Concurrent validity and reliability of an artificial intelligence–assisted physiotherapy assessment system for maximal mouth opening, cervical, and shoulder range of motion in patients with oral cancer

PURPOSE: Accurate assessment of joint range of motion (ROM) and maximal mouth opening (MMO) is essential in the rehabilitation of patients with oral cancer. This study aimed to evaluate the concurrent validity and reliability of an artificial intelligence (AI)-assisted physiotherapy assessment system (AIMS) for measuring MMO, cervical, and shoulder ROM in patients with oral cancer, with a healthy control group for comparison. METHODS: Twenty patients with oral cancer and 20 healthy individuals were included. ROM measurements obtained using AIMS were compared with clinician-based assessments using an electrogoniometer and a TheraBite ROM scale. Two human raters performed manual measurements. In the healthy group, all three raters assessed each movement, whereas in the patient group, AIMS and Rater 1 performed the assessments. Concurrent validity between the AIMS and clinician-based measurements was evaluated using intraclass correlation coefficients (ICC), standard error of measurement (SEM), minimal detectable change (MDC95), and Bland-Altman analysis. Within-session repeatability of the AIMS and human raters was evaluated using repeated measurements. RESULTS: The AIMS demonstrated variable concurrent validity across movement tasks. Concurrent validity was highest for shoulder abduction (ICC 0.68-0.80) followed by cervical lateral flexion (ICC 0.44-0.57), whereas lower agreement was observed for cervical rotation and MMO in the healthy group (ICC 0.03-0.35). In contrast, concurrent validity was generally higher in the patient group across all movement tasks (ICC 0.44-0.76). The AIMS demonstrated high within-session repeatability across repeated measurements. Bland-Altman analysis demonstrated relatively small mean bias across most movement tasks, although agreement varied according to movement and no consistent directional bias was observed. CONCLUSION: The AIMS demonstrated variable concurrent validity across movement tasks and high within-session repeatability, with the strongest performance observed for shoulder abduction and cervical lateral flexion. These findings support its potential as an objective tool for movement assessment in patients with oral cancer, while further refinement of cervical rotation and MMO is warranted.