Objective:

Develop and validate a pre-alert model that predicts 90-day poor outcome (mRS >2) using information available at stroke alert, including time-of-day and weekend context, and compare it with common clinical tools.

Background:

Key workflow and communication decisions often occur before CT or CTA is obtained or interpreted. Tools such as NIHSS and SPAN-100 help stratify severity, but do not provide calibrated probabilities or incorporate timing and operational context at the moment of alert.

Design/Methods:

Consecutive stroke alerts from a single health system with 90-day mRS (N = 928) were analyzed retrospectively. To emulate deployment, the newest 20% by alert time comprised the temporal Test set. Pre-treatment predictors included: age, NIHSS, transfer/LVO flags, hour-of-day (sin/cos), weekend, and nonnegative intervals from LKW/CSA to CT/CTA. Outcome/treatment fields were excluded by name. A logistic regression model with isotonic calibration produced calibrated probabilities. Comparators were NIHSS-only, age+NIHSS, and SPAN-100 (rule and calibrated). The frozen model was tested in a deduplicated clinical replication cohort from the same system and window (N = 386). Primary metrics were ROC AUC, PR-AUC, and Brier score; decision-curve analysis assessed net benefit. EVT was not a predictor; used only for subgroup reporting.

Results:

Temporal Test (N = 185, prevalence 0.465): AUC 0.793, PR-AUC 0.726, Brier 0.1854. Replication (N = 386, prevalence 0.521): AUC 0.749, PR-AUC 0.766, Brier 0.2288. On the Test set, the model outperformed NIHSS-only (AUC 0.770, PR-AUC 0.702, Brier 0.1945) and SPAN-100 (calibrated AUC 0.770, PR-AUC 0.746, Brier 0.1957; rule AUC 0.638, PR-AUC 0.590, Brier 0.3405), and matched age plus NIHSS on discrimination (AUC 0.799) while improving probability accuracy. Decision-curve analysis showed positive net benefit.

Conclusions:

A time-aware, imaging-independent pre-alert model provides calibrated 90-day risk estimates and improves on widely used scores. It supports faster readiness and clear prognosis before the scan. Findings justify a silent prospective run and multi-site validation.