Abstract: Anomaly detection is a key technique for medical insurance fund supervision.Most existing approaches rely on complex feature engineering and domain expertise to characterize suspicious behaviors,making rules costly to build and maintain and difficult to adapt to evolving fraud patterns.Meanwhile,labels are also scarce,delayed,and noisy,further limiting the reliable deployment of supervised methods.To this end,this paper presents SeqRecon-AD(Sequence Reconstruction for Anomaly Detection),an unsupervised anomaly detection framework in medical insurance that models each account as a time-ordered sequence of reimbursed items and measures account risk by its deviation from normal transition patterns.Specifically,an autoregressive Transformer is trained with a next-item reconstruction objective to capture regular item transitions.Token-level negative log-likelihood losses are then aggregated into an account-level anomaly score via Top-k loss aggregation,which emphasizes sparse abnormal segments rather than average behavior.Experimental results on a real-world city-scale dataset show that SeqRecon-AD outperforms classical unsupervised baselines,representative sequence models as well as reconstruction-based autoencoders.SeqRecon-AD provides an effective and deployable unsupervised solution for medical insurance anomaly detection without relying on anomaly labels for training,improving AUC by 29.21% over the best unsupervised baseline.

Key words: Medical insurance anomaly detection, Unsupervised learning, Sequence reconstruction, Autoregressive Transformer, Top-k loss aggregation