The challenge of long-term stroke outcome prediction and how statistical correlates do not imply predictive value

Abstract

Personalized prediction of stroke outcome using lesion imaging markers is still too imprecise to make a breakthrough in clinical practice. We performed a combined prediction and brain mapping study on topographic and connectomic lesion imaging data to evaluate (i) the relationship between lesion-deficit associations and their predictive value and (ii) the influence of time since stroke. In patients with first-ever ischaemic stroke, we first applied high-dimensional machine learning models on lesion topographies or structural disconnection data to model stroke severity (National Institutes of Health Stroke Scale 24 h/3 months) and functional outcome (modified Rankin Scale 3 months) in cross-validation. Second, we mapped the topographic and connectomic lesion impact on both clinical measures. We retrospectively included 685 patients [age 67.4 ± 15.1, National Institutes of Health Stroke Scale 24 h median(IQR) = 3(1; 6), modified Rankin Scale 3 months = 1(0; 2), National Institutes of Health Stroke Scale 3 months = 0(0; 2)]. Predictions for acute stroke severity (National Institutes of Health Stroke Scale 24 h) were better with topographic lesion imaging (R² = 0.41) than with disconnection data (R² = 0.29, P = 0.0015), whereas predictions at 3 months (National Institutes of Health Stroke Scale/modified Rankin Scale) were generally close to chance level. In the analysis of lesion-deficit associations, the correlates of more severe acute stroke (National Institutes of Health Stroke Scale 24 h > 4) and poor functional outcome (modified Rankin Scale 3 months ≥ 2) were left-lateralized. The lesion location impact of both variables corresponded in right-hemisphere stroke with peaks in primary motor regions, but it markedly differed in left-hemisphere stroke. Topographic and disconnection lesion features predict acute stroke severity better than the 3-months outcome. This suggests a likely higher impact of lesion-independent factors in the longer term and highlights challenges in the prediction of global functional outcome. Prediction and brain mapping diverge, and the existence of statistically significant associations-as here for 3-months outcomes-does not imply predictive value. Routine neurological scores better capture left- than right-hemispheric lesions, further complicating the challenge of outcome prediction.

Keywords: disconnection; imaging biomarker; lesion mapping; machine learning; recovery.