WCLC 2025 - Posters & ePosters-P1.17.54 Long-Term Outcomes From Pembrolizumab in Patients With Advanced PD-L1 =50% NSCLC and Poor PS: A Transformer-Based AI Approach

P1.17.54 Long-Term Outcomes From Pembrolizumab in Patients With Advanced PD-L1 =50% NSCLC and Poor PS: A Transformer-Based AI Approach

Back to course

Pdf Summary

This study investigates long-term outcomes of pembrolizumab monotherapy in a challenging population: patients with advanced non-small cell lung cancer (NSCLC), PD-L1 tumor proportion score (TPS) ≥50%, and poor performance status (ECOG PS 2). Using real-world data from the global Pembro-Real 5Y registry comprising 1,050 patients (161 with ECOG PS 2), the authors analyzed survival over five years. Median overall survival (OS) for patients with PS 2 was 5.4 months, but notably, 13% survived at least five years, indicating a subset achieves durable benefit despite frailty.<br /><br />Traditional statistical and AI approaches were used to identify prognostic factors. Elastic Net regression highlighted high tumor mutational burden (TMB) and KRAS mutations as significant predictors of 5-year survival, though estimations were unstable due to data limitations. A transformer-based AI model (NAIM) was applied to capture complex, time-dependent interactions among variables. NAIM found that bone metastases and baseline corticosteroid use strongly predicted early mortality, while factors like body mass index (BMI) increase and systemic health markers (e.g., hypertension, dyslipidemia) correlated with long-term survival. However, NAIM suffered from overfitting, limiting its predictive robustness.<br /><br />The findings reaffirm that poor performance status alone should not exclude patients from receiving pembrolizumab first-line immunotherapy, as some can achieve long-term survival. They also underscore difficulties in predicting long-term outcomes in this heterogeneous group due to limitations of static baseline variables and modeling approaches. This real-world evidence supports personalized treatment considerations beyond performance status and highlights the need for improved predictive models incorporating dynamic and multifactorial patient data.<br /><br />Overall, the study demonstrates the prognostic complexity in advanced NSCLC with poor PS and suggests that machine learning models may provide insights but require refinement to be clinically reliable for long-term survival prediction in immunotherapy-treated patients.

Asset Subtitle

Alessio Cortellini

Meta Tag

Speaker Alessio Cortellini

Topic Global Health, Health Services, and Health Economics

Keywords

pembrolizumab monotherapy

advanced non-small cell lung cancer

NSCLC

PD-L1 tumor proportion score ≥50%

poor performance status ECOG PS 2

real-world data

long-term survival

tumor mutational burden

KRAS mutations

machine learning prognostic models