1. Patient Data
Our study retrospectively collected 175 lung cancer patients treated with IMRT at Xiangya Hospital from January 2011 to January 2022, including 87 RP patients and 88 non-RP patients. The independent external validation cohort consisted of 13 RP patients and 11 non-RP patients from Changsha Central Hospital (Figure S1). All patients underwent IMRT, administered with 1.5-3 Gy per day, 5 days a week, 10–35 fractions with a total dose of 30–70 Gy. Patients will be enrolled with the following criteria: (i) Primary lung cancer diagnosed with pathology. (ii) Patients completed radiotherapy and recorded the clinical information. Clinical characteristics included age, sex, SI, BMI, tumor-node-metastasis (TNM) stage, related primary disease (chronic obstructive pulmonary disease (COPD), diabetes, and hypertension), pathology, tumor location, carcinoembryonic antigen (CEA), SII, NLR, PLR, LMR, chemotherapy regimens and total dose. SI is calculated as cigarettes per day multiply duration of smoking (years). SII, NLR, PLR and LMR are defined as follows: SII = platelet x neutrophil/lymphocyte, NLR = neutrophil/lymphocyte, PLR = platelet/lymphocyte, LMR = lymphocyte count/monocyte count. (iii) High-quality images and volumetric dose can be obtained from the radiation planning CT. (iv) The RP patients were diagnosed via radiology or symptoms, and RP was classified into mild-RP (≤ 2 grade) and severe-RP (> 2 grade) groups based on Criteria for Adverse Events (CTCAE) and RTOG criteria.
2. Image collection and ROIs delineation
All patients were scanned and located by CT analogue positioning system (SOMATOM Definition AS) with scanning voltage, tube current, helical sweep pitch, and slice thickness of 140 kV, 271 mAs, 0.6 and 3mm, respectively. Four ROIs were selected for analysis in the planning CT, including GTV, PTV, TL-GTV, and TL-PTV. The GTV was confirmed by a senior oncologist and radiologist.
3. Radiomics features extraction
The images were normalized to avoid variation, and a total of 1960 radiomics features, including shape (n = 72), texture (n = 1452), and intensity features (n = 436), were extracted from four ROIs by an open-source software, IBEX. Specifically, shape features describe geometric properties, such as diameters, volumes, and surface areas (Fig. 1). Texture features reflect the degree or abruptness of gray-level intensity fluctuations, which include the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), et al. Intensity features represent the intensity distribution of individual voxels.
4. Dosiomics and dosimetric features extraction
The planning CT-based radiation dose was classified into two major types: dosiomics and dosimetric. The dosiomics features were extracted from the dose distributions of the four ROIs. A total of 160 dosiomics features, including GLCM (n = 96) and GLRLM (n = 64), were analyzed in our study by the 3D Slicer software. The dosimetric factors were extracted from the dose-volume histogram (DVH). The factors consisted of whole lung V5, V10, V15, V20, V25, V30, V35, V40, V45, V50. Additionally, the maximum dose, the minimum dose and the average dose in the total lung volumes, GTV and PTV were also included into the analysis.
5. Feature selection
The feature selection strategy was designed as follows: (i) Spearman’s correlation analysis was used to screen the extracted features, where a correlation larger than 0.85 is considered as strongly correlated, and one of the features was removed. (ii) The variance was used to select the features, and the two features with a variance difference of less than 0.01 were regarded as solid correlations, and one of the features was removed. (iii) The Least absolute shrinkage and selector operation (LASSO) Cox regression model with 10-times repeated 5-fold cross-validation was used to select the most valuable features correlated to RP status.
6. Model validation and performance evaluation
The subject were divided into a training/validation set (n = 150, 85.7%) and a testing set (n = 25, 14.3%). To select the optimal machine learning algorithm, there used multiple algorithms, such as Support Vector Classification (SVC), Logistic Regression (LR), Gaussian Naive Bayes (GNB), Bernoulli Naive Bayes (BNB), K Nearest Neighbor (KNN), Decision Tree (DT) and Random Forest (RF) for predicting RP status. To make the distribution of the training set more reasonable, the overall data set was randomly divided into a training/testing set in a 6:1 ratio over 100 times. The 10-fold stratified cross-validation method was carried out to train classifiers. We have trained more than 5000 times to select the best performing model. The established model was evaluated in the independent external validation set (n = 24). The predictive performance of the machine learning model was evaluated using the AUC. All calculation was completed via python (version 3.6).
7. Survival analysis
According to the CTCAE and RTOG criteria, all patients were divided into two groups, which were severe-RP (> 2 grade) and mild-RP (≤ 2 grade). Kaplan–Meier method was used to compare the survival curves of non-RP and RP patients (severe-RP patients and mild-RP patients). The difference in OS was analyzed using the log-rank test and Cox regression model in all patients. Survival analysis was done with R (version 4.2.1). A two-tailed p-value < 0.05 was considered statistically significant.