News
Article
Novel Deep-Learning Model May Predict Immune Checkpoint Inhibitor Responses in Patients With Non−Small Cell Lung Cancer
Author(s):
Key Takeaways
- Deep-IO AI model predicts ICI response in NSCLC, outperforming traditional biomarkers like PD-L1, TMB, and TILs.
- The model demonstrated superior predictive accuracy, with higher sensitivity and specificity, enhancing patient stratification.
The model, Deep-IO, shows improved specificity and sensitivity compared with existing predictive biomarkers.
Researchers developed an AI model capable of predicting response to immune checkpoint inhibitors (ICIs) in patients with advanced non–small cell lung cancer (NSCLC), according to data published in JAMA Oncology. Use of the model, called Deep-IO, could help health care professionals enhance treatment precision and improve patient selection for different therapies.1
Immune checkpoint inhibitors blocking cancer cells | Image Credit: © Papisut - stock.adobe.com
Prior studies and clinical trials showed the efficacy, safety, and benefit of ICIs in treatment of patients with advanced or metastatic NSCLC; however, only 25% to 30% of patients respond to treatment. Some patients with lower PD-L1 levels, a predictive biomarker for ICI monotherapy response, do respond to treatment, but many patients with high levels do not. Although the FDA did approve tissue-derived tumor mutational burden (TMB) as a predictive biomarker for ICI in various solid tumors, TMB is costly with limited sensitivity and specificity. This underscores the need for improved predictive tools and the development of deep learning models.1,2
The study compared the predictive capabilities of the Deep-IO model to PD-L1 biomarkers, TMB, and tumor-infiltrating lymphocytes (TILs). To develop and validate this model, the researchers analyzed a total of 295,581 whole-slide hematoxylin and eosin–stained images from 958 patients (mean [SD] age, 66.0 [10.6] years; 456 [48%] females and 502 [52%] males) treated with ICI for NSCLC. The participants were divided into a United States-based developmental cohort (n=614) and a Europe-based validation cohort (n=344). The researchers measured performance of the model based on clinical end points and objective response rate (ORR).1,2
"The deep learning model has the capability to predict ICI responses directly from a single image of [a hematoxylin and eosin] stained slide," the researchers concluded, according to MedPage Today. "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.”2
According to the data, patients in the developmental cohort achieved an ORR to ICI of 26% compared with 28% in the validation group. The researchers also reported an area under the receiver operating characteristic curve (AUC) for ORR of 0.75 (95% CI, 0.64-0.85) and 0.66 (95% CI, 0.60-0.72) in the internal test sets and validation groups, respectively.1,2
In a survival analysis, the researchers categorized the model’s probability scores into median and tertile cutoffs from the validation cohort. They reported that higher Deep-IO scores are associated with significantly longer progression-free survival (PFS; 6.2 vs 3.0 months P<.001) and overall survival (OS; 13.7 vs 8.9, P<.001). Additionally, a multivariable analysis showed the learning model’s score was an independent predictor of ICI response in the validation cohort for both PFS (hazard ratio [HR], 0.56; 95% CI, 0.42-0.76; P < .001) and OS (HR, 0.53; 95% CI, 0.39-0.73; P < .001).1,2
The model outperformed TMB, TILs, and PD-L1 in the internal set and had the highest AUC (0.75 vs 0.57-0.70). In the validation cohort, it outperformed TILs and was comparable to PD-L1 (AUC, 0.67; 95% CI, 0.60-0.74), with a 10% improvement in specificity. Additionally, the model had the highest sensitivity in the test set (0.91 vs 0.54-0.83). Combining the deep learning model with PD-L1 scores in the validation cohort resulted in an AUC of 0.70 (95% CI, 0.63-0.76), surpassing each marker individually and showing a higher response rate (51% vs. 41% for PD-L1 ≥50%).1,2
Unlike traditional biomarkers such as PD-L1, TMB, and TILs, the Deep-IO model demonstrated superior predictive capabilities, offering a more accessible and cost-effective alternative. Its ability to predict ICI response based on a single image could serve as an important complementary biomarker, potentially improving patient stratification and optimizing therapeutic decisions.
