Data-driven prognostic stratification in Guillain-Barré syndrome: a machine learning approach

Current classifications of Guillain-Barré syndrome (GBS) cannot adequately predict the prognosis and response to immunotherapy of patients, leading to treatment uncertainty and suboptimal outcomes. This study aimed to establish a novel, data-driven classification to identify clinically distinct GBS subtypes and improve outcome prediction. Clinical, laboratory, and electrodiagnostic data from 155 patients with GBS who underwent plasmapheresis were retrospectively analyzed. A clustering algorithm, mixture model, was applied to identify distinct patient subgroups. Subsequently, a streamlined, interpretable decision tree model was developed using the most crucial features. Three distinct clusters were identified. Cluster 1 ( n = 39) represented the most severe presentation, with a median admission GBS disability scale of 5 and the lowest rate of post-treatment independent ambulation at 30 days after symptom onset (12.8%). Cluster 2 ( n = 61) demonstrated the baseline severity (median of GBS disability scale: 3) at admission and the most favorable prognosis, with 54.1% having independent ambulation. Cluster 3 ( n = 55), characterized by prominent demyelinating features, showed a notably higher response rate to plasmapheresis (85.5% of patients improving by ≥ 1 GBS disability scale), despite presenting with intermediate severity (median of GBS disability scale: 4). The decision tree model, incorporating five common clinical and electrophysiological variables, achieved excellent classification fidelity to practically approximate the clusters (accuracy: 0.877). Machine learning identified three GBS subtypes with distinct prognoses and clinical trajectories to plasmapheresis. Derived from these findings, a simple five-variable decision tree provides a practical tool for early prognostic stratification, facilitating early prognostic outcome estimation and risk-adapted clinical monitoring.