The use of Deep Learning (DL) in healthcare holds immense potential to assist clinicians and revolutionize patient care. Its application can significantly decrease reporting delays and enhance overall workflows. However, coming up with the right DL model is always challenging for researchers and depends on quality, quantity, and type of data. In this paper, Motor Activity Data is utilized for the classification of psychotic disorders. Datasets contain activity signals from depressive (unipolar and bipolar depression), schizophrenic, and control subjects. Depresjon Datasets contain sensor data from 20 female and 35 male subjects, and the Psykose Dataset contains sensor data from 40 male and 13 female subjects. Daytime distribution of data is different from nighttime distribution across all the samples. To cope with this problem, we are using a Multi-branch DL architecture that can capture features at different scales, allowing it to handle patterns of varying sizes. Further, the concatenating output of branches is followed by Attention Mechanisms to focus on the most important features. The Grad-CAM (Gradient-weighted Class Activation Mapping) technique is used to understand the decision-making process of the proposed model. The benchmark datasets were used to validate the model, which exhibited a remarkable performance accuracy of 0.93 for classifying depressive episodes from control samples, an accuracy of 0.92 for classifying schizophrenic episodes from control samples, and an accuracy of 0.80 for classifying depressive and schizophrenic subjects from control samples. This accuracy further increases when combining the control samples from both datasets, to 0.96 for depression and 0.98 for schizophrenia. This accuracy surpasses that of the state-of-the-art approach, indicating that the multi-branch CNN model demonstrates superior performance on Motor Activity Data when properly combined with the Attention Mechanism.
