Explainable machine learning model for predicting functional outcomes in posterior circulation stroke after thrombectomy

Clinical Snapshot

Key Findings

• P Value: Not reported for primary discrimination metric

• Effect Size: AUC 0.81-0.90 depending on feature set and validation cohort

• Primary Outcome: Good functional outcome (mRS 0-3) at 3 months

• Nnt Or Sensitivity: Random Forest model with 7 preoperative features achieved AUC 0.83 (internal) and 0.81 (external validation)

• Confidence Interval: Not reported

Clinical Application

High - web-based calculator available, uses readily available clinical variables Relevant to Australian comprehensive stroke centers performing EVT; aligns with National Stroke Foundation guidelines for prognostic assessment Patients with posterior circulation stroke undergoing endovascular thrombectomy in tertiary stroke centers

Abstract

BACKGROUND: The early prediction of functional outcomes in patients with posterior circulation stroke (PCS) is crucial for timely interventions and optimizing treatment plans. We have developed and validated a machine learning (ML) model for predicting 3-month functional outcomes in patients with PCS undergoing endovascular thrombectomy (EVT). METHODS: The derivation cohort, consisting of 202 patients with PCS who underwent EVT at four medical centers from January 2020 to December 2023, was separated for training and internal validation, and an external dataset of 54 patients admitted from January 2020 to July 2023 was used for external validation. The target outcome was a good functional outcome, defined as a modified Rankin Scale score of 0-3 at 3 months. Seven ML models were trained using preoperative features, with the primary evaluation metric being the area under the receiver operating characteristic curve (AUC). The top performing model was further trained using intraoperative and postoperative features. Model interpretations were generated using the Shapley additive explanations (SHAP) method. RESULTS: The Random Forest model demonstrated the best discriminative ability among the models considered. After feature selection, the final preoperative model used seven features, achieving an AUC of 0.83 in the test set and 0.81 in the external validation cohort. The inclusion of intraoperative and postoperative features further enhanced the model's performance, resulting in an AUC of 0.84 and 0.90 in the test set and 0.83 and 0.90 in the external validation cohort, respectively. These models have been incorporated into a publicly accessible web-based calculator (https://zhelvyao-123-60-basilarz.streamlit.app). CONCLUSION: The interpretable ML models provide dynamic accurate predictions of functional outcomes in patients with PCS after EVT, offering valuable insights for personalized risk stratification and optimizing perioperative management, with potential for integration into clinical workflows.

References

1. 1.Yao, Z., Ji, Q., Zang, X., Yun, W., Luo, Y., Cao, J., Xu, J., Ke, Z., Xie, Z., Mao, C., Guan, Q., Lv, W., Zhu, Z., Huang, Y., Peng, Y., & Xu, Y. (2026). Explainable machine learning model for predicting functional outcomes in posterior circulation stroke after thrombectomy. Journal of Neurointerventional Surgery. https://doi.org/10.1136/jnis-2025-023624