A Machine Learning Decision-Support Model for Selective Use of Online Adaptive SBRT In Left Adrenal Tumors
Abstract
Purpose
Online adaptive stereotactic body radiation therapy (SBRT) reduces organ at risk (OAR) dose for abdominal targets with highly variable anatomy but significantly increases treatment time and clinical workload. Adaptive workflows are routinely assigned to left adrenal tumor patients due to gastrointestinal OAR proximity. This approach may overlook patients who could safely receive conventional SBRT. This study aims to develop a simulation-based, machine learning decision-support model to identify patients unlikely to require online adaptation and streamline the ART workflow.
Methods
A cohort of 52 patients treated for left-adrenal tumors with magnetic resonance or computed tomography-guided adaptive SBRT was compiled from one institution. Geometric and dosimetric input parameters were derived from 44 patients’ simulation scans, including target volumes, OAR to target distance, and clinical dose objectives. Ordinary least squares (OLS) regression was performed for baseline multivariate prediction and LASSO to evaluate regularization and feature selection benefits. Both models initially predicted the number of fractions expected to require adaptation. To increase clinical applicability, the continuous predictions were subsequently converted to a conservative binary decision. The model’s performance was assessed through internal validation using LOOCV and an independent external testing set of 8 patients.
Results
Both models demonstrated high sensitivity for identifying patients requiring adaptive SBRT. Internal validation yielded a sensitivity of 0.97 for both models, with specificity ranging from 0.60 to 0.67. External validation maintained a sensitivity of 1.00 with a specificity of 0.50, indicating conservative behavior which preserved patient safety while identifying patients unlikely to benefit from adaptation. OLS and LASSO demonstrated comparable predictive capabilities.
Conclusion
This simulation-based decision-support model provides a conservative framework for optimizing resource allocation in online ART. Using a limited set of planning-stage features, the model enables standardized, data-driven identification of patients who may safely forgo adaptive SBRT, reducing unnecessary clinical burden while preserving treatment quality.