Abstract
BACKGROUND: Carpal tunnel syndrome (CTS) is the most common neuropathy affecting the upper limbs. Carpal tunnel release (CTR) surgery effectively alleviates symptoms; however, patient satisfaction and clinical outcomes post-surgery can vary significantly. This study aimed to develop a machine learning (ML) model to predict post-CTR patient satisfaction and outcomes, serving as a preoperative screening tool. METHODS: Data from 303 participants diagnosed with CTS who underwent minimally invasive open carpal tunnel release (CTR) under local anesthesia were analyzed. These participants were referred to our center between January 2015 and July 2023. Outcomes evaluated included symptom severity and functional status, measured using the Boston Carpal Tunnel Questionnaire (BCTQ), as well as patient satisfaction post-CTR. Based on BCTQ scores, patients were categorized into desirable and undesirable groups: for the symptom severity scale (SSS), scores of 11–24 were classified as desirable, while scores of 25–55 were classified as undesirable; for the functional status scale (FSS), scores of 8–16 were considered desirable, and scores of 17–40 as undesirable. Four machine learning algorithms—random forest (RF), gradient boosting machine (GBM), support vector machine (SVM), and k-nearest neighbors (k-NN)—were used to predict these outcomes. The models’ performance was assessed based on accuracy, sensitivity, specificity, and area under the curve (AUC). RESULTS: The mean participant age was 50.8 years (± 7.76), with 76.2% female. Among the cohort, 207 patients (68%) reported satisfaction post-surgery, and 212 (70%) achieved desirable SSS outcomes. Additionally, 249 patients (82.2%) attained desirable FSS outcomes. The RF model identified grip strength, age, BMI, electrophysiological results, and symptom duration as key factors influencing clinical outcomes and satisfaction. The RF model demonstrated the highest accuracy in predicting patient satisfaction at 0.901 (95% CI: 0.869, 0.933), followed by SVM at 0.876 (95% CI: 0.844, 0.908), GBM at 0.872 (95% CI: 0.840, 0.904), and k-NN at 0.859 (95% CI: 0.826, 0.892). Additionally, RF demonstrated excellent sensitivity of 0.852 (95% CI: 0.817, 0.887), specificity of 0.941 (95% CI: 0.909, 0.973), and AUC of 0.894 (95% CI: 0.860, 0.927) for patient satisfaction.Regarding SSS prediction, SVM achieved the greatest accuracy of 0.892 (95% CI: 0.859, 0.925), sensitivity of 0.824 (95% CI: 0.790, 0.858), specificity of 0.961 (95% CI: 0.948, 0.974), and AUC of 0.894 (95% CI: 0.860, 0.927).For FSS prediction, RF achieved the highest accuracy of 0.933 (95% CI: 0.921, 0.945), with k-NN following at 0.890 (95% CI: 0.854, 0.925), GBM at 0.870 (95% CI: 0.837, 0.903), and SVM at 0.802 (95% CI: 0.769, 0.835). RF also recorded the best sensitivity of 0.901 (95% CI: 0.889, 0.913), specificity of 0.960 (95% CI: 0.951, 0.969), and AUC of 0.932 (95% CI: 0.920, 0.944) for FSS predictions. CONCLUSIONS: The strong performance of these machine learning models underscores their potential to complement clinical judgment by providing insights into expected surgical outcomes. Such tools could assist in personalized patient counseling and help reduce unnecessary surgeries.