Clinical Context¶
Narcolepsy type 1 (NT1) is underdiagnosed due to reliance on costly and invasive diagnostic procedures. This project explores whether machine learning models trained on cataplexy questionnaires, optionally combined with HLA-DQB1*06:02, can enable scalable screening.
Dataset¶
| Group | N |
|---|---|
| NT1 Patients | 280 |
| Controls | 927 |
All cases were reclassified under ICSD-3-TR criteria.
Feature Sets Evaluated¶
- Emotional triggers (laughter, joking, anger)
- Muscle weakness localization
- Epworth Sleepiness Scale (ESS)
- Optional HLA-DQB1*06:02 genotype
Model Architecture¶
Six classifiers were evaluated using nested cross-validation and Optuna-based hyperparameter optimization:
- SVM
- Elastic Net
- LDA
- LightGBM
- XGBoost
Optimization prioritized specificity, reflecting real-world screening constraints.
Performance Summary¶
| Feature Set | AUROC | Specificity |
|---|---|---|
| ESS only | 0.86 | 92.1% |
| Reduced + HLA | 0.995 | 99.0% |
| Full + HLA | 0.996 | 99.2% |
A post hoc HLA veto rule further eliminated false positives.
Impact¶
This approach enables:
- Low-cost population screening
- Reduced unnecessary referrals
- Integration into large biobank-scale studies
The framework is directly transferable to other low-prevalence neurological conditions.
References¶
- Ricciardiello et al., Optimizing Machine Learning Classification of Narcolepsy Type 1, Manuscript in preparation.