The multi-dimensional equity of predictive models in acute ischaemic stroke

Accurate outcome prediction in acute ischaemic stroke is critical to clinical decision-making and optimal trial design. Predictive performance may, however, vary inequitably across multidimensionally defined patient subpopulations shaped by the intersection of diverse demographic, clinical, and anatomical factors such as the lesion distribution. Here we investigate the variation in achievable predictive performance— the epistemic equity of the model—across a multi-site cohort of 4083 richly characterised stroke episodes. Flexible machine learning models of demographic, clinical, and diffusion-weighted imaging data were used to predict an array of care needs and clinical outcomes. To identify systematic multidimensional subpopulations, deep representation learning was used to derive low-dimensional latent projections of high-dimensional clinical and imaging data. The distinct, consistent, interpretable multidimensional subpopulations thereby identified were used to quantify model epistemic equity across the population. Predictive performance varied substantially across subpopulations, revealing disparities invisible to simple, unidimensional measures of equity. Combining deep representation learning with stratified performance evaluation, we demonstrate the need for expressive, multidimensional representations in assuring the equity of predictive models in acute stroke care and research.