1632P Predicting real-world overall survival in advanced melanoma using machine learning

1L aMel pts were identified from the ConcertAI EHR database (Jan 2016 – Oct 2023). Index date was the start of 1L therapy. Data was split 80:20 into training and test sets, respectively. A gradient-boosted survival ML model (XGBoost) was trained to predict all-cause OS using baseline clinical and patient characteristics. Model performance was evaluated using the concordance index (c-index). Top predictors were visualized in a cox proportional hazard nomogram of OS at 6 months, 1 yr, and 2 yrs. Pts with ≥median of predicted risk scores were classified as high-risk.

The study cohort had 2,038 pts (mean age, 65.1 yrs; 64.8% male) with median OS of 22.9 months (1,091 events). The XGBoost achieved a c-index of 0.71. Higher risk of death was associated with ECOG-PS ≥ 2, presence of brain or liver metastases, elevated lactate dehydrogenase (LDH) and white blood cell (WBC) counts, and age ≥ 70 yrs at 1L initiation (Table). ECOG-PS ≥ 2 increased the risk of death by 119%, brain and liver metastases by 98% and 45%, respectively, and elevated LDH by 29%. Median probability of 6-month, 1-yr, and 2-yr OS for the high-risk group was 71.3%, 55.8%, and 37.8% vs. 85.3%, 75.9%, and 63.1% for lower-risk group, respectively.

We developed a melanoma-specific model to predict OS using a large real-world dataset derived from diverse oncology practices across the US that may have implications across drug development and clinical practice decisions in the future.