A global effort to benchmark predictive models and reveal mechanistic diversity in long-term stroke outcomes

Abstract

Stroke remains a leading cause of mortality and long-term disability worldwide, with variable recovery trajectories posing substantial challenges in anticipating post-event care and rehabilitation planning. To address these challenges, we established the NeuralCup consortium to benchmark predictive models of stroke outcome through a collaborative, data-driven approach. This study presents findings from 15 international teams who used a comprehensive dataset including clinical and imaging data, to identify and compare predictors of motor, cognitive, and emotional outcomes one year post-stroke. Our analyses integrated traditional statistical approaches and novel machine learning algorithms to uncover 'optimal recipes' for predicting each domain. The differences in these 'optimal recipes' reflect distinct brain mechanisms in response to different tasks. Key predictors across all domains included infarct characteristics, T1-weighted MRI sequences, and demographic factors. Additionally, integrating FLAIR imaging and white matter tract analysis significantly improved the prediction of cognitive and motor outcomes, respectively. These findings support a multifaceted approach to stroke outcome prediction, underscoring the potential of collaborative data science to develop personalized care strategies that enhance recovery and quality of life for stroke survivors. To encourage further model development and validation, we provide access to the training dataset at http://neuralcup.bcblab.com.

Figure 1

Visualization of the process for obtaining the theoretically optimal feature combination for predictions: inversing the randomise analysis yields heatmaps for each analyzed feature. After binarizing these maps, we investigated the overlap of the local maxima of the clinical tests with the binarized feature maps (1 representing the presence of the feature, and 0 representing the absence of a feature in the final combination ‘optimal recipe’).

Figure 2

Summary of participating teams and the approaches taken for all predictions. a Locations of the teams’ affiliated labs. b Summary of different inputs (A: age, B: gender, C: DWI, D: T1, E: lesion, F: FLAIR, G: tracts, H: atlases, I: disconnectome) and methods (A: clustering, B: artificial neural networks, C: regression, D: feature selection, E: dimensionality reduction, F: parcellation, G: cross-validation, H: bootstrap) used for each prediction. (Figure modified from).

Figure 3

a Mean R² comparison for all submitted predictions (motor, cognitive, and psychological outcomes) of five neuropsychological scores (FM total, MoCA, IST, HAD-A, HAD-D) sorted ascendingly from the highest score to the lowest score across all teams whose number is indicated on the x-axis. The stars indicate the prediction number, the whiskers indicate the standard error of the mean (SEM). b Median R² comparison for all submitted predictions of the same scores, sorted ascendingly with whiskers indicating the interquartile range (IRQ). c T-statistic maps for each clinical outcome test, obtained from the FSL randomise analysis. The yellow regions indicate a significant t-value, the purple regions indicate a non-significant t-value. The UMAP distribution of all teams is plotted on the t-statistic maps. d Local maximum of the t-statistic map for each analyzed outcome score.