3471 - Differential Dose-Volume Indices-Based Machine Learning Model for Severe Radiation-Induced Lymphopenia in Patients with Upper Abdominal Malignancies
Presenter(s)
Y. Li1, M. Li1, A. Liu2, Y. Zhu3, X. Wang1, Y. Li4, D. Liu5, and L. Guo6; 1Shandong Cancer Hospital and Institute, Jinan, China, 2Qilu Hospital of Shandong University, Jinan, China, 3Hunan Xianghui Human Resources Service Co., Ltd., Changsha, China, 4Shandong Cancer Hospital and Institute, Jinan, None, China, 5Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, Shandong, China, 6Dongying People’s Hospital, Dongying, China
Purpose/Objective(s): Severe radiation-induced lymphopenia (SRIL) impairs immunotherapy efficacy and predicts adverse outcomes in the immunotherapy era. However, the availability of satisfactory predictive models and modeling methodologies remains limited. This study aims to develop a high-performance prediction model for SRIL in patients with upper abdominal malignancies, addressing this urgent and critical clinical challenge.
Materials/Methods: A retrospective analysis was performed on 545 patients with upper abdominal malignancies recruited from two institutions. Three machine learning (ML) algorithms (XGBoost, Random Forest, Logistic Regression) were employed to develop SRIL prediction models integrating clinical features, differential Dose-Volume (DV) indices, and other dosimetric information. We constructed the ML model using 20-fold cross-validation. SHapley Additive exPlanations (SHAP) were applied to interpret the optimal model. A simplified decision tree algorithm was subsequently derived for rapid and accurate assessment of patients’ SRIL risk.
Results: Among 482 patients, 108 (22.4%) experienced SRIL, compared to 30.2% in the external validation cohort. The XGBoost model outperformed other prevalent models and demonstrated robust generalization performance within the independent cohort (accuracy 0.78, ROC-AUC 0.89, PR-AUC 0.78, F1 score 0.72, specificity 0.91, precision 0.69 and recall 0.89). The study demonstrates that the splenic volume exposed to 0-5Gy radiation is positively correlated with an increased risk of SRIL (SHAP = 0.639, SHAP direction = 0.041). Additionally, baseline lymphocyte count emerged as the most protective feature (SHAP = 0.559, SHAP direction = –0.119).
Conclusion: The XGBoost model performed well in predicting SRIL and the decision tree enhances the applicability of the model. Additionally, the differential DV indices concept can be extended to dose-response models that are currently based on DV indices.