Predictive Modeling of Healthcare Arbitration: Evidence, Methodological Challenges, and Machine-Learning Research Directions for Federal IDR

The federal Independent Dispute Resolution (IDR) process established under the No Surprises Act creates an important setting for machine-learning research on administrative decision-making. However, descriptive evidence about arbitration outcomes should not be confused with evidence that a predictive system is accurate, generalizable, or beneficial. This targeted narrative review synthesizes federal reporting, peer-reviewed outcome studies, and relevant machine-learning methodology to develop a research agenda for IDR prediction. The available evidence identifies substantial outcome imbalance, repeated interactions among disputing organizations, heterogeneity in payment benchmarks, and important restrictions in public-use data. We distinguish prediction of escalation, procedural disposition, contested offer selection, payment, and resolution time, each requiring a different population and information set. A proposed evaluation framework emphasizes release-aware temporal validation, dispute-level dependence, probability calibration, uncertainty quantification, and prospective assessment of decision utility. The discussion addresses selective labels, regulatory change, strategic feedback, legally permissible evidence, and unequal access to analytic resources. Eight testable hypotheses concern stage-specific models, relational learning, document-grounded systems, temporal adaptation, uncertainty-aware review, tabular foundation models, federated learning, and decision-focused deployment. No new predictive experiments are reported. The central contribution is a falsifiable and reproducible agenda for establishing when machine learning adds value beyond simple baselines without mistaking predictive associations for legal justification or causal policy effects.