발표연제 검색
- 연제번호 : P 2-39 북마크
-
제목 Multimodal MRI correlates of motor outcome after stroke using machine learni 소속 Veterans Health Service Medical Center, Seoul, Department of Physical Medicine and Rehabilitation1 저자 Jong Hyuk Lee1*, Dae Hyun Kim1† Objective: The objective of this study is to apply machine learning regression to predict motor outcome after stroke based on multimodal magnetic resonance imaging.
Methods: Fifty-four stroke patients, who underwent T1 weighted (T1), diffusion tensor (DTI), and resting state functional magnetic resonance imaging (rfMRI) were retrospectively included. All patients were assessed using the Fugl-Meyer motor assessment (FM) score for motor outcome after stroke (Table 1). The kernel rigid regression machine algorithm was applied to gray and white matter maps in T1, fractional anisotropy and mean diffusivity maps in DTI, and two motor related independent component analysis maps in rfMRI to predict FM scores with the covariate as the onset duration after stroke. The results were validated using the leave-one-subject-out cross-validation method. To the best of our knowledge, this is the first attempt at applying machine learning in this area using multimodal MRI data and forms the main novelty of this study.
Results: We found that multimodal magnetic resonance imaging correctly predicted the FM score in 72% cases with a normalized mean squared error of 5.93 (p value = 0.0020, Fig 1). The ipsilesional premotor, periventricular, and contralesional cerebellar areas were shown to be of relatively high importance in the prediction (Fig 2).
Conclusion: Machine learning using multimodal magnetic resonance imaging data after stroke enables a prediction of motor outcome.File.1: figure1.jpgScatter plot of predictions of Fugl-Meyer motor assessment scores of 54 patients modeled using a kernel ridge regression. This plot represents the predicted values (x-axis) against the real values (y-axis). A perfect prediction is represented by the diagonal on this plot.File.2: figure2.jpgWeights map at the voxel level for the prediction of Fugl-Meyer motor assessment. The cross-hair was positioned at the region with high contribution toward the outcome. The T1 weighted images are presented in the gray box while the x, y, and z-axes are based on the Montreal Neurological Institute space.File.3: table1.jpgGeneral characteristics of included patients
