An Artificial Intelligence-Assisted Decision Support Tool Developed Using Real Patient Data: Classification of Normal...

Area of Science:

• Pediatric Neurology
• Machine Learning in Medicine
• Clinical Electrophysiology

Background:

• Establishing normal nerve conduction study (NCS) values in children is challenging due to a lack of defined pediatric norms.
• This gap hinders accurate diagnosis of pediatric neuropathies.

Purpose of the Study:

• To develop and evaluate a machine learning (ML)-based decision support tool for differentiating normal from neuropathic NCS findings in pediatric patients.
• To establish age-based normative values for NCS parameters in children.

Main Methods:

• Retrospective analysis of NCS data from 1007 children across six years.
• Training and evaluation of Logistic Regression, Random Forest, and XGBoost models using motor and sensory nerve conduction parameters.
• Performance assessment using accuracy, recall, specificity, AUC, and ΔF1 scores; SHapley Additive exPlanations (SHAP) for feature importance.

Main Results:

• Tree-based models (Random Forest, XGBoost) generally outperformed Logistic Regression.
• XGBoost demonstrated high accuracy (≥0.91) and specificity (≥0.98) across datasets.
• Random Forest achieved superior accuracy for peroneal and tibial nerves (≥0.94) with high recall (≥0.88).
• SHAP analysis identified amplitude and conduction velocity as key predictors; age was more influential in sensory nerve models.

Conclusions:

• A practical ML-based decision support tool was developed, accurately differentiating normal from neuropathic NCS findings, particularly for motor nerves.
• The study highlights the potential of ML tools to support clinical decision-making in pediatric neurology.
• External validation in independent cohorts is recommended due to the single-center design.