Participants and clinical assessment
The human study was approved by the local ethics committee, and written informed consent was obtained from each participant. Patients were recruited from the neurological wards. According to the third version of the International Classification of Headache Disorders (ICHD-3) [6], nighty-seven patients was diagnosed as migraine. Diagnostic criteria of migraine with aura were then used to classify migraine patients into MwoA and MwA groups. Patients with probable migraine, additional neurological disease other than migraine, severe head injury, drug abuse, other major medical illness, brain vascular disease, or hydrocephalus, as well as failing to finish the MR examination were excluded from the study. After screening, 88 migraine patients were finally enrolled into the training sample, including 56 MwoA and 32 MwA (22 of them have visual or retinal symptoms, 8 of them have sensory symptom, 4 of them have speech and/or language symptoms, one of them has motor symptom) patients. 44 healthy control subjects (HC) who were matched to patients in terms of age, sex and education were also enrolled into our study. They were recruited from the local population and had no personal or family history of migraine, or any other types of headaches. To minimize hormonal influences on cortical excitability, all female subjects were included at mid-cycle and excluded if being pregnant or breast-feeding. Migraine patients and HC were all right-handers according to self-report. Moreover, 30 migraine patients (10 MwA) were included into the testing sample.
All patients completed a neuropsychological assessment including the Self-Rating Anxiety Scale (SAS), Self-Rating Depression Scale (SDS), Montreal Cognitive Assessment (MoCA), Headache Impact Test-6 (HIT-6), and Migraine Disability Assessment Score (MIDAS).
Image acquisition and preprocessing
After at least 4 h fasting, all subjects underwent MR examinations at one of two different 3.0 Tesla MRI scanners (uMR 780, United Imaging Healthcare, Shanghai, China; Ingenia, Philips Medical Systems, Best, Netherlands) for the patients in the training and testing samples, respectively. All subjects were scanned with a protocol including a high-resolution three-dimensional fast-echo T1-weighted MR image (resolution 1 × 1 × 1 mm3, TR/TE = 8.1/3.7 mm, slices = 170, FA= 8°, acquisition matrix = 256 × 256, FOV = 256 mm × 256 mm) and a three-dimensional pseudo-continuous ASL image (TR = 4000 ms, label duration = 1650 ms, TE = 11 ms, FA = 90°, post-label delay = 1600 ms, FOV = 240 mm × 240 mm, thickness = 4 mm, gap = 0.4 mm, acquisition matrix = 64 × 64, axial slices = 20). Finally, each subject contained 60 volumes used as 30 label-control image pairs.
The ASL data was preprocessed using the Statistical parameter mapping software (SPM12) (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/) and the toolbox ASLtbx (https://cfn.upenn.edu/~zewan). The procedure for obtaining CBF maps was detailed in our previous study [20]. The major steps included removing skull and cropping the gap, correcting motion artifacts, acquiring frame-wise displacement (FD) between groups and calculating CBF map. The CBF images were linearly co-registered in the native space to their corresponding T1-weighted images, which were non-linearly registered to the standard MNI space (the ICBM152 template). Then, each CBF image underwent spatial smoothing using a Gaussian kernel of FWHM of 8 mm. Afterwards, CBF map was normalized by dividing the value of cerebral blood flow in each voxel (2 mm × 2 mm × 2 mm) with the mean value of the whole brain CBF.
Voxel- and ROI- based comparisons
The voxel-based comparison of normalized CBF was conducted using a two-sample t-test to identify CBF variations between MwA and MwoA. Statistical threshold was set at t > 3.0 and p < 0.05, false discovery rate (FDR) corrected at cluster level. The brain regions showing significant differences were extracted as ROIs and the mean normalized CBF value in each ROI was calculated as an imaging feature and further pair-wise compared among MwA, MwoA and HC.
Model construction and evaluation
There were 88 and 30 migraine patients in the training and testing sets, with nearly the same percentage of MwoA and MwA patients (p = 0.76 in a chi-squared test). Based on the identified imaging features from the training set, support vector machine (SVM) models were established to differentiate MwA and MwoA under five-fold cross validation. The predictive ability of the SVM models was further evaluated in the testing set using a receiver operating characteristic (ROC) curve.