Toward Explainable Detection of Hospital Cost-Allocation Anomalies: Evidence and Research Opportunities Using Medicare Cost Reports

Each Medicare-certified hospital in the United States files an annual Medicare Cost Report (MCR, Form CMS-2552-10), in which total expenses are apportioned to standard cost centers and overhead is allocated to patient care departments using a statistical step-down procedure. While most Medicare inpatient reimbursements are made prospectively, the MCR nonetheless underpins cost finding and allocation through cost-to-charge ratios, graduate medical education payments, and disproportionate share payments now based on reported uncompensated care. Nevertheless, independent comparisons have found missing items, discrepancies across sources, and computational errors in MCR data. This narrative review synthesizes three literatures that have developed largely independently: investigations into the structure and data quality of the MCR itself; machine-learning approaches to the detection of anomalies and fraud in Medicare claims and in provider data; and explainable artificial intelligence (XAI) in auditing and accounting. Sources were identified by targeted database searches and citation tracking rather than through a systematic search. Fraud detection in Medicare is an established field, and XAI in auditing is emerging, but we identified no application of explainable anomaly detection to hospital cost-allocation data at the level of cost centers. Based on the synthesis, the review proposes a domain-specific conceptual framework that maps the structure of the MCR onto auditable anomaly indicators, including allocation bases, administrative and general costs, employee benefit allocations, cost-to-charge ratios, reclassifications, and year-on-year changes, and links each to potential detection methods (robust z-scores, Isolation Forest, and the local outlier factor) and explanation methods such as SHAP. Because outliers may reflect differences in hospital size, case mix, facility organization, or reporting practices, the framework proposes that flagged data points be treated as investigative leads for auditors and regulators.