Foundation Model Forecasting of Headache Days in People With Persisting Postconcussion Symptoms.

Abstract

Objectives: We aimed to forecast headache in individuals with persisting postconcussion symptoms using foundation machine learning (ML) models and multimodal longitudinal data.

Methods: This was an ML analysis of data from the Digital Solutions for Concussion (DiSCo) study, a research project assessing the usability and feasibility of 2 mobile health apps for individuals with persisting postconcussion symptoms. The participants completed daily symptom registrations and daily biofeedback sessions measuring heart rate variability, peripheral skin temperature, and upper trapezius muscle tension. Two foundation ML models, TabPFN and longitudinal TabPFN, were recruited and evaluated to predict moderate-to-severe (at least 4 on the 11-point numeric rating scale) headache days occurring in the subsequent 24-hour window using prior physiological and symptom data. The top-performing model was further evaluated for forecasting days with fatigue, which was rated as the second to headache most bothersome symptom, within the subsequent 24-hour window. Models were trained and optimized on a train-set and evaluated on a hold-out test-set with the area under the receiver operating characteristics curve (AUC) and 95% confidence intervals (CIs).

Results: Twenty individuals were included in the forecasting models, with data collected over a planned 28-day registration period, yielding a total of 338 days. The most stable model, longitudinal TabPFN, achieved a test-set AUC of 0.69 (95% CI: 0.62-0.74). The conventional TabPFN model achieved an AUC of 0.68 (95% CI: 0.55-0.80). In the case of fatigue, the longitudinal TabPFN model achieved an AUC of 0.81 (95% CI: 0.75-0.86). Both physiological measurements and symptomology information seemed to be important predictors.

Conclusions: Moderate-to-severe postconcussion headache days can be predicted with modest accuracy 24 hours before their occurrence from multimodal physiological and self-reported symptomatology data. Incorporating temporal dynamics using longitudinal foundation models may improve such forecasting.