Development and validation of domain-specific clinical prediction models of post-stroke cognitive impairment

Background Post-stroke cognitive impairment (PSCI) is highly prevalent across multiple cognitive domains. Individualised PSCI prognosis has mainly been researched using global cognitive outcomes. Here, we develop and externally validate clinical prediction models for overall and domain-specific PSCI, including language, memory, attention, executive function, numeracy, and praxis. Methods N =430 stroke survivors completed the Oxford Cognitive Screen (OCS) in acute care and at 6-month follow-up (binarized outcome; impaired vs unimpaired). Logistic regression models were fitted comprising both mandatory clinically-relevant (age, sex, stroke severity, education, stroke hemisphere, acute PSCI) and data-driven (acute mood difficulties, length of stay in acute care, multimorbidity) predictors using backward elimination ( p < 0.10) on multiply imputed data. Internal validation used bootstrapping to obtain optimism-adjusted performance estimates. External validation used the optimism-adjusted C-Slope as a uniform shrinkage factor. Results Compared to the overall PSCI model (C-Statistic=0.76 95% CI=0.71–0.80), comparable or improved optimism-adjusted performance was observed in models of language (C-Statistic=0.77 95% CI=0.72–0.81) memory (C-Statistic=0.72 95% CI=0.65–0.75), and attention (C-Statistic=0.74 0.69–0.78). Numeracy (C-Statistic=0.69 95% CI=0.63–0.74), executive function (C-Statistic=0.71 95% CI=0.65–0.76), and praxis (C-Statistic=0.60 95% CI=0.53–0.65) models showed weaker performance. In external validation, the overall PSCI model was comparable to development data (C-Statistic=0.74 95% CI=0.67–0.79). Conclusions Domain-specific prediction models have the potential to offer more meaningful PSCI prognoses compared to overall PSCI models. External performance of overall PSCI models show promise in different stroke severity cohorts. Future recalibration of memory, numeracy, executive function, and praxis models would be beneficial.