Identification of Cognitive Deficits Based on T1-Weighted Magnetic Resonance Imaging

Machine learning methods are widely used to classify cognitive deficits. However, high data processing speed is required for their successful application in clinical settings. In this study, we developed a method for cognitive deficit classification that addresses the problem of processing data from different sources with limited computational resources. The method is based on a random forest model with a parameter optimizer trained using morphometric data of patients with cognitive deficit and cognitively normal people from the control group obtained from the database of the State Budgetary Health Care Institution of Moscow “N. A. Alekseev Psychiatric Clinical Hospital No. 1 of the Department of Health Care of Moscow”. Morphometric data were obtained from T1-weighted MRI images of the brain by neuroanatomical segmentation and parcellation of the cortex based on deep learning and subsequent diffeomorphic registration to measure cortical thickness. The Montreal Cognitive Assessment Scale (MoCA) was selected as a tool to assess cognitive deficits. The model was trained on four subsets of data and evaluated using cross-validation. Our model achieved an accuracy of 0.73–0.77 with ROC-AUC metric values of 0.73–0.78. The average processing time per person was 30–35 min. These results demonstrate the potential of the method to be used as a rapid tool for classifying cognitive deficits.