KEY POINTS
- This prediction-model study used the prospective ProPED-LUNG cohort from University Medical Center Groningen, the Netherlands, including 452 patients with stage III NSCLC treated with curative-intent chemoradiotherapy between January 2017 and January 2024. Median age was 68 years, 30% were current smokers, and 18% had baseline WHO performance status ≥2. Treatment included 158 patients receiving proton therapy (35%) and 294 receiving photon VMAT (65%).
- Chemoradiotherapy generally consisted of 60 Gy in 25 fractions with concurrent or sequential platinum-based chemotherapy. Proton selection followed institutional normal-tissue complication probability criteria rather than random allocation. After CRT, durvalumab was generally considered in patients without progression and with adequate clinical performance status, typically beginning within six weeks. The model's endpoint was actual immunotherapy initiation, rather than an independently adjudicated determination of physiological eligibility.
- Approximately 48% of the full cohort initiated post-CRT immunotherapy. The manuscript reports 216/452 patients in its abstract, while the detailed modeling results describe 198 initiations among 416 patients in the complete-data model population. This distinction matters because model performance and treatment proportions should be interpreted according to the relevant analytic population, not assumed to apply to every enrolled patient without missing data.
- Lung V1 was the strongest and most consistently selected dosimetric predictor of subsequent immunotherapy initiation. The final logistic model included lung volume receiving ≥1 Gy, age, current smoking, clinical T stage, and baseline WHO performance status. Higher lung V1 was associated with lower likelihood of immunotherapy initiation, with an apparent regression coefficient of −0.03 per percentage point of lung V1 (95% CI −0.04 to −0.02). Lung V1 was retained in all 1,000 bootstrap predictor-selection samples, while age and smoking were selected in 910 and 672 samples, respectively.
- Model discrimination was moderate and declined after internal validation. Apparent AUC was 0.75 (95% CI 0.71–0.80), decreasing to 0.71 (95% CI 0.67–0.76) after optimism correction. Calibration was considered acceptable, with a corrected slope of approximately 0.89. An alternative model replacing V1 with lung V5 had slightly worse overall performance, although V5 may be easier to incorporate into conventional photon treatment-plan optimization.
- Predictive performance was weaker when assessed separately by radiation modality. In available-data analyses, AUC was 0.67 (95% CI 0.56–0.77) among proton-treated patients and 0.66 (95% CI 0.59–0.72) among VMAT-treated patients, without significant discrimination differences between modalities (p=0.87). The apparent predictive contribution of V1 may therefore partly reflect modality selection and differences in the low-dose bath, rather than an independent effect of radiation dose distribution.
- The model does not prove that lowering V1 increases immunotherapy eligibility. Proton therapy was offered using model-based selection, the cohort spans the introduction of durvalumab into clinical practice, and baseline contraindications to immunotherapy were incompletely documented. The unexpected positive association between advanced clinical T stage and immunotherapy initiation further suggests unmeasured treatment-selection effects. In addition, the study did not evaluate whether plan modifications based on model predictions improve durvalumab initiation, toxicity, or survival.
CLINICAL TAKEAWAY
A model incorporating lung V1 and routinely available clinical variables showed moderate ability to predict post-chemoradiotherapy immunotherapy initiation in stage III NSCLC. The findings suggest that low-dose lung exposure may be relevant when evaluating treatment plans, but do not establish that reducing V1 will causally increase durvalumab use. External validation across different fractionation schedules and prospective evaluation of model-guided plan optimization are required before clinical implementation.