Artificial Intelligence-Assisted Peer Review in Radiation Oncology

Objective Peer review is an essential part of the patient treatment process that examines and, where necessary, recommends revisions to clinical data, therapeutic parameters, and potential alternative approaches to treatment. Our hypothesis is that artificial intelligence (AI) and machine language technologies can enhance peer-review efficacy by screening cases for potential treatment interruptions caused by re-planning and treatment cessation. Materials and Methods Fifty-five features of clinical and therapeutic parameters from 3881 radiotherapy patients (7142 plans) treated from 2014 to 2021 were used as input for two AI models: a multivariable least absolute shrinkage and selection operator (LASSO) logistic regression model and a pattern recognition feed-forward neural network (NN). The dataset was split into 70% training and 30% testing, with the training set divided into five groups for cross-validation. Analysis was performed on the full cohort and on subsets based on treatment site. Performance metrics of accuracy, sensitivity, and specificity were calculated. Results Overall, 8.1% of all cases had treatment interruptions, most commonly in the head and neck region compared to other sites (19% vs 6%-9%, P <.01). For the LASSO model, test set sensitivity, specificity, and accuracy ranged from 37%-70%, 59%-78%, and 60%-76%, respectively, with higher specificity than sensitivity for site subsets. For the NN model, test set sensitivity, specificity, and accuracy ranged from 41%-68%, 53%-79%, and 53%-78%, respectively. Both models demonstrated the highest accuracy in the brain subset. For the full cohort, NN accuracy (58%) was similar to LASSO (60%). The largest accuracy differences between LASSO and NN were in the lung/breast/chest (LASSO: 71% vs NN: 57%) and spine/extremity (LASSO: 66% vs NN: 54%) subsets. Conclusion Our results provide proof-of-concept that AI- and ML-based technologies have potential as screening tools to aid peer review in radiation oncology. Early identification of patients at risk for radiation therapy interruptions using these tools could translate into higher treatment completion rates. The study is being continued to include more clinical features and to optimize model hyperparameters.