Machine learning models predict adverse events and survival in patients receiving immunotherapy

Two poster presentations at the ESMO Targeted Anticancer Therapies Congress 2025 (Paris, 3–5 March) showed that some machine learning algorithms accurately predicted immunotherapy-related adverse events (irAEs) in melanoma patients and survival outcomes in patients with metastatic renal cell carcinoma (mRCC) receiving first-line immunotherapy, thus supporting the use of artificial intelligence (AI) techniques to inform clinical decision-making for immune-checkpoint inhibitors (ICIs).

In a first study, six machine learning algorithms were developed using 80% of the data from 455 datasets for patients with melanoma initiated on ICIs between 2014-2024, and 20% for validation (Abstract 3P). The model predicted irAEs with the most accuracy of 92%. In particular, key predictors included increasing age, female gender and exposure to combination therapy or pembrolizumab.

In a second study, data from 4895 patients with mRCC who received first-line immunotherapy since 2015 were collected from a National Cancer Database, and split into training (70%) and testing (30%) sets (Abstract 4P). Fifteen features were selected based on the univariate Cox regression for OS, including demographics, Charlson-Deyo Score, tumour side, grade, lymph vascular invasion, and prior surgery or radiotherapy. When comparing classification models to predict overall survival (OS), Ensemble methods, particularly CatBoost, showed superior performance with an area under curve (AUC) of 0.87. In particular, key predictors were tumour grade, surgery and patient age.

“Taken together, these data contribute to consolidate the role of artificial intelligence in ICIs administration,” says Prof. Umberto Malapelle, University of Naples Federico II, Italy, commenting on the findings. “However, future work will aim to validate these models and incorporate them into ICI drug development trials to prospectively predict whose patients will benefit from these agents or develop adverse events.”
He concludes: “The introduction of ICIs has significantly modified the clinical outcomes in patients with advanced solid tumours. Despite the efficacy proven across different tumour types, however, much still needs to be done to better select patients who can really benefit from these treatments. In this scenario, artificial intelligence and machine learning may be promising tools to support informed decision-making between clinicians and patients considering starting, stopping or continuing ICI therapy.”