In a study, a machine learning approach outperformed cancer-specific models and revealed universal biomarkers to guide end-of-life care decisions
A machine learning-based pan-cancer model showed to outperform cancer-specific models in predicting 30-day mortality in patients with various advanced cancers according to findings published on the ESMO Real World Data and Digital Oncology (ESMO Real World Data and Digital Oncology, Volume 8, 100146).
In a Danish study, researchers analysed clinical data from 8,690 patients across 10 cancer types (lung, breast, pancreatic, ovarian) referred to the Department of Oncology at Aalborg University Hospital and who died between 2010 and 2021. In the study, the XGBoost algorithm was trained on rich time-series electronic health record (EHR) data, covering patients’ clinical history, and used to predict 30-days mortality.
In 8 out of 10 cancer types, the pan-cancer model delivered equal or superior performance compared to single-cancer models with an average precision (AP) of 0.56 versus 0.51 respectively, with the highest AP observed in urinary cancer (0.70). Brain cancer was the only type where the cancer-specific model performed better, and according to researchers it was likely due to data quality issues.
Through explainable AI techniques using Shapley Additive Explanations (SHAP), the analysis also revealed three key biomarkers as the most influential predictors of 30-days mortality: plasma albumin levels (reflecting nutritional status and systemic inflammation), white blood cell count (indicating immune response and infection risk), and lactate dehydrogenase levels (associated with tumour burden and tissue damage).
“These findings suggest that cross-cancer prognostic modeling could enhance clinical decision-making by identifying universal biomarkers of terminal decline in advanced malignancies”, explained Dr Rodrigo Dienstmann, from Oncoclínicas Group, São Paulo, Brazil, and Vall d’Hebron Institute of Oncology, Barcelona, Spain, and Editor-in-Chief of the ESMO’s peer-reviewed journal dedicated to the transformation of oncology care with real-world evidence and digital technologies.
Systemic anticancer therapy (SACT) near the end of life (EOL) reduces the quality of the patient’s remaining life without clinical benefit, and artificial intelligence holds promise to optimise end-of-life care decisions. “Machine learning-based prognostic models show potential to assist clinicians in guiding systemic anti-cancer therapy decisions for palliative care patients,” he concludes.
Submit your research to ESMO Real World Data and Digital Oncology
ESMO Real World Data and Digital Oncology is dedicated to publishing high-quality data science and educational content on the transformation of oncology care with real-world evidence and digital technologies that physicians can trust and understand.
The journal publishes innovative research that provides actionable insights from real-world data sources and digital devices, such as generalisable observational and health services research, clinical informatics solutions, algorithm-powered diagnostics, and decision support systems and direct-to-patient mobile applications.
A special issue will be published in concomitance with the ESMO AI & Digital Oncology Congress 2025 (12-14 November 2025) and submissions are currently open.
