AI Model Has Promise for Predicting Checkpoint Inhibitor Activity in NSCLC

Artificial intelligence (AI)-aided analysis of pathology slides showed potential as a predictive biomarker for non-small cell lung cancer (NSCLC) response to immune checkpoint inhibitors (ICIs), a multicenter study showed.

The deep-learning model (called Deep-IO) achieved an area under the receiver operating characteristic curve (AUC) of 0.75 for objective response in a U.S.-based developmental cohort and 0.66 in a Europe-based validation cohort. By multivariable analysis of the validation cohort results, the model’s score proved to be an independent predictor of ICI response for progression-free (PFS) and overall survival (OS).

The tuned model outperformed tumor mutational burden (TMB), tumor-infiltrating lymphocytes (TILs), and PD-L1 expression in the developmental set, and proved superior to TILs and comparable to PD-L1 in the validation cohort. Combining the model with PD-L1 expression improved specificity by 10 percentage points, outperforming either marker alone, reported Mehrdad Rakaee, PhD, of Oslo University Hospital in Norway, and co-authors in JAMA Oncology.

“The deep learning model has the capability to predict ICI responses directly from a single image of [a hematoxylin and eosin] stained slide,” the authors concluded. “This analysis could serve as an auxiliary biomarker alongside PD-L1 immunohistochemistry for advanced NSCLC, potentially enhancing patient stratification and improving selection of tailored therapy for each patient while optimizing the benefit-cost balance in ICI treatment.

“Further validation of the clinical utility of Deep-IO or a similar method for predicting response to various treatment regimens in NSCLC will be of interest.”

The study reflects the current direction of oncology and use of ICIs, said Roy Herbst, MD, PhD of Yale Cancer Center in New Haven, Connecticut.

“This is exactly where we need to go to start personalizing therapy: who to treat, how long to treat,” he told MedPage Today. “It’s a good first step and it generates a path for further study.”

Separate studies of the model will likely have to be conducted in different types of cancer to define its utility.

“Renal cancer might behave differently from lung or melanoma,” Herbst said. “That’s just my gut sense, without any data to back it up.”

Multiple studies have shown that patients with advanced/metastatic NSCLC benefit from treatment with ICIs, but only 25-30% of patients respond. The primary predictive marker, PD-L1 expression, is an imperfect measure, given that even patients with low levels of positivity benefit from ICI therapy, the authors noted in their introduction.

In 2020, the FDA approved tissue-derived TMB as a predictive marker of ICI response in various solid tumors, including NSCLC. However, use of TMB has been limited by a variety of factors, including cost, assay variability, optimal cutoff values, and limited sensitivity and specificity, Rakaee and colleagues continued. At the same time, interest has grown in identifying additional biomarkers of response to ICIs.

Advances in AI have facilitated analysis of digital pathologic images. The authors have developed several AI-based pathologic classification systems to identify immune phenotypes, TILs, and tertiary lymphoid structures from standard histologic digital images.

Immune biomarkers derived from machine learning have shown an association with response to ICIs and overall survival in NSCLC and melanoma. Additionally, immune biomarkers have been associated with recurrence risk in early lung cancer.

“Deep learning models have been developed by several groups to interpret complex spatial patterns in histologic images and to predict factors such as survival and genomic alterations, at a level of sophistication beyond that of most human experts,” the authors stated. “The capability of deep learning to fully analyze image features without prior constraint or bias, enables comprehensive assessment of many histopathologic patterns, potentially leading to more accurate predictions of clinical outcomes.”

Continuing their investigation of AI-derived biomarkers, Rakaee and colleagues developed a response stratification model to predict ICI efficacy from digital images of pathology specimens from patients with advanced NSCLC. They performed internal and external validation studies and compared results with PD-L1, TMB, and TILs for predicting response to ICI treatment.

The Deep-IO model was designed to predict responses to ICI monotherapy directly from histologic images. The model was trained on the basis of objective response rate to ICI as defined by RECIST criteria.

The analytic exercise involved 295,581 image tiles from 958 patients with advanced NSCLC, 614 from the U.S., and 344 from Europe. The objective response rate to ICIs was 26% in the U.S.-based development cohort and 28% in the European validation cohort.

Deep-IO probability scores were categorized into median and tertile cutoffs from the validation cohort for the survival analysis. The results showed that a higher Deep-IO score was associated with significantly longer PFS (6.2 vs 3.0 months P0.001) and OS (13.7 vs 8.9, P

In a multivariable analysis of the validation cohort, the machine model’s predictive score for ICI response proved to be an independent predictor of PFS (HR 0.56, 95% CI 0.42-0.76, PP

An analysis of biomarkers’ accuracy for predicting response to ICI showed that Deep-IO had the highest AUC in the test set (0.75 vs 0.57-0.70 for others) and was comparable to PD-L1 in the validation cohort (0.66 vs 0.67), but had a 10% higher specificity for identifying nonresponders. The model also had the highest sensitivity in the test set (0.91 vs 0.54-0.83). Combining Deep-IO and PD-L1 improved response classification and produced the highest positive predictive value and negative predictive value as compared with individual biomarkers.