Despite the proven efficacy of immune checkpoint inhibitors (ICIs) in treating patients with non–small-cell lung cancer (NSCLC), only a subset of patients achieves the desired treatment goals.1 To date, identifying reliable and robust biomarkers to predict the benefit of ICI treatment remains challenging.2 The objectives of this study were to evaluate the ability to predict response to ICI therapy by integrating medical imaging, histopathologic, and genomic characteristics to develop a multimodal biomarker for immunotherapy response.
The investigators used baseline data from a diagnostic clinical workup at a single center.2 The multimodal data set included DNA alterations from next-generation sequencing, computed tomography scan images, and digitized PD-L1 immunohistochemistry (IHC). A workflow was designed to extract data for each patient and used an attention-gated machine learning approach to integrate the features into a risk prediction model.
The study included 247 patients with advanced NSCLC who received immunotherapy and completed radiology, pathology, genomics, and clinical evaluations. The patient cohort was 54% female with a median age of 68 years (range, 38-93 years), and 88% of patients had a smoking history.
A radiomics approach was used, and the average individual lesion predictions were aggregated to construct patient-level response predictions, which resulted in an overall AUC (area under the curve) = 0.65 (95% confidence interval [CI], 0.57-0.73), where AUC = 0.0 is 100% wrong and AUC = 1.0 is 100% correct. In addition, formalin-fixed, paraffin-embedded tissue slides of pretreatment PD-L1 IHC staining of tumor specimens were digitized. Overall, 52% of slides showed a PD-L1 tumor proportion score (TPS) ≥1% and were used to extract IHC-texture, a novel spatial characterization of PD-L1 staining. Logistic regression modeling on IHC-texture resulted in prediction accuracy of AUC = 0.62 (95% CI, 0.51-0.73), which was inferior to the pathologist-assessed PD-L1 TPS (AUC = 0.73; 95% CI, 0.65-0.81).
A dynamic, deep attention-based, multiple-instance learning model was implemented with masking to evaluate the impact of combining features from all modalities. The multimodal model (AUC = 0.80; 95% CI, 0.74-0.86) was superior to unimodal measures, including tumor mutational burden (AUC = 0.61; 95% CI, 0.52-0.70) and PD-L1 TPS (AUC = 0.73; 95% CI, 0.65-0.81).
In conclusion, this proof-of-concept study demonstrated that multimodal approaches using expert-guided machine learning are superior to conventional unimodal strategies in predicting response to immunotherapy in patients with NSCLC.
Learn more about our family of publications.
View Our Publications