Task-adaptive multi-view learning with ordinal graph and continuous-score paths for public hospital performance stratification

Introduction Public reporting provides repeated measures of hospital quality, spending, safety, and patient experience. Outcomes are ordered, and feature availability can change across releases. We evaluated Health–Insurance–Operational–Proxy synergy-aware Ordinal Graph Learning (HIOP–GL) for next-release performance stratification among publicly owned US hospitals. Methods Twelve Centers for Medicare & Medicaid Services (CMS) snapshots from 2023 to 2026 were aligned without concurrent targets. For the next-release overall hospital-rating task (PH-Star), development used 1,849 hospital-release observations from 341 hospitals, followed by 1,236 post-development observations from 338 hospitals. Because annual Hospital Value-Based Purchasing total performance scores (TPS) repeated within fiscal years, the corresponding PH-VBP task used FY2024 development ( n = 321) and FY2025 tuning ( n = 301) for selection. After architecture, loss, cutoff, and epoch choices were fixed using FY2024–FY2025 data, the model was refitted on both years (622 hospital-year records) and evaluated in FY2026 ( n = 308). PH-Star used a three-relation ordinal graph path specified before evaluation, whereas PH-VBP used a graph bypass, continuous TPS head, and tier cutoffs fixed from FY2024. Results Across five fits, PH-Star quadratic weighted kappa (QWK) was 0.816 ± 0.009 on the temporal test and 0.757 ± 0.013 after development. The two-year-refit PH-VBP model reached QWK 0.678 ± 0.021 in FY2026; ensemble QWK was 0.717 (95% CI 0.657–0.773). Hospital-disjoint and Census-region-disjoint cross-fitting yielded QWK 0.689 and 0.636. A PH-Star ensemble held constant after development achieved QWK 0.637 (95% CI 0.546–0.718; n = 229) in zero-overlap California non-public hospitals. Independent-authority checks found concordance with 2023 California multi-measure ratings ( n = 243, ρ = 0.247, 95% CI 0.125–0.366; Holm p = 0.0006) and 2024 New York all-payer mortality composites in 57 exact-linked hospitals (ρ = 0.451, 95% CI 0.195–0.663; Holm p = 0.0010). Concordance varied by criterion, with no supported California mortality association. Discussion HIOP-GL's clearest advantage was ordinal agreement in later-release evaluation. Macro-F1 and low-tier recall remained metric-dependent, and hospital-grouped PH-Star QWK was 0.661 vs. 0.730 for ordinal logistic. Ownership transport and multi-state independent-criterion concordance extend the assessment beyond development. The results support outcome-specific modeling for retrospective ordinal stratification; prospective same-outcome validation across independent systems remains necessary.