Development and Preliminary Validation of RP-WX: A WeChat Mini- Program-Based Prediction Model for Radiation Pneumonitis in Patients Undergoing Concurrent Chemoradiotherapy for Locally Advanced Squamous Cell Lung Cancer

Background Predicting the development of grade II or higher radiation pneumonitis in locally advanced squamous cell lung cancer (LASCLC) patients prior to concurrent chemoradiotherapy remains challenging, as traditional indicators based on dose-volume histograms (DVHs) or biological markers typically lack data or assessed post-treatment. In this study, we explored the potential of utilizing multi-omics (radiomics, dosimetric, clinical, and radiobiology features) as novel biomarkers to predict the Radiation pneumonitis of grade 2 or higher (RP2+) in LASCLC patients undergoing concurrent chemoradiotherapy. Methods A total of 129 patients with locally advanced small cell lung cancer (LASCLC) from four institutions were included in the training and validation cohort, with an additional 34 patients allocated to an independent test set. Four distinct feature categories—radiomics, dosimetry, clinical, and radiobiological—were employed to develop and validate the predictive model. A four-step feature selection algorithm was applied for dimensionality reduction. The three machine learning algorithms demonstrating the highest predictive performance were integrated into an ensemble model. Model interpretability was achieved using Shapley Additive Explanations (SHAP) values. Finally, a user-friendly graphical user interface (GUI) was developed to facilitate clinical translation. Findings: RP2 + occurred in 51.3% of enrolled patients. Univariate analysis revealed statistically significant differences between RP2 + and non-RP2 + patients in smoking status, radiotherapy position (RTP), Lungs_V5, PTV volume, Heart_V30, LEUD(=0.3)_SICK, LEUD(=0.3)_TOTAL, NTCP_LEUD_SICK, and NTCP_LKB_SICK. Nine features—comprising three dosimetric, three radiomic, and three radiobiological variables—were ultimately selected for model training and validation. Across all nine machine learning algorithms, four features consistently demonstrated strong predictive performance for RP2+: two dosimetric parameters (Lung_V5 and Lung_V20), one radiobiological metric (NTCP_LEUD_SICK), and one radiomic feature (glcm_InverseVariance_PGTV), each achieving mean AUC values > 0.70. The combined radiomic and radiobiological signature (RM + RB) yielded the highest model generalization accuracy (MGA), exceeding 0.92 across all three ensemble models, closely followed by the radiobiological-only model (MGA > 0.90). Within the RM + RB signature, three features contributed positively and three negatively to RP2 + prediction. Notably, lower SHAP values for NTCP_LEUD_SICK were associated with a reduced probability of RP2+. Interpretation: A user-friendly graphical user interface (GUI) was developed to facilitate the clinical implementation of the predictive model, thereby supporting clinical decision-making in routine practice.