OSAS Detection System
Leakage-aware transformer research across 961,357 physiological observations
An end-to-end research pipeline for detecting and classifying sleep-apnea events from vital signs, ECG/PPG waveforms, and polysomnography signals using multimodal transformer architectures.
The 15-second brief
At a glance
- Outcome
- Converted roughly 961,357 physician-labelled observations from 30 stroke patients into about 31,986 training windows for binary, five-class, and multitask OSAS experiments.
- Delivery
- Research, 2025
- Focus
- ML, Research
- Scope
- 7 core technologies
Ownership and evidence
What I built
The product decisions and engineering work that shaped the final result.
Fused five vital-sign features, four ECG/PPG waveform channels, and five PSG signal groups through modality-specific encoders.
Generated 60-second windows with 50% overlap and discarded windows exceeding 50% missing data.
Prevented patient leakage with patient-aware train/validation/test splits across the 30-person cohort.
Addressed an approximately 87/13 normal-to-anomaly imbalance with weighted losses, focal loss options, stratification, and macro-F1 model selection.
Supported binary anomaly detection, five-class event classification, and joint multitask training with checkpointed evaluation outputs.
Product evidence
See it in action
Interface views and demonstrations from the working product.
Visual assets are not published for this project. The verified delivery and engineering scope are documented above.