JMIR Publications

ABSTRACT

Background:

Timely hospital admission is a prerequisite for effective acute stroke management, yet a substantial proportion of patients fail to reach medical facilities within the optimal therapeutic window. Existing prediction models often lack temporal robustness and clinical interpretability, limiting their utility in real-world, evolving healthcare systems.

Objective:

The objective of our study was to develop and temporally validate machine learning and deep learning models using multicentre clinical data to predict early hospital admission (≤24 hours) after acute stroke.

Methods:

A total of 1,327 patients were included, of whom 821 were assigned to the Train Set and 506 to the independent Temporal Testing Set. Among the six candidate models, the MLP showed the best overall performance in the independent Temporal Testing Set, achieving an AUC of 0.9020 (95% CI 0.8718–0.9283), sensitivity of 91.5%, specificity of 75.6%, and F1-score of 0.9033. Formal statistical comparisons showed that the MLP achieved significantly higher AUC values than logistic regression, support vector machine, random forest, and 1D-CNN after false discovery rate correction, with a smaller but still statistically significant improvement over the LSTM. Calibration analysis further showed that the MLP had the most favorable overall calibration profile among the candidate models.

Results:

Deep learning models demonstrated superior temporal generalizability compared with traditional classifiers. The MLP achieved the most stable performance in the testing cohort (AUC = 0.90; sensitivity = 91.5%), while conventional models showed marked performance decay. SHAP analysis revealed that dysphagia was consistently associated with delayed admission , whereas acute physiological abnormalities—particularly hypertension and diabetes—were dominant drivers of early presentation.

Conclusions:

In this multicentre Chinese cohort, the MLP showed favorable temporal performance for predicting early hospital admission after stroke. The model may support future risk stratification and targeted public health interventions, although further external validation and calibration refinement are needed before deployment-oriented use.