Combined use of gray matter volume and neuropsychological test performance for classification of individuals with bipolar I disorder via artificial neural network method.

Abstract

Diagnosis of patients with bipolar disorder may be challenging and delayed in clinical practice. Neuropsychological impairments and brain abnormalities are commonly reported in bipolar disorder (BD); therefore, they can serve as potential biomarkers of the disorder. Rather than relying on these predictors separately, using both structural and neuropsychiatric indicators together could be more informative and increase the accuracy of the automatic disorder classification. Yet, to our information, no Artificial Intelligence (AI) study has used multimodal data using both neuropsychiatric tests and structural brain changes to classify BD. In this study, we first investigated differences in gray matter volumes between patients with bipolar I disorder (n = 37) and healthy controls (n = 27). The results of the verbal and non-verbal memory tests were then compared between the two groups. Finally, we used the artificial neural network (ANN) method to model all the aforementioned values for group classification. Our voxel-based morphometry results demonstrated differences in the left anterior parietal lobule and bilateral insula gray matter volumes, suggesting a reduction of these brain structures in BD. We also observed a decrease in both verbal and non-verbal memory scores of individuals with BD (p < 0.001). The ANN model of neuropsychiatric test scores combined with gray matter volumes has classified the bipolar group with 89.5% accuracy. Our results demonstrate that when bilateral insula volumes are used together with neuropsychological test results the patients with bipolar I disorder and controls could be differentiated with very high accuracy. The findings imply that multimodal data should be used in AI studies as it better represents the multi-componential nature of the condition, thus increasing its diagnosability.