Abstract 683: A machine learning based clinical decision support system (CDSS) for anti-PD-1 therapy using non-invasive blood marker and clinical information for lung cancer patients

Abstract

Abstract Cancer immunotherapy with anti-PD-1 therapy (αPD-1) has greatly improved the survival outcome of patients with non-small cell lung cancer (NSCLC). However, the response rates of αPD-1 are around 20-30% based on clinical trials. The presence of PD-L1, tumor infiltrating lymphocyte or tumor mutation burden may be used as indicators of response, but are still limited to predict αPD-1 response. In this study, we established the machine learning based clinical decision support system (CDSS) to predict the αPD-1 response by comprehensively combining clinical information and blood-based data which are easily assessable in routine practice.We enrolled 126 patients with NSCLC treated with the αPD-1 at Yonsei cancer center. Clinical data including patient characteristics, mutation, treatment outcomes, and adverse events were collected and analyzed. Forty patients additionally had blood-based immune data by flow cytometry. There were two data sets; clinical data set (n=126) with 15 variables, and immune data set (n=40) with 37 variables. We found that 27 variables out of 52 variables are selected by recursive feature elimination. The responders are defined as PR or SD ≥6 months and the non-responders are defined as the others. Supervised learning algorithms such as the LASSO, Ridge, Elastic Net, SVM, ANN, and RF were applied to each data set for predicting the αPD-1 response. The performances of each model were evaluated according to the ROC curve or cross-validation errors. Variable importance was measured by using the random forest and gradient boosting.Patient characteristics included male (69.8%), age ≥ 60 years (66.2%), ECOG 0/1 (77.7%), adenocarcinoma (69.8%), EGFR mutations (15.1%), and PD-L1 positive (61.2%). We classified the patients into responders (38%) and non-responders (62%) in total 126 patients. A result from the clinical data set of 126 patients demonstrated that the Ridge regression model (AUC: 0.78) can more accurately predict the αPD-1 response than others. Of 15 clinical variables, some are considered to be important in the following order; tumor burden,age, PD-L1, ECOG PS, and irAE based on the random forest. When we performed the machine learning process with clinical and immune data, The Ridge regression model (AUC:0.82) showed the good performance to predict αPD-1 response, compared to the single clinical model. The machine learning based CDSS for aPD-1 to NSCLC patients has benefit for predicting aPD-1 responses. Our prediction model could be easily accessible and fast processed in routine practice. The supervised machine learning based non-invasive predictive score (NIPS) demonstrates the rate of aPD-1 response on NSCLC patients. We will validate NIPS in independent patient cohort and currently are establishing the NIPS as web-based software. Citation Format: Kyoung-Ho Pyo, Beung-Chul Ahn, Chun-Feng Xin, Dongmin Jung, Chang Gon Kim, Min Hee Hong, Byoung Chul Cho, Hye Ryun Kim. A machine learning based clinical decision support system (CDSS) for anti-PD-1 therapy using non-invasive blood marker and clinical information for lung cancer patients [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 683.