AI Application for Movement Recognition in Assessing Practical Skills of Healthcare Student - A Feasibility Study

The primary objective was to explore the feasibility of using an AI-based video analysis system to support assessment of healthcare students’ practical skills. A further objective was to examine whether model outputs could be aligned with educator rubrics and used to reduce manual video review. This feasibility study combined a person detection model (Faster R-CNN) with a SlowFast network (ResNet‑50 backbone), pre-trained on Kinetics 400 and finetuned on the AVA action‑detection benchmark, to recognise practical skill movements. The model was evaluated on anonymised, educator‑annotated video clips from undergraduate physiotherapy simulation sessions, labelled with action categories and rubric‑based judgements. The Faster R‑CNN person detector achieved an accuracy of 0.74 on educator‑annotated clips, while the AVA‑trained SlowFast model achieved an accuracy of 0.75 for action recognition, with its strongest performance for common actions such as walking and standing. Model summaries correlated well with educator rubric scores and indicated that around 30 - 40% of clips could be selectively reviewed rather than watched in full, substantially reducing manual video‑review effort. This study shows that a SlowFast model trained on AVA, combined with person detection, can provide reliable, interpretable movement recognition that aligns with educator assessments of practical skills. It functions effectively as a decision-support tool, pre-screening video while preserving educator control over more complex and high-stakes decisions. Integrating AI-based movement recognition into simulation-based teaching could help educators monitor large cohorts more efficiently, focus their attention on complex or concerning clips and provide more timely, consistent feedback.