Estimating temporal treatment-effect patterns of radiotherapy and chemotherapy in lower-grade gliomas using causal machine learning

Treatment decisions for lower-grade gliomas (WHO grades 2–3) rest on trial averages, which lack temporal resolution. We applied Causal Analysis of Survival Trajectories (CAST), a causal-machine-learning method that builds treatment-effect trajectories from horizon-specific estimates, to 776 adults from The Cancer Genome Atlas (TCGA, n = 512) and the Chinese Glioma Genome Atlas (CGGA, n = 264) across six radiotherapy and alkylating-chemotherapy scenarios on overall (OS) and progression-free survival (PFS). Elastic-net propensity scores with overlap weighting (target: average treatment effect on the overlap population, ATO) balanced age, sex, grade, IDH, 1p/19q, and extent of resection. Chemotherapy showed adjusted survival-probability gains peaking at 0.34 (95% CI -0.32 to 1.00) at 84 months (TCGA OS) and 0.48 (0.04 to 0.92) at 108 months (CGGA OS); E-values of 5.1–27.6 indicate robustness to unmeasured confounding. Radiotherapy estimates were mixed (E-values 1.1–5.1) and are reported as adjusted associations sensitive to residual confounding from missing extent-of-resection and performance-status data, not as evidence of treatment-induced effect. Age drove most heterogeneity (46–52% of splits); refutation tests supported the chemotherapy findings.