Objective:

To externally validate eight prediction algorithms and assess the added value of complex machine-learning models over logistic regression in predicting prehospital delays in acute ischemic stroke.

Approach:

• Study Design: Retrospective cohort study involving two hospitals in Kashgar, China, with Hospital 1 as the development cohort and Hospital 2 for independent validation.
• Data Collection: Consecutive patients with acute ischemic stroke treated between January 2019 and December 2024 were included, with prehospital delay defined as an onset-to-door interval exceeding 4.5 hours.
• Model Evaluation: Eight algorithms were validated, with assessments of discrimination, calibration, and decision curves, including sensitivity analyses with and without SMOTE.

Key Findings:

• Prehospital delay affected 80.9% of patients in Hospital 1 and 83.2% in Hospital 2.
• The logistic model without SMOTE showed improved discrimination (AUC 0.774) compared to the SMOTE-trained model (AUC 0.756).
• XGBoost achieved the highest validation AUC (0.771; 95% CI, 0.734–0.807), but its advantage over logistic regression was not statistically significant (difference, 0.015; 95% CI, −0.003 to 0.034; p = 0.107).

Limitations:

• Some predictors are only available after hospital arrival, limiting the models' applicability for pre-arrival prediction.
• The study's retrospective design may introduce biases related to data collection and patient selection.