Classification of narcolepsy type 1 using machine learning on Stanford Cataplexy Questionnaire responses and...

Area of Science:

• Neurology
• Sleep Medicine
• Computational Biology

Background:

• Narcolepsy Type 1 (NT1) presents with sleepiness and cataplexy, but diagnosis is difficult due to low prevalence and reliance on polysomnography.
• Cataplexy, a sudden loss of muscle tone triggered by emotions, is a key symptom of NT1.
• Developing scalable screening tools for NT1 is crucial for accurate diagnosis in large populations.

Purpose of the Study:

• To investigate the efficacy of a brief cataplexy questionnaire in supporting scalable screening for Narcolepsy Type 1 (NT1).
• To evaluate the performance of machine learning (ML) classifiers using cataplexy features and HLA typing for NT1 diagnosis.

Main Methods:

• Trained six ML classifiers on data from 280 NT1 patients and 927 controls, utilizing 10- and 27-item cataplexy feature sets and the Epworth Sleepiness Scale (ESS).
• Incorporated HLA-DQB106:02 typing and employed nested cross-validation with Optuna optimization.
• Implemented a post hoc veto rule to enhance specificity by reclassifying subjects lacking the HLA-DQB106:02 allele as non-NT1, prioritizing specificity due to NT1's low prevalence.

Main Results:

• ESS alone achieved an AUC of 0.863 with 38.9% sensitivity and 92.1% specificity.
• ML models using cataplexy features (with or without HLA typing) achieved high AUCs of 0.995-0.996.
• Inclusion of HLA typing significantly improved specificity, reaching up to 99.2%, effectively reducing false positives.

Conclusions:

• Machine learning models integrating cataplexy questionnaire data and HLA typing provide a highly accurate and scalable method for NT1 screening.
• Further validation in larger, population-based studies is recommended to confirm these promising findings.