| 제목 |
Prediction of arm impairment by machine learning algorithm from kinematic measures in stroke |
| 소속 |
Chung-Ang University Hospital, Chung-Ang University College of Medicine, Department of Physical Medicine and Rehabilitation1, Neofect, Data Science & Rehab Research Team2 |
| 저자 |
Jun Mo Jo1*, Yeongchang Jo1, Dokyeong Ha2, Hyoseok Yi2, Dongseok Yang2, Si Hyun Kang1, Don-Kyu Kim1, Kyung Mook Seo1, Jaewon Beom1† |
Objective: To predict hemiparetic upper-limb impairment by machine learning algorithm from kinematic measures in a 2-dimensional planar device in stroke patients.
Methods: In the multi-center observational study, 63 subacute and chronic stroke patients with hemiparetic arm of Brunnstrom stage 3, 4, or 5 were enrolled. Hemiparetic arm function was evaluated and trained with 3 kinds of tasks (free exploration, point-to-point reaching, and round shape drawing) using the RAPAEL Smart BoardTM (Neofect, Korea) (Figure 1). The device has 2-dimensional planar board and position sensors.
Results: Among 63 subjects, the ratio of patients of Brunnstrom stage 3 was 61%, whereas stage 4 was 17%, and stage 5 was 22%. Among the kinematic variables, zero crossings in acceleration, mean arrest period rate (Figure 2), hand path ratio, and duration time in point-to-point reaching task had significant correlation with Fugl-Meyer assessment scale. Those variables showed higher correlation in right hemiparesis than in left. In the patients who showed much improvement in the Fugl-Meyer scale, zero crossings in acceleration and duration time decreased in point-to-point reaching task. Zero crossings in acceleration, reaction time, and duration time revealed correlation with box and block test as well as pegboard test. Bias in X-axis (Figure 3) had negatively correlated with Fugl-Meyer scale in round shape drawing task. From these features, the mean absolute error for prediction of Fugl-Meyer scale using 5-fold cross validation in artificial neural network was 14.25 points per 66 (cross-validated R2=0.51).
Conclusion: Upper-limb impairment in stroke patients can be predicted by machine learning algorithm from main kinematic variables in a 2-dimensional planar device. Various kinematic measures were correlated with clinical parameters. An accurate machine learning algorithm needs to be drawn with big data. |
|
Three tasks in a 2-dimensional planar device
Kinematic measures for movement smoothness
Bias for accuracy assessment
|
