Enhancing the early diagnosis of bipolar disorder with machine learning models

Bipolar Disorder (BD) remains difficult to diagnose early due to overlapping symptoms, heterogeneity in presentation, and the absence of definitive clinical biomarkers. Single-modality approaches often fail to capture the complex interplay of neurological, cognitive, and clinical factors. To overcome this limitation, this study uses a multimodal dataset consisting of fMRI neuroimaging, neurocognitive performance measure, and structured clinical records of a binary classification task differentiating between BD patients and healthy controls. A new model, BDiagNet-3D, is suggested, which includes a 3D Convolutional Neural Network (3D-CNN) used to extract spatial-temporal patterns of brain activations and a Deep Neural Network (DNN) used to combine these features with clinical and cognitive data. The structured modalities are also processed using RF and SVM models providing an interpretable result in addition to the deep-learning results. The resulting fusion process can make the model take advantage of complementary interrelationships between imaging and non-imaging features, which boosts predictive strength. Experimental evaluation in Python shows RF achieving 64% accuracy, SVM achieving 73.2% accuracy, and BDiagNet-3D achieving 99.1% accuracy, with precision, recall, and F1-scores following a similar performance trend. These findings highlight the value of integrating high-dimensional neuroimaging data with structured clinical and cognitive inputs within a unified architecture. The proposed method demonstrates strong potential for advancing precision psychiatry by delivering accurate, interpretable, and early BD detection, offering a foundation for AI-assisted tools in clinical decision-making.