Conference Agenda

The use of clinical artificial intelligence (AI) models in healthcare systems has prompted concerns about algorithmic fairness, especially for marginalized populations who have been underrepresented in medical data. Simulated data reflecting rural Honduran demographics and illness incidence patterns is used to examine demographic and distributional biases in clinical prediction models. Gradient boosting (XGBoost) and neural network classifiers (Multilayer Perceptron with 3 hidden layers) were trained for cardiovascular risk prediction using synthetic clinical datasets generated by Monte Carlo simulation. Fairness metrics including equalized odds difference, demographic parity ratio, and calibration error assessed model performance discrepancies across protected factors like geographic location, socioeconomic status, and indigenous heritage. Sample reweighting, adversarial debiasing, and decision threshold optimization were evaluated. Baseline models showed significant performance differences, measuring AUROC differences of up to 0.089 between urban and remote rural populations (p < 0.001) and equalized odds differences reaching 0.15 across socioeconomic groupings. Adversarial debiasing reduced equalized odds differences by 62.4% while maintaining clinically acceptable discrimination (AUROC = 0.823) after mitigation. Sample reweighting increased demographic parity by 41.3% without degrading predictive performance. These findings demonstrate that clinical AI models developed without explicit fairness constraints can exacerbate existing healthcare disparities, and that simulation-based auditing frameworks combined with bias mitigation techniques can substantially enhance algorithmic equity while maintaining predictive utility. The findings recommend required algorithmic auditing measures before deployment in resource-limited healthcare settings.