Abstract WP102: Risk Score Estimation For The Prognostication Of Post-stroke Functional Independence Using Data-driven Score Assignment With Clinical And Imaging Biomarkers

Objective: Accurate prognostication of functional outcomes after an acute ischemic stroke (AIS) is essential for the planning of post-stroke treatments. We sought to provide an improved risk score for the prognostication by leveraging an interpretable machine learning and neuroradiologic biomarker. Data: We evaluated 913 patients with available 90-day mRS and admission NIHSS from the MRI-GENIE study. Age, sex, hypertension, atrial fibrillation, diabetes, coronary artery disease, smoking and prior stroke were available. We quantified WMH volume (WMHv) on FLAIR images. We divided patients into development and validation sets from multiple sites, and an external validation set from an individual cohort (Figure (a)). Methods: We measured the unaccounted WMHv (uWMH) by the residual of the log-linear regression model of WMHv regarding the clinical characteristics. Excessive uWMH (eWMH) was defined as present vs. absent based on a uWMH higher than +1 standard deviation. The clinical characteristics, admission NIHSS, and presence of eWMH were used as the inputs of a constrained logistic regression model with variable selection (RiskSLIM) that estimated a risk score for the prediction of post-stroke functional independence (mRS 0-2). Results: Admission NIHSS, age, sex, diabetes, prior stroke, and eWMH were selected with assigned risk scores (Figure (b)). The estimated risk score outperformed the conventional THRIVE and SPAN100 in the development, validation, and external validation sets regarding the area under curve (AUC) and the brier score (Figure (c)). Conclusion: This proposed risk score leveraging the interpretable machine learning and imaging biomarkers improved the prognostication of the post-stroke functional independence. Since the proposed risk score system does not require high computational resources, it is expected to be widely used in daily clinical practices for centers with MR imaging capability.