0804 Multimodal Detection of Narcolepsy Type 1 Using Genetic and Nocturnal Polysomnography Data

Abstract Introduction Efficient screening methods with high specificity which can be applied to large sample sizes are needed to improve significantly underdiagnosed people with narcolepsy type 1 (NT1). To address this unmet need, we combined polygenetic risk scores (PRS), HLA typing, and nocturnal polysomnography (PSG) data. These methods are developed and validated using a dataset significantly larger than those used in previous studies, ensuring their reliability and generalizability. Methods People with diagnosed narcolepsy were studied across 7 different global locations, involving multiple collaborators (coauthors limited by guidelines). We developed a custom U-Sleep model for variable-resolution sleep staging using electroencephalography, electrooculography, and electromyography data from 19,381 PSG recordings from the National Sleep Research Resource and the Stanford Sleep Clinic. Sleep staging was performed across epochs from 0.25 to 3600 seconds, leveraging novel multiscale transition matrix (MTM) features to capture NT1-specific transitions and overlapping sleep stages. Next, a Gaussian processes classifier was trained on features from 21,846 PSGs across 14 cohorts (NT1: n=327; controls: n=21519) and tested on 634 separate PSGs (NT1: n=317; controls: n=317). An ensemble model combined multiple time resolutions, weighted by prediction accuracy. Classifier performance was evaluated using 5-fold cross-validation. We developed a multi-information model integrating new PRS scores with HLA DQB1*06:02 typing, where HLA- and PRS+ conditions map to control and NT1 prediction, respectively, with PSG used when neither applies. Results Adding PRS to HLA increased the specificity from 81.9% to 100.0%, with a sensitivity of 25.9%, enabling potential screening of NT1 using genetics alone. When combining PSG with HLA, our approach yielded 99.4% specificity and 94.6% sensitivity, surpassing State-of-the-Art (SOTA). Adding PRS further rescued HLA+ cases missed based off PSG, and increased sensitivity to 95.9% while maintaining 99.4% specificity. MTM features meaningfully boosted classification performance (approximately 1% gain) compared to SOTA features, which themselves significantly outperformed standard clinical PSG measures. Ensemble models showed that combining multiple resolutions improved performance, notably specificity. Additionally, post-hoc analyses show that genetic information noticeably improved model robustness. Conclusion These findings suggest that the proposed multimodal model is a robust framework for NT1 classification, providing more specific and sensitive screening methods. Support (if any) Funded by Takeda Development Center Americas, Inc.