Unveiling psychotic disorder patterns: A deep learning model analysing motor activity time-series data with explainable AI

• Published in: Biomedical signal processing and control Vol. 91; p. 106000
• Main Authors: Mehraj Misgar, Muzafar, Bhatia, M.P.S.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2024
• Subjects: Depression; Explainable AI; IoMT; Motor Activity; Schizophrenia; Schizophrenia; Depression; Explainable AI; Motor Activity; IoMT
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2024.106000

•Wearable Internet of Medical Things (IoMT) devices, such as wrist actigraphs, canprovides a more objective and effective approach to early-stage mental health diagnosis.•The methodology includes a custom imputation method, a novel combinatorial sampling method for sample expansion•a multi-branch deep learning architecture is employed with one branch fed with nighttime data and another branch with daytime data•The use of Gradient weighted Class Activation Map (Grad-CAM) technique aids in visualizing and comprehending the model's decision-making process.•The model achieved a commendable accuracy of 0.94 for distinguishing depressive and schizophrenic samples from control samples individually. Deep Learning (DL) holds immense potential in revolutionizing healthcare, offering robust support to clinicians and enhancing patient care. However, coming up with the right DL model is always challenging and depends on quality, quantity, and type of data. In this paper, two motor activity datasets are utilized: “Depresjon” dataset includes activity recordings from 32 healthy individuals (402 days) and 23 individuals with unipolar and bipolar depression (291 days) and the “Psykose” dataset consists of 22 schizophrenia subjects (285 days) and 32 healthy subjects (402 days). The motor activity data, represented as time-series signals, poses challenges due to variable lengths and non-uniform starting timestamps for each subject. Additionally, the daytime and nighttime distributions differ across samples, requiring explicit handling for Convolutional Neural Network models. To address this issue, a multi-branch DL architecture is employed with one branch fed with nighttime data and another with daytime data, capable of capturing features across various scales, accommodating patterns of different sizes. Moreover, the combined outputs of these branches are subjected to a self-attention-mechanism (MultiHeadAttention), which prioritizes essential features. The use of Gradient weighted Class Activation Map (Grad-CAM) technique aids in comprehending the model's decision-making process. The benchmark datasets were used to validate the model, which exhibited an accuracy of 0.94 for both classifying depressive and schizophrenic episodes from control subjects. An accuracy of 0.81 for classifying depressive episodes, and schizophrenics from control samples. This accuracy further increases when combining the control samples from both datasets, to 0.97 for depression and 0.98 for schizophrenia.