Prediction of hematotoxicity using CT-derived body composition in rectal cancer patients receiving immunotherapy-based total neoadjuvant therapy.

Abstract

3599 Background: Immunotherapy-based total neoadjuvant therapy (iTNT) is a promising strategy to enhance complete response rates and facilitate organ preservation in locally advanced rectal cancer (LARC). However, it is linked to a considerable incidence of severe hematotoxicity, while reliable predictive biomarkers are lacking. Body composition, reflecting nutritional status, has been linked to treatment-related toxicity, but its relationship with iTNT toxicity remains unclear. This study aimed to investigate the association between body composition and hematotoxicity of iTNT and to develop a predictive model. Methods: A cohort of 204 LARC patients treated with iTNT was included for model development, with internal training and validation via five-fold cross-validation. Additionally, 43 metastatic rectal cancer patients treated with the same regimen served as the external validation cohort. Body composition parameters were obtained from pretreatment CT images at L3. LASSO regression was applied to select variables and four machine learning (ML) algorithms including logistic regression, random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) were developed. Results: 33.8% of the discovery cohort experienced grade 3-4 hematotoxicity. Among baseline characteristics, skeletal muscle index (SMI) showed the most significant correlation with toxicity. The cut-offs for SMI, determined by optimal stratification, were 50.0 cm²/m² for males and 39.6 cm²/m² for females. Patients with low SMI had a higher incidence of severe toxicities ( P < 0.001). Seven features were included in the ML models: age, body mass index (BMI), SMI, subcutaneous adipose tissue index (SATI), subcutaneous adipose tissue attenuation (SAT_HU), visceral adipose tissue index (VATI), and skeletal muscle attenuation (SM_HU). The predictive performance of the models is summarized in the following table. Overall, the SVM model showed the highest discriminative performance and robust generalizability. Conclusions: SMI was significantly associated with hematotoxicity during iTNT. A CT-based machine learning model was developed and validated to predict hematotoxicity, with the SVM model showing the most robust performance (AUC > 0.85) to guide clinical decision-making. Cohort Model Accuracy Sensitivity Specificity AUC Internal validation SVM 0.850 ± 0.024 0.849 ± 0.021 0.833 ± 0.027 0.881 ± 0.019 RF 0.755 ± 0.037 0.749 ± 0.037 0.777 ± 0.032 0.803 ± 0.033 Logistic regression 0.704 ± 0.048 0.672 ± 0.055 0.720 ± 0.048 0.730 ± 0.033 XGBoost 0.796 ± 0.024 0.769 ± 0.038 0.788 ± 0.036 0.814 ± 0.027 External validation SVM 0.821 ± 0.034 0.829 ± 0.026 0.800 ± 0.019 0.858 ± 0.029 RF 0.713 ± 0.030 0.723 ± 0.051 0.734 ± 0.043 0.750 ± 0.038 Logistic regression 0.666 ± 0.034 0.652 ± 0.044 0.649 ± 0.027 0.671 ± 0.037 　 XGBoost 0.751 ± 0.046 0.714 ± 0.038 0.744 ± 0.031 0.779 ± 0.044