Frontiers in Public Health · Nov 11, 2026
Bayesian Network Analysis of Cognitive Function and Its Associated Factors among Older Adults in Hubei, China
By Xiaoxuan Xia, Ruikang Liu, Yi Luo et al.
Background: Cognitive impairment is increasingly prevalent with population aging. Using Tabu-search-based Bayesian networks (BNs), we examined factors associated with MMSE-defined cognitive status and their probabilistic relationships to identify candidate domains for joint assessment and hypotheses for longitudinal research among older adults. Methods: This cross-sectional study recruited adults aged ≥60 years from seven healthcare and residential care institutions in Hubei Province between June 2025 and June 2026 using convenience sampling. Participants were classified as MMSE-normal (≥27) or MMSE-abnormal (<27). Logistic regression and BN modeling were performed. BN structure and parameters were learned using Tabu search and Bayesian estimation. The hypothesis-driven network was compared with a no-whitelist reference network. Selected-profile distributions, prevalence ratios (PRs), and reverse probabilities were examined, and discrimination was evaluated using the area under the curve (AUC). Results: A total of 864 participants were included (normal MMSE, n=395; abnormal MMSE, n=469). Multivariable logistic regression identified high education (OR=0.508, 95% CI 0.353-0.728) and regular exercise (OR=0.469, 95% CI 0.333-0.658) as associated with lower odds of abnormal MMSE, whereas diabetes (OR=1.454, 95% CI 1.043-2.029) and high depressive risk (OR=2.676, 95% CI 1.910-3.759) were associated with higher odds. The main BN contained 21 directed edges, including 17 mandatory edges. Education, exercise, depressive symptoms, and diabetes were specified as direct parents of MMSE. No direct BADL-MMSE edge was retained. The no-whitelist reference network contained 10 edges, indicating dependence of the selected structure on prior constraints. Among selected profiles, normal MMSE was observed in 59/84 participants with favorable characteristics and 5/39 with unfavorable characteristics. Irregular exercise and high depressive risk had PRs of 1.742 and 1.801, respectively. Reverse probability estimates for irregular exercise, IADL limitation, and high depressive risk were 80%, 75%, and 64%, respectively. The BN achieved AUCs of 0.740 (95% CI 0.700-0.779) in the training set and 0.801 (95% CI 0.748-0.855) in the test set. Conclusion: The BN complemented multivariable logistic regression by representing conditional dependencies among the included characteristics and distinguishing imposed relationships from non-mandatory connections. Comparison with the no-whitelist reference network revealed differences in connection presence and orientation, providing exploratory hypotheses concerning functional, psychological, behavioral, and metabolic relationships.