Evaluating the Trade-offs Between Model Accuracy, Patient Equity, and Operational Cost-Reduction in Publicly Funded Integrated Delivery Networks

Abstract

Decentralized Publicly funded Integrated Delivery Networks (IDNs) face increasing pressure to leverage predictive

analytics for operational efficiency, yet must remain vigilant to unintended consequences regarding patient equity.

This study evaluates the inherent trade-offs between maximizing model accuracy, promoting patient equity, and

achieving operational cost-reduction within these resource-constrained systems. Employing a mixed-methods

research design, the study combines quantitative analysis of synthetic IDN operational data (simulating 500,000

patient records across three networks) with qualitative thematic analysis of 45 semi-structured interviews with

healthcare administrators, data scientists, and frontline clinicians. The quantitative phase involved developing three

predictive models (High-Accuracy, Equity-Constrained, and Hybrid) to forecast emergency department high-

utilization and hospital readmission. Results demonstrate that a strict pursuit of model accuracy (AUC 0.91) leads to

significant algorithmic bias (disparate impact ratio of 0.62 for minority populations and 0.71 for low-income groups),

generating cost savings of 18.2% but at substantial equity costs. Conversely, an equity-constrained model reduced

bias (disparate impact ratio > 0.85) yet increased operational costs by only 6.8% while decreasing accuracy modestly

(AUC 0.84). The Hybrid model, employing adversarial debiasing and equalized odds post-processing, achieved a

balanced performance (AUC 0.87, cost reduction 12.5%, disparate impact ratio 0.79). The findings imply that no

single objective can be perfectly optimized; instead, IDNs must adopt dynamic, transparent trade-off frameworks

guided by community-defined fairness metrics and incremental implementation cycles, acknowledging that

operational cost-reduction cannot ethically supersede equity.