Abstract text (incl. figure legends and references)

Objective: Previous research carried out in our group demonstrated that patients with migraine exhibit changes in hemodynamic variables, suggesting a dysregulation of the autonomic nervous system (ANS). We aim to evaluate whether hemodynamic variables measured by cutting-edge wearable devices can predict migraine pain onset.

Methods: We performed a prospective study including patients with migraine in which we recorded real-time hemodynamic signals, including skin distal temperature (T), heart rate (HR) and electrodermal activity (EDA), obtained from a 24-hours wrist wearable device. Personalized prediction models were generated using the Artificial Recurrent Neural Networks long short-term memory (LSTM) to compute on a one-minute basis if the pain was going to appear in the next 120 minutes. Data were balanced in time periods of pain-no pain to train the models.

Results: A total of 8 patients with episodic migraine were included in the study. Most patients were women 7/8 (87.5%) and median age was 46 (IQR:34-48) years. Median duration of migraine was 27 (IQR: 18-35) years. The algorithm was able to predict migraine attacks with 95% sensitivity in the whole sample. The model predicted 23/24 (95%) of the pain attacks. In 7/8 (85.7%) of patients" migraine attacks were predicted, based on the 60-minute model, with no false negatives among them.

Conclusions: This study confirms that it is possible to predict a migraine attack using hemodynamic variables recorded by an easy-to-use wrist wearable. This research opens new possibilities to study the effect of early treatment on the evolution of the pain in a migraine crisis.