EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC 4.0

Provider-Adaptive Graph–Tabular Fusion for Healthcare Fraud Screening: Reliability Gating and Shrinkage in a Controlled Study

Sophie Andrews

Academic Journal of Applied Sciences · 03.11.2026 · 3 (2) · 59-67

DOI: 10.54097/s3bwbf98 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Academic Journal of Applied Sciences, online erschienen am 03.11.2026. Verfasst von Sophie Andrews. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Relational evidence can improve provider-level fraud screening, yet neighbor messages become unreliable when shared beneficiaries or physicians reflect legitimate encounters rather than suspicious coordination. We evaluate provider-adaptive graph–tabular fusion (PAGTF), which starts from the EIGF fusion design—an inverse-entity-degree projected provider graph whose mean-aggregating GraphSAGE branch is fused with a tabular expert (here histogram gradient boosting) through a validation-selected global mixture—and replaces the global mixture with a provider-wise convex gate derived from label-free neighborhood agreement and weighted degree. In ten experiments spanning two synthetic graph-reliability regimes and five independently generated populations per regime, each population contains 2,400 simulated providers. Mean test area under the precision–recall curve (PR-AUC) in the mixed-reliability setting was 0.6471 for PAGTF versus 0.6462 for EIGF global fusion; in the noisy setting, it was 0.6339 versus 0.6393. Adaptive fusion consequently offered no stable ranking improvement, despite its greater local flexibility. Degree-only and agreement-only ablations, self-message inference, review-budget measures, and held-out subgroup analyses identify validation instability as a substantive design risk. A subsequent exploratory shrinkage check that halves the gate logit modulation increases the mixed/noisy mean PR-AUCs to 0.6502/0.6411, but paired variation precludes claims of reliable superiority. The study is a reproducible synthetic-data stress test, not a clinical or Medicare benchmark.

Zitieren

APA

Andrews, S. (2026). Provider-Adaptive Graph–Tabular Fusion for Healthcare Fraud Screening: Reliability Gating and Shrinkage in a Controlled Study. Academic Journal of Applied Sciences, 3(2), 59-67. https://doi.org/10.54097/s3bwbf98

Vancouver

Andrews S. Provider-Adaptive Graph–Tabular Fusion for Healthcare Fraud Screening: Reliability Gating and Shrinkage in a Controlled Study. Academic Journal of Applied Sciences. 2026;3(2):59-67. doi:10.54097/s3bwbf98

BibTeX
@article{Andrews2026,
  author = {Andrews, Sophie},
  title = {Provider-Adaptive Graph–Tabular Fusion for Healthcare Fraud Screening: Reliability Gating and Shrinkage in a Controlled Study},
  journal = {Academic Journal of Applied Sciences},
  year = {2026},
  volume = {3},
  number = {2},
  pages = {59--67},
  doi = {10.54097/s3bwbf98}
}
Kostenlos starten →