Migraine attack prediction using wearable biosensor data.

Abstract

BackgroundMany migraine sufferers report difficulties in managing their attacks, leading to missed workdays, impaired cognitive function and reduced participation in social activities. This challenge underscores the importance of identifying reliable indicators of an impending attack. The present study aimed to develop and evaluate machine learning models for predicting migraine attacks based on continuous, non-invasive monitoring of autonomic nervous system activity.MethodsData from a wrist-worn biosensor device the Empatica EmbracePlus (Empatica Inc., Boston, MA, USA) capturing electrodermal activity, heart rate, skin temperature and metabolic equivalent of task were analysed using machine learning models to detect prodromal autonomic nervous system alterations and predict impending migraine headaches. Personalised models were developed using participant-specific model selection, whereas generalised transferability was assessed using a leave-one-participant-out approach.ResultsData from 27 participants were analysed, with an average monitoring period of 30.8 ± 12.9 days, during which 96 migraine attacks were recorded, corresponding to 3.6 ± 1.0 episodes per individual. Personalised prediction models, each reflecting a participant-specific best-selected configuration, achieved good overall predictive performance, with mean values of 0.82 ± 0.10 for accuracy, 0.75 ± 0.16 for precision, 0.72 ± 0.12 for recall and 0.73 ± 0.12 for the F1-score. The mean area under the receiver operating characteristic curve was 0.84 ± 0.10. Gaussian Naive Bayes was most frequently selected, accounting for 44% of valid participant-specific configurations. Generalised models provided above-chance performance in only one-third of participants.ConclusionsThis study supports the feasibility of migraine attack prediction using continuous physiological data from a wrist-worn biosensor analysed with personalised machine learning approaches.