An evolving machine-learning-based algorithm to early predict response to anti-CGRP monoclonal antibodies in patients with migraine

Background The present study aimed to determine whether machine-learning (ML)-based models can predict 3-, 6, and 12-month responses to the monoclonal antibodies (mAbs) against the calcitonin gene-related peptide (CGRP) or its receptor (anti-CGRPmAbs) in patients with migraine using early predictors (up to one month) and to create an evolving prediction tool. Methods In this prospective cohort study, data from patients with migraine who had received anti-CGRP mAbs for 12 months were collected. Demographic and monthly clinical variables were collected, including monthly headache days (MHDs), days with acute medication use, number of analgesics and Headache Impact Test-6. Response rates were categorized as <25%, 26–50%, 51–75% and >75% reduction in MHDs. ML models were trained using random forest algorithm and optimized to maximize the F1 score. ML model performance was also evaluated using standard evaluation metrics, including accuracy, precision and area under the receiver operating characteristic curve (AUC-ROC). Sequential backward feature selection was employed to identify the most relevant predictors for each model. Each model was given 11 baseline data inputs and month-based predictors for months 1, 3 and 6. Each model was then validated against an external test cohort of patients who had received anti-CGRP mAbs for 12 months. Results Three hundred thirty-six patients treated with anti-CGRP mAbs were included. The external cohort included 93 patients treated with anti-CGRP mAbs. We developed six models to predict 3- 6- and 12-month responses using early predictors. ML-based models yielded predictions with an accuracy score in the range 0.40–0.73 and an AUC-ROC score in the range 0.56–0.76 during internal testing and yielding predictions with an accuracy in the range 0.39–0.64 and an AUC-ROC score in the range 0.52–0.78 when tested against an external test cohort. Shapley Additive explanations summary plots were generated to interpret the contribution of each feature for each model. Based on these findings, a response prediction tool was developed. Each model was run through a backward feature selection to find the most relevant features for the models. The MHDs reduction of the previous data point tends to be the most relevant, while the migraine with aura indicator tends to be the least effective predictor. Conclusions The response prediction tool utilizing evolving ML-based models holds promise in the early prediction of treatment outcomes for patients with migraine undergoing anti-CGRP mAbs treatment.