Unsupervised Anomaly Detection in Hospital Financial Statements: A Multi-Algorithm Consensus Approach With Explainable AI
Yongjun Kim, Eun Jung Sun · IEEE Access · 2026
Monitoring the financial health of hospitals is critical for healthcare sustainability, yet traditional approaches rely on supervised learning methods that require labeled datasets—an assumption rarely met in practice. This study presents a multi-algorithm consensus framework for unsupervised anomaly detection in hospital financial statements. Using a balanced panel of 165 Korean general and tertiary hospitals over five fiscal years (2019–2023, $N = 827$ ), we apply five complementary unsupervised algorithms: Isolation Forest, Local Outlier Factor, One-Class SVM, DBSCAN, and an autoencoder based on reconstruction error, using 10 financial and structural features (excluding the calendar year to prevent temporal leakage). A consensus mechanism identifies “strong anomalies” as observations flagged by three or more algorithms, yielding 59 consensus anomalies (7.1%). Mann-Whitney U tests reveal that consensus anomalies exhibit significantly higher labor cost ratios ( $p \lt 0.001$ ), administrative cost ratios ( $p \lt 0.001$ ), and subsidy dependence ( $p \lt 0.001$ ) compared to normal hospitals. Temporal analysis reveals increasing financial heterogeneity from 4.2% anomaly rate in 2019 to a peak of 9.6% in 2022, before moderating to 6.7% in 2023. SHAP-based explainability is applied both to the Isolation Forest and to a consensus-level surrogate logistic regression, identifying administrative cost ratio, hospital size, and ownership type as the primary drivers. Notably, only 47.5% of consensus anomalies overlap with accounting-deficit hospitals ( $\chi ^{2} = 0.72$ , $p = 0.395$ ); the difference is not statistically significant, indicating that unsupervised anomaly detection captures structurally unusual financial patterns beyond simple profitability. Using deficit status as a proxy reference, the framework achieves precision = 0.475, recall = 0.082, F1 = 0.140, and a false positive rate of 0.064 (6.4% of non-deficit hospitals flagged), reflecting the fundamental complementarity between anomaly detection and deficit prediction.