Development of a prognostic score for stage IV NSCLC using advance machine learning algorithms: An international comparative study across South Korea, US, and France.

Abstract

e20615 Background: Patients with Stage IV NSCLC have a poor prognosis, with approximately 25-30% succumbing to the disease within the first three months. However, survival outcomes can vary significantly depending on individual factors and advancements in treatment options. Therefore, it is crucial to identify patients at high risk of poor survival outcomes to tailor personalized treatment strategies and optimize clinical management. An international study has been conducted to generate prognostic scores for patients with Stage IV NSCLC. Methods: A retrospective observational study was conducted using real-world data (RWD) from South Korea (ROOT-HEALTH database from Samsung Medical Center), the USA (COTA NSCLC dataset) and France (CIAN lung cohort from Centre Léon Bérard). Patients with stage IV NSCLC between January 2008 and December 2023 were included. The study's primary objective was to learn from patients ‘baseline characteristics affecting overall survival (OS) and progression-free survival (PFS) to derive a prognostic score reflecting relative risk of death for given patient. A partition strategy was used, with 70% of the cohort for training and 30% for validation. Data-driven feature elimination has been applied to avoid overfitting of the models. Nested cross-validation reduced selection bias in hyperparameter tuning and model selection. Results: A total of 7,805 patients with NSCLC were included in the final analysis (USA: n=610, France: n=271, Korea: n=6,924). In the Korean dataset, the model achieved a C-index of 0.872 for OS using 35 selected variables. For the USA dataset, the model, based on 15 selected variables, demonstrated a C-index of 0.707 for OS. For PFS, the Korean dataset demonstrated a C-index of 0.823, the USA dataset achieved a C-index of 0.664, and the French dataset reported a C-index of 0.64, using 15 selected variables. The important predictors for OS included albumin, neutrophil/lymphocyte ratio, bilirubin, and druggable mutations. For PFS, important predictors included LDH, albumin, hemoglobin, MET mutation, bilirubin, and BMI. The prognostic scores derived from these models allowed to identify patients with higher risk of death, than average, in each country. Conclusions: This study generates prognostic scores for OS and PFS in stage IV NSCLC using longitudinal international data from the USA, France, and South Korea throughout the cancer journey. It has the potential to support better treatment strategy for severe patients in clinical practice and early decision-making in clinical studies. Further research utilizing diverse international cohorts is essential to refine prognostic scores and enhance personalized treatment strategies for stage IV NSCLC patients.