Multitask AI Models for the Joint Prediction of Overall Survival, Progression-Free Survival, and Death without Progression as a Composite Endpoint for LA-NSCLC Patients Treated with Chemoradiotherapy

Abstract

Prior methods model the risk of endpoints separately. Herein, we construct a composite AI model that considers multiple endpoints jointly, including overall survival (OS), progression-free survival (PFS), and death without progression (DWP). Our hypothesis is that the composite model potentially improves predictive performance for patients with locally advanced non-small cell lung cancer (LANSCLC) treated with chemoradiotherapy (CRT). A total of 335 LANSCLC patients treated with definitive CRT, including all evaluable patients accrued from Oct 2017 to Dec 2021, were randomly split into training/test subsets (n = 234/101). Cardio-pulmonary substructures (CPSs) were autocontoured, manually reviewed, and edited if necessary. A total of 1093 non-independent dosimetric parameters were extracted, including GTVp, GTVn, GTV, PTV, esophagus, lungs minus IGTV, left/right lung, 15 CPSs, and the overlapping volume of each OAR with PTV and the distance from each OAR to GTVp/GTVn. Other clinical parameters included age, consolidation immunotherapy (CI), ECOG score, Charlson comorbidity index, coronary heart disease, histology, PD-L1 expression, and clinical stage (AJCC 8). Within training, censored time-to-event data were imputed based on conditional event distributions derived from Kaplan-Meier estimators for casting survival analysis as a regression problem and training neural additive model (NAM) regressors. Features were selected by LASSO regression for a single endpoint (OS, PFS, DWP) and multi-task (MT) LASSO regression for four separate composite endpoints (OS-PFS, OS-DWP, PFS-DWP, OS-PFS-DWP). The performance of MT NAMs in the test set that jointly predicted the composite endpoints was evaluated using the C-index and compared to that of a single task (ST) NAM that predicted each endpoint separately. The best testing performance in predicting OS and DWP was attained by the MT NAM that jointly predicted all endpoints (c-index = 0.65, 95% CI 0.58-0.71 for OS; c-index = 0.78, 95% CI 0.69-0.87 for DWP). The best model to predict PFS was also MT between PFS and DWP (c-index = 0.59, 95% CI 0.52-0.65). The c-indices of all ST NAMs were less than 0.56. The best MT NAMs significantly outperformed ST NAMs in predicting OS (p = 0.001) and DWP (p = 0.01) except for PFS (p = 0.32). The best MT NAM in predicting OS and DWP included ECOG score, atria-PTV overlap volume, D75% [Gy] to the left atrium (LA), pulmonary arterial volume, histology (adenocarcinoma), D65% [Gy] to the descending aorta (DA), V10 Gy [%] of the LA and CI in order of overall importance. ECOG score consistently ranked as the most important feature for all four MT NAMs. An increase of ECOG score from 0 to 2 indicated a 6-month earlier risk of mortality and DWP. Atria-PTV overlap volume and D65% [Gy] to the DA were included in all four MT NAMs. MT AI models improved outcome prediction in patients with LANSCLC treated with CRT by jointly learning commonalities between the primary and auxiliary endpoints.