Abstract
BACKGROUND: Myasthenia gravis (MG) is a rare disorder characterized by fluctuating muscle weakness with potential life-threatening crises. Timely interventions may be delayed by limited access to care and fragmented documentation. Our objective was to develop predictive algorithms for MG deterioration using multimodal telemedicine data. METHODS: In this 12-week randomized controlled study, 30 MG patients recorded symptoms using patient-reported outcome measures (PROMs) and patient-performed measures via a mobile app, alongside data from wearables. MG deterioration was defined as a ≥ 3-point worsening in the Quantitative Myasthenia Gravis score, occurrence of MG-related hospitalization or exacerbation. A machine learning linear classifier was trained to predict deterioration and cross-validated. The area under the receiver operator characteristic curve (AUROC) was calculated, accepting 1-2 false alarms (FAs) per week. RESULTS: The model achieved the best predictive performance when using all input signals (AUROC 0.85 (95% confidence interval 0.77-0.91)) and remained stable across look-back windows of 4-10 days. Model sensitivity was 0.65 (0.48-0.83) to 0.82 (0.60-1.00) (1 and 2 FAs per week, respectively). PROMs reflected worsening symptoms before deterioration; wearables alone showed higher AUROCs. INTERPRETATION: Multimodal self-monitoring via MyaLink predicted MG deterioration with good performance at acceptable FA rates. This approach may enable earlier clinical interventions of MG worsening. TRIAL REGISTRATION: The study was registered under the German Clinical Trial Registry (DRKS00029907).