External validation of a deep learning CT biomarker to predict first-line immune checkpoint inhibitor monotherapy-associated survival in PD-L1–high metastatic non–small cell lung cancer.

Abstract

2534 Background: Immune checkpoint inhibitor (ICI) monotherapy is a standard first-line treatment for patients with metastatic non–small cell lung cancer (mNSCLC) with high PD-L1 expression. However, 60% of patients treated with ICI monotherapy progress within 1 year of treatment. Imaging-based biomarkers from routine pretreatment computed tomography (CT) scans may provide a noninvasive approach to refine patient selection and guide treatment decisions. Methods: Enhanced CT Response Score (eCTRS v0.0.2; Sako et al, JCO CCI, 2024) is an imaging-based biomarker that stratifies patient survival among mNSCLC patients receiving ICI monotherapy. eCTRS is derived from pre-treatment CT scans, using deep-learning extracted imaging features, lesion features, and clinical variables of age and sex. eCTRS was previously trained on a diverse, real-world multi-institutional dataset of 1,058 mNSCLC patients. This retrospective study externally validated eCTRS using a deidentified, EHR-derived longitudinal database with imaging from Flatiron Health. Patients with PD-L1–high (PD-L1 tumor proportion score ≥ 50%) mNSCLC without actionable mutations who received first-line ICI monotherapy and had imaging from 12 weeks before to 2 weeks after treatment start were included. Patients were stratified into eCTRS High and eCTRS Low groups using a pre-determined threshold. Survival analyses were conducted by an external, independent group. Kaplan-Meier and Cox proportional hazards analyses evaluated progression-free survival (PFS) and overall survival (OS) stratification. PFS was derived from RECIST 1.1 assessments by a centralized multi-reader radiologist adjudication process, and OS was defined as time from treatment start to death by any cause. Results: 205 patients met all inclusion criteria (median age 72 years, 48% female, 20% non-White). Eighty (39%) patients were classified as eCTRS Low and 125 (61%) as eCTRS High. PFS was improved among eCTRS High patients (Hazard Ratio [HR], 0.71; 95% CI, 0.51–1.00; p=0.048), with median PFS of 231 (95% CI: 133, 350) days versus 88 (95% CI: 57, 179) days in eCTRS Low patients. eCTRS High patients demonstrated significantly improved OS (HR, 0.56; 95% CI, 0.39–0.80; p=0.001), with median OS of 484 (95% CI: 361, NA) days for eCTRS High, versus 155 (95% CI: 75, 295) days for eCTRS Low. Conclusions: A deep learning–based imaging biomarker derived from routinely acquired pretreatment CT imaging identified survival benefit in patients with PD-L1–high metastatic NSCLC treated with first-line ICI monotherapy. These findings suggest that imaging-based biomarkers may serve as complementary, noninvasive tools to identify patients most likely to derive benefit from ICI monotherapy and to support risk-adapted treatment strategies.