A novel data-driven algorithm to predict anomalous prescription based on patient's feature set
This work addresses patient safety in radiotherapy by automating anomaly detection to assist physicians, though it is incremental as it builds on existing data-driven methods for a specific medical domain.
The researchers tackled the problem of manual, error-prone quality assurance in radiotherapy by developing a data-driven algorithm to predict anomalous prescriptions, achieving an F1 score between 75% and 94% across different treatment technique groups and demonstrating a lower type 2 error rate compared to manual peer review.
Appropriate dosing of radiation is crucial to patient safety in radiotherapy. Current quality assurance depends heavily on a peer-review process, where the physicians' peer review on each patient's treatment plan, including dose and fractionation. However, such a process is manual and laborious. Physicians may not identify errors due to time constraints and caseload. We designed a novel prescription anomaly detection algorithm that utilizes historical data 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 defining how far a new patient's prescription is from historical prescriptions. F represents how far away a patient's feature set is from the group with an identical or similar prescription. We flag prescription if either metric is greater than specific optimized cut-off values. We used thoracic cancer patients (n=2356) as an example and extracted seven features. Here, we report our testing f1 score, between 75%-94% for different treatment technique groups. We also independently validate our results by conducting a mock peer review with three thoracic specialists. Our model has a lower type 2 error rate compared to manual peer-review physicians. Our model has many advantages over traditional machine learning algorithms, particularly in that it does not suffer from class imbalance. It can also explain why it flags each case and separate prescription and non-prescription-related features without learning from the data.