Bilateral idiopathic carpal tunnel syndrome (CTS) is a neuromuscular condition involving the compression of the median nerve at both wrists, leading to pain, neurological symptoms, and loss of function. This paper proposes a robust machine-learning framework for a randomized crossover clinical trial comparing two physiotherapeutic treatment regimens: stretching followed by myofascial mobilization (S/M) and the reverse sequence (M/S). Instead of making inferences about the superiority of one treatment over another, the treatment regimen serves as a structured analytical label for investigating predictive separability, feature representation, and model stability within a controlled experimental setting. The clinical dataset of 73 patients underwent rigorous preprocessing, including strength feature aggregation and principal component analysis (PCA). Various classifiers were evaluated, with CatBoost achieving an ROC-AUC of 0.985 and a test accuracy of 96.5%, while Random Forest demonstrated strong adversarial robustness with an adversarial accuracy of 96.83%. To assess robustness, clinically constrained perturbations were introduced into the PCA feature space, simulating realistic input variability. The findings indicate that ensemble learning algorithms can capture structured patterns in crossover clinical datasets and remain stable under low-magnitude adversarial perturbations. The study underscores the importance of robustness evaluation and interpretability when applying machine learning models to biomedical data, particularly in small and well-structured clinical cohorts.
Kharat et al. (Tue,) studied this question.