AI FOR PHYSICAL THERAPY EDUCATION

Multi-Stage AI Pipeline for Movement Analysis

The core of our solution is a sophisticated AI pipeline designed for robust movement analysis. It begins with Person Detection using RTMDet, followed by precise Pose Estimation with RTMPose, which extracts 17 standard 2D keypoints. Multi-Object Tracking and Role Identification (via ByteTrack and cosine similarity) ensures consistent labeling of therapist and patient. Subsequently, Feature Engineering generates time-series datasets, including spatial normalization, derived kinematic/postural features, and scale-invariant relational features. Finally, a Compact Gated Recurrent Unit (GRU) Model is trained for sequence classification, distinguishing correct from incorrect executions. This integrated architecture ensures comprehensive and reliable assessment.

Validated Keypoint Detection for PROM Assessment

The pose estimation component of our pipeline achieved remarkable accuracy in detecting key joint landmarks crucial for PROM exercises. For the Left Elbow Extension exercise, most keypoints (including therapist's shoulders, elbows, and patient's shoulders, elbow) achieved 100% identification rates, with the therapist's right wrist and patient's left wrist showing slight variations (96.7% and 90%, respectively) due to minor occlusions. Similarly, for Left Shoulder Flexion, identification rates were consistently high, ranging from 97.6% to 100% for therapist keypoints, and 92.9% to 100% for patient shoulders/elbow, except for the patient's left wrist (81%) where occlusions were more pronounced. Overall, this demonstrates the pipeline's strong foundation for accurate movement analysis.

High-Fidelity Classification of Exercise Correctness

Our Gated Recurrent Unit (GRU) model demonstrated strong performance in classifying the correctness of PROM exercises. For Elbow Extension, the model achieved an impressive 96.8% accuracy with a 94.9% precision, 97.4% recall, and an F1-score of 96.1%. For Shoulder Flexion, performance was also robust, yielding 93.5% accuracy, 89.2% precision, 94.3% recall, and a 94.7% F1-score. These metrics confirm the model's high reliability in movement analysis, proving its capability to accurately discern between correctly and incorrectly performed exercises, a crucial step for automated feedback in physical therapy education.

Movement Analysis Model Performance

Robust End-to-End System Performance for PT Education

The complete end-to-end pipeline, integrating pose estimation and movement analysis, demonstrated robust performance in practical application. For Elbow Extension, the system achieved an overall 96.8% accuracy, with a 94.9% precision, 97.4% recall, and a 96.1% F1-score. Similarly, for Shoulder Flexion, the system maintained high reliability with an overall 95.7% accuracy, 91.9% precision, 97.1% recall, and a 94.4% F1-score. This consistent high performance across both exercises validates the integrated system as a reliable and practical tool for real-world PROM assessment in physical therapy education, providing objective, automated feedback for students.

Future Directions for Advanced PT Solutions

To further enhance the system's capabilities, several strategic next steps are recommended. A crucial advancement involves transitioning to 3D pose estimation via multi-view triangulation to achieve true volumetric analysis and eliminate view-dependent distortions. Fine-tuning perception models (RTMDet, RTMPose) on custom clinical datasets, rather than relying solely on general datasets, will significantly improve robustness and generalization. Expanding the training dataset and exploring more complex sequence architectures like LSTMs will allow the system to learn more intricate temporal patterns, supporting analysis of a wider range of PT procedures and complex movements.