Human-AI Collaborative Decision System for Insurance Claims Fraud through Bayesian Deep Learning with Expert Prior

Insurance claim fraud detection is a critical task in financial risk control, where traditional machine learning methods suffer from limited uncertainty quantification and insufficient utilization of domain knowledge. This paper proposes a Bayesian deep learning-based human-AI collaborative decision system that integrates expert prior knowledge to improve both detection accuracy and decision reliability. Expert experience rules are encoded as informative prior distributions over Bayesian neural network weights, while variational inference is employed for efficient posterior approximation and uncertainty estimation. Based on predictive entropy, an adaptive human-AI routing mechanism is designed to automatically process high-confidence cases and refer uncertain cases to human experts. Experiments on a real-world insurance dataset containing 127,543 claim records demonstrate that the proposed method achieves an F1 score of 0.7980 and an AUC of 0.9218, outperforming an ensemble deep learning baseline by 7.5% and 2.1%, respectively. The recall rate is improved to 0.8143, representing an 8.4% relative gain. Moreover, the expected calibration error is reduced to 0.028, indicating substantially improved confidence calibration. Ablation studies confirm that expert priors, uncertainty quantification, and human-AI collaboration each contribute significantly to performance gains. The proposed system can automatically handle 68.4% of cases while maintaining high accuracy, providing a practical and trustworthy solution for insurance fraud detection.