Exploring a Hybrid Deep Learning Approach for Anomaly Detection in Mental Healthcare Provider Billing: what it means for business leaders

A 2025 study shows how pairing Isolation-Forest pseudo-labels with LSTM-Transformer ensembles flags fraudulent mental-health claims earlier, giving insurers and compliance teams an efficient, scalable fraud-screening tool.

1. What the method is

The pipeline creates pseudo-labels with Isolation Forest and autoencoders, then trains an LSTM and a Transformer on these labels. A gradient-boosted meta-learner fuses their outputs to score each billing sequence’s fraud risk.

2. Why the method was developed

Manual fraud labels are scarce, biasing supervised models toward legitimate claims. The hybrid approach enlarges the minority class automatically and exploits both temporal patterns (LSTM) and cross-field interactions (Transformer) to surface new scams.

3. Who should care

• Special-investigation units at health insurers
• Revenue-cycle & compliance managers at provider groups
• Government auditors policing over-billing
• AI product leads building payment-integrity tools

4. How the method works

Cleaned claim tables are windowed into 50-record sequences. Outlier scorers flag the top 0.4 % as anomalies. These pseudo-labels train an LSTM and a Transformer; their logits feed a boosted tree meta-classifier that outputs final fraud probabilities.

5. How it was evaluated

On two Dutch datasets (71 k declarations, 1.59 M operations) with 80 / 20 splits, models were judged by recall, precision, F1, and AUROC. Five-fold validation set hyper-parameters.

6. How it performed

Declaration-level recall hit 96.3 % (F1 0.84). At the operation level, the full hybrid reached 74.4 % recall, a 10-point gain over the best single model, enabling earlier, broader fraud capture. (Source: arXiv 2507.01924, 2025)