P097 | SEQUENTIAL MACHINE LEARNING MODELS FOR PREDICTING ONABOTULINUMTOXINA RESPONSE IN CHRONIC MIGRAINE: A COHORT STUDY
G. Vigani1, M. Romozzi2|3, A. Burgalassi1, A. Lokhawandala4, C. Vollono2|3, P. Calabresi2|3, F. Tascedda5, A. Chiarugi1|6, F. De Cesaris1, L. Francesco Iannone7 | 1Headache Center and Clinical Pharmacology Unit, Careggi University Hospital, Florence, Italy; 2Dipartimento Universitario di Neuroscienze, Università Cattolica del Sacro Cuore, Rome, Italy; 3Neurologia, Dipartimento di neuroscienze, Organi di Senso e Torace, Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Rome, Italy; 4Drexel University, Philadelphia, USA; 5Department of Life Sciences, University of Modena and Reggio Emilia, Italy; 6Section of Clinical Pharmacology and Oncology, Department of Health Sciences, University of Florence, Italy; 7Department of Biomedical, Metabolic and Neural Sciences, University of Modena and Reggio Emilia, Italy
Abstract
Background. Chronic migraine (CM) is associated with high disability and heterogeneous response to onabotulinumtoxinA (BoNT-A). Baseline clinical predictors of response remain inconsistent, limiting individualized treatment decisions. Machine learning may improve prediction by integrating longitudinal treatment data. This study aimed to develop sequential machine-learning models to predict BoNT-A response in CM after successive treatment cycles under real-world conditions.
Methods. We performed a post-hoc analysis of a prospectively collected cohort of adults with CM treated with BoNT-A according to the PREEMPT protocol at a tertiary headache center. Response was defined as a ≥30% reduction in monthly headache days (MHDs) from baseline. Four sequential supervised classification models were trained to predict response after cycles 1-4. Model 1 used baseline demographic and clinical variables only; models 2-4 progressively incorporated post-treatment MHDs and acute medication use after preceding cycles. Multiple algorithms were compared using stratified 5-fold cross-validation, with balanced accuracy as the primary performance metric. Model performance was assessed using accuracy, sensitivity, specificity, positive and negative predictive values, macro-F1 score, and AUC-ROC. Feature importance was evaluated using SHapley Additive exPlanations.
Results. The baseline cohort included 385 patients with CM; mean age was 48.1±13.1 years, 82.9% were female, baseline MHDs were 25.4±5.9, and medication overuse was present in 326/383 patients (85.1%). The number of patients included decreased across models with 385 in model 1 and 161 in model 4. Baseline-only prediction showed limited performance, with balanced accuracy of 0.60±0.03 and AUC-ROC of 0.62±0.04. Incorporating early post-treatment data markedly improved discrimination. Model 2 achieved balanced accuracy of 0.82±0.04 and AUC-ROC of 0.88±0.01, while model 3 showed the highest balanced accuracy, 0.84±0.04, and AUC-ROC of 0.87±0.03. Across longitudinal models, on-treatment MHDs were the dominant predictors of response.
Conclusions. Sequential machine-learning models incorporating early post-treatment headache-frequency data outperformed baseline-only prediction of BoNT-A response in CM. These findings support dynamic, monitoring-informed prediction models and require prospective external validation before clinical implementation.
Methods. We performed a post-hoc analysis of a prospectively collected cohort of adults with CM treated with BoNT-A according to the PREEMPT protocol at a tertiary headache center. Response was defined as a ≥30% reduction in monthly headache days (MHDs) from baseline. Four sequential supervised classification models were trained to predict response after cycles 1-4. Model 1 used baseline demographic and clinical variables only; models 2-4 progressively incorporated post-treatment MHDs and acute medication use after preceding cycles. Multiple algorithms were compared using stratified 5-fold cross-validation, with balanced accuracy as the primary performance metric. Model performance was assessed using accuracy, sensitivity, specificity, positive and negative predictive values, macro-F1 score, and AUC-ROC. Feature importance was evaluated using SHapley Additive exPlanations.
Results. The baseline cohort included 385 patients with CM; mean age was 48.1±13.1 years, 82.9% were female, baseline MHDs were 25.4±5.9, and medication overuse was present in 326/383 patients (85.1%). The number of patients included decreased across models with 385 in model 1 and 161 in model 4. Baseline-only prediction showed limited performance, with balanced accuracy of 0.60±0.03 and AUC-ROC of 0.62±0.04. Incorporating early post-treatment data markedly improved discrimination. Model 2 achieved balanced accuracy of 0.82±0.04 and AUC-ROC of 0.88±0.01, while model 3 showed the highest balanced accuracy, 0.84±0.04, and AUC-ROC of 0.87±0.03. Across longitudinal models, on-treatment MHDs were the dominant predictors of response.
Conclusions. Sequential machine-learning models incorporating early post-treatment headache-frequency data outperformed baseline-only prediction of BoNT-A response in CM. These findings support dynamic, monitoring-informed prediction models and require prospective external validation before clinical implementation.
Metrics
Citations and online attention from Dimensions, Altmetric and PlumX.
License
Copyright (c) 2026 The Author(s)
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
PAGEPress has chosen to apply the Creative Commons Attribution NonCommercial 4.0 International License (CC BY-NC 4.0) to all manuscripts to be published.