Infarct-volume prognostic value depends on outcome ascertainment and validation design in public stroke MRI datasets

To determine how participant-level outcome missingness and center-level endpoint absence influence estimates of the incremental prognostic value of infarct volume in public stroke magnetic resonance imaging datasets.

We analyzed two public cohorts separately. In the Stroke Outcome Optimization Project (SOOP), 948 of 1106 S records had a valid acute lesion mask and core predictors; 620 had discharge modified Rankin Scale (mRS). Models containing age, sex, and admission National Institutes of Health Stroke Scale, with or without log lesion volume, underwent repeated nested cross-validation and inverse-probability weighting. In ISLES'24, models developed in center 1 (n = 87) were evaluated in center 2 (n = 39) for 3-month mRS and compared with pooled random cross-validation. The primary measure was the ranked probability score (RPS) difference.

In SOOP, observed outcomes were associated with larger lesions than missing outcomes (median, 15.08 vs 3.68 mL; standardized mean difference for log volume, 0.486). Adding volume improved RPS by 0.0609 (95 % confidence interval [CI], 0.0314-0.0893); the weighted improvement was 0.0538 (95 % CI, 0.0267-0.0799). In ISLES'24, discharge mRS was absent from center 2. For 3-month mRS, the center-held-out RPS improvement was 0.0488 (95 % CI, -0.1228 to 0.2473), compared with 0.1351 in pooled random validation. Both models overpredicted disability in center 2.

Outcome ascertainment and validation design changed the apparent prognostic contribution of infarct volume. Endpoint-by-center auditing should precede interpretation of multicenter imaging performance.