An Artificial Intelligence-Assisted Decision Support Tool Developed Using Real Patient Data: Classification of Normal and Neuropathic Findings in Pediatric Nerve Conduction Studies.

BACKGROUND: Normal values for nerve conduction studies (NCS) have not been clearly defined in the pediatric population, creating challenges in clinical practice. This study aimed to develop a machine learning (ML)-based decision support tool to distinguish normal from neuropathic values in childhood.

METHODS: We retrospectively analyzed NCS data collected from 1007 children over a 6-year period at a single pediatric neuromuscular center. Parameters derived from median, ulnar, peroneal, and tibial motor studies along with median and sural sensory studies were used to train Logistic Regression (LR), Random Forest, and XGBoost models. Model performance was evaluated using accuracy, recall, specificity, receiver operating characteristic area under the curve, and ΔF1 scores. SHapley Additive exPlanations analysis was applied to identify influential features. Age-based normative values were calculated for all nerves.

RESULTS: Tree-based models generally outperformed LR, except in the median motor nerve dataset. XGBoost achieved accuracy ≥0.91 and recall ≥0.67 across all datasets, with specificity ≥0.98. Random Forest showed the highest accuracy in peroneal (0.98) and tibial nerves (0.94), with recall values exceeding 0.88 in these motor nerve datasets. ΔF1 values remained within ±0.02 for all models except ulnar LR, indicating strong generalization and minimal overfitting. SHapley Additive exPlanations analyses showed that amplitude and conduction velocity were the most influential predictors for neuropathy classification across all nerves. In sensory nerve models, age had a greater impact than in motor nerve models, particularly in the median sensory nerve, while latency contributed less than age in the sural nerve model, which is consistent with the known age-dependent maturation of sensory conduction parameters in children.

CONCLUSIONS: We developed a practical ML-based decision support tool that accurately differentiates normal from neuropathic NCS findings. The tool performed especially well regarding motor nerves. This study suggests that ML-based tools may offer meaningful support in clinical decision-making. However, due to its single-center design of the study, external validation in independent cohorts is required.