Use of Al-based models integrating serum proteomics and clinicopathologic features to predict survival outcomes in NSCLC patients undergoing immunotherapy.

Abstract

e14577 Background: Responses to immune checkpoint inhibitors in patients with advanced NSCLC present a wide range of outcomes. Integrating serum proteomics with clinicopathologic data offers a promising approach to personalized outcome prediction. Methods: We retrospectively analyzed 148 consecutive patients with advanced-stage NSCLC who received either ICI monotherapy or ICI combined with chemotherapy as frontline treatment. Demographic and disease-related parameters (age, sex, smoking history, stage, performance status, tumor mutational burden, PD-L1 expression, and treatment regimen) were collected. Serum proteomic profiles were obtained via mass spectrometry, generating high-dimensional intensity data. We performed variance-based feature selection, imputed and scaled data, extracted peak features, and applied dimensionality reduction before training a Naive Bayes model for 1-year survival prediction. To enhance performance, we applied feature engineering and utilized a nested 5-fold cross-validation framework. Naive Bayes classifier machine learning model was used to predict 1-year overall survival and progression-free survival rate. Results: The median age was 70.6 years (IQR: 63.1-76.0), and 52.0% were female. 85.8% of patients were current or former smokers, and 14.2% never smoked. Cancer types were distributed with 68.2% adenocarcinoma and 27.7% squamous cell carcinoma. Most patients (75.0%) had a performance status of 0–1. The majority (82.1%) presented with stage IV disease. The median TMB was 5.4 mut/Mb (IQR: 3.2–7.8). PD-L1 expression was ≥1% in 56.8%. Treatment consisted of ICI monotherapy (35.8%), dual ICI (14.2%), or ICI plus chemotherapy (50.0%). The median OS was 19.5 months (95% CI: 16.1–22.7), and the median PFS was 7.1 months (95% CI: 5.3–9.2). For the 1-year OS prediction model, the area under the curve (AUC) was 0.61 for proteomic-only data and 0.67 for combined clinicopathologic and proteomic data, respectively. For the 1-year PFS prediction model, the AUC were 0.62 for proteomic-only data and 0.69 for combined data, respectively. The AI-based models integrating serum proteomic and clinicopathologic factors demonstrated higher predictive power than ones using proteomic data alone for both OS and PFS. Additionally, risk stratification based on serum proteomic features revealed two groups (low vs. high risk) with differential survival outcomes mOS 15.6 mo, 9.1 mo, HR = 1.71(1.21–2.24), p < 0.05, mPFS 10.2 mo, 3.7 mo, HR = 2.76(1.14–4.42), p < 0.05, respectively. Conclusions: An AI-powered model using mass spectrometry-based serum proteomics features may predict survival outcomes in advanced-stage NSCLC patients receiving immunotherapy. Further larger studies are warranted to validate serum proteomics' role in predicting response in immunooncology.