Multi-class severity classification of peripheral nerve impairment using multimodal electrodiagnostic features and...

Peripheral nerve injuries impose substantial functional morbidity and long-term neurological impairment, yet electrodiagnostic severity assessment remains dependent on specialized expertise and susceptible to inter-examiner variability. This study proposes a supervised machine learning framework for automated four-class severity classification of upper-extremity peripheral nerve impairment using integrated electromyography (EMG) and nerve conduction study (NCS) features.

Five classifiers were evaluated using stratified five-fold cross-validation: support vector machine (SVM), random forest, LightGBM, XGBoost, and tabular prior-data fitted network (TabPFN). Feature selection combined Kruskal-Wallis and Chi-square statistical testing with Fisher score ranking. For the integrated EMG-NCS configuration, Gaussian Copula-based synthetic augmentation addressed the scarcity of paired electrodiagnostic samples (28 records).

On the EMG dataset (n = 912), the ensemble of SVM, random forest, and TabPFN achieved the highest accuracy (73.77%) and macro-F1 (72.64%). On the NCS dataset (n = 1020), TabPFN achieved 98.04% accuracy and a macro-F1 of 98.60%, with the ensemble model producing comparable performance. On the augmented training distribution, SVM and TabPFN achieved 100% accuracy and macro-F1 on an independent held-out test set. Given the limited test size (n = 9), these findings are presented as proof of concept pending validation on larger paired clinical cohorts. SHAP-based explainability identified MUAP polyphasia percent as the primary EMG predictor, and sensory latency and motor conduction velocity as the key NCS predictors.

These findings demonstrate the feasibility of interpretable machine learning for multimodal electrodiagnostic analysis and support its potential as a scalable decision-support tool for objective peripheral nerve impairment severity grading.