Frontiers in Neurology · Published 2026-07-29 · DOI 10.3389/fneur.2026.1859051
Peng Hu, Jingming Liu, Yan Tang, Jiantao Huang, Xinyi Chen, Chang Liu, Yuxing Wang
BackgroundFunctional recovery after Guillain-Barré syndrome (GBS) varies substantially. Prediction models that combine routinely available acute-phase clinical information with traditional Chinese medicine (TCM) syndrome features have not been adequately reported or validated. This study developed an acute-phase dynamic prediction model for 12-month functional outcome in GBS and evaluated discrimination, calibration, and potential clinical utility.MethodsWe retrospectively screened 601 patients admitted between January 2015 and December 2023; 525 met eligibility criteria and had 12-month outcome data. The prediction index time was the end of acute in-hospital assessment/discharge; no post-discharge follow-up variables were used as predictors. Candidate predictors included demographics, acute clinical severity, laboratory and electrophysiological findings, treatment-related acute-care variables, and standardized TCM syndrome elements. Poor outcome was defined as Hughes Functional Grading Scale (HFGS) score > = 3 at 12 months. Patients were stratified into a training cohort (n = 368) and a hold-out validation cohort (n = 157). Preprocessing, imputation, standardization, feature selection, and synthetic oversampling were performed within the training data only. Logistic regression, random forest, and a fully connected neural-network model were compared. Performance was assessed with 95% confidence intervals, calibration plots, Brier score, decision-curve analysis, and formal AUC comparisons.ResultsPoor 12-month outcome occurred in 148 of 525 patients (28.2%). Mechanical ventilation, admission HFGS score, AMSAN subtype, qi deficiency, and blood stasis were independently associated with poor outcome. In the hold-out validation cohort, AUCs were 0.744 (95% CI 0.655–0.833) for logistic regression, 0.843 (95% CI 0.780–0.906) for random forest, and 0.887 (95% CI 0.831–0.943) for the neural network. The neural network had the lowest Brier score (0.123) and showed acceptable calibration (calibration intercept 0.03; slope 0.94). DeLong testing showed that the neural network outperformed logistic regression (p = 0.006) but was not statistically superior to random forest (p = 0.174). Decision-curve analysis suggested net benefit for the machine-learning models across clinically plausible thresholds.ConclusionIn this single-center retrospective cohort, acute-phase clinical variables and TCM syndrome elements were associated with 12-month functional outcome in GBS. The neural-network model had the highest numerical AUC and lowest Brier score, but it was not statistically superior to random forest; therefore, no superiority over random forest is claimed. External validation, comparison with established prognostic scores when required variables are available, and prospective implementation studies are required before clinical use.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →
Hu, P., Liu, J., Tang, Y., et al. (2026). Predicting 12-month functional outcome in Guillain-Barré syndrome by combining acute-phase clinical data and traditional Chinese medicine syndrome features: a retrospective machine learning study. Frontiers in Neurology. https://doi.org/10.3389/fneur.2026.1859051