Yoav N. Nygate, MS1 • Sam Rusk, BS1 • Chris R. Fernandez, MS1 • Zac Winzurk, BS1 • Emerson M. Wickwire, PhD2 • Emmanuel Mignot, MD, PhD3 • Nathaniel F. Watson, MD, MS4
Introduction
Accurate diagnosis of Type 1 narcolepsy (T1N) is cumbersome – involving clinical, biological, and electrophysiological components. Multiple sleep latency tests (MSLT) are central to diagnosis; however, current medications, sleep schedule, drug use, and testing environment can all compromise the interpretation of an MSLT. Incorrect MSLT results have devastating consequences for patients, including prolonged suffering from untreated T1N or adverse effects from unnecessary testing and therapies. Simple, accessible, accurate and cost-effective diagnostic solutions are needed. We evaluate automated methods to expand the diagnosis of T1N based on single night polysomnography (PSG).
Methods
The dataset included N=225 T1N patients and N=455 negative controls from 6 different cohorts. We evaluated two separate machine learning (ML) models for T1N detection: a supervised PSG-EEG based deep learning (PSG-DL) model focused on sleep microarchitecture, and a random forest (Sleep-RF) model based on traditional PSG report data. To measure performance, we calculated the area under the receiver operating characteristic curve (AUC-ROC) for each model and performed a feature importance analysis for the Sleep-RF model.
Results
AUC-ROC values were 0.960 and 0.878 for the PSG-DL and Sleep-RF models, respectively. Furthermore, the PSG-DL model produced a sensitivity and specificity of 84% and 94%, while the Sleep-RF model resulted in 75% and 83%. The feature importance analysis for the Sleep-RF model revealed these most important features: REM latency, sleep latency, total N3 time, and ratio of arousals in REM vs. non-REM.
Conclusions
Machine Learning [EW1] methods automatically detected T1N in PSG-EEG with promising degrees of accuracy. These methods overcome common barriers to accurate diagnosis that can compromise the interpretability of an MSLT. Broad implementation of this method has potential to supplement the MSLT, increasing diagnostic detection rates and accuracy for T1N. Access to care can be broadened if T1N diagnosis were to occur anywhere a sleep EEG can be obtained, including the home setting, rather than being limited to in-center testing. This work has potential to further the developments of T1N detectors and assist in their widespread clinical adoption. Additional research is underway to help improve accuracy and ensure these methods are generalizable across platforms and clinical datasets.
