Background: The efficacy of front-line pembrolizumab has been established in studies that limit treatment duration to 2 years, but decision to stop pembrolizumab after 2 years is often at physician's discretion. ATHENA is a retrospective cohort study using a comprehensive administrative database aimed firstly at exploring the optimal duration of pembrolizumab and secondly real-life prognosis factors in patients with advanced non-small cell lung cancer (NSCLC). Methods: Using the French National Health Insurance database (SNDS), we identified patients with incident lung cancer in France from 2015 to 2022. Treatments and patients' characteristics were extracted or inferred from hospital, outpatient care, pharmacy delivery reports. The duration's hazard ratio (HR) was estimated with Cox model weighted by inverse of propensity score to account for confounding. Prognostics factors in first line population were identified with Cox model selected by a LASSO procedure. Findings: 391,106 patients with lung cancer were identified, of whom 43,359 received up-front pembrolizumab for an advanced disease. There were 67% (29,040/43,359) of male and the median age at diagnosis was 65 years old. After a median follow-up time of 25.9 months (min-max, [0-97.6]), the median overall survival (OS) after pembrolizumab initiation in first line was 15.7 [CI 95, 15.3-16.0] months. In multivariable analysis, several covariables were independently associated with worse OS, including male sex with chemo-immunotherapy, age, hospital category, high deprivation index, inpatient hospitalization for first pembrolizumab, and history of diabetes, diuretic, beta blocker, painkiller prescription. At landmark time of 29 months after pembrolizumab initiation, continuation beyond 2 years was not associated with better OS than a fixed 2-year treatment, HR = 0.97 [0.75-1.26] p = 0.95. Interpretation: This study supports the notion that stopping pembrolizumab after 2 years could be safe for patients with advanced NSCLC. However, because observational studies are prone to confounding and selection bias, causality cannot be affirmed.