Abstract / Summary
BACKGROUND: Existing models for discharge to long-term institutional care after stroke are developed within single health systems and pool to an area under the curve near 0.80, but none is accompanied by an estimate of how much of the outcome is a property of the health system rather than the patient. METHODS: Secondary analysis of the International Stroke Trial, a randomised trial in 19,435 patients with suspected acute ischaemic stroke recruited in 1991-1996 across 36 countries. Among 14,885 survivors to six months with residence recorded, we developed and internally validated a logistic prognostic model for residential or nursing-home residence using 17 admission-day terms, and compared it against a pre-specified benchmark of age and stroke syndrome alone. Between-country variation was estimated by a method-of-moments decomposition with a cluster bootstrap. RESULTS: Institutional residence was recorded for 1,984 of 14,885 survivors (13.3%); the model was fitted on 14,121 patients containing 1,935 events. In 10-fold cross-validation with the whole procedure refitted per fold, the out-of-fold c-statistic was 0.789, calibration slope 0.987, calibration-in-the-large 0.1371 predicted against 0.1370 observed, and Brier score 0.1013. A benchmark of age and stroke syndrome alone reached an optimism-corrected 0.774 against 0.790 for the full model, a corrected difference of 0.0160. Between-country variation gave an intraclass correlation of 0.0409 (0.0191 to 0.0673), and in leave-one-country-out validation the c-statistic ranged from 0.635 to 0.823 across the 14 countries with enough events to estimate it. CONCLUSIONS: Institutional residence after stroke can be stratified at admission with discrimination comparable to existing models and calibration that holds in unseen patients, but almost all of the signal is carried by age and stroke syndrome. Performance varies substantially between countries, pointing to placement systems as the limit on what a pooled model can do.