Machine Learning Classification of Migraine Using fNIRS During a Postural Task

Migraine diagnosis and severity staging rest on clinical interviews according to the ICHD-3 criteria, with no validated objective biomarker. Functional near-infrared spectroscopy (fNIRS) is a portable and non-invasive method, but existing fNIRS migraine classification studies remain relatively small and rely on cognitive tasks. We assessed whether prefrontal hemodynamic responses with a head-down-to-knees maneuver separate migraine patients from controls, and high- from low-severity migraine.

Prefrontal fNIRS, including short separation channels, was recorded during the maneuver in 50 interictal migraine patients and 52 controls screened for the absence of migraine, serious chronic conditions and hypertension. Nine hemodynamic parameters per chromophore (HbO, Hb and HbT) across thirty channels entered a hypothesis-neutral pipeline of 270 candidate pipelines (3 chromophores × 3 feature selection strategies × 30 classifiers) with no predefined region of interest.

Deoxyhemoglobin features selected by embedded L1 regularization with shrinkage-regularized linear discriminant analysis separated the groups with 92% balanced accuracy on the internal hold-out (92% sensitivity, 92% specificity, ROC-AUC 0.99; permutation p = 3 × 10^-4), against 87% in development-set cross-validation and 88% under nested selection cross-validation, with a selection bias of +0.03. High- vs. low-severity classification (19 high, 31 low) did not exceed chance under nested validation (45%; permutation p = 0.45).

A wide-scale pipeline achieved a robust, validated separation of migraine from screened controls, carried by a distributed venous-weighted deoxyhemoglobin signature; specificity against other headache disorders remains to be established. Attack frequency severity was not separable above chance, indicating that scalar hemodynamic descriptors are sufficient for a categorical but not a graded contrast.