Machine learning based on blood test biomarkers predicts fast progression in advanced NSCLC patients treated with immunotherapy

Abstract

Objective Fast progression (FP) represents a desperate situation for advanced non-small cell lung cancer (NSCLC) patients undergoing immune checkpoint inhibitor therapy. We aimed to develop a predictive framework based on machine learning (ML) methods to identify FP in advanced NSCLC patients using blood test biomarkers. Methods and analysis We extracted data of 1546 atezolizumab-treated patients from four multicentre clinical trials. In this study, patients from the OAK trial were taken for model training, whereas patients from the other trials were used for independent validations. The FP prediction model was developed using 21 pretreatment blood test variables in seven ML approaches. Prediction performance was evaluated by the receiver operating characteristic (ROC) curve. Results The prevalence of FP was 7.6% (118 of 1546) in all atezolizumab-treated patients. The most important variables for the prediction model were: C reactive protein, neutrophil count, lactate dehydrogenase and alanine transaminase. The Support Vector Machine (SVM) algorithm applied to these four blood test parameters demonstrated good performance: the area under the ROC curve obtained from the training cohort (OAK), validation cohort 1 (BIRCH) and cohort 2 (merged POPLAR and FIR) were 0.908, 0.666 and 0.776, respectively. In addition, the absolute difference in median survival between the SVM-predicted FP and non-FP groups was significant in both progression-free survival and overall survival (p<0.001). Conclusion SVM trained using a 4-biomarker panel has good performance in predicting the occurrence of FP regardless of programmed cell death ligand 1 expression, hence providing evidence for decision-making in single-agent atezolizumab immunotherapy for patients with advanced NSCLC. Data are available upon reasonable request. Qualified researchers may request access to individual patient level data through the clinical study data request platform (Vivli, [https://vivli.org][1]/). For further details please refer to Roche’s Global Policy on the Sharing of Clinical Information and how to request access to related clinical study documents; see and. Our source codes for the prediction of FP are available at. [1]: https://vivli.org/