Semester of Graduation
Summer 2026
Degree Type
Thesis
Degree Name
Masters of Artificial Intelligence
Department
Computer Science
Committee Chair/First Advisor
M. Rasel Mahmud
Second Advisor
Chen Zhao
Third Advisor
Md. Shazibul Islam Shamim
Abstract
Rehabilitation assessment often relies on periodic observation, while many XR prototypes show scripted rather than recorded-motion evidence. iD-MORE Vision is an offline pipeline trained on KIMORE and IRDS and linked through JSON packets to a two-mode Unity desktop prototype. Both datasets include controls and rehabilitation participants with neurologic, musculoskeletal, or mobility impairments. This improves relevance but does not clinically validate the system.
Under fixed subject-wise splits, the primary five-seed Random Forest predicted KIMORE clinician scores with MAE 6.087 ± 0.044 cTS and R² 0.568 ± 0.006; the subject-level R² interval crossed zero. The primary IRDS five-run CUDA GRU averaged 0.877 accuracy and 0.814 macro-F1; a secondary full-sequence ensemble reached 0.899 and 0.842 with 0.669 incorrect recall. A validation-selected threshold raised recall to 0.992 while reducing incorrect precision to 0.504. The runtime model reached only 0.322 incorrect recall after aggregation and is not ready for autonomous use. Calibration did not help, and learned anomaly detection did not consistently beat amplitude.
Python packets contain recorded joints, model scores, geometry, and provenance; Unity performs no inference. Twelve valid packets demonstrated desktop playback. No live sensing, HMD, clinician usability, or prospective study was conducted. The result is a reproducible decision-support prototype, not a clinical system.
Included in
Computational Engineering Commons, Other Analytical, Diagnostic and Therapeutic Techniques and Equipment Commons