Presentation (CTI)

OBJECTIVES: To quantify how insurance‑level resource allocation affects inpatient outcomes for Parkinson’s disease, identify socioeconomic and clinical differences in treatment response, measure inequities in financial protection, and design budget‑neutral reallocation strategies that improve both efficiency and equity within Iran’s health insurance system.
METHODS: Administrative inpatient claims from the Iranian Health Insurance Organization (2023-2025), covering six provinces. The dataset included 21,121 service‑line transactions aggregated into 1,813 hospitalization episodes.A six‑stage econometrically grounded machine‑learning framework was used. Gradient‑boosted models predicted recovery, readmission, and costs. Double Machine Learning estimated average treatment effects, and Causal Forests captured socioeconomic and clinical heterogeneity. Vertical and horizontal inequities were assessed using the Concentration Index with Wagstaff decomposition. Budget‑neutral, welfare‑maximizing reallocations were identified through constrained integer linear programming. Sensitivity to unmeasured confounding was evaluated using Rosenbaum bounds.
RESULTS: Predictive performance was strong (AUC recovery = 0.805; AUC readmission = 0.850; cost R² = 0.804). Prolonged length of stay reduced readmission by 8.7 percentage points (95% CI: −0.124 to −0.051) but increased costs by roughly 67 percent. Higher insurance coverage independently reduced readmission by 4.5 percentage points and total costs (ATE = −0.078 log‑units). Treatment effects varied substantially: coverage benefits were minimal for rural patients, while prolonged stay was most protective for younger and low‑severity patients. Recovery was pro‑rich concentrated (Erreygers CI = +0.092). Out‑of‑pocket burden showed pronounced horizontal inequity (HI = −0.101), driven primarily by rural residence (132.7 percent of total inequity). Budget‑neutral optimization corrected misaligned severity targeting and improved projected recovery by up to 0.87 percentage points without increasing expenditure. Sensitivity analysis indicated strong robustness to unmeasured confounding.
CONCLUSIONS: Machine‑learning-guided, equity‑constrained reallocation of existing insurance resources can improve clinical outcomes and reduce socioeconomic disparities in PD inpatient care. However, coverage generosity alone is insufficient for rural populations; complementary supply‑side investment is needed to translate insurance benefits into equitable health and financial protection.