Machine learning research of bulimia nervosa based on diffusion tensor image

Abstract Background Bulimia nervosa is a type of eating disorder characterized by recurrent, uncontrollable, impulsive binge eating. No previous study selected diffusion tensor imaging as data features to conduct machine learning research on the diagnosis of bulimia nervosa. We tried to use machine learning methods to study the diagnosis of this disease and explore neurobiological markers . Methods This retrospective study examined 34 patients with bulimia nervosa and 34 healthy subjects. The selected characteristics were Fractional Anisotropy (FA), Axial Diffusivity (AD), Radial Diffusivity (RD) and Mean Diffusivity (MD). we used machine learning methods of support vector machines to distinguish bulimia nervosa and healthy controls. Results A total of 5 machine learning models were constructed. Classification effect of FA model and FA + MD + AD + RD model were acceptable. FA model classification effect was the best. The machine learning results of the five models were as follows: the area under the Receiver Operator Characteristic (ROC) curve of the FA model was 0.821, and the different brain regions were brainstem, temporal lobe, frontal lobe, inferior occipital gyrus, midbrain, middle frontal gyrus and caudate nucleus; MD model curve The lower area was 0.689, and the difference brain areas were posterior cerebellar lobe, frontal lobe, precentral gyrus, middle frontal gyrus, parietal lobe, superior frontal gyrus and paracentral lobule; the area under the AD model curve was 0.621, and the difference brain areas were cerebellar tonsil, cerebellar Stem, inferior frontal gyrus, midbrain, frontal lobe, precentral gyrus, postcentral gyrus, middle frontal gyrus and medial frontal gyrus; the area under the curve of the RD model was 0.625, and the difference brain regions were posterior cerebellum, midbrain, and middle frontal gyrus, precentral gyrus, frontal lobe, postcentral gyrus, superior frontal gyrus and medial frontal gyrus; the area under the curve of the FA + MD + AD + RD model was 0.739. Conclusions This study suggested that using diffusion tensor magnetic resonance imaging machine learning, it could distinguish between bulimia nervosa and healthy subjects and find neurobiological markers.