Cervical cancer is a malignant neoplasm at the junction of the squamous epithelial cells of the vaginal or transitional zone of the cervix and the endocervical canal columnar epithelial cells. Cervical cancer is the fourth most common cancer worldwide (Sung et al., 2021). Radiotherapy is one of the most effective methods for treating pelvic malignancies and has an irreplaceable role in treating cervical cancer at all stages. The main complication of radiotherapy for pelvic malignancies is radiation proctitis (Yeung et al., 2020). Fifty percent of patients with cervical cancer or endometrial cancer who received postoperative intensity-modulated radiotherapy developed acute rectal toxicity, and 5%–10% developed chronic rectal toxicity (Zelefsky et al., 2008; Yeung et al., 2020).

Unlike simple radiation damage to other organs, radiation proctitis is a complex disease closely related to the microbiota. A study using a rectal radiation mouse model showed that radiation affected both host and intestinal microbiota (Gerassy-Vainberg et al., 2018). Radiation therapy could induce local microbial ecological dysbiosis, and the dysbiosis microbiota could exert a direct pro-inflammatory effect on epithelial cells. In another study of 32 female patients with chronic radiation proctitis, differential patterns of dysbiosis were observed after analyzing the gut microbiota of patients with or without hematochezia (Liu et al., 2021). Gut microbiota could offer a set of biomarkers for radiation enteritis prediction, disease activity evaluation, and treatment selection (Wang et al., 2019).

However, the current prediction models of radiation proctitis were focused mainly on clinical and radiotherapy dose features. Several univariate and multivariate analyses showed that features, including tumor size, pathological characteristics, and radiological parameters, were significantly correlated with post-radiotherapy comorbidities in patients undergoing pelvic radiotherapy (Albert et al., 2008; Schmidt et al., 2022). A study by Fiorino et al. showed that rectal function was significantly correlated with treatment volume, PTV margins, radiation therapy dose, hemorrhoids presence, anticoagulant use during follow-up, and relative (%) and absolute (cm³) values of rectal V38Gy and V40Gy correlated with rectal bleeding (Fiorino et al., 2008). A study by Mahal et al. also noted that total radiation dose, dose per fraction, radiotherapy techniques, and treatment volume affected the rectum of patients (Mahal et al., 2014).

Another review also suggested features associated with radiation proctitis, including vascular diseases such as smoking, diabetes, hypertensive diabetes and atherosclerosis, collagen vascular disease, comorbid inflammatory bowel disease, and human immunodeficiency virus infection. Also, the review noted that specific underlying genetic changes could affect patients’ sensitivity to radiation. There was a correlation between genes and higher risks of gastrointestinal and genitourinary tract radiotoxicity after radiotherapy (Shadad et al., 2013).

The above studies have shown a strong correlation between patients’ oncologic features, pathologic features, and radiologic dose and the appearance of radiation proctitis in postoperative radiotherapy for pelvic cancer. However, the conclusions of these studies are inconsistent, and the accuracy of the prediction of radiation proctitis is unsatisfactory.

In recent years, computer-aided diagnosis, especially machine learning methods, has also been used for postoperative radiotherapy side effects and comorbidities prediction in oncology patients. Lee et al. applied machine learning methods such as random forest and bioinformatics to genome-wide data to predict and interpret advanced urogenital toxicity (Lee et al., 2018). They designed more robust predictive models and identified plausible biomarkers and biological processes associated with late urogenital toxicity. A study by Lewis & Kemp et al. showed that the integration of machine learning and genome-scale metabolic modeling identified multi-group biomarkers of radiation resistance (Lewis and Kemp, 2021).

However, it should be noted that the machine learning models above were developed using clinical features only. It ignored the large number of features embedded in computed tomography images (CT) that are imperceptible to the naked eye. Moreover, in the process of treatment plan determination, physicians are more focused on extracting focal area information and lack awareness of pelvic rectal information. Therefore, a comprehensive model is urgently needed to deepen the understanding of patient rectal image information to accurately assess radiotherapy treatment plans and reduce severe complications of radiation proctitis.

Deep learning, as well as dynamical modeling, is demonstrating powerful feature extraction and modeling capabilities in various medical fields (Li et al., 2021; Qian et al., 2021; Chen et al., 2022; Hu et al., 2022; Li Y. et al., 2022; Li X. et al., 2022). In data-driven disease research, a graph neuro network was used to predict the potential associations of disease-related metabolites (Sun et al., 2022). Deep learning can also be used to explore the identification of circRNA-disease associations (Wang et al., 2021) and predict the potential human lncRNA interactions (Zhang et al., 2021; Jingxuan et al., 2022; Wang et al., 2022). In drug metabolism research, deep learning can be used to predict the ability of a compound to permeate across the blood–brain barrier (Tang et al., 2022) and drug response (Kuenzi et al., 2020). At the same time, deep learning is also a popular tool for radiotherapy research. Zhong et al. developed a deep learning-based radiomic nomogram that could predict the prognosis of patients with different treatment regimens (Zhong et al., 2021). Qiang et al. established a prognosis prediction system based on deep learning for locoregionally advanced nasopharyngeal carcinoma (Qiang et al., 2021). Although deep learning has been widely applied in the analysis and prediction of various diseases, the poor interpretability of the deep learning model makes it difficult for clinicians to understand and trust these tools (Huff et al., 2021).

Radiomics can extract biomedical images containing information reflecting the underlying pathophysiology and reveal the relationships through quantitative image analysis (Le et al., 2021; Lam et al., 2022). In previous studies, radiomics has been used to predict postoperative radiotherapy-induced toxicity in prostate cancer patients. Mostafaei et al. showed that models based on CT radiomics, clinical features, and dose-volume parameters could predict radiation toxicity. The combination of imaging and clinical features could improve the performance of radiotoxicity prediction models (Mostafaei et al., 2020). However, no study has been performed to predict postoperative radiotherapy comorbidity in cervical cancer patients using radiomic features. Due to the limited resolution, the information on microbiota is almost impossible to extract directly by radiomics in theory, and no relevant studies have been reported. However, it is feasible that radiomics can indirectly reflect the effect of microbiota on the rectum.

Therefore, this study uses radiomics to extract the rectal information from medical images and improve the model performance and diagnostic accuracy through quantitative image analysis. Moreover, this study creatively introduces SHapley Additive exPlanation (SHAP) values to explore the interpretability of nomogram to improve the clinicians’ understanding of the model and its radiomic features, which facilitates later clinical promotion.

With the advancement of radiotherapy techniques, the postoperative survival rates of cancers such as cervical cancer have increased dramatically (Citrin, 2017). However, complications and side effects caused by postoperative radiotherapy or chemotherapy are difficult to avoid. Radiation proctitis is one of the most common complications of postoperative radiotherapy in patients with pelvic tumors (Rustagi et al., 2015; Qian et al., 2021), with mild diarrhea or mild rectal exudate in mild cases and even intestinal necrosis or bleeding in severe cases, endangering patients’ lives (Citrin, 2017). In clinical practice, doctors currently rely on the dose-volume features of radiotherapy plans to assess the risk of radiation proctitis. However, there is a lot of valuable information in pathology and clinical imaging that is not considered by clinicians. Moreover, the sensitivity of the rectum to radiotherapy radiation also varies significantly between individuals.

To further refine the assessment of radiation proctitis, we selected radiomic features associated with radiation proctitis by univariate regression and the LASSO algorithm. Radiotherapy techniques (OR = 0.000 (0.000–0.086), p = 0.005), Maximum rectal dose (OR = 1.006 (1.001–1.011), p = 0.020), Contrast (OR = 0.000 (0.000–0.002), p = 0.046) were independent risk factors for radiation proctitis. Finally, we developed an integrated prediction model based on clinical and radiomic features [AUC = 0.6855 (0.5174–0.8535)]. Current studies of radiation proctitis had mainly focused on local radiotherapy dose limits rather than comprehensive predictive models (Snyder et al., 2001; Huang et al., 2004). There was only one study using radiomics to build a predictive model for radiation proctitis (Mostafaei et al., 2020). In gastrointestinal toxicities modeling, the AUC of radiomic model of their study was 0.71, which was relatively higher compared with our study. However, the study was conducted based on data from only 64 patients and was only suitable for patients with prostate cancer.

The radiomic features of the model potentially incorporated the effect of microbiota on rectal radiosensitivity. The model without radiomic features showed lower validity, while the model containing both radiomic features and clinical features showed better performance on the ROC curve. The change of net benefit in Figure 3D suggested that radiomic features had played a supporting role in predictive models. And as a measure of the local intensity variation, a larger contrast correlated with a greater disparity in intensity values among neighboring voxels. In our study, the contrast suggested that a lower tissue density compared to the surrounding tissue was associated with higher radiosensitivity.

In most cases, PyRadiomics followed the image biomarker standardization initiative (IBSI)'s definition of features. PyRadiomics development was also involved in the standardization effort by the IBSI team. Still, there were some differences between PyRadiomics and feature extraction as defined in the IBSI documents. Most notably were the differences in gray value discretization (just for the fixed bin size type) and resampling. In summary, the definitions of PyRadiomics and IBSI were slightly different, but did not represent one over the other. Moreover, IBSI was only an initiative, not a standard. For these reasons, IBSI would not significantly impact the reproducibility and validity of this study.

While SHAP was often used to explain features in machine learning algorithms and neural network models (Bang et al., 2021; Park et al., 2022; Shaji et al., 2022; Shi et al., 2022), SHAP analysis of logistic regression models had not yet been mentioned. Although logistic regression algorithms were simpler and more explicit than other machine learning algorithms and neural networks, logistic regression models were more challenging to understand than they may seem. Users could not directly measure the importance of features between continuous and categorical variables through odds ratio (OR) or coefficients (Table 2). In particular, for radiomic models, the significant variation in the magnitude of radiomic features made it more challenging to understand the actual decision-making process of the model through the coefficients and OR values of logistic regression. We wanted to help users better understand each feature’s role in the model. In subsequent clinical treatment, model users can further quantify the contribution of radiomic features in each model output.

To address this issue, we introduced SHAP for the first time to a radiomics-based logistic regression model, which further revealed the model’s decision-making mechanism (Figure 4A). The total contribution of SHAP for each feature included in the model was analyzed (Figure 5B). Radiotherapy techniques and the maximum rectal dose occupied vital positions in the model contribution. Notably, the SHAP value of the radiomic feature was the lowest. It suggested that the radiomic feature was weaker than the clinical feature and dose-volume feature. The SHAP could also analyze correlations between variables (Figures 5C,D). Correlations in SHAP values were observed between the three features. It may suggest an inter-collaborative relationship between variables in the model. However, this can only indicate a correlation between SHAP values, not between the values of the variables. In the subplot of therapy (Figure 5C), we can find that the most relevant variable was the maximum rectal dose. There was a harmful effect of maximum rectal dose in the decision-making process of these VMAT therapy samples. However, no fixed pattern was observed in the subplot of contrast (Figure 5D).

SHAP had a unique role in radiomics-based logistic models as a game-theoretic approach. SHAP helped us understand radiomic features that vary significantly in magnitude. Furthermore, SHAP provided a quantitative and visual representation of the decision mechanisms within the model for each patient.

We recommend that clinicians can reduce the value of the maximum rectal dose by modifying the plan when the model suggests that the current radiotherapy plan has a high probability of radiation proctitis. Clinicians can rely on interpretable models to precisely control the risk of the final plan to an acceptable level. Patients with cervical cancer can reduce unnecessary radiation doses and the incidence of radiation proctitis with the help of the comprehensive model.

We successfully developed and validated an integrated radiomic model containing rectal information in this study. The integrated radiomic model enables the accurate quantitative assessment of the probability of radiation proctitis in postoperative cervical cancer patients, addressing the limitations of the current qualitative assessment based on dose-volume parameters only. Based on the model output and SHAP values analysis, we suggest that clinicians can adjust the radiation dose to minimize the occurrence of severe radiation proctitis while not compromising the effectiveness of radiation therapy.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethics Committee in Clinical Research (ECCR) of the First Affiliated Hospital of Wenzhou Medical University. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

CW, XX, and CX conceived the project, developed the prediction method, designed and implemented the experiments, analyzed the result, and wrote the manuscript. XZ, SR, and XJin implemented the experiments and analyzed the result. QZ, WD, HaizL, SW, and YZ analyzed the result. HaiL, QH, YL, XJia, and JS contributed to the interpretation of the results. All authors contributed to the article and approved the submitted version.

This work was supported by Zhejiang Engineering Research Center for Innovation and Application of Intelligent Radiotherapy Technology in the Second Affiliated Hospital of Wenzhou Medical University, Wenzhou key Laboratory of radiotherapy and Translational Research of Cancer (2021100848), and Wenzhou Science and Technology Bureau Y2020733. This work was also supported by the Ministry of Science and Technology of the People’s Republic of China under grant 2021ZD0201900, the National Natural Science Foundation of China under grant numbers 12090052 and 11874310, and Major Projects in Fujian Province under grant 2020Y4001.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.