Sleep Stage Classification Using Non-Invasive Signals from Thermal Imagery

Poster Presentation

(click on the image for higher resolution)

📌 Motivation

Accurate identification of sleep stages is crucial for health assessment and early disease detection. However, manual scoring is labor-intensive and invasive. While current state-of-the-art ML/DL models achieve high accuracy, they rely on EEG, EOG, and EMG signals which are invasive and cannot be derived from thermal imagery. This highlights the need for non-invasive alternatives for sleep stage classification

This project aims to:

1. Evaluate whether machine learning models using non-invasive signals can approach the performance of EEG-based models.

2. Analyze key EEG features (e.g., K-complexes, sleep spindles, α-rhythms, sawtooth waves) to understand their critical role as defined by AASM standards.

Exploratory Data Analysis (EDA)

Distribution of sleep stages in the dataset

Sleep Stages for Subject 5 over time (a) 6 sleep stages (b) 2 sleep stages

Flow and Ribcage Signals(Non-invasive) for subjects

We observe the need to normalize the signals and apply smoothening

Preprocessing using Savitsky-Goalsy filter with various hyperparameters for window length and polynomial order

Normalizing and smootening data

⚙️ Methodology

Preprocessing: Normalization, smoothing (Savitzky–Golay filter).

Feature Extraction (3 ways are experimented here):

1. Using TSFEL
2. CNNs
3. hand‑crafted statistical & physiological features.

Modeling: Decision Tree based(XGboost, Random Forests), sequence‑based (BiLSTM), and attention‑based (Transformer) architectures.

Evaluation: Accuracy and specificity on 2‑, 3‑, and 4‑class sleep staging tasks.

📊 Why Specificity Matters

High specificity means fewer false positives in medical diagnostics:

Better resource allocation: Focus interventions (e.g., CPAP therapy) on truly affected patients.

Greater clinical confidence: Avoid unnecessary treatments caused by false alarms.

✅ Conclusions

Non-invasive signals can be used for simple Sleep/Wake classification in smartwatches, etc as accuracy is \textbf{0.83}.

Feature engineering improves specificity of the classes significantly. Specificity is really important especially in medical domain, as it tells us how confident we are that the negatives predicted for a class are actually negative. For example:

Medical Diagnosis: High specificity means that when a model says that a patient is not having a particular sleep disorder (e.g., obstructive sleep apnea), clinicians can confidently exclude that diagnosis. This prevents unnecessary treatments, medications, or interventions that carry their own side effects and costs.

Resource Allocation: In sleep medicine, treatments like CPAP therapy are resource-intensive. High specificity reduces false positives, ensuring these resources are directed to patients who truly need them. EEG signals are critical for multi-stage classification (REM vs NREM, Light vs Deep, etc). but XGboost with appropriate feature extraction can be promising.

📚 References

AASM Scoring Manual (2017)

Mohammadi Foumani et al., Data Mining and Knowledge Discovery, 2023

Lee et al., IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2024