Speaker
Description
Objective
To develop an explainable artificial intelligence framework integrating radiotherapy dosimetry and quantitative imaging to identify survival-related patient phenotypes and inform personalised dose optimisation in locally advanced non-small cell lung cancer (LA-NSCLC).
Methods
This multicentre retrospective study included 388 patients treated with curative-intent radiotherapy across six institutions between 2014 and 2022. Clinical and treatment data were combined with quantitative features automatically extracted from planning CT and dose distributions. These included dose-volume metrics from cardiac substructures, estimated dose to circulating immune cells (EDIC), CT-derived body composition (visceral fat, skeletal muscle and intermuscular fat), coronary and aortic calcium scores, aortic diameters, and lung low-attenuation areas (LAA) as a surrogate of emphysema.
After dimensionality reduction and feature selection using Variance Inflation Factor, mutual information and permutation testing, three machine-learning classifiers (LightGBM, XGBoost and Random Forest) were trained to predict 2-year overall survival (OS2) using a 70/30 train-test split. Model discrimination was assessed using AUC, while SHAP analysis provided patient- and feature-level explainability. To identify clinically interpretable patient phenotypes beyond binary prediction, the multidimensional feature space was projected using UMAP and clustered using HDBSCAN.
Results
OS2 was 56.7%. Six features emerged as the strongest predictors of survival: EDIC, visceral fat volume, planning target volume (PTV), aortic Agatston calcium score, LAA, and equivalent uniform dose (EUD, n=0.1) to the left anterior descending coronary artery. Random Forest achieved the highest predictive performance, with AUC values of 0.83 in training and 0.65 in testing.
UMAP-HDBSCAN analysis identified eight distinct patient clusters characterised by combinations of two radiation-dose metrics, one tumour-related variable and three quantitative imaging biomarkers. These clusters showed different survival profiles, indicating that radiation exposure interacts with patient-specific cardiovascular, pulmonary and metabolic characteristics rather than acting as an isolated risk factor.
Conclusion
An explainable AI pipeline combining routinely available CT imaging and radiotherapy dosimetry can identify clinically meaningful survival phenotypes in LA-NSCLC. The integration of supervised prediction, SHAP-based interpretation and unsupervised patient stratification provides a potential framework for moving from population-based dose constraints towards patient-specific radiotherapy optimisation, in which acceptable organ-at-risk exposure may be adapted according to individual cardiopulmonary and metabolic vulnerability.