Optical hand kinematics for machine learning-based screening of carpal tunnel syndrome

Carpal tunnel syndrome is usually assessed by clinical examination and electrodiagnostic testing, but these methods may not always be practical for rapid screening. In this work, it was explored whether hand motion recorded with a low-cost optical camera could add useful information to routine clinical data. Demographic variables, symptom-related findings obtained after the Phalen test, and five thumb-to-finger opposition tasks were collected. Hand landmarks were extracted with MediaPipe and used to derive kinematic descriptors, which were then analyzed together with clinical variables by machine-learning and hybrid deep-learning models. To reduce the risk of overly optimistic estimates, the main evaluation was carried out at hand level with leave-one-group-out cross-validation. In the primary leakage-safe comparison, the best Dataset 4 multimodal model yielded the strongest binary screening profile. Compared with the clinical Dataset 1 baseline, accuracy increased from 82.6% (95% CI 75.0%-90.2%) to 87.0% (79.3%-93.5%), sensitivity from 70.0% (54.4%-85.1%) to 80.0% (64.7%-93.4%), specificity from 88.7% (80.3%-95.6%) to 90.3% (83.1%-96.8%), and ROC-AUC from 0.845 (0.745-0.934) to 0.855 (0.752-0.945). In the same-architecture tests, the kinematic variables did not behave uniformly across model families. The clearest gain among the classical models was seen with light gradient boosting machine. Feature-level results were also consistent with the model findings, as selected thumb-opposition coordinates differed across stages and appeared to capture a functional component not represented by symptom scores. In the four-class severity task, the improvement was small: accuracy increased from 81.5% (72.8%-89.1%) to 82.6% (73.9%-90.2%), and macro-1 increased from 0.706 (0.544-0.823) to 0.748 (0.588-0.850). In this cohort, the optical hand-kinematic measures mainly strengthened the screening analysis. Their use for finer severity staging will need confirmation in larger external datasets and in participant-level analyses.