Unsupervised anomaly detection for discrete sequence healthcare data
share
Fraud in healthcare is widespread, as doctors could prescribe unnecessary treatments to increase bills. Insurance companies want to detect these anomalous fraudulent bills and reduce their losses. Traditional fraud detection methods use expert rules and manual data processing. Recently, machine learning techniques automate this process, but hand-labeled data is extremely costly and usually out of date. That is why unsupervised fraud detection system in healthcare is also of great importance. However, there are almost no applications of unsupervised anomaly detection based on the processing of sequential data. To process sequential data, we propose two deep learning approaches: LSTM neural network for prediction next patient visit and a seq2seq model. We assume that errors of predictions correspond to anomality for both cases and compare different ways to aggregate errors and detect abnormality of the whole sequence. For normalization of anomaly scores, we consider Empirical Distribution Function (EDF) approach: the algorithm can work with high class imbalance problems during aggregation of errors. We use real data on sequences of patients' visits data from a major insurance company. The results show that both models outperform a baseline for unsupervised anomaly detection. Our EDF approach improves the quality of LSTM model. Moreover, both models provide new state-of-the-art results for unsupervised anomaly detection for fraud detection in healthcare insurance.
READ FULL TEXT
