💤 Sleep Stage Prediction Using Consumer Wearable Data

📌 Project Overview

This project explores whether consumer-grade smartwatches (specifically the Apple Watch Gen 1) can reliably emulate the sleep detection accuracy of clinical-grade devices such as the Phillips Actiwatch. Using accelerometer-derived ENMO data, a custom transformation and weighted classification algorithm were used to determine sleep vs. wake states.

Goal: Convert ENMO data into a clinical-equivalent format to classify sleep accurately using a transparent, interpretable ML pipeline.

---

🧠 Problem Context

Clinical sleep monitoring is effective but costly, making it inaccessible to many. This project applies data science and machine learning to create a low-cost, scalable solution using data from commercial wearables — opening up possibilities for future real-time health insights.

---

👥 Authors

• Alex Conroy
• Nebojsa Ajdarevic
• Karen Hau

Group Project — IFN646 Biomedical Data Science, Queensland University of Technology (QUT)

---

🧾 Data Source

• Study Name: ENMO-based Sleep Classification
• Participants: 14 individuals across 27 nights
• Epoch Duration: 15 seconds
• Features:
  • Timestamp
  • Clinical Activity Count
  • Sleep/Wake Label (ground truth)
  • ENMO Value from Apple Watch

🗂 Dataset: DOI via QUT

---

🔬 Methods

Data Transformation

• Applied linear regression to convert ENMO values into Phillips Actiwatch-equivalent activity counts.

Sleep Classification

• Implemented a weighted moving average algorithm mimicking clinical-grade classification.
• Adjusted thresholds to accommodate the ENMO-transformed values.

Evaluation

• Compared predicted sleep labels to ground truth using classification metrics.

---

📈 Results

Metric	Score
Accuracy	96.4%
Recall	83.5%
Precision	89.8%
F1 Score	86.5%
Specificity	98.5%

✅ Shows ENMO-based sleep classification is highly accurate for normal sleep behaviour.
⚠️ Less generalisable to fragmented or non-nighttime sleep sessions.

---

🗂️ Project Structure

sleep-stage-prediction/ ├── data/ │ └── combined_csv.csv # Aligned dataset with clinical labels & ENMO features │ ├── notebooks/ │ └── Wearables_v5.ipynb # End-to-end exploratory analysis and evaluation │ ├── src/ │ └── model.py # Classification model logic and evaluation functions │ ├── utils/ │ ├── processing.py # Data cleaning and ENMO-to-activity transformation │ └── visualisation.py # All visualisation functions (e.g., time series, histograms) │ ├── report/ │ ├── Project report FINAL.pdf # Full research report │ └── Presentation.pdf # Project summary slides │ ├── requirements.txt # Python dependencies └── README.md # Project summary and documentation