0686 Further Development of Photoplethysmography (PPG) Based Narcolepsy Type-1 Detection for HSATs

Abstract Introduction Narcolepsy Type-1 (NT1) is a chronic, rare, debilitating neurological disorder characterized by sleepiness, cataplexy, sleep disturbance, and other symptoms. Previous studies demonstrated sleep staging and sleep architectural parameters from home sleep apnea test (HSAT) photoplethysmography (PPG) could identify NT1 with reasonable ROC-AUC. We seek to build on prior work using recent developments in deep learning architectures and training methods to improve the performance and generalizability for identifying NT1 patients with PPG in routine HSATs. Methods We applied deep learning (DL) models to the Mignot Nature Communications (MNC) dataset available through NSRR. This dataset utilized cerebrospinal fluid (CSF) orexin deficiency as the gold-standard for NT1 diagnosis. Two AI models were applied: one using multi-channel EEG inputs and another using only PPG input. 2-Fold cross-validation was applied. First, we tested models on the MNC studies where PPG was available, which included N=37 NT1 patients, N=14 healthy controls, and N=38 non-narcoleptic hypersomnia controls. Second, we tested models on an out-of-distribution sample of multi-night PPG-based HSATs, which included N=19,074 patients with age (mean: 49.3-years, SD=13.6), BMI (mean: 31.3, SD=6.6), and gender (35% female, 65% male). N=105 patients had ≥2 SOREMPS and a Mean Sleep Latency 8 minutes, which were used to establish the NT1-positive labels based on the standard criteria in multiple sleep latency tests (MSLTs). Results The EEG model had a receiver-operating-characteristic (ROC) area-under-the-curve (AUC) of 0.841 in MNC. The PPG model had a ROC-AUC of 0.717 in the same MNC data. Performance of the PPG-only DL model in MNC was further characterized with a 54.5% Sensitivity at the 75% Specificity threshold, 48.6% Sensitivity at the 85% Specificity threshold, and 21.6% Sensitivity at 95% Specificity threshold. In the out-of-sample analysis of N=19,074 multi-night HSATs, an ROC-AUC of 0.629 was observed, corresponding to 46% Sensitivity at the 75% Specificity threshold, and 32.2% Sensitivity at 85% Specificity threshold. While the EEG model achieved the greatest overall performance, PPG-DL appeared to generalize preferably to simpler ML approaches (e.g. Random Forest) in the large-scale retrospective home sleep testing dataset. Conclusion The PPG-based deep learning model showed moderate overall performance for NT1 detection in gold-standard PSG and large-scale HSAT datasets. Support (if any)