Empirical Assessment of Bias in Machine Learning Models for Healthcare Resource Allocation in Nigerian Hospitals

Machine learning models informing healthcare resource allocation decisions may encode socioeconomic and geographic biases present in historical health data, and empirically assessing these biases in Nigerian hospital contexts is an important step toward responsible health AI deployment. This study empirically assessed bias in machine learning models trained for bed demand forecasting and patient triage prioritisation at two tertiary hospitals in Oyo State (SW) and Anambra State (SE). Electronic health record data for 48,000 patient encounters were used to train gradient boosting models for each task. Fairness analysis examined differential model performance across patient socioeconomic status, insurance coverage status, geographic origin (urban versus rural), and gender. Results showed significant performance disparities for triage models: AUC was 0.86 for insured patients versus 0.74 for uninsured patients. Rural patients' triage scores were systematically underestimated relative to their clinical acuity. Bed demand forecasting showed smaller but significant biases correlated with seasonal patterns predominantly affecting rural admissions. Equalised odds fairness constraints reduced performance disparity by 62 percent at a cost of 3.4 percent overall accuracy on triage prediction. The study fills an empirical gap in health AI fairness for Nigerian hospitals and recommends embedding fairness constraints in all hospital ML models, regular demographic performance audits, and mandatory reporting of fairness metrics to hospital ethics committees alongside clinical accuracy statistics.

Keywords: machine learning bias, healthcare resource allocation, Nigerian hospitals, fairness in AI, electronic health records