EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Risk-Based Audit Prioritization for Automobile Insurance Fraud: A Comparative Evaluation of Machine Learning and Deep Learning Under Class Imbalance

Re-Mi Hage, Giovanni Choufani, Leonie Nassif

Risks · 06.11.2026 · 14 (11) · 258

DOI: 10.3390/risks14110258 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Risks, online erschienen am 06.11.2026. Verfasst von Re-Mi Hage und 2 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Insurance fraud detection is a rare-event classification problem in which fraudulent claims represent a small fraction of reported claims. Deep learning is increasingly advocated for fraud analytics, but its added value is unclear when the available information is structured, tabular, and of moderate dimension. Using a publicly available portfolio of 15,420 automobile claims, of which 923 are confirmed fraudulent, we compare logistic regression, decision trees, support vector machines, isolation forests, XGBoost extreme gradient boosting (XGBoost), autoencoders, variational autoencoders (VAEs), and VAE–XGBoost hybrids within a unified rare-event framework combining Boruta feature selection, the synthetic minority oversampling technique (SMOTE), and random undersampling. Grounded in the actuarial theory of optimal claim auditing, fraud detection is framed as risk-based audit prioritization, with scores converted into investigation decisions through prevalence-calibrated thresholds and assessed by the area under the receiver operating characteristic curve (ROC-AUC), the area under the precision–recall curve (PR-AUC), the area under the precision–recall curve (PR-AUC) and its lift, recall, specificity, and misclassification counts, with 95% confidence intervals (CIs). Reconstruction-based deep models rank claims close to random (ROC-AUC 0.52–0.54), indicating that fraudulent claims are not simple reconstruction anomalies, while hybrid models raise recall at a substantial cost in false positives. XGBoost with Boruta selection is strongest overall (ROC-AUC 0.926, 95% CI 0.904–0.947; PR-AUC 0.559, 95% CI 0.500–0.617; recall 0.830), and its ROC-AUC advantage over every other model family is statistically significant (p < 0.001); repeated cross-validation and a 1996 out-of-time test confirm the ranking. For structured claims data, gradient boosting with disciplined feature selection outperforms the reconstruction-based and VAE-hybrid architectures evaluated here. Because the data are a 1994–1996 public benchmark, the results concern the relative performance of model families rather than current fraud levels.

Zitieren

APA

Hage, R., Choufani, G., & Nassif, L. (2026). Risk-Based Audit Prioritization for Automobile Insurance Fraud: A Comparative Evaluation of Machine Learning and Deep Learning Under Class Imbalance. Risks, 14(11), 258. https://doi.org/10.3390/risks14110258

Vancouver

Hage R, Choufani G, Nassif L. Risk-Based Audit Prioritization for Automobile Insurance Fraud: A Comparative Evaluation of Machine Learning and Deep Learning Under Class Imbalance. Risks. 2026;14(11):258. doi:10.3390/risks14110258

BibTeX
@article{Hage2026,
  author = {Hage, Re-Mi and Choufani, Giovanni and Nassif, Leonie},
  title = {Risk-Based Audit Prioritization for Automobile Insurance Fraud: A Comparative Evaluation of Machine Learning and Deep Learning Under Class Imbalance},
  journal = {Risks},
  year = {2026},
  volume = {14},
  number = {11},
  pages = {258},
  doi = {10.3390/risks14110258}
}
Kostenlos starten →