ESTRO 2024 - Abstract Book

S4482

Physics - Machine learning models and clinical applications

ESTRO 2024

Development of an AI predictive model for patient selection for adaptive head and neck radiotherapy.

Mark Ashburner 1 , Omer Ali 1 , Gill Dobbie 2 , Xinyi Guo 2 , Yun Sing Koh 2

1 Waiakto Ragional Cancer Centre, Medical Physics, Waikato, New Zealand. 2 University of Auckland, Technology, Auckland, New Zealand

Purpose/Objective:

Adaptive radiotherapy (ART) offers a tailored approach to treatment and has been shown to be beneficial to patients undergoing treatment for head and neck carcinoma. The challenge lies in prospectively identifying patients who will benefit from ART intervention at the planning stage. This study presents the development and assessment of an AI based predictive model aimed at addressing this challenge.

Material/Methods:

Retrospective data from 100 head and neck patients was analysed, encompassing various patient features, including weight, neck dimensions, body volume, and target volumes. The training phase began with a single decision tree algorithm compared to a selection of other possibly suitable classifiers, being: random forest, bagging, adaBoost and gradient boosting. In addition to accuracy, F1 score, and cross-validation (CV), post-hoc Nemenyi testing was used to make inferences about which differences, if any, were significant. Initial features in the classifier were selected based on expert (RO) opinion, then feature engineering was done to refine our final model. The final model and features were tested on retrospective unseen clinical data from 115 patients, 25 of which required adaptive re-planning. The final model's performance was assessed using precision, recall, specificity, and sensitivity, providing a more thorough understanding of its capabilities.

Results:

The initial single decision tree model achieved F1 = 65%, accuracy = 60% and CV = 72%.

Classifier comparison showed that a random forest classifier consistently achieved better scores for accuracy (0.73) and F1 score (0.73), with ADA boost showing best performance for cross validation (0.8).

It was decided to proceed with a random forest classifier, which then had feature engineering to determine which features give best classifier prediction. The final classes to be used were: Weight at treatment start, Neck width& depth, body volume and equivalent Sphere diameter measured on planning CT, and primary CTV and PTV volume. Testing was carried out on the unseen data , using the described features in a random forest classifier. Initial results showed a Low precision (0.49), high recall and Sensitivity (0.83 , 0.83 respectively) and F1 of 0.62. Out of the 115 tested patients 21 were identified as a false positive – ie the classifier identified these patients as needing a replan, when they didn’t have one during their course at all.

Further analysis of these cases revealed a spectrum of adaptive interventions, from minor physical adjustments to the treatment shell, to intricate plan alterations without rescanning. A reclassification of these instances, indicating