Abstract 4626: A machine learning model for predicting hyperprogressive disease risk in cancer immunotherapy: Integrating clinical and molecular data

Abstract Background: Immunotherapy has revolutionized cancer treatment, offering significant efficacy across various cancer types. However, hyperprogressive disease (HPD)—a severe and unexpected acceleration of tumor growth—poses a critical challenge, as some patients experience this adverse effect. Early and accurate identification of HPD risk is essential for optimizing immunotherapy strategies and improving patient outcomes. To address this, we developed a machine learning-based predictive model for HPD risk across cancer types, integrating clinical and molecular data. Methods: This study analyzed data from 1, 005 patients with solid tumors treated with at least one dose of immune checkpoint inhibitors between 2018 and 2024. After stringent data preprocessing, feature selection was performed using chi-square and Fisher’s exact tests to identify significant variables associated with HPD. Logistic regression evaluated feature importance, accounting for multicollinearity, correlation, and interpretability. A random forest algorithm was then employed to construct a predictive model based on selected features. Results: From an initial pool of 44 variables, eight features were identified as significant predictors of HPD, including HPD-associated genes, immunotherapy resistance genes, cancer type, bone metastasis, the number of metastatic sites, lactate dehydrogenase levels, neutrophil-to-lymphocyte ratio, and treatment regimen. Patients lacking pre-immunotherapy genetic and blood test data for the selected features were excluded, resulting in a final cohort of 363 patients across 17 cancer types, including 52 HPD cases and 311 non-HPD cases. The model demonstrated strong predictive performance, achieving a cross-validated AUC of 0.87, sensitivity of 0.90, and specificity of 0.78. On an independent test set, the model attained an AUC of 0.71, sensitivity of 0.70, specificity of 0.73, and overall accuracy of 73%. Using the same eight features, a traditional logistic regression model was constructed (AUC = 0.66) and found to be less effective than the machine learning model. Conclusions: We identified eight statistically and clinically significant predictors of HPD and developed the first high-sensitivity pan-cancer predictive tool combining molecular and clinical features. This machine learning-based approach enhances HPD risk stratification, assisting clinicians in tailoring immunotherapy and mitigating risks, ultimately improving patient outcomes across cancer types. Citation Format: Haitao Wang, Lili Wang, Yu Zhang, Hongzheng Li, Nan Li, Qijie Jian, Yaxian Wang, Zhijian Song, Fei Pang, Kai Wang. A machine learning model for predicting hyperprogressive disease risk in cancer immunotherapy: Integrating clinical and molecular data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4626.