Explainable ultrasound-based multimodal deep learning for outcome prediction after needle release in carpal tunnel syndrome: a multicenter study.
Authors
Affiliations (6)
Affiliations (6)
- Department of Trauma and Orthopedics, Peking University People's Hospital, No.11, Xizhimen South Street, Xicheng District, Beijing 100044, PR China. Electronic address: [email protected].
- Department of Biomedical Engineering, College of Chemistry and Life Science, Beijing University of Technology, No.100 Pingleyuan, Chaoyang District, Beijing 100124, PR China. Electronic address: [email protected].
- Department of Ultrasound, Tianjin Hospital, Tianjin University, No. 406, Jiefang South Road, Hexi District, Tianjin 300211, PR China. Electronic address: [email protected].
- Department of Ultrasound, The First Affiliated Hospital of Guangxi University of Traditional Chinese Medicine, No. 89, Dongge Road, Qingxiu District, Nanning, Guangxi 530023, PR China. Electronic address: [email protected].
- Department of Ultrasound, Shijiazhuang People's Hospital, No. 30, Fanxi Road, Changan District, Shijiazhuang, Hebei 50011, PR China. Electronic address: [email protected].
- Department of Trauma and Orthopedics, Peking University People's Hospital, No.11, Xizhimen South Street, Xicheng District, Beijing 100044, PR China. Electronic address: [email protected].
Abstract
Ultrasound (US)-guided percutaneous needle release is an effective treatment for carpal tunnel syndrome (CTS) before proceeding to surgery. However, conventional methods cannot be used to effectively predict prognosis.We aimed to develop explainable multimodal deep learning (DL) models to enhance individualized decision-making. Pre-therapy sonographic features (cross-sectional area, B-mode in transverse and longitudinal views, and color Doppler imaging), electrophysiological parameters, and clinical information of patients with CTS (n = 449) were retrospectively collected from four clinical centers for model training, validation, and internal testing. A multimodal DL model (CTSNet-P) was established after ablation experiments to predict the 6-month therapeutic outcome (effective or ineffective). Two additional cohorts (n = 169) were prospectively recruited for external testing. Shapley values were used to display feature contributions, while Grad-CAMs were used to elucidate the model's focus. Outcome predictions from six radiologists with varying levels of expertise were compared with those obtained using CTSNet-P in the test sets. Artificial intelligence (AI)-aided strategies were then conducted in the external test set to assess the model's impact on the radiologists' predictive performance. The patient-reported outcomes at 6 months after treatment served as the reference standard. The areas under the receiver operating characteristic curve (AUCs) in the test sets using CTSNet-P were 0.943 and 0.931, respectively. Higher AUCs were obtained with the model than the mean values of the six radiologists (0.943 vs. 0.798 and 0.931 vs. 0.793, respectively；all p < 0.001). The average AUC was significantly improved with assistance from CTSNet-P (0.793 vs. 0.862; p < 0.001). US consistently played the most significant role in the prediction process, and the heatmaps indicated a trend towards alignment with the attention patterns exhibited by the human radiologists. The interpretable CTSNet-P exceeded the average level of radiologists in prognosis prediction of needle release for CTS, and significantly augmented the radiologists' performance. However, these findings necessitate further investigations in a larger cohort.