Abstract: Objective To develop a screening model for ischemic stroke by relying on large-scale electronic health records and combining the advantages of Bayesian network uncertainty inference. Methods The screening model derivation cohort was derived from the Cheeloo Lifespan Electronic Health Research Data-library(Cheeloo LEAD)and divided into training and testing sets in a 7∶3 ratio. The external validation cohort was sourced from the Boxing Collaboration Center Database of the National Healthcare Big Data Research Institute(Boxing Database). The univariate Logistic regression analysis was used to screen for factors significantly associated with the ischemic stroke. These associated screening factors were used to develop the Bayesian network. The tabu search algorithm was employed for structure learning, while Bayesian estimation algorithm was used for parameter learning, ultimately leading to the development of the ischemic stroke screening model. The performance of the model was evaluated in terms of both discrimination and calibration abilities, and compared with the traditional Logistic regression model in screening for ischemic stroke. Results The derivation cohort included 1,067,609 individuals, among whom 31,019 suffered from ischemic stroke. The external validation cohort included 386,773 individuals, among whom 13,393 suffered from ischemic stroke. After the univariate screening, 67 screening factors were identified. The final Bayesian network model included 68 nodes and 440 directed edges. The parent nodes of the ischemic stroke node included age, hypertensive diseases, ischemic heart diseases, chronic lower respiratory diseases, other cerebrovascular diseases, episodic and paroxysmal disorders, and the symptoms and signs involved cognition, perception, emotional state and behavior. The AUC for the training set, testing set, and external validation cohort were 0.840(95%CI: 0.838-0.843), 0.839(95%CI: 0.836-0.843), and 0.811(95%CI: 0.808-0.814), respectively, indicating good discrimination ability, and calibration ability also performed well. Our newly developed screening model continued to outperform the traditional Logistic regression screening model, even in the presence of missing data. Conclusion This study developed the ischemic stroke screening model with the advantage of Bayesian network uncertainty inference. The model has good discrimination and calibration abilities, providing a convenient and efficient method for early ischemic stroke screening.

Key words: Electronic health records, Bayesian network, Logistic regression, Ischemic stroke, Screening model