Integrating machine-learning-enabled hand tracking and rule-based automated assessment into ICSI simulation training: development and validation of a browser-based platform

Intracytoplasmic sperm injection (ICSI) is technically demanding and difficult to practice outside specialist assisted-reproduction laboratories. This study developed a browser-based simulator designed for use with consumer-grade hardware, using machine-learning-enabled hand tracking and predefined rule-based assessment, and evaluated feasibility, preliminary known-groups validity, short-term within-simulator performance change, and usability. MediaPipe Hands detected hand landmarks through a standard webcam, while predefined gestures controlled virtual holding and injection instruments. Graphical deformation and aspiration were represented using simplified two-dimensional effects. Five experienced ICSI practitioners and 20 medical students completed five consecutive scored trials under standardized conditions. The primary analysis compared expert and novice Trial 1 scores. Novice repeated scores were analyzed using repeated-measures analysis of variance, and repeated aspiration-completion-threshold status using Cochran’s Q. Twenty novices completed the System Usability Scale (SUS) and an exploratory eight-item questionnaire. Trial 1 total scores were higher for experts than novices (92.8 ± 1.3 vs. 42.5 ± 2.9; Welch t(14.85) = 58.16, P < 0.001; mean difference 50.3, 95% CI 48.46–52.14). The difference remained after excluding the Safety component (93.0 ± 1.3 vs. 68.0 ± 1.2; P < 0.001). Novice scores increased from 42.5 ± 2.9 to 79.2 ± 4.9 across five trials (Greenhouse–Geisser-corrected F(1.41, 26.85) = 2413.01, P < 0.001). The repeated-trial effect remained significant after excluding Safety (F(2.54, 48.17) = 89.44, P < 0.001). Failure to reach the predefined aspiration-completion threshold decreased from 19/20 (95%) to 0/20 (0%) (Cochran Q(4) = 52.51, P < 0.001). The mean SUS score was 82.5 ± 5.4. The platform distinguished two extreme experience groups and was associated with short-term improvement on repeated simulator trials. These findings constitute preliminary known-groups evidence and within-simulator performance change; they do not establish reliability, retention, transfer to biological ICSI, superiority over conventional training, or a causal effect of automated feedback. Not applicable. This was a validation study of a simulation-training platform and did not involve a clinical intervention, patient recruitment, or therapeutic outcome.