PO-66 | Predicting the evolution to resistant and refractory migraine: a machine learning analysis from the REFINE Study

Chiara Rosignoli,¹ Raffaele Ornello,¹ Daniele Lozzi,² Enrico Mattei,² Federico De Santis,¹ Agnese Onofri,¹ Mark Braschinsky,³ Christian Lampl,⁴ Isabel Pavão Martins,⁵ Paolo Martelletti,⁶ Dimos Mitsikostas,⁷ Raquel Gil-Gouveia,⁸ Maria Pia Prudenzano,⁹ Kristina Ryliskiene,¹⁰ Margarita Sanchez del Rio,¹¹ Patricia Pozo-Rosich,¹² Aynur Özge,¹³ Fabrizio Vernieri,¹⁴ Marta Waliszewska-Prosół,¹⁵ Zaza Katsarava,¹⁶ Giuseppe Placidi,² Simona Sacco¹ | ¹Dept. of Biotechnological and Applied Clinical Sciences, University of L'Aquila, Italy; ²A²VI-lab, Dept. MeSVA University of L’Aquila; ³Dept. of Neurology and Neurosurgery, Institute of Clinical Medicine, Faculty of Medicine, University of Tartu, Estonia; ⁴Headache Medical Center, Seilerstaette Linz, Austria. Center; ⁵Headache Center, Hospital da Luz, Lisbon, Portugal; ⁶School of Health, Unitelma Sapienza University of Rome, Italy; ⁷First Dept of Neurology, Aeginition Hospital, National and Kapodistrian University of Athens, Greece; ⁸Centro de Estudos Egas Moniz, Faculty of Medicine, University of Lisbon, Portugal; ⁹Headache Center, Amaducci Neurological Clinic, Policlinico General Hospital, Bari, Italy; ¹⁰Center of Neurology, Vilnius University, Lithuania; ¹¹Dept. of Neurology, Clínica Universidad de Navarra, Madrid, Spain; ¹²Neurology Department, Hospital Universitari Vall d'Hebron, Barcelona, Spain; ¹³Dept. of Neurology, Mersin University Medical Faculty, Mersin, Turkey; ¹⁴Università Policlinico Campus Bio-Medico, Rome, Italy; ¹⁵Dept. of Neurology, Wroclaw Medical University, Wrocław, Poland; ¹⁶Dept of Neurology, Christian Hospital Unna and University of Duisburg-Essen, Ruhr Metropolitan, Germany

Background: Resistant and refractory forms of migraine represent a significant clinical challenge and are often associated with a poor quality of life and a limited response to treatment. Identifying patients at risk of progression early on may facilitate more timely and personalised interventions. This study explores whether machine learning techniques can identify clinical predictors of progression from non-resistant, non-refractory migraine (NRRM) to resistant (RM) or refractory (RefM) migraine.

Methods: The data were derived from the multicentre REFINE study, which was conducted prospectively across 15 tertiary headache centres in Europe. At baseline, participants were categorised into three diagnostic groups according to the EHF criteria: NRNRM, ResM and RefM. Clinical progression was monitored over 3 to 6 months. The dataset included demographic, clinical, and treatment-related variables. Data preprocessing included handling of missing values, outlier detection, and undersampling of the majority class. Feature selection was performed through both classical statistical approaches (e.g., correlation analysis) and machine learning methods (e.g., Random Forest feature importance). Selected features were used to train and compare various classification models. Performance metrics included accuracy and F1-score. SHAP (SHapley Additive exPlanations) analysis was applied to the best-performing model for interpretability.

Results: Several variables were identified as key predictors of diagnostic transition. The strongest correlations with diagnostic class were found for EHF baseline classification (r = 0.63), presence of ≥8 days/month with debilitating headache (r = 0.55), HIT-6 score (r = 0.35), and HURT item 4 (r = 0.33). Random Forest-based feature selection highlighted additional predictors such as average number of migraine days in the past 3 months, number of treatment failures, HALT score, and number of triptan doses used. Gradient Boosting was the best-performing model based on F1-score. SHAP analysis confirmed the relevance of these variables in predicting diagnostic evolution.

Conclusion: This study highlights the potential of machine learning approaches in identifying clinical variables that influence migraine progression to resistant or refractory forms. These findings could help clinicians in the early recognition of patients at high risk for preventive treatment failures, allowing for targeted and timely interventions.