Classification of narcolepsy type 1 using machine learning on Stanford Cataplexy Questionnaire responses and HLA-DQB1∗06:02.

BACKGROUND

Narcolepsy type 1 (NT1) is characterized by sleepiness, disturbed sleep, and cataplexy-episodes of sudden muscle tone loss triggered by emotions. Accurate diagnosis in large-scale studies remains challenging due to NT1's low prevalence and reliance on polysomnography. Prior studies have shown that cataplexy-related questions can help identify genuine cataplexy. This study investigates whether a brief cataplexy questionnaire can support scalable NT1 screening.

METHODS

We analyzed responses from 280 NT1 patients and 927 controls from the Stanford narcolepsy database. We trained six ML classifiers using 10- and 27-item cataplexy feature sets (emotional triggers, affected muscles), Epworth Sleepiness Scale (ESS) with and without HLA-DQB106:02 typing. Model was trained with nested cross-validation and Optuna optimization. A post hoc veto rule reclassified any subject predicted as NT1 but lacking the HLA-DQB106:02 allele as non-NT1 to reduce false positives when HLA was not part of the feature set. Given the low prevalence of NT1, optimization prioritized specificity.

RESULTS

ESS obtained an AUC of 0.863 (95% CI: 0.840-0.885) with sensitivity of 38.9% and specificity of 92.1%, at τ=0.5. The reduced feature set with (k = 11) and without (k = 10) HLA obtained similar AUCs of 0.995 (95% CI: 0.993-0.998), at τ=0.5. The full feature set with (k = 28) and without (k = 27) HLA obtained a similar AUCs of 0.996 (95% CI: 0.995-0.998), at τ=0.5. Inclusion of HLA corrected false positives increasing specificity up to 98.8%, 99.0% and 99.2%, respectively.

CONCLUSIONS

ML applied to cataplexy questionnaire and HLA typing enables scalable NT1 screening. Further population-based validations are needed to confirm these findings using larger samples of controls.