Deep learning models for diagnosing mood disorders using integrated MRI and genetic data

Background Brain disorders are conditions that affect brain structure, function, or chemistry, causing various symptoms and impairments. Brain disorders are categorized into neurodegenerative disorders, including Alzheimer’s disease and Parkinson’s disease, which involve progressive neuronal degeneration; mental health disorders, including depression, anxiety, bipolar disorder (BD), and schizophrenia; and traumatic brain injuries resulting from external force, causing temporary or permanent brain damage. Mood disorders, including major depressive disorder (MDD) and BD, are frequently underdiagnosed, thereby contributing to a significant clinical burden. To address this challenge, we introduce a novel computational framework that uses multimodal data integration by combining patient-specific magnetic resonance imaging (MRI) with whole-exome sequencing data. Methods Our dataset consisted of brain imaging and genetic data from 321 East Asian individuals, comprising 147 diagnosed with MDD, 78 with BD, and 96 healthy controls, along with corresponding single-nucleotide polymorphism (SNP) data containing 212 features per subject. Further, we used a child MRI dataset for external validation. We initially prepared and preprocessed our data for the adult dataset. SNP data were then loaded from a Excel file; features were normalized, and MRI images stored in Neuroimaging Informatics Technology Initiative (NIFTI) format were preprocessed by resizing and augmenting them. Various deep learning models ( e.g ., InceptionV3, ResNet) were employed to extract features from MRI data. SNP characteristics were extracted from the preprocessed genetic data. The number of samples was aligned between the SNP and MRI feature sets; these features were concatenated to form a combined feature set, and the combined features were normalized. Results The combined features were input into machine learning classifiers ( e.g ., support vector machine (SVM), K Nearest Neighbors) for final classification, yielding the best accuracy of 74.2% on a linear SVM classifier for detecting mood disorders. Further, two more results were considered, with the second being the classification of the child MRI dataset into abnormal and normal categories, which achieved an exceptional accuracy of 99.8% on the cubic SVM classifier. Conclusion Our approach supports the diagnostic evaluation of patients with psychiatric disorders by incorporating additional neuroimaging modalities and genomic information into routine clinical workflows.