1048 - Bridging Prediction and Causality: Deciphering Clinical Outcomes in Locally Advanced Non-Small Cell Lung Cancer Treated with Chemoradiation via Multitask Machine Learning and Causal Inference of Proto

S. H. Lee¹, N. Yegya-Raman¹, C. Friedes¹, M. Iocolano¹, R. Caruana², J. D. Bradley¹, G. D. Kao¹, S. J. Feigenberg¹, and Y. Xiao³; ¹Department of Radiation Oncology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, ²Microsoft Research, Redmond, WA, ³Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA

Purpose/Objective(s): This study aims to improve the identification of dose-volume constraints (DVCs) and the predictability of clinical outcomes (COs) by using advanced multitask machine learning (MT ML), augmented by causal inference (CI) to analyze treatment effects. We hypothesize that incorporating MT ML with CI can effectively predict COs and differentiate the impacts of proton vs photon chemoradiotherapy (CRT). Materials/

Methods: We evaluated 663 patients with locally advanced non-small cell lung cancer (LA-NSCLC), treated with either proton (n=260) or photon (n=403) CRT, including a subset of 179 who received consolidation immunotherapy (CIO). These patients were randomly split into training (n=464) and testing (n=199) cohorts. A total of 1,201 features were extracted: 37 clinical, 1,093 geometric/dose-volume histogram (DVH), 50 immuno-hematological DVH, and 21 absolute lymphocyte/neutrophil count (ALC/ANC) kinetic parameters, including those derived from mathematical models, to describe the radio-immune response during and post-CRT. In training, MT LASSO used five-fold cross-validation (5FCV) to select features for 5 COs: overall survival (OS), death without progression (DWP), progression-free survival (PFS), and locoregional/distant failure-free survival (LFFS/DFFS). MT gradient boosting machine (GBM) and MT neural additive model (NAM) were trained with these features to jointly predict the COs, validated by c-index via 5FCV and ensemble predictions on the test dataset. The causal impact of proton vs photon CRT was estimated through propensity score matching across all patients for each CO, using the same features identified as confounders or covariates in a directed acyclic graph. To verify the robustness of the estimated CI effect on each CO, a refutation test was carried out using a randomly selected subset of the original data.
Results: For CO prediction, 25 features were selected: 7 clinical, 12 geometric/DVH, and 6 ALC/ANC kinetics. ECOG status, the top predictor in both MT GBM and MT NAM, was associated with an increase in OS risk by 20 months for MT NAM and 30 months for MT GBM with a rise from 0 to 2. Only MT GBM identified distinct risk shifts at specific DVC or feature thresholds, indicating that the risk for all COs increases with a maximum ANC to ALC ratio >20, age >70 and atria-PTV overlap >15 cc, and D35%[Gy] >20 in the left anterior descending coronary artery, while CIO receipt lowers this risk. Comparative tests showed that MT GBM outperformed MT NAM in all COs, with c-indices: OS (0.7 vs 0.69), DWP (0.8 vs 0.76), PFS (0.64 vs 0.63), LFFS (0.71 vs 0.7), and DFFS (0.65 vs 0.63). CI revealed improvements in all COs with proton CRT: OS by 4.59, DWP by 1.52, PFS by 1.38, LFFS by 1.8, and DFFS by 2.98 months. Refutation tests confirmed the robustness of causal estimates: OS (p=0.16), DWP (p=0.6), PFS (p=0.4), LFFS (p=0.28), and DFFS (p=0.42).
Conclusion: MT GBM enhances the identification of DVCs and the prediction of COs, with proton CRT demonstrating superior COs over photon CRT in LA-NSCLC.