A Dangerous Complication. Radiation pneumonitis (RP) is a serious lung inflammation that occurs weeks to months after radiotherapy for lung cancer. Patients with pre-existing interstitial lung disease (ILD) are at especially high risk because ILD already causes diffuse parenchymal inflammation and fibrosis, making the lungs far more sensitive to radiation damage.
The Gap in Existing Tools. Most available predictive models are limited to small, single-institution studies and rely on only one type of data -- either dosimetric parameters (how radiation is delivered) or hematological markers (blood-based indicators). No validated model integrates all three data domains: clinical features, dose delivery data, and inflammatory biomarkers.
Why This Matters. Without reliable individualized risk tools, clinicians cannot accurately determine which ILD patients will develop RP after radiotherapy. This leads to suboptimal treatment planning, inadequate monitoring, and preventable severe complications in a vulnerable population.
Retrospective Cohort Design. Researchers enrolled 424 lung cancer patients with pre-existing ILD who received radiotherapy at Shandong Cancer Hospital between September 2020 and September 2024. All patients had histologically confirmed lung cancer, radiographic evidence of ILD before radiotherapy, and complete clinical and dosimetric records.
Comprehensive Data Collection. The study extracted demographic data, tumor characteristics, treatment details, and dosimetric parameters from hospital systems. Baseline laboratory values -- including complete blood counts and serum albumin -- were collected within four weeks before radiotherapy began. From these, inflammatory indices such as monocyte-to-lymphocyte ratio (MLR), neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), systemic immune-inflammation index (SII), and prognostic nutritional index (PNI) were calculated.
ILD Classification. Two thoracic radiologists independently scored the extent and pattern of ILD using high-resolution CT scans. CT patterns were classified as usual interstitial pneumonia (UIP), probable UIP, or indeterminate/alternative diagnosis following 2022 ATS/ERS/JRS/ALAT guidelines. Any discrepancies between readers were resolved by a third senior radiologist.
Machine Learning Feature Selection. Three complementary ML algorithms were applied to select the most important predictors: LASSO regression (penalized variable selection), random forest (RF), and extreme gradient boosting (XGBoost). Only variables identified by all three methods were retained, ensuring robust and consistent feature selection across both dosimetric and inflammatory predictor sets.
Creating the D Score. From the dosimetric feature selection process, four key parameters emerged: the clinical target volume to lung volume ratio (CV), mean heart dose (MHD), percentage of both lung volumes receiving at least 20 Gy (BV20), and percentage of heart volume receiving at least 30 Gy (HV30). These were weighted using logistic regression coefficients and combined into a composite dosimetric risk score called the D score, which was then standardized using Z score normalization.
Creating the Inflamm Score. Similarly, four inflammatory markers were selected by all three ML algorithms: monocyte count, MLR, total leukocyte count, and PNI. These were combined into a composite inflammatory risk score (Inflamm score) using the same logistic regression coefficient weighting approach, followed by Z score standardization.
Building the Nomogram. Clinical predictors identified from multivariate logistic regression, along with the D score and Inflamm score, were integrated into a final multivariable logistic regression model. A user-friendly nomogram was constructed from this integrated model to enable individualized risk estimation for each patient before radiotherapy begins.
Model Evaluation Framework. Performance was assessed using area under the ROC curve (AUC) for discrimination, calibration plots and the Hosmer-Lemeshow test for calibration, and decision curve analysis (DCA) for clinical utility. DeLong's test was used to formally compare AUC values between models.
High Incidence in This Population. Of the 424 patients studied, 200 (47%) developed radiation pneumonitis -- a strikingly high rate. Patients who developed RP were more likely to have advanced disease, poorer performance status, greater ILD severity, and more aggressive treatment regimens including immunotherapy and concurrent chemoradiotherapy.
Independent Clinical Risk Factors. Multivariate analysis confirmed seven independent clinical predictors of RP: higher performance status, higher Charlson Comorbidity Index (CCI), UIP pattern of ILD (odds ratio 6.11), receipt of immunotherapy (odds ratio 4.15), concurrent chemoradiotherapy (odds ratio 3.13), 28 or more radiotherapy fractions (odds ratio 2.13), and lower lung volume (odds ratio 2.33).
Diabetes as a Novel Risk Factor. The integrated model identified type 2 diabetes mellitus (T2DM) as an independent RP risk factor (odds ratio 4.75) in this specific population -- the first study to establish this association in lung cancer patients with ILD. Chronic hyperglycemia is thought to amplify inflammatory cytokines like IL-6 and increase reactive oxygen species production, worsening the inflammatory response to radiation.
Dosimetric and Inflammatory Scores Both Predict RP. Both the D score (AUC 0.758) and the Inflamm score (AUC 0.910) independently predicted RP. The Inflamm score substantially outperformed the D score, suggesting that systemic inflammation carries more predictive information than radiation dose distribution alone in patients with underlying ILD.
Exceptional Discrimination. The final integrated nomogram achieved an AUC of 0.929 (95% CI: 0.905-0.952), representing excellent ability to distinguish patients who will develop RP from those who will not. This substantially outperformed the clinical-only model (AUC 0.86) and the D score (AUC 0.758), both with p values below 0.001.
Comparison with Individual Scores. While the Inflamm score alone (AUC 0.910) approached the nomogram's performance, the integrated model still showed numerically superior results. The AUC improvement over the Inflamm score was 0.019 (p = 0.054), suggesting the added clinical and dosimetric information provides marginal but consistent benefit beyond inflammation markers alone.
Strong Calibration. Calibration curves showed close agreement between the nomogram's predicted RP probabilities and actual observed outcomes. The Hosmer-Lemeshow test confirmed good model fit. This means the model does not systematically overestimate or underestimate risk.
Clinical Utility Confirmed. Decision curve analysis demonstrated that using the nomogram to guide clinical decisions provides greater net benefit than treating all patients or no patients across a wide range of threshold probabilities. Subgroup analyses showed especially strong performance in patients under 65 and those with stage III-IV disease.
High vs. Low Risk Groups. For practical application, patients were divided into high-risk and low-risk groups using a nomogram total score cutoff of 111 points. Patients scoring above 111 were classified as high-risk. Representative case reviews showed that all high-risk patients in the sample developed RP during follow-up, while only rare cases appeared in the low-risk group.
What High-Risk Looks Like. High-risk patients consistently showed more adverse clinical comorbidities, greater radiation dose exposure reflected in higher D scores, and elevated systemic inflammation reflected in higher Inflamm scores. CT imaging in these patients revealed clear radiographic changes consistent with RP.
Guiding Treatment Decisions. The nomogram offers actionable guidance: for patients with high D scores, clinicians can optimize treatment planning to reduce dosimetric exposure. For patients with high Inflamm scores, more intensive monitoring and potentially preventive interventions can be initiated before radiotherapy begins, rather than waiting for symptoms to develop.
Future Validation Needed. The authors note that the cutoff of 111 may need recalibration for different clinical centers and patient populations. External multicenter validation studies are required to determine how best to adapt the threshold to local practice contexts.
Challenging Conventional Thresholds. Standard dosimetric limits -- such as keeping V20 below 30% and mean lung dose below 20 Gy -- were insufficient to prevent RP in this ILD cohort. A 47% RP incidence occurred despite most patients meeting these conventional criteria, demonstrating that dosimetric thresholds developed for patients without ILD cannot be directly applied to this high-risk group.
Why Inflammation Matters More. In ILD patients, the inflammatory score outperformed the dosimetric score, suggesting a different pathogenic mechanism drives RP in this population compared to lung cancer patients without ILD. Radiation may not only cause direct alveolar damage but also trigger aberrant immune responses that are amplified by the pre-existing inflammatory state of ILD.
The Role of Monocytes. The study identified elevated monocyte count as closely associated with RP -- a novel finding. ILD lungs already have increased CCL2 (a chemokine), which radiation further induces, leading to massive recruitment of inflammatory monocytes that worsen lung injury. This mechanism helps explain why baseline inflammatory status is such a strong RP predictor.
Nutritional Status as a Predictor. The prognostic nutritional index (PNI), reflecting both albumin levels and lymphocyte counts, was identified as a novel RP predictor. Low albumin reflects a pro-inflammatory state that facilitates cytokine cascades implicated in radiation-induced lung injury. Radiation-induced lymphopenia further disrupts immune regulation, delaying recovery from lung damage.
A First-of-Its-Kind Tool. This study represents one of the first large-sample systematic analyses comparing clinical, dosimetric, and hematological predictors specifically in ILD patients undergoing thoracic radiotherapy. The resulting integrated nomogram provides a practical, evidence-based tool using routine pretreatment data that clinicians can readily access.
Limitations to Address. The retrospective single-center design may introduce selection bias. Antifibrotic therapy -- which recent evidence suggests may reduce radiation-induced lung injury -- was not included due to incomplete records. Multi-omics approaches including radiomics and genomics were also not incorporated in this iteration.
The Path Forward. The research team plans multicenter external validation aggregating data across institutions, integration of radiomics features with standardized extraction methods, and application of SHAP (SHapley Additive exPlanations) analysis to improve model interpretability. These steps would enable prospective clinical deployment of the nomogram.
Broader Significance. By enabling accurate individualized RP risk assessment before radiotherapy begins, this tool could improve treatment outcomes for one of oncology's most challenging patient populations -- lung cancer patients whose pre-existing lung disease makes every treatment decision more consequential and more complex.