Predicting upper limb motor prognosis in stroke patients with a graph embedding model using EEG sign
Poster No:
1662
Submission Type:
Abstract Submission
Authors:
Seoyeon Kim1, Yunjeong Jang2, Yunhee Kim3, Minji Lee4
Institutions:
1Department of Artificial Intelligence, The Catholic University of Korea, Bucheon, South Korea, Bucheon, Seoul, 2Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore, Seoul, Seoul, 3Department of Physical and Rehabilitation Medicine, Sungkyunkwan University School of Medicine, Suwo, Seoul, Seoul, 4Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore, Bucheon, Gyeonggi
First Author:
Seoyeon Kim
Department of Artificial Intelligence, The Catholic University of Korea, Bucheon, South Korea
Bucheon, Seoul
Department of Artificial Intelligence, The Catholic University of Korea, Bucheon, South Korea
Bucheon, Seoul
Co-Author(s):
Yunjeong Jang
Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore
Seoul, Seoul
Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore
Seoul, Seoul
Yunhee Kim
Department of Physical and Rehabilitation Medicine, Sungkyunkwan University School of Medicine, Suwo
Seoul, Seoul
Department of Physical and Rehabilitation Medicine, Sungkyunkwan University School of Medicine, Suwo
Seoul, Seoul
Minji Lee
Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore
Bucheon, Gyeonggi
Department of Biomedical Software Engineering, The Catholic University of Korea, Bucheon, South Kore
Bucheon, Gyeonggi
Introduction:
Stroke is increasingly occurring in modern society, and its severity is getting worse day by day. Patients suffer from poor motor and cognitive functions, which are directly related to the quality of life of survivors. Therefore, in order to present proper rehabilitation training, a system that predicts the motor and cognitive function of patients is needed.
Recent studies have notably illuminated the intricate relationship between electroencephalography (EEG) signal patterns and motor recovery in stroke patients [1, 2]. However, few studies have used these patterns to present prediction of motor recovery as a biomarker from a machine learning perspective. In this study, we proposed the deep learning framework in subacute stroke patients to predict the upper limb motor function.
Recent studies have notably illuminated the intricate relationship between electroencephalography (EEG) signal patterns and motor recovery in stroke patients [1, 2]. However, few studies have used these patterns to present prediction of motor recovery as a biomarker from a machine learning perspective. In this study, we proposed the deep learning framework in subacute stroke patients to predict the upper limb motor function.
Methods:
Ten chronic stroke patients (60.6 ± 7.9 years; Female = 5; Infarction = 8) were participated in this study. In the motor function using an Fugl-Meyer upper extremity (FMA-UE), they had 38.0 ± 11.4 scores in subacute phase, and 55.3 ± 5.78 scores after two months. As a result, we proposed to predict the prognosis of motor function by dividing the improvement score of FMA-UL into good recovery group and low recovery group based on 20 points [3, 4]. The good group (class 0) changed from 30.0 ± 8.0 to 55.6 ± 5.0, whereas the low group (class 1) changed from 46.0 ± 11.2 to 55.0 ± 6.4. In summary, each group was assigned five subjects.
EEG data were measured at rest-state with eyes closed for 5 min and the sampling rate was 4,000 Hz. The 27 channels using 10-20 international system were used. Continuous data were downsampled to 500 Hz and segmented into 5-sec epochs. In addition, bandpass filtering was performed from 1 to 45 Hz in preprocessing step.
The feature used the normalized mutual information (NorMI) [5], and as a result, the matrix of 27×27 was calculated. To achieve subject-balance while preserving the characteristics of each subject, generative adversarial networks symmetry was used. Pairwise subjects from each class are considered in the wasserstein distance calculation to account for the influence of multiple subjects on each other and the individual characteristics of each subject. We utilized the 3-layer neural network to classify the masked training and test sets (Figure 1). We performed leave-one-subject-out cross-validation, because of independence on a specific stroke patient.
EEG data were measured at rest-state with eyes closed for 5 min and the sampling rate was 4,000 Hz. The 27 channels using 10-20 international system were used. Continuous data were downsampled to 500 Hz and segmented into 5-sec epochs. In addition, bandpass filtering was performed from 1 to 45 Hz in preprocessing step.
The feature used the normalized mutual information (NorMI) [5], and as a result, the matrix of 27×27 was calculated. To achieve subject-balance while preserving the characteristics of each subject, generative adversarial networks symmetry was used. Pairwise subjects from each class are considered in the wasserstein distance calculation to account for the influence of multiple subjects on each other and the individual characteristics of each subject. We utilized the 3-layer neural network to classify the masked training and test sets (Figure 1). We performed leave-one-subject-out cross-validation, because of independence on a specific stroke patient.
Results:
The mean accuracy was approximately 0.93±0.07 excluding when test subject was 8 and 10, whereas 0.78±0.23 including them. A significant decrease was observed with the case when test subject is 8 and 10. The t-distributed stochastic neighbor embedding (t-SNE) result show the feature distribution of NorMI matrices. Subject 8 and 10 show similar distributions and these are mixed at the boundary of the classes. If one of them falls into the test set, the class boundary may not fit the test set (Figure 2).
We designed another framework to evaluate the effectiveness of our approach. When training the designed model, wasserstein distance was calculated with all subjects in each class at once. In results, when test subject was in class0, the mean accuracy was 0.74±0.27, but for subjects in class 1, the accuracy was 0. In the cases where the subject in class 0 as test set, the highest accuracy was 1.0 and the lowest was 0.28. In other words, due to some strong subject NorMI pattern, subject independence was reduced.
We designed another framework to evaluate the effectiveness of our approach. When training the designed model, wasserstein distance was calculated with all subjects in each class at once. In results, when test subject was in class0, the mean accuracy was 0.74±0.27, but for subjects in class 1, the accuracy was 0. In the cases where the subject in class 0 as test set, the highest accuracy was 1.0 and the lowest was 0.28. In other words, due to some strong subject NorMI pattern, subject independence was reduced.
Conclusions:
In this study, we proposed the classification framework using normalized mutual information for predicting upper limb motor prognosis. These results would be of great help in guiding rehabilitation training based on stroke prognosis, and should be challenged not only in the motor function but also in the cognitive function.
Brain Stimulation:
Non-Invasive Stimulation Methods Other 2
Modeling and Analysis Methods:
EEG/MEG Modeling and Analysis 1
Keywords:
Electroencephaolography (EEG)
Other - Functional connectivity, Graph embedding, Normalized mutual information, Wasserstein distance, Deep learning, Pattern classification, Stroke, Prediction, Rehabilitation
1|2Indicates the priority used for review
Provide references using author date format
[2] Milani, G., Antonioni, A., Baroni, A., Malerba, P., and Straudi, S. (2022), ‘Relation Between EEG Measures and Upper Limb Motor Recovery in Stroke Patients: A Scoping Review’, Brain Topography, vol 35, no. 5-6, pp. 651-666.
[3] Gladstone, D.J., Danells, C.J., and Black, S.E. (2002), 'The Fugl-Meyer Assessment of Motor Recovery after Stroke: A Critical Review of Its Measurement Properties', Neurorehabil Neural Repair, vol. 13, pp. 232–240.
[4] Fugl-Meyer, A.R., Jääskö, L., Leyman, I., Olsson, S., Steglind, S. (1975), 'The post-stroke hemiplegic patient. 1. a method for evaluation of physical performance', Scandinavian Journal of Rehabilitation Medicine, vol. 7, no. 1, pp. 13-31.
[5] Jin, J., Sun, H., Daly, I., Li, S., Liu, C., Wang, X., and Cichocki, A. (2022), 'A Novel Classification Framework Using the Graph Representations of Electroencephalogram for Motor Imagery Based Brain-Computer Interface', IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol.30, pp.20-29.