Application of CT-based foundational artificial intelligence and radiomics models for prediction of survival for lung cancer patients treated on the NRG/RTOG 0617 clinical trial

Abstract

Objectives: To apply CT-based foundational artificial intelligence (AI) and radiomics models for predicting overall survival (OS) for patients with locally advanced non-small cell lung cancer (NSCLC).

Methods: Data for 449 patients retrospectively treated on the NRG Oncology/Radiation Therapy Oncology Group (RTOG) 0617 clinical trial were analyzed. Foundational AI, radiomics, and clinical features were evaluated using univariate cox regression and correlational analyses to determine independent predictors of survival. Several models were fit using these predictors and model performance was evaluated using nested cross-validation and unseen independent test datasets via area under receiver-operator-characteristic curves, AUCs.

Results: For all patients, the combined foundational AI and clinical models achieved AUCs of 0.67 for the Random Forest (RF) model. The combined radiomics and clinical models achieved RF AUCs of 0.66. In the low-dose arm, foundational AI alone achieved AUC of 0.67, while AUC for the ensemble radiomics and clinical models was 0.65 for the support vector machine (SVM). In the high-dose arm, AUC values were 0.67 for combined radiomics and clinical models and 0.66 for the foundational AI model.

Conclusions: This study demonstrated encouraging results for application of foundational AI and radiomics models for prediction of outcomes. More research is warranted to understand the value of ensemble models toward improving performance via complementary information.

Advances in knowledge: Using foundational AI and radiomics-based models we were able to identify significant signatures of outcomes for NSCLC patients retrospectively treated on a national cooperative group clinical trial. Associated models will be important for application toward prospective patients.

Keywords: cross-validation; feature selection; foundational model; machine learning; non-small cell lung cancer; prognosis; radiomics.

Figure 1.

Summary of the various radiomics workflow: schematic steps includes collection of patient clinical information and image (CT and dose) dataset, extracting clinical, DVH, radiomics and foundational AI feature, feature selection, and machine learning model construction for the development of the predictive models.

Figure 2.

(A) Performance of models based on foundational AI features (FD), hand-crafted radiomics (HR) features, clinical features (CF) and ensembles for the 3 different patient sub-cohorts. Results are shown for the models with the best prognostic performance based on AUC. For the “all dose” and “high dose” cohorts, the RF algorithm yielded highest performance. For the low dose cohort, the support vector machine achieved highest accuracy. Experiments were performed using 5-repeat 10-fold nested cross-validation and are reported for the unseen independent test datasets. (B) The ROC curve for the highest performing model in each arm on an unseen independent test set. RF = random forest; SVM = support vector machine.