Automatic Prescription Anomaly Detection Tool Assisting Peer Review Chart Rounds in Radiotherapy

Appropriate dosage of radiation is crucial in patient safety in radiotherapy. The current quality assurance heavily depends on a peer-review chart-round process, where the physicians reach a consensus on each patient’s dosage. However, such a process is manual and laborious. Physicians might not catch an error because of their limited time and energy. We designed a novel anomaly detection algorithm that utilized historical data from the past to predict anomalous cases. Such a tool can serve as an electronic peer who will assist the peer-review process providing extra safety to the patients.

In our primary model, we created two dissimilarity metrics, “R” and “F”. R defines how far a new patient’s prescription is from the historical patients’ prescriptions. F represents how far away a patient’s feature set is away from that of the group who has an identical or similar prescription. We flag the patients if either metric is greater than certain optimized cut-off values. We used thoracic cancer patients (n=3125) as an example and extracted seven features. Here, we report our testing F1 score which is between 73%-94% for different treatment technique groups.  

We also independently validate our results by conducting a mock peer review with three thoracic specialty MDs. Our model has the lowest miss rate (false negative rate) and outperformed MDs in the recall, precision, F1, and accuracy scores.  

Our model has many advantages over traditional ML algorithms, such as that it does not suffer from the class-imbalance problem. It can also explain why it flags each case and can impose the separation between prescription features and non-prescription-related features without having to be learned from the data.