Interpretable machine-learning models for Guillain-Barré syndrome subtype assignment using acute-phase nerve conduction study: a multicenter study

Guillain-Barré syndrome (GBS) comprises clinically and electrophysiologically heterogeneous subtypes whose assignment may vary according to nerve conduction study (NCS) timing, serological availability, and examiner expertise. Interpretable machine-learning models may help standardize subtype assignment from acute-phase NCS data, with and without serology. In a multicenter cohort of 99 patients across nine Korean hospitals (73 internal, 26 external), NCS parameters harmonized across acquisition systems and stimulation protocols were used to train gradient-boosting models (XGBoost, LightGBM, CatBoost) to classify three mechanism-oriented operational subtypes (axonal-predominant, demyelinating, and Miller Fisher spectrum) under with-serology and without-serology conditions, with the without-serology model as the primary analysis. Discrimination was assessed by the area under the receiver operating characteristic (ROC) curve (AUC) and the area under the precision-recall curve (PRC) over repeated internal splits, with the external set serving as an independent validation set. SHapley Additive exPlanations (SHAP) analysis was used to examine whether model attributions were consistent with conventional electrophysiological patterns. In the primary without-serology analysis, macro-averaged AUC was 0.87 internally and 0.80 on external validation, increasing to 0.92 and 0.90 with serology. SHAP attributions identified subtype-relevant NCS features and supported model face validity, although performance estimates were limited by the small external sample. Interpretable machine-learning models showed preliminary ability to reproduce expert-assigned GBS subtype assignments from acute-phase NCS. Feature attributions were consistent with expected electrophysiological patterns, supporting model face validity. Larger prospective cohorts with independent adjudication and outcome linkage are needed before these models can be used for clinical decision-making or mechanistic inference.