AI can enhance migraine diagnosis using easy-to-measure clinical data

Machine-learning algorithms could predict migraine based on clinical parameters that are not being used for formal diagnosis. Adding genotyping data further improved the predictive value for diagnosis.

Dr Antonios Danelakis and colleagues (Norwegian University of Science and Technology, Norway) investigated the predictive value of machine-learning algorithms in diagnosing migraine [1]. They used data from the 2nd and 3rd Nord-Trøndelag Health Study (HUNT), including genotype (genetic dosages of migraine-related risk variants), clinical data, and headache phenotype in 43,197 individuals. General clinical data from the HUNT study included demographics, socioeconomic data, comorbidity, blood pressure, physical characteristics, stimulant/medication use, and non-headache complaints.

The study population included data from 10,286 individuals with headache (24% of the overall population) and 32,911 (76%) individuals as control; 90% of the initial study population was used to train the model, while the remaining 10% was used for testing. In total, 3 datasets were employed: a dataset containing only genotype data (7,840 genetic variants), a dataset containing only clinical data (74 variables), and a combined dataset containing both.

Of all the models in the study, the Light Gradient Boosting Machine (LGBM) model showed the top performance for all datasets. In the genotype dataset, the LGBM model led to an area under the curve (AUC) of 65% in the training population and 63% in the testing population, while in the clinical dataset, the AUC values were 80% and 79%, respectively. In the combined genotype and clinical dataset, the LGBM model led to an AUC of 81% when using the training data and an AUC of 80% for the testing dataset. These numbers are “a good indication of the generalisability of this model,” said Dr Danelakis.

In summary, “migraine could be predicted by machine learning surprisingly well if we consider that none of the typical information that is used for migraine diagnosis was included here,” said Dr Danelakis. However, he added that “predicting migraine based on clinical data provides better results than just using individual genotype data, but the latter still adds some additional information.”

1. Danelakis A, et al. Machine learning can predict migraine from genotype and non-headache clinical data with high accuracy. 18^th European Headache Congress, 4–7 December 2024, Rotterdam, the Netherlands.

Copyright ©2024 Medicom Medical Publishers



Posted on 6 January 20256 January 2025