Kaoru Sakatani1,2, Katsunori Oyama3, Lizhen Hu1. 1. Department of Human and Engineered Environmental Studies, Graduate School of Frontier Sciences, The University of Tokyo, Tokyo, Japan. 2. Institute for Healthcare Robotics, Future Robotic Organization, Waseda University, Tokyo, Japan. 3. Department of Computer Science, College of Engineering, Nihon University, Tokyo, Japan.
Abstract
Background: In order to develop a new screening test of cognitive impairment, we studied whether cognitive function can be estimated from basic blood test data by applying deep learning models. This model was constructed based on the effects of systemic metabolic disorders on cognitive function. Methods: We employed a deep neural network (DNN) to predict cognitive function based on subject's age and blood test items (23 items). We included 202 patients (73.48 ± 13.1 years) with various systemic metabolic disorders for training of the DNN model, and the following groups for validation of the model: (1) Patient group, 65 patients (73.6 ± 11.0 years) who were hospitalized for rehabilitation after stroke; (2) Healthy group, 37 subjects (62.0 ± 8.6 years); (3) Health examination group, 165 subjects (54.0 ± 8.6 years) admitted for a health examination. The subjects underwent the Mini-Mental State Examination (MMSE). Results: There were significant positive correlations between the predicted MMSE scores and ground truth scores in the Patient and Healthy groups (r = 0.66, p < 0.001). There were no significant differences between the predicted MMSE scores and ground truth scores in the Patient group (p > 0.05); however, in the Healthy group, the predicted MMSE scores were slightly, but significantly, lower than the ground truth scores (p < 0.05). In the Health examination group, the DNN model classified 94 subjects as normal (MMSE = 27-30), 67 subjects as having mild cognitive impairment (24-26), and four subjects as having dementia (≤ 23). In 37 subjects in the Health examination group, the predicted MMSE scores were slightly lower than the ground truth MMSE (p < 0.05). In contrast, in the subjects with neurological disorders, such as subarachnoid hemorrhage, the ground truth MMSE scores were lower than the predicted scores. Conclusions: The DNN model could predict cognitive function accurately. The predicted MMSE scores were significantly lower than the ground truth scores in the Healthy and Health examination groups, while there was no significant difference in the Patient group. We suggest that the difference between the predicted and ground truth MMSE scores was caused by changes in atherosclerosis with aging, and that applying the DNN model to younger subjects may predict future cognitive impairment after the onset of atherosclerosis.
Background: In order to develop a new screening test of cognitive impairment, we studied whether cognitive function can be estimated from basic blood test data by applying deep learning models. This model was constructed based on the effects of systemic metabolic disorders on cognitive function. Methods: We employed a deep neural network (DNN) to predict cognitive function based on subject's age and blood test items (23 items). We included 202 patients (73.48 ± 13.1 years) with various systemic metabolic disorders for training of the DNN model, and the following groups for validation of the model: (1) Patient group, 65 patients (73.6 ± 11.0 years) who were hospitalized for rehabilitation after stroke; (2) Healthy group, 37 subjects (62.0 ± 8.6 years); (3) Health examination group, 165 subjects (54.0 ± 8.6 years) admitted for a health examination. The subjects underwent the Mini-Mental State Examination (MMSE). Results: There were significant positive correlations between the predicted MMSE scores and ground truth scores in the Patient and Healthy groups (r = 0.66, p < 0.001). There were no significant differences between the predicted MMSE scores and ground truth scores in the Patient group (p > 0.05); however, in the Healthy group, the predicted MMSE scores were slightly, but significantly, lower than the ground truth scores (p < 0.05). In the Health examination group, the DNN model classified 94 subjects as normal (MMSE = 27-30), 67 subjects as having mild cognitive impairment (24-26), and four subjects as having dementia (≤ 23). In 37 subjects in the Health examination group, the predicted MMSE scores were slightly lower than the ground truth MMSE (p < 0.05). In contrast, in the subjects with neurological disorders, such as subarachnoid hemorrhage, the ground truth MMSE scores were lower than the predicted scores. Conclusions: The DNN model could predict cognitive function accurately. The predicted MMSE scores were significantly lower than the ground truth scores in the Healthy and Health examination groups, while there was no significant difference in the Patient group. We suggest that the difference between the predicted and ground truth MMSE scores was caused by changes in atherosclerosis with aging, and that applying the DNN model to younger subjects may predict future cognitive impairment after the onset of atherosclerosis.