Due to the high morbidity and potential mortality of RP in radiotherapy for thoracic tumors, it is particularly important to explore efficient and convenient predictive variables to reduce the incidence of RP and mitigate radiotherapy toxicity. Clinical parameters [15], cytokines [16], and dosimetric parameters [17] have been used in previous studies for prediction, which have poor predictive efficacy and stability. Moreover, traditional dosimetric parameters ignore the detailed features of dose information, and the accuracy of predicting RP by applying dosimetric features alone is only 60% ~ 70% [18], and it is still difficult to predict occurring of RP efficiently in many patients in the clinic via the physical dose of lung tissues.
In this study, in the process of training the model, we used the cross-validation method to group the patients and divided them into the training set and validation set according to the ratio of 7:3, and the statistical results showed that the sample distribution had a good homogeneity between two groups. Previous studies on clinical parameters had shown that patients' age, gender, physical condition, stage, type of pathology, and smoking status may be associated with the occurrence of RP [19–23]. In the present study showed that advanced age was found to be significantly associated with the risk of developing RP, which may be related to lower resistance, poorer physical condition, and poorer resistance of lung tissue to radiation in the elderly. The level of lymphocyte ratio could effectively reflect the functional status of the patient's immune system, and the results of our study suggest that abnormal pre-radiotherapy lymphocyte ratio is associated with radiation pneumonitis, and the study of Zhang X et al[24] also suggests that pre-radiotherapy lymphocyte ratio can predict the occurrence of radiation pneumonitis in patients undergoing radiotherapy for thoracic tumors, and the results of our study are in line with them. Among the dosimetric parameters, lung irradiation volume [25], total irradiation dose [26], MLD [27], lung volume [28], and lung FDG-PET uptake [29] are influential factors for the occurrence of radiation pneumonitis. The more commonly used dosimetric factors in clinical studies are V20Gy [3] and MLD [27], and their correlation with radiation lung injury has been confirmed by numerous researchers. Furthermore, this study showed that lung V20Gy and MLD were correlated with the occurrence of RP, and the difference in V20Gy was more significant, suggesting that the volume of biased dose is more likely to lead to the occurrence of RP, so we included the V20Gy and MLD parameters in the model construction. Meanwhile, this also reminds us that minimizing the radical dose of radiotherapy during radiotherapy implementation can help prevent the occurrence of RP under the condition of ensuring the therapeutic effect.
Radiomics is a rapidly emerging discipline, which is imaging mining analysis of the heterogeneity and biological features of tumors and normal tissues has largely compensated for the insufficiency of the predictive efficacy of purely clinical factors or dose-volume parameter models [30–31]. In practice ,certain second-order or higher-order eigenvalue changes were found to be significantly correlated with the signs of RP occurrence in some studies where CT images of patients with lung cancer before and after radiation therapy were collected for radiomics feature extraction,which confirmed the possibility of predicting RP by radiomic methods[32]. The peritumor microenvironment plays a major role in carcinogenesis, recurrence, and metastasis. Tumor cells and the tissues in which tumors grow are extremely varied. Several previous studies had shown that model construction based on intratumor and peritumor radiomic has played an important role in the prediction of lymph node metastasis in lung adenocarcinoma [33], the identification of lung adenocarcinoma and sarcoidosis [34], and the prognosis determination of NSCLC [35]. However, there hasn't been as much study done on how to forecast radiation pneumonitis in NSCLC. Thus, the extraction and analysis of peritumor features were carried out in this work in accordance with the tumor ROI outreach of 0.5 cm to construct a peritumor model, so as not to overlook the hidden information offered by CT images surrounding the tumor, and then, the intratumor and peritumor portions of the patient's pre-radiotherapy CT were screened for comparison of the radiomic features through LASSO, to effectively search for a subset of the radiomic features that can better predict RP independently in early stage.It is found that in the construction of intratumor and peritumor models with different ranges of CT image features, a subset of features can be extracted independently of each other that are closely related to the RP, and choosing the better models to participate in joint model,which has a good innovation.
The results suggested that the predictive efficacy of the intratumoral model was significantly higher than that of the peritumoral model (AUC 0.798 vs. 0.714), and the reason for the analysis was that the intratumoral ROI was the high-dose area of the radiotherapy dose, and its dose distribution was significantly higher than that of the peritumoral ROI, and the dose of radiotherapy and the size of the area would directly affect the dose parameters such as V20Gy and MLD, which were significantly related to the RP. Therefore, we chose to construct an effective combined model by combining the intratumoral model Rad -score with other five indicators such as whether the age is ≥ 60 years old, whether smoking, whether the lymphocyte ratio is in the normal range, MLD, V20Gy, etc. Excitingly, the joint model’ AUC in the training and validation sets amounted to 0.928 and 0.765, which were significantly improved over the predictive efficacy of single parameters.Huang Y et al [36] based on the characteristics of dosimetry and radiomic constructed a hybrid model outperformed a single-parameter model and improved the prediction performance of RP after SBRT (AUC of 0.82). In a study of RP prediction based on planned CT radiomic combined with dosimetry parameters and clinical parameters, Jiang W et al [37] obtained an AUC of 0.83 based on dosimetry features, and the AUC of the combined model was more than 0.9, which was at the same level of predictive efficacy as the training set in our study. This suggests that the joint Nomogram model constructed by combining radiomic parameters, clinical parameters and dosimetry parameters has better predictive efficacy for the occurrence of RP. Most importantly, these parameters are all necessary for the radiotherapy consultation of patients, which has the advantages of accessibility, affordability, efficiency and convenience.
Of course ,there are also some shortcomings in this study. First, the AUC of the joint model we built in this study was as high as 0.928 in the training set but only reached 0.765 in the validation set, and the rationale behind this result could be that the model's limited sample size causes overfitting of the data, which lowers the model's AUC in the validation set. Secondly, because this study was retrospective in nature and was carried out by a single institution, there may have been selection bias due to a lack of a more rigorous multicenter study design or a superior external validation set. As a result, future research involving multiple centers and large sample sizes will be necessary to further validate our findings.In addition, only two of the most studied parameters, MLD and lung V20Gy, were included in the dosimetric factors, and other factors—like cardiac dose—were not given enough weight. While prior research has also shown that MLD and lung V20Gy are the main risk factors for RP, the dose to the heart also influences the pulmonary toxicity that results in RP, and some researchs shown that radiation therapy-induced pulmonary toxicity is worse when there is decreased cardiorespiratory function[38–39]. Consequently, future research could be considered for incorporate data like cardiac dosage in order to further enhance predictive performance of the model. Furthermore, this study simply employed a machine learning methodology, and in future work, radiomics can also be combined with deep learning artificial intelligence methods to deeply mine better machine learning modes and build better prediction models.
In conclusion,the study's findings suggest that combined modeling based on radiomics along with clinical and dosimetry characteristics may enhance the prediction of RP even more,which can be used by clinical doctors to guide clinical work.