Assessing the Impact of Sociodemographic Factors on Artificial Intelligence models in Predicting Dementia: Retrospective Cohort Study (Preprint)

BACKGROUND Artificial intelligence (AI) is increasingly applied to healthcare, yet concerns about fairness persist, particularly in relation to sociodemographic disparities. Prior studies suggest that socioeconomic status (SES) and sex may influence AI model performance, potentially exacerbating existing health inequalities. OBJECTIVE To investigate how SES and sex intersect with AI model performance in predicting dementia risk, and to assess whether these factors contribute to algorithmic bias. METHODS Data from two population-based cohorts—the Rochester Epidemiology Project and the Mayo Clinic Study on Aging—were utilized. SES was quantified at both group and individual levels using the Area Deprivation Index and the HOUSES index, respectively. Four AI models (Random Forest, Logistic Regression, Support Vector Machine, and Naïve Bayes) were trained to predict dementia risk. Balanced error rate (BER) was used as a fairness metric to evaluate model bias across SES and sex strata. Additionally, an oversampling technique was implemented to improve minority SES group representation in the training data. RESULTS The study found substantial disparities in model performance, with low SES groups consistently showing higher BERs across all models, indicating reduced accuracy and potential bias. These disparities were present despite the use of different model architectures. Application of the oversampling method demonstrated potential for reducing these biases. CONCLUSIONS This research highlights the importance of incorporating sociodemographic context into AI modeling in healthcare. The HOUSES index, as a validated, nationwide individual-level SES measure, offers a promising tool for bias assessment. Future AI development should integrate strategies like oversampling to promote equity and ensure models do not reinforce existing disparities in health outcomes.