benefits

• Non-invasive and low-cost long-term sleep monitoring
• Higher accuracy through multimodal ECG–PPG–IMU fusion
• Real-time abnormal breathing detection
• Suitable for continuous home monitoring

---

Example Repository Structure