A Telemedicine App for Nonrigid Facial Rehabilitation Training Enhanced by Efficient Fully Convolutional Neural Network With Residual Network (EffiFCNN-ResNet) to Improve Accessibility for Patients With Nasopharyngeal Carcinoma Cancer: Randomized Controlled Trial

BACKGROUND: Resource limitations in public hospitals may hinder timely monitoring and management of rehabilitation in patients with nasopharyngeal carcinoma (NPC) after radiotherapy. OBJECTIVE: This study developed and evaluated the telemedicine app "Open Care," which integrates the Efficient Fully Convolutional Neural Network with Residual Network (EffiFCNN-ResNet) model and computer vision to monitor facial training exercises and provide real-time feedback, aiming to improve outcomes in patients with restricted mouth opening. METHODS: Initially, the EffiFCNN-ResNet model underwent 5-fold cross-validation, expert validation, and robustness testing to assess its reliability and clinical applicability in complex real-world environments. Subsequently, to evaluate the telemedicine app, a parallel-group, 2-arm randomized controlled trial was conducted with 109 patients, who were randomly assigned to either the intervention group (n=55) or the control group (n=54). The intervention group performed mouth-opening exercises under the supervision and guidance of the telemedicine app, whereas the control group followed traditional video-based instructions. Primary outcome measures included maximum mouth opening, mouth-opening symmetry, exercise frequency, and rehabilitation-related health beliefs. Secondary outcomes included fatigue (Brief Fatigue Inventory), health-related quality of life (Assessment of Quality of Life-6 Dimensions), and system usability scores. Data were analyzed using 2-tailed (unpaired) independent-samples t tests and chi-square tests, and the Mann-Whitney U test was used to assess intra- and inter-group differences before and after the intervention. RESULTS: The "Open Care" system leverages a lightweight fully convolutional neural network (FCNN) depth model integrated with network communication to enable real-time capture, recognition, and correction of nonrigid facial training movements. It also provides visual feedback and supports automated rehabilitation assessment. The model demonstrated strong generalization ability (macro-averaged F1-score, mean 0.96, SD 0.01) and clinical-grade stability (performance degradation: mean 5.2%, SD 0.6%, under lighting disturbances and challenging pathological cases; n=160 video segments). Compared with the control group, the intervention group showed significant improvements in maximum mouth opening (P=.04), exercise frequency (P=.001), perceived severity (P=.007), perceived benefits (P=.04), perceived barriers (P=.001), self-efficacy (P=.04), cues to action (P=.001), health behavior (P=.03), and fatigue (P=.04). Participants also reported favorable training experiences, with a mean system usability score of 74.3 out of 100. CONCLUSIONS: This telemedicine approach was more effective than traditional methods, improving patient engagement and rehabilitation outcomes while providing a more objective and precise monitoring tool. Future apps may benefit patients with NPC and other head and neck cancers. TRIAL REGISTRATION: Chinese Clinical Trial Registry ChiCTR2400090305; https://www.chictr.org.cn/showprojEN.html?proj=235073.