Deep Learning · Time Series · 2026
Temporal Characterization and Deep Learning-Based Modeling of Human Sleep
Modelled temporal sleep physiology from polysomnography signals and compared deep learning against gradient-boosted sequence models for sleep-stage prediction.
PythonBi-LSTMXGBoostTime Series AnalysisVARSleep-EDF
84%
Bi-LSTM accuracy
0.65
Bi-LSTM F1
74%
XGBoost accuracy
0.57
XGBoost F1
Model performance: Bi-LSTM vs. XGBoost
Problem / Research Question
Sleep staging from raw physiological signals is a sequence-classification problem where cross-signal timing matters as much as signal amplitude. The project asked whether a sequence-aware deep learning model could out-perform a strong gradient-boosted baseline on this task.
Dataset
Multivariate EEG, EOG, and EMG channels from polysomnography recordings in the Sleep-EDF dataset.
Methodology
- 01Analyzed temporal sleep physiology using multivariate EEG, EOG, and EMG signals.
- 02Modelled cross-signal interactions and temporal dependencies using Vector Autoregression (VAR).
- 03Built and compared a Bi-LSTM sequence model against an XGBoost baseline for sleep-stage classification.
Models
Bi-LSTMXGBoost
Key Findings
- The sequence-aware Bi-LSTM outperformed the XGBoost baseline on both accuracy and F1, consistent with sleep-stage transitions carrying temporal structure that a non-sequential model can't fully capture.
- VAR modelling of cross-signal dependencies (EEG/EOG/EMG) supported the case that sleep stages are a multivariate temporal phenomenon rather than a per-timestep classification problem.
Limitations
- F1 scores (0.65 and 0.57) are meaningfully lower than raw accuracy, indicating class imbalance across sleep stages affects both models, so accuracy alone would overstate performance.
- Results are reported on Sleep-EDF; generalization to other recording setups or populations wasn't evaluated in this project.