Daniele Martinelli (Pavia/ IT), Maria Magdalena Pocora (Pavia/ IT), Roberto De Icco (Pavia/ IT), Marta Allena (Pavia/ IT), Grazia Sances (Pavia/ IT), Gloria Castellazzi (Pavia/ IT), Cristina Tassorelli (Pavia/ IT)
Abstract text (incl. figure legends and references)
OnabotulinumtoxinA (BoNT-A) reduces the frequency of migraine but the clinical profile of those patients who might benefit from it is still missing.
Objective: In this single-centre, real-life study, we applied machine learning (ML) algorithms to a database of patients who underwent treatment with BoNT-A to identify baseline clinical characteristics capable to predict response to treatment.
Methods: We collected baseline demographic and clinical data of consecutive patients who started BoNT-A at IRCCS Mondino Foundation from January 2017 to March 2022. All patients had a diagnosis of chronic migraine or high-frequency episodic migraine and underwent at least one treatment cycle with BoNT-A according to the PREEMPT paradigm. Patients were primarily classified according to the monthly migraine days reduction in the 12-week period after the fourth BoNT-A treatment, as compared to a 28-day baseline period. Patients who early terminated the treatments after 1 or 2 consecutive administrations without any effect were profiled as well. Other classifications were obtained using secondary endpoints like migraine disability assessment test (MIDAS) and abortive drug use reduction.
Collected data were used as input features to run different kinds of supervised and unsupervised ML algorithms.
Results: Of the 212 patients included in the evaluation, 35 qualified as responders to BoNT-A administration and 91 as non-responders. Not a single, or panel, of anamnestic characteristics, proved capable to discriminate responders from non-responders. All ML models coherently reached good accuracy but underperformed and lacked specificity.
Conclusions: Overall, ML findings suggest that routine anamnestic features acquired in real-life settings cannot accurately predict the patients that will benefit from BoNT-A treatment. A deeper phenotyping of patients" features, possibly combined with multimodal parameters, is probably required to identify features predictive of response to BONT-A.