Machine learning sorts Guillain-Barré syndrome subtypes from nerve conduction data

Researchers in South Korea have built interpretable machine-learning models that can assign Guillain-Barré syndrome (GBS) patients to electrophysiological subtypes using nerve conduction study (NCS) data collected during the acute phase of illness, according to a multicenter study published in BMC Medical Informatics and Decision Making.

GBS is not a single uniform disease but a family of clinically and electrophysiologically heterogeneous subtypes, and assigning a patient to the correct category can vary depending on when the nerve conduction study is performed, whether serological tests are available, and the expertise of the examiner. The study set out to test whether machine-learning models could help standardize this subtype assignment from acute-phase NCS data, both with and without serology.

The cohort comprised 99 patients drawn from nine Korean hospitals, split into an internal training set of 73 patients and an external validation set of 26. NCS parameters were harmonized across acquisition systems and stimulation protocols before modeling, an important step given that measurement differences between machines can otherwise distort model training.

The team trained gradient-boosting models—XGBoost, LightGBM, and CatBoost—to classify three mechanism-oriented operational subtypes: axonal-predominant, demyelinating, and Miller Fisher spectrum. The without-serology condition served as the primary analysis, reflecting the frequent clinical reality where serological results are unavailable at the time of early assessment.

Performance was assessed using the area under the receiver operating characteristic curve (AUC) and the area under the precision-recall curve (PRC) over repeated internal splits, with the external set held out as an independent validation set. In the primary without-serology analysis, macro-averaged AUC reached 0.87 internally and 0.80 on external validation. Adding serology improved discrimination to 0.92 internally and 0.90 externally.

Beyond raw discrimination, the researchers applied SHAP (SHapley Additive exPlanations) analysis to examine whether the models’ feature attributions matched conventional electrophysiological patterns. The attributions identified subtype-relevant NCS features and supported what the authors describe as model face validity—meaning the models appeared to rely on physiologically sensible signals rather than artifacts of the data.

The authors are careful to frame these results as preliminary. Performance estimates were limited by the small external sample, and the conclusions emphasize that larger prospective cohorts with independent adjudication and outcome linkage are needed before such models can be used for clinical decision-making or mechanistic inference. The models reproduced expert-assigned subtype assignments; they did not independently establish subtype diagnoses against a gold standard beyond expert assignment.

The study is published as an open-access article under a share-first arrangement, meaning the version released on 26 October 2026 is citable and carries a permanent DOI but remains subject to further edits before the final Version of Record replaces it.

For clinicians, the appeal of the approach lies in standardization: an interpretable model applied to harmonized acute-phase NCS data could reduce variability in subtype assignment across centers and examiners, particularly in settings where serology is not immediately available. For researchers, the SHAP-based consistency with expected electrophysiological patterns suggests the models capture known disease mechanisms rather than spurious correlations, though this remains a validity check rather than proof of clinical utility.

The work represents an early step toward algorithmic support for GBS classification. Its reported performance—0.87 internal and 0.80 external AUC without serology, rising to 0.92 and 0.90 with serology—demonstrates feasibility in a modest multicenter cohort, while the authors’ own caveats about sample size and the need for prospective validation define the path forward before any bedside use.

Subject of Research: Interpretable machine learning for Guillain-Barré syndrome subtype classification from acute-phase nerve conduction studies

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

Article References: Yoon, H., Kim, Y., Kim, Y., Park, J.-H., Baek, S.-H., Park, J.-W., Kim, J. K., Oh, S.-I., Kim, S., Seok, H. Y., Kim, S., Kang, S.-Y., Cho, E. B., Kim, S., Kim, J., Sohn, E., Yoon, B.-A., Kim, B.-J., & Lee, H. (2026). Interpretable machine-learning models for Guillain-Barré syndrome subtype assignment using acute-phase nerve conduction study: a multicenter study. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03920-4

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03920-4

Keywords: Guillain-Barré syndrome, machine learning, nerve conduction study, XGBoost, SHAP, electrophysiology, multicenter cohort, Miller Fisher syndrome, demyelinating neuropathy, clinical decision support, South Korea, BMC Medical Informatics and Decision Making