A risk model that added baseline inflammation measures and an early sign of radiation-related toxicity improved prediction of grade 2 or higher radiation pneumonitis, an inflammatory lung reaction after treatment, in lung cancer patients receiving thoracic radiotherapy, according to a retrospective single-center study. The strongest version of the study's explainable nomogram, a clinical risk-assessment tool, combined clinical and treatment factors with baseline LDH, a systemic inflammation score and acute radiation esophagitis.
Radiation pneumonitis occurred in 86 of 726 patients, or 11.8%, and the median time to diagnosis was 121.5 days. The primary endpoint was grade 2 or higher radiation pneumonitis; deaths unrelated to pneumonitis were handled as a competing event in the analysis.
How the risk model was built
Researchers examined the 726-patient cohort treated at Jilin Provincial Cancer Hospital between January 2015 and December 2025. The three sequential Cox models were built in stages. Model 1 used clinical and treatment factors, Model 2 added baseline LDH in 50 U/L increments and the systemic inflammation score, and Model 3 added early acute radiation esophagitis.
Acute radiation esophagitis was assessed at a prespecified 30-day landmark and then entered as a fixed covariate. No radiation pneumonitis or death occurred before that landmark in eligible patients. Feature selection used LASSO-penalized Cox regression with an alpha of 1.0, followed by clinical and statistical refinement to reduce collinearity and preserve a parsimonious model.
What the analysis found
In the primary multivariable Cox model, the analysis reported independent associations for alcohol history, TOMO rather than VMAT, baseline LDH, a systemic inflammation score of 0 and acute radiation esophagitis. The reported hazard ratios were 1.63, 2.16, 1.24 for each 50 U/L increase in LDH, 0.59 and 1.85, respectively. A hazard ratio compares the modeled rate of an event over time with a reference group, so these figures describe associations within this cohort, not cause and effect.
Prediction improved as variables were added. The C-index, a measure of how well a model separates patients with and without the outcome, rose from 0.625 for Model 1 to 0.673 for Model 2 and 0.688 for Model 3. At six months, the time-dependent AUC, another measure of separation at a fixed time, rose from 0.646 to 0.694 to 0.713, with Model 3 highest.
Compared with Model 1, Model 3 produced a continuous net reclassification improvement of 0.274 and an integrated discrimination improvement of 0.037, two measures used to assess whether a new model sorts risk better. The reported P value was less than 0.001.
The researchers also assessed calibration, which compares predicted with observed risk, and decision utility. Model 3 had a Brier score of 0.108, compared with 0.112 for the null model. At a 10% intervention threshold, its net benefit was 0.047, versus 0.045 for Model 2 and 0.032 for Model 1. That utility comparison was tied to the 10% threshold, rather than being a general measure of clinical benefit.
Interpretation and limits
To make the model easier to inspect, the researchers used SHAP, an interpretability method that estimates how much each feature contributes to a prediction. The systemic inflammation score ranked first by mean absolute SHAP value at 0.296, followed by alcohol history at 0.229, treatment modality at 0.192, LDH at 0.165 and acute esophagitis at 0.159. The ranking describes how the model made predictions, not proof that any feature is a biological cause.
The competing-risk sensitivity check gave a similar overall result. Non-radiation-pneumonitis competing events occurred in 3.9% of the cohort, and estimates from a Fine-Gray competing-risk analysis were described as concordant with the standard Cox estimates.
The evidence has important limits. The cohort was retrospective and single-center, and the real-world sample was heterogeneous in patient selection, radiotherapy modalities and systemic treatments. The dataset lacked radiotherapy dosimetric parameters such as mean lung dose, V20 and V5. The authors called for prospective multicenter validation with standardized protocols.
The model was internally assessed with 1,000 bootstrap resamples. Taken together with the study design and missing dosimetry, that means the reported performance remains an internal prediction result. The study shows associations and model performance within this cohort; it does not establish that alcohol history, radiotherapy modality, LDH, the inflammation score or esophagitis causes radiation pneumonitis.
Paper data and sources
Original title: An explainable nomogram incorporating systemic inflammatory indicators and radiation esophagitis for risk assessment of radiation pneumonitis in lung cancer
Authors: Li J, Li C, Liu J et al.
Journal/Repository: Frontiers in immunology
Status: Peer-reviewed
First online: 2026-10-10
DOI: Not available
Original paper · Full text