Virtual Library
Start Your Search
Hiroshi Yoshida
Author of
-
+
FP07 - Pathology (ID 109)
- Event: WCLC 2020
- Type: Posters (Featured)
- Track: Pathology, Molecular Pathology and Diagnostic Biomarkers
- Presentations: 1
- Moderators:
- Coordinates: 1/28/2021, 00:00 - 00:00, ePoster Hall
-
+
FP07.04 - Predictive Efficacy of Morphological Biomarkers Based on Digital Pathology for ICI Therapy of Non-Small Cell Lung Cancer (ID 3436)
00:00 - 00:00 | Author(s): Hiroshi Yoshida
- Abstract
Introduction
PD-L1 IHC is a widely used biomarker for immune checkpoint inhibitor therapy (ICI), but it does not have high specificity. It is essential to establish more accurate biomarkers for modern medicine. Our previous preliminary study (presented at 2019 WCLC) indicated the morphological feature's substantial value as a biomarker for ICI therapy. The morphological biomarkers (MBM) using digital whole-slide images can be tested from archived FFPE specimens. Here, we report our study on the prediction potency of morphological biomarkers for lung cancer patients treated by anti-PD1 inhibitors.
Methods
255 NSCLC who received ICI therapy were recruited. Digital images of H&E and PD-L1 (22C3) IHC stained slides of pre-treatment biopsied or resected materials were examined by previously reported image analysis techniques (NEC, Japan). The morphological characteristics of cancer cells were also evaluated by the pathologist's eyeball (PPI, pathological prediction index score 1-3). PD-L1 IHC (22C3) and tumor mutation burden (TMB) by the NGS-based target sequence (NCC oncopanel ®) were examined. Using morphological characteristics of HE images, we build a prediction model using the decision tree method (MBM-DT) first, then applied a deep learning framework (MBM-DNN) to response prediction. We compared the prediction potency for the ICI-therapeutic response of each score. The relative area proportion of the seven-index spider plot using test quality indicators was measured. The logistic regression test was calculated (SPSS). A p-value of less than 0.05 was defined as statistically significant.
Results
Pathological prediction index (PPI) scores showed superior prediction ability to that of the PD-L1 test. The accuracy of PPI-score 3, PPI-score 2+3, PD-L1 with 1% cutoff (PD-L1-1%), PD-L1 with 50% cutoff (PD-L1-50%), TMB-high, and driver gene mutation-negative was 0.71, 0.64, 0.60, 0.66, 0.57 and 0.57, respectively. MBM_DNN had the highest accuracy, specificity, positive prediction value (PPV), and the lowest false-negative rate (FNR) among the tested biomarkers. PPI-(2+3) had the highest sensitivity and negative prediction value (NPV), and false-negative rate (FNR). The seven-index spider plot showed the superiority of PPI-score 3, PPI-score 2+3 to PD-L1. Both PPI and MBM showed superior accuracy to PD-L1, TMB, and gene mutation status. The area proportion of seven test-index spider plots was 50, 56, 45 for PPI-score 3, MBM_DNN, and PD-L1-50%. Accuracy of training, validation, and a test set of MBM-DNN resulted in 0.85, 0.61, and 0.64, respectively. The logistic regression analysis revealed that males, smokers, the absence of driver gene mutation, positive/high expression of PD-L1, high PPI-score, the positive MBM-DNN, and MBM-DT are likely to be non-responder by univariate analysis. PPI-score 2+3(0.31, <0.001), MBM-DNN (0.23, <0.001), and MBM-DT (0.32, <0.001) are the significant factors for prediction of ICI response, but others are not by multivariate analysis. Conclusion
Our results showed the superior value of the morphological biomarker for ICI response prediction, compared to PD-L1 IHC and TMB. The morphological biomarker can be a useful biomarker for clinical therapeutic decisions.
