Using AI to Predict Patient's Atypical Prescription Based on Individual's Feature-Set

Abstract

<h3>Purpose/Objective(s)</h3> The goal of this project is to train unsupervised machine learning (ML) models to cluster patients into different groups based on their individual features (such as morphology, cancer stage, treatment intent, site location, etc.). Then, we use the distribution of the prescription - Rx (number of fractions times dose per fraction) within each cluster to establish typical Rx values as well as less frequent values. In clinical setting, a new patient's Rx can then be flagged if it does not fall into the typical range of Rx for the individual patient's feature-set. Additionally, the value of Rx is predicted with supervised ML models. This AI computed value can be compared against the clinical Rx to create an additional flag if there is a significant relative deviation. We aim to implement an automatic mechanism based on this ML model along with visualization tools to assist peer review during weekly chart rounds in order to provide extra safety for patients by flagging the infrequent used Rx. <h3>Materials/Methods</h3> We queried all radiation oncology patients from the electronic patient information management system treated between 01/01/2007 to 01/01/2021 (14 years of retrospective data) at our institution. Based on their diagnostic code (ICD9/ICD10 code), we categorize the patients into different disease groups (i.e., prostate, H&N, CNS patients and etc.) for various disease specific model training. Several clustering models (<b>k</b>-means, hierarchical) were applied to cluster patients into different groups based on their individual features, which includes morphology code, treatment intent, treatment techniques, treatment type (initial or cone down), TNM stage, treatment site and etc.). Then, supervised ML models (Random Forest Regression) were applied to predict each group's Rx distribution, which gives an estimate of the most frequent and least frequent Rx range within the group. A data pipeline was developed and then applied to prostate patients for initial testing. <h3>Results</h3> We applied agglomerative clustering to the prostate group and found natural partitions of the patients into <b>k</b> = 2-5 clusters. Observation of the Rx in different clusters showed that each cluster had distinct characteristic Rx's. We plotted the Rx distribution for each feature, the scatterplot of Rx against each feature, and Rx distribution for selected feature-sets for visualization. The supervised prediction results are pending. <h3>Conclusion</h3> This initial prostate ML model along with visualization tools can flag the Rx that are atypical for a particular cluster for further investigation in order to improve the efficiency of peer review process. This further investigation could result in detecting anomaly Rx therefore providing extra safety for patients.