Beung-Chul Ahn1, Jea-Woo So2, Chun-Bong Synn3, Tae Hyung Kim2, Jae Hwan Kim4, Yeongseon Byeon4, Young Seob Kim5, Seong Gu Heo4, San-Duk Yang4, Mi Ran Yun6, Sangbin Lim4, Su-Jin Choi3, Wongeun Lee6, Dong Kwon Kim3, Eun Ji Lee3, Seul Lee3, Doo-Jae Lee7, Chang Gon Kim1, Sun Min Lim1, Min Hee Hong1, Byoung Chul Cho1, Kyoung-Ho Pyo8, Hye Ryun Kim9. 1. Division of Medical Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Republic of Korea. 2. TheragenBio, Seongnam, Republic of Korea. 3. Severance Biomedical Science Institute, Yonsei University College of Medicine, Seoul, Republic of Korea; Brain Korea 21 Plus Project for Medical Science, Yonsei University College of Medicine, Seoul, Republic of Korea. 4. Severance Biomedical Science Institute, Yonsei University College of Medicine, Seoul, Republic of Korea. 5. Department of Research Support, Yonsei Biomedical Research Institute, Yonsei University College of Medicine, Seoul, Republic of Korea. 6. JEUK Institute for Cancer Research, JEUK Co., Ltd., Gumi-City, Republic of Korea. 7. Wide River Institute of Immunology (WRII) Seoul National University, Hongcheon-gun, Gangwon-do 250-812, Republic of Korea. 8. Division of Medical Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Republic of Korea; Severance Biomedical Science Institute, Yonsei University College of Medicine, Seoul, Republic of Korea. Electronic address: pkhpsh@yuhs.ac. 9. Division of Medical Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Republic of Korea. Electronic address: nobelg@yuhs.ac.
Abstract
OBJECTIVE: Anti-programmed death (PD)-1 therapy confers sustainable clinical benefits for patients with non-small-cell lung cancer (NSCLC), but only some patients respond to the treatment. Various clinical characteristics, including the PD-ligand 1 (PD-L1) level, are related to the anti-PD-1 response; however, none of these can independently serve as predictive biomarkers. Herein, we established a machine learning (ML)-based clinical decision support algorithm to predict the anti-PD-1 response by comprehensively combining the clinical information. MATERIALS AND METHODS: We collected clinical data, including patient characteristics, mutations and laboratory findings, from the electronic medical records of 142 patients with NSCLC treated with anti-PD-1 therapy; these were analysed for the clinical outcome as the discovery set. Nineteen clinically meaningful features were used in supervised ML algorithms, including LightGBM, XGBoost, multilayer neural network, ridge regression and linear discriminant analysis, to predict anti-PD-1 responses. Based on each ML algorithm's prediction performance, the optimal ML was selected and validated in an independent validation set of PD-1 inhibitor-treated patients. RESULTS: Several factors, including PD-L1 expression, tumour burden and neutrophil-to-lymphocyte ratio, could independently predict the anti-PD-1 response in the discovery set. ML platforms based on the LightGBM algorithm using 19 clinical features showed more significant prediction performance (area under the curve [AUC] 0.788) than on individual clinical features and traditional multivariate logistic regression (AUC 0.759). CONCLUSION: Collectively, our LightGBM algorithm offers a clinical decision support model to predict the anti-PD-1 response in patients with NSCLC.
OBJECTIVE: Anti-programmed death (PD)-1 therapy confers sustainable clinical benefits for patients with non-small-cell lung cancer (NSCLC), but only some patients respond to the treatment. Various clinical characteristics, including the PD-ligand 1 (PD-L1) level, are related to the anti-PD-1 response; however, none of these can independently serve as predictive biomarkers. Herein, we established a machine learning (ML)-based clinical decision support algorithm to predict the anti-PD-1 response by comprehensively combining the clinical information. MATERIALS AND METHODS: We collected clinical data, including patient characteristics, mutations and laboratory findings, from the electronic medical records of 142 patients with NSCLC treated with anti-PD-1 therapy; these were analysed for the clinical outcome as the discovery set. Nineteen clinically meaningful features were used in supervised ML algorithms, including LightGBM, XGBoost, multilayer neural network, ridge regression and linear discriminant analysis, to predict anti-PD-1 responses. Based on each ML algorithm's prediction performance, the optimal ML was selected and validated in an independent validation set of PD-1 inhibitor-treated patients. RESULTS: Several factors, including PD-L1 expression, tumour burden and neutrophil-to-lymphocyte ratio, could independently predict the anti-PD-1 response in the discovery set. ML platforms based on the LightGBM algorithm using 19 clinical features showed more significant prediction performance (area under the curve [AUC] 0.788) than on individual clinical features and traditional multivariate logistic regression (AUC 0.759). CONCLUSION: Collectively, our LightGBM algorithm offers a clinical decision support model to predict the anti-PD-1 response in patients with NSCLC.