Machine learning classification of functional neurological disorder using structural brain MRI features.

Abstract

Brain imaging studies investigating grey matter in functional neurological disorder (FND) have used univariate approaches to report group-level differences compared with healthy controls (HCs). However, these findings have limited translatability because they do not differentiate patients from controls at the individual-level. 183 participants were prospectively recruited across three groups: 61 patients with mixed FND (FND-mixed), 61 age-matched and sex-matched HCs and 61 age, sex, depression and anxiety-matched psychiatric controls (PCs). Radial basis function support vector machine classifiers with cross-validation were used to distinguish individuals with FND from HCs and PCs using 134 <i>FreeSurfer</i>-derived grey matter MRI features. Patients with FND-mixed were differentiated from HCs with an accuracy of 0.66 (p=0.005; area under the receiving operating characteristic (AUROC)=0.74); this sample was also distinguished from PCs with an accuracy of 0.60 (p=0.038; AUROC=0.56). When focusing on the functional motor disorder subtype (FND-motor, n=46), a classifier robustly differentiated these patients from HCs (accuracy=0.72; p=0.002; AUROC=0.80). FND-motor could not be distinguished from PCs, and the functional seizures subtype (n=23) could not be classified against either control group. Important regions contributing to statistically significant multivariate classifications included the cingulate gyrus, hippocampal subfields and amygdalar nuclei. Correctly versus incorrectly classified participants did not differ across a range of tested psychometric variables. These findings underscore the interconnection of brain structure and function in the pathophysiology of FND and demonstrate the feasibility of using structural MRI to classify the disorder. Out-of-sample replication and larger-scale classifier efforts incorporating psychiatric and neurological controls are needed.