0627 Improved Primary CNS Hypersomnia Diagnosis With Statistical Machine Learning

Abstract

Introduction: The MSLT is the current gold standard for diagnosis of the primary CNS hypersomnias. While current thresholds for defining a positive MSLT are sufficient for a diagnosis of NT1, the arbitrary diagnostic thresholds result in poor test-retest reliability and differentiation of other hypersomnias. In order to determine whether better differentiation of the primary CNS hypersomnias - narcolepsy type 1 (NT1), narcolepsy type 2 (NT2), idiopathic hypersomnia (IH) - is possible, data from the preceding polysomnogram (PSG) was added to the multiple sleep latency test (MSLT), with newly defined thresholds. Methods: Cases from the world's largest hypersomnia database at the Stanford Narcolepsy Center were combined with a control population derived from the Wisconsin Sleep Cohort. NT1 vs NT2 status was defined strictly by CSF hypocretin (<110 pg/mL considered diagnostic for NT1). Adjustment for confounders was performed, where possible. Multiple, machine-learning models (stepwise multinomial logistic regression, decision trees, random forest, boosting, and recursive portioning and regression trees) were compared using total and category-specific classification accuracies. Results: For classification accuracies in the validation set, stepwise regression performed the best (0.95 vs 0.83-0.88 for other models), and was the only model that had good category specific accuracy for NT1 (0.91 vs 0.58-0.66 for other models). All models performed poorly in classifying IH (accuracies 0.5-0.67). In addition to expected MSLT features, new features of interest from the preceding PSG (e.g. total sleep time, N2 percent) improved the ability to differentiate hypersomnias. Conclusion: All of these models perform excellently at categorization (well above the 25% accuracy expected for chance, with 4 categories), by incorporating more clinical information that is already part of the workup, but at different thresholds than are the current standard. Persistent challenges with differentiating IH from controls and NT2 highlights that better diagnostics and/or thresholds are needed going forward.