Assessing the Impact of Sociodemographic Factors on Artificial Intelligence Models in Predicting Dementia: Retrospective Cohort Study.

Background: Artificial intelligence (AI) is increasingly applied to health care, yet concerns about fairness persist, particularly in relation to sociodemographic disparities. Previous studies suggest that socioeconomic status (SES) and sex may influence AI model performance, potentially affecting groups that are historically underserved or understudied. Objective: This study aimed to (1) assess algorithmic bias in AI-driven dementia prediction models based on SES and sex (biological sex), (2) compare the utility of an individual-level SES measure (the Housing-Based Socioeconomic Status [HOUSES] Index) versus an area-level measure (the Area Deprivation Index) for bias detection, and (3) evaluate the effectiveness of an oversampling technique (the Synthetic Minority Oversampling Technique for Nominal and Continuous features) for bias mitigation. Methods: This study used data from two population-based cohorts: the Mayo Clinic Study on Aging (n=3041) and the Rochester Epidemiology Project (n=19,572). Four AI models (random forest, logistic regression, support vector machine, and Naïve Bayes) were trained using a 5-year observation window of structured electronic health record data to predict dementia onset within the subsequent 1-year window. Fairness and model performance were assessed using the balanced error rate (BER) across intersecting SES-sex subgroups. The Synthetic Minority Oversampling Technique for Nominal and Continuous features algorithm was applied to the training data to balance the representation of SES groups. Results: Across both cohorts, individuals with lower SES generally exhibited higher BERs (worse performance) than high SES groups, confirming the presence of bias. In the MCSA cohort, males with high SES, as indicated by the HOUSES Index, consistently exhibited the lowest BERs across all evaluated models. Balancing the training data based on a specific SES measure showed a trend toward reducing the BER disparity when evaluated using that same measure. However, this targeted improvement demonstrated nonuniversal benefits; in some cases, it exacerbated disparities when evaluated using other, unbalanced SES measures. This pattern suggests that fairness interventions are not universally beneficial across different definitions of the protected attribute. While the balancing approach improved fairness in model performance for lower SES groups, it often came at the cost of a slight reduction in overall model performance. However, an exception was observed in the MCSA cohort when balancing based on the HOUSES Index using logistic regression, support vector machine, and Naïve Bayes, where the performances of both the high and low SES groups improved. Conclusions: This research highlights the importance of incorporating sociodemographic context into AI modeling in health care. The choice of SES measure may lead to different assessments of algorithmic bias. The HOUSES Index, as a validated individual-level SES measure, may be more effective for bias mitigation than area-level measures. Future AI development should integrate bias mitigation strategies to ensure models do not reinforce existing disparities in health outcomes.