0822 AI-Enabled Narcolepsy Type-1 Screening with PPG: A Proof-of-Concept Study Free

Abstract

Introduction

While highly specific, the current diagnostic paradigm for type 1 narcolepsy (NT1), defined largely by sleep-onset REM period event (SOREMP) observations in PSGs and MSLTs, remains limited by its NT1 disorder sensitivity and procedural complexity. Machine learning methods have shown promise to accurately detect NT1 from PSG via comprehensive assessment of EEG biomarkers. Photoplethysmography (PPG), used for home sleep apnea testing, can determine sleep stages through machine learning without EEG. We explore whether such PPG-based sleep stages are robust enough to detect sleep architectural abnormalities specifically associated with NT1.

Methods

The dataset included a total of N=110 patients, N=35 positive NT1 patients and N=75 negative NT1 subjects. The negative NT1 patients were composed of N=61 confounding hypersomnolence disorders and N=14 negative controls. We evaluated four separate input data types, trained with stratified 10-fold cross-validation with supervised random forest machine learning (ML) models for NT1 detection. The following 4 input data types are features derived from ML models applied for automated sleep staging of EEG and PPG signals extracted respectively from overnight PSG: EEG-based sleep stage report indices (EEG-Stage), PPG-based sleep stage report indices (PPG-Stage), EEG-based hypnodensity derived features (EEG-Hypno), and PPG-based hypnodensity derived features (PPG-Hypno). To measure performance, we calculated the area under the receiver operating characteristic curve (ROC-AUC) for each model and performed a feature importance analysis for all models.

Results

ROC-AUC values were 0.889 and 0.843 for the EEG-Hypno and PPG-Hypno models, respectively. Furthermore, ROC-AUC values were 0.813 and 0.842 for the EEG-Stage and PPG-Stage models. Feature importance analyses for the EEG-Stage model revealed the highest-ranking features: sleep latency, total N3 time, N3 prevalence, total sleep time, and REM latency. Feature importance analysis for the PPG-Stage model revealed these highest-ranking important features comparatively: sleep latency, REM latency, total N3 time, and N3 prevalence.

Conclusion

ML methods automatically detected NT1 in PPG with comparable degrees of accuracy to EEG. The PPG sensor offers a simple and accessible modality in ecologically valid home settings. This method demonstrates potential extensions for screening of NT1 in HSATs, whereby patients with NT1-associated sleep architectural characteristics may be flagged for further hypersomnolence disorders evaluation and testing.

Support (if any)