A Deep Learning Model for Predicting Immune Checkpoint Inhibitor Efficacy in NSCLC
Researchers developed and validated a deep learning model, Deep-IO, designed to predict the efficacy of immune checkpoint inhibitors (ICIs) in treating advanced non-small cell lung cancer (NSCLC), according to a study published in JAMA Oncology. The model analyzed histologic images from pathology specimens to predict ICI response based on objective response rate (ORR), classifying patients into responders or nonresponders. The model was trained using data from a cohort at the Dana-Farber Cancer Institute (DFCI) and tested across 3 European Union centers. In both the developmental cohort (958 patients) and validation cohort (344 patients), Deep-IO demonstrated high performance, with an F1-score of 0.71, recall of 0.70, and precision of 0.73 in the test set, and similar results in the validation cohort.
Deep-IO's predictions were aligned with key clinical outcomes such as progression-free survival (PFS) and overall survival (OS), with higher Deep-IO scores correlating with better survival outcomes. When compared to established biomarkers like PD-L1 expression, tumor mutational burden (TMB), and tumor-infiltrating lymphocytes (TILs), Deep-IO outperformed TILs and TMB, and showed comparable performance to PD-L1 in predicting ICI response. Combining Deep-IO with PD-L1 further enhanced predictive accuracy. Subgroup analyses revealed that Deep-IO was particularly effective in patients with high or moderate PD-L1 expression and performed better in lung adenocarcinoma cases compared with squamous cell carcinoma.
The study also explored the clinical utility and explainability of Deep-IO. In multivariable analysis, Deep-IO emerged as an independent predictor of both PFS and OS. To improve the interpretability of the model, Gradient-weighted Class Activation Mapping (GradCAM) was used, highlighting regions of interest in the histologic images. The model focused on tumor epithelial areas and inflammatory regions. Overall, Deep-IO demonstrated the potential to be a valuable tool for predicting ICI response, offering a novel, image-based approach that could complement existing biomarkers and contribute to more personalized treatment decisions in cancer therapy, according to the researchers.
Reference
Rakaee M, Tafavvoghi M, Ricciuti B, et al. Deep learning model for predicting immunotherapy response in advanced non−small cell lung cancer. JAMA Oncol. 2024. doi:10.1001/jamaoncol.2024.5356
