A Multi-Institutional External Validation of a Deep-Learning Based Platform for Prediction of Outcomes following SBRT Treatment for Early-Stage Non-Small Cell Lung Cancer

Abstract

Despite favorable tumor control of NSCLC with SBRT, some patients remain at high risk for treatment failure. Our group previously developed a deep learning (DL) platform that integrated clinical, radiographic, and dosimetric data to predict outcomes in patients with early-stage NSCLC at a single institution. This study aimed to externally validate the generalizability of DL based predictions on a blinded cohort of patients at a geographically distinct institution and assess model performance versus traditional statistical models. A DL algorithm was trained on a cohort of 527 patients with cT1-cT2 NSCLC treated with SBRT at a single-institution. The DL algorithm incorporated feed-forward neural network architecture to model clinical co-variates, 2D convolutional neural network (CNN) and long-short term memory recurrent neural network architectures to model radiographic data, and 1D CNN to model dosimetric data. Outcomes of interest were all-cause mortality as well as local, regional, distant, and overall treatment failures at 2-years. The DL model was tested with a blinded cohort of 62 patients treated at an external institution. Discriminatory ability of the DL model on the validation set was assessed using receiver operating characteristic curves (ROC) and compared to predictions generated from a traditional Cox Proportional Hazard model. Statistically significant differences in ROC curves were tested using the Delong method. The training and blinded validation cohorts showed no statistically significant differences in disease presentation, workup, or outcomes. When tested on the blinded external validation cohort, the DL platform maintained strong discriminatory ability in predicting all-cause mortality (AUC 0.73) as well as local (AUC 0.79), regional (AUC 0.69), distant (AUC 0.72), and overall (AUC 0.85) treatment failures. Discriminatory ability of the DL platform significantly outperformed the predictions generated from a traditional Cox Proportional Hazard model for all outcomes of interest (Table 1). A DL platform that effectively integrates clinical, imaging, and dosimetric data successfully models outcomes in early-stage NSCLC patients treated with SBRT on a blinded external validation cohort. The DL platform shows strong generalizability across multiple institutions and consistently outperforms predictions generated from a traditional Cox Proportional Hazard model. Integration of similar DL models that effectively analyze multiple data streams may improve clinical outcome prediction within clinical practice and guide personalization of therapy.Abstract 1103; Table2-Year OutcomesAUC DL Model95% Confidence IntervalAUC Cox Model95% Confidence IntervalDelong TestAny Treatment Failure.85.70-1.00.65.52-.78p<.001Local Failure.79.62-.96.58.47-.69p<.001Distant Failure.72.56-.88.60.47-.73p = .008Regional Failure.69.52-.86.61.51-.71p = .041All-Cause Mortality.73.61-.85.60.48-.72p<.001 Open table in a new tab