Multicentre development and validation of a risk model integrating immunotherapy and coagulation biomarkers for thrombosis in autoimmune neurological disorders

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: AUROC 0.949 for XGBoost model vs traditional Padua score

• Primary Outcome: Thrombosis prediction accuracy in patients with autoimmune neurological disorders

• Nnt Or Sensitivity: AUROC 0.949, AUPRC 0.544, Brier score 0.045

• Confidence Interval: Not reported

Clinical Application

Requires implementation of machine learning infrastructure and standardised biomarker collection; three-tier risk stratification system appears practical for clinical use Relevant to Australian neurologists and immunologists managing conditions like multiple sclerosis, myasthenia gravis, and autoimmune encephalitis; could inform PBS thromboprophylaxis prescribing decisions and align with RACGP prevention guidelines Patients with autoimmune neurological disorders receiving immunomodulatory therapy, particularly those with elevated D-dimer, prolonged immobilisation, or advanced age

Abstract

PURPOSE: Patients with Autoimmune Neurological Disorders (ANDs) require routine immunomodulatory therapy, which inherently increases thrombosis risk. Despite this recognised association, reliable tools for thrombosis prediction remain limited, highlighting the need for accurate risk stratification to inform targeted prevention strategies. PATIENTS AND METHODS: This multicentre retrospective study developed a machine learning-based thrombosis risk prediction model using data from 786 patients with ANDs receiving immunomodulatory therapy. The model was trained on data from three centres and externally validated on an independent fourth centre, evaluating six algorithms based on the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and Brier score. Feature selection and Shapley Additive Explanations (SHAP) analysis were applied to ensure model interpretability. RESULTS: Among all algorithms tested, XGBoost (XGB) demonstrated the best performance. On the external validation set, the XGB model achieved an AUROC of 0.949, AUPRC of 0.544, and Brier score of 0.045, significantly outperforming the traditional Padua score. SHAP analysis identified the top three thrombotic risk factors: elevated D-dimer levels, prolonged immobilisation, and advanced age. The total doses of intravenous immunoglobulin and glucocorticoid were also confirmed as significant contributors. The model further enabled stratification into low-, intermediate-, and high-risk groups, offering a practical tool for individualised thromboprophylaxis. CONCLUSION: This study presents a high-performing, interpretable model for early thrombosis risk prediction in patients with ANDs, which facilitates timely clinical decision-making and targeted preventive strategies to reduce thromboembolic complications.