Therapeutics
Yongxin Luo, Master of Medicine
Ms.
School of Rehabilitation Science, Shanghai University of Traditional Chinese Medicine; Engineering Research Center of Traditional Chinese Medicine Intelligent Rehabilitation
Shanghai, Shanghai, China (People's Republic)
Li Ding, PhD
phD
Department of Rehabilitation Medicine, Huashan Hospital Fudan University, Shanghai, China
Shanghai, Shanghai, China (People's Republic)
Jian Lu, PhD
Professor
School of Rehabilitation Science, Shanghai University of Traditional Chinese Medicine, Shanghai, China
Shanghai, Shanghai, China (People's Republic)
The degree of embodiment is a critical determinant of mirror visual feedback (MVF) efficacy variations, yet remains challenging to quantify through conventional psychological paradigms. This study optimizes the upper-limb MVF task protocol and establishes a machine learning-powered embodiment prediction model, thereby advancing the standardization and clinical application of MVF techniques.
Design:
A randomized observational crossover study enrolled 69 unilateral stroke patients with upper limb motor dysfunction (mean age 53.19±15.06 years; 47M/22F; hemorrhagic/ischemic: 28/41; simplified Fugl-Meyer: 27.69±26.50; modified Barthel: 77.72±22.26). A 40×60 cm mirror was fixed at the body midline, with the unaffected limb on the reflective side. Patients completed two trial rounds (2-hour interval) of randomly ordered tasks: simple motor tasks (SMT1-3: wrist dorsiflexion, fist unfolding, finger pairing) and object-based tasks (OBT1-3: wooden block grasping, bouncing ball holding, paper clip manipulation).
Task-specific visual illusion frequency (F) and latency (LT) were recorded. Post-trial embodiment questionnaires (EQ) assessed symmetry (S1, S2), ownership (O1, O2), agency (A1, A2), and deafference (D1, D2). Statistical analysis used SPSS 26.0. Clinical scales and behavioral data were normalized in PyCharm2023, split 7:3 into training/validation sets. Feature selection and model training employed Random Forest (RF) and Support Vector Machine (SVM).
Results: There were significant differences in behavioral (F, LT) and EQ scores between SMT and OBT on the motor task by Wilcoxon signed rank sum test (see Figure 1).
< ![if !vml] >
< ![endif] >
Figure 1.
The average accuracy of the RF-based algorithm prediction model was 70.83% and the average accuracy of the SVM-based algorithm prediction model was 74.48% .
Conclusion: The behavior results demonstrate that combined object-movement training protocols increased visual illusion frequency with reduced latency, effectively eliciting embodied experiences of symmetry, agency, and deafference in patients. For the degree of embodiment prediction, the SVM algorithm exhibits superior classification accuracy.
