Fusion of Clinical and Deep Learning Features for Predicting Pembrolizumab Monotherapy Response in Advanced Non-Small Cell Lung Cancer.
Authors
Affiliations (7)
Affiliations (7)
- Basic and Translational Research, BC Cancer Research Institute, 675 West 10th Avenue, Vancouver, BC V5Z Il3, Canada.
- Department of Pathology, Faculty of Medicine, University of British Columbia, 2329 West Mall, Vancouver, BC V6T IZ4, Canada.
- Interdisciplinary Oncology Program, Faculty of Medicine, University of British Columbia, 2329 West Mall, Vancouver, BC V6T IZ4, Canada.
- BC Cancer, Vancouver Center, 600 West 10th Avenue, Vancouver, BC V5Z 4E6, Canada.
- Department of Medical Oncology, Faculty of Medicine, University of British Columbia, 2329 West Mall, Vancouver, BC V6T IZ4, Canada.
- Department of Respirology, Faculty of Medicine, University of British Columbia, 2329 West Mall, Vancouver, BC V6T IZ4, Canada.
- Department of Radiology, Faculty of Medicine, University of British Columbia, 2329 West Mall, Vancouver, BC V6T IZ4, Canada.
Abstract
<b>Objective</b>: Pembrolizumab monotherapy is an anti-PD-1 immunotherapy that is approved as a first-line treatment for non-small cell lung cancer (NSCLC) patients with high PD-L1 expression (≥50%). However, approximately 55% of these patients do not respond. Early identification of likely non-responders is critical to enable timely transition to alternative treatments. <b>Materials</b>: This study analyzed a retrospective cohort of NSCLC patients treated with first-line PD-L1 monotherapy, divided into a discovery training set (<i>n</i>: 97; 27 non-responders) and a preliminary test set (<i>n</i>: 17; 9 non-responders). Treatment response was assessed using baseline and follow-up CT scans in accordance with the response evaluation criteria in solid tumors (RECIST v1.1). <b>Methods</b>: Our objective was to extract deep learning (DL) features from the two groups of patients and apply transfer learning techniques to identify patients at risk of progression on pembrolizumab monotherapy. A nonparametric statistical test (Mann-Whitney U) was employed to rank the discriminative power of the 128 features from these training groups. Two types of support vector machine (SVM-RBF and SVM-Polynomial) classifiers were employed to investigate the discriminating power of the highest-ranked features as measured by F1 score and AUC values over ROC curves at the three levels of the data (slice, lesion, and patient) with and without clinical descriptors. <b>Results</b>: SVM-RBF performed best when trained on the 10 highest-ranked DL features and five clinical descriptors, achieving AUC of 0.742 (CI 95% 0.47-1.00), SN of 88.9%, SP of 75% and F1 score of 84.2% on preliminary test set patients, whereas an AUC of 0.902 ± 0.031, SN of 81.5%, SP of 81.4% and F1 score of 71% were observed for the discovery training set. <b>Conclusions</b>: Integrating CT-based DL features with clinical descriptors demonstrated balanced performance, offering a promising tool to identify patients at risk of progression on pembrolizumab monotherapy to support first-line treatment decisions in PD-L1-high NSCLC.