ClaimShield AI for Predictive Insurance Fraud Detection and Financial Risk Resilience Using Machine Learning

Insurance fraud imposes direct claim losses, investigation costs, operational delays, and wider pressure on insurers’ financial resilience. This study develops ClaimShield AI, a machine-learning decision-support framework that combines fraud classification, claim-level risk scoring, explainability, and financial-exposure measurement. The analysis uses the supplied benchmark dataset of 10,000 insurance claims, containing 9,200 legitimate claims (92%) and 800 fraudulent claims (8%). The source document explicitly identifies these values and model outputs as synthetic and illustrative; accordingly, the present paper treats the evidence as a controlled simulation study rather than as observed insurer field data. The design uses an 80:20 stratified train-test split, with class-imbalance treatment restricted to the training data. Six classifiers are compared: logistic regression, decision tree, random forest, support vector machine, artificial neural network, and XGBoost. XGBoost provides the strongest operational balance, with 93.60% accuracy, 60.67% precision, 56.88% recall, 96.79% specificity, and an F1-score of 58.71%. Logistic regression achieves the highest recall (83.12%), ROC-AUC (0.932), and PR-AUC (0.670), demonstrating that the preferred model depends on the cost assigned to missed fraud versus unnecessary review. Fraud is concentrated in claims with longer reporting delays, more prior claims, larger claim amounts, absent witnesses, absent police reports, theft incidents, and online submissions. The ClaimShield risk bands further concentrate 69.5% of all fraudulent claims within only 7.41% of the portfolio at high or critical risk. In the test portfolio, ClaimShield flags $3.232 million of $5.357 million in fraudulent claim value, yielding a 60.32% fraud-value capture rate. The study concludes that machine learning can improve fraud triage and financial-risk visibility when combined with human review, transparent thresholds, and external validation on real insurer data.