AI-Powered Prognostic Index Predicts Immunotherapy Response in NSCLC Using PD-L1 and Clinical Data

Background: FDA-approved immunohistochemical<u> </u>assessment of programmed cell death ligand 1 (PD-L1) expression serves as a companion diagnostic to select patients for immune checkpoint inhibitor (ICI) therapy and predict responses to anti-PD-1/PD-L1 treatment in non-small cell lung cancer (NSCLC). However, the predictive and prognostic value of PD-L1 expression is debatable due to inconsistencies across tumor types and clinical settings. This variability, along with tumor heterogeneity and differences in observer interpretation, reduces the accuracy of patient selection for ICI and decreases treatment efficacy. In this academic-industry collaboration, we compared an AI-powered PD-L1 quantification algorithm (AIM-PD-L1, PathAI) to pathologist scoring and evaluated their association with ICI treatment outcomes and survival. We also developed a novel prognostic index for NSCLC patients by incorporating PD-L1 assessment with clinical and pathological features. Design: Digitized slides from an institutional cohort of 692 NSCLC patients (of whom 310 were ICI-treated and 230 had treatment response data) were stained for PD-L1 (DAKO 22C3, PharmDx) at LabCorp. Manual tumor proportional scores, categorized as <1% (negative), 1-49% (low), or >=50% (high), were then compared to AI-generated quantitative scores. Survival probability and Cox proportional hazards were assessed using Kaplan-Meier analysis and uni-/multivariable Cox regression. Results: Although AIM-PD-L1-derived scores and pathologist assessments were highly correlated (r=0.87, p<0.05), only AIM-PD-L1 predicted clinical endpoints. Specifically, AIM-PD-L1 scores were significantly associated with improved overall survival (OS) in both univariate (HR=0.52, p=0.003) and multivariate analyses (HR=0.59, p=0.017), unlike pathologist assessments. This predictive strength extended to early-stage (AJCC I/II) NSCLC, where AIM-PD-L1 positive scores in multivariate analysis also correlated with improved OS (HR 0.48, p=0.043). These insights guided the development of a new prognostic index which better assesses true risk of shorter survival in the patients undergoing immunotherapy. Cox regression analysis showed that each unit increase in the index was associated with significantly poorer survival outcomes (HR=1.34, p<0.001), while achieving a mean concordance index of 0.67 in 5-fold internal cross-validation. Conclusion: Our study provides compelling evidence that AIM-PD-L1 holds significant potential to enhance the accuracy and objectivity of PD-L1 assessment in NSCLC, offering the possibility of improved personalized prognostication and treatment decisions compared to traditional approaches.

Learning Objectives: 

1. Understand the limitations of traditional PD-L1 assessment and the need for improved prognostic tools in NSCLC.
2. Learn how AI-powered PD-L1 quantification enhances prognostic accuracy over manual scoring.
3. Be introduced to the development and performance of a novel AI-integrated prognostic index for immunotherapy in NSCLC.