1151 - Dynamic Hematologic Trajectories and Machine Learning to Predict Survival and Radiotherapy Benefit In Unresectable Esophageal Squamous Cell Carcinoma Receiving Immunotherapy: A Multicenter, Real-World Study
Presenter(s)
W. Shen1, X. Zhao2, Y. Zhang3, Y. Liu4, X. Wang5, J. Cao6, Y. Xu7, X. Ye8, M. Hou9, W. Zhou10, Y. Song11, J. Lv12, Z. Zeng13, and B. Aray14; 1The Fourth Hospital of Hebei Medical University, Shijiazhuang, Hebei, China, 2Department of Radiation Oncology, Fourth Hospital of Hebei Medical University,, Shijiazhuang, China, 3The Affiliated Anyang Tumor Hospital of Henan University of Science and Technology, Anyang, China, 4Nanjing Medical University Affiliated Cancer Hospital, Nanjing, Jiangsu, China, 5Cancer Hospital Chinese Academy of Medical Sciences, Beijing, China, 6Shanxi Province Cancer Hospital/ Shanxi Hospital Affiliated to Cancer Hospital, Chinese Academy of Medical Sciences/Cancer Hospital Affiliated to Shanxi Medical University, Taiyuan, China, 7Fujian Cancer Hospital, Fuzhou, Fujian, China, 8Department of Radiation Oncology, The First Affiliated Hospital, Zhejiang University School of Medicine, Hangzhou, China, 9Affiliated Hospital of North Sichuan Medical College, Nanchong, Sichuan, China, 10Department of Radiation Oncology, Chongqing University Cancer Hospital, Chongqing, China, China, 11Huai’an First People’s Hospital, Huai'an, Jiangsu, China, 12Sichuan Clinical Research Center for Cancer,Sichuan Hospital Cancer & Institute, Sichuan Cancer Center, University of Electronic Science and Technology of China, Chengdu, China, 13the second affliated hospital of Nanchang University, Nanchang, China, 14Lun Lun Research Institute of AI FOR SCIENCE, Beijing Diji Technology Co., Ltd., Beijing, China, Beijing, Beijing, China
Purpose/Objective(s): Immunotherapy has revolutionized unresectable esophageal squamous cell carcinoma (ESCC) treatment; however, identifying candidates for combined radiotherapy remains a critical clinical challenge. We hypothesized that a machine learning (ML) model integrating dynamic hematologic features would superiorly predict survival and identify patients who benefit from radiotherapy compared to static markers. This study leveraged ML on dynamic hematologic features to predict survival and stratify beneficiaries of radiotherapy intervention.
Materials/Methods: In this multicenter study of 1,607 patients, the cohort was divided into a training set (n=884), an internal testing set (n=222), and an independent external validation set (n=501). Latent Class Growth Modeling (LCGM) mapped longitudinal trajectories of inflammatory indices. Several ML algorithms were compared to build a survival model. Radiotherapy efficacy and specific benefit thresholds were evaluated using Inverse Probability of Treatment Weighting (IPTW) and Subpopulation Treatment Effect Pattern Plot (STEPP).
Results: Longitudinal trajectory analysis demonstrated that dynamic patterns of inflammatory indices—such as the resolving pattern of SIS and SII—were associated with significantly better survival outcomes compared with the rebound and delayed increase patterns. The Cox proportional hazards (CoxPH) model showed the best predictive performance, achieving a C-index of 0.701 in the training cohort and 0.670 in the validation cohort, and incorporated 11 key features. The risk stratification system derived from this model effectively discriminated survival outcomes, with patients in the high-risk group exhibiting a 2.7-fold increased risk of death (HR 2.70, 95% CI 2.24–3.25, P < 0.0001). IPTW analysis indicated that radiotherapy was associated with an overall survival benefit in the entire population; however, this benefit was not significant in the high-risk subgroup. STEPP analysis further revealed a non-linear pattern of radiotherapy benefit: potential survival benefits were observed only in patients with a Risk Score < 1.19 (validated at < 1.26 in the external cohort), whereas no apparent benefit was detected beyond this cutoff. The model has been made open-source.
Conclusion: This study validates a robust model demonstrating the superiority of dynamic hematologic trajectories over static baselines. By defining a quantifiable threshold, this cost-effective tool identifies specific beneficiaries of combined radiotherapy, optimizing efficacy while minimizing overtreatment in high-risk populations