Predicting end-of-life risk in patients with cancer: A multicenter cohort study

ObjectiveManaging end-of-life (EOL) patients with cancer has always been a major challenge in healthcare. Previous studies have highlighted the need for individualized EOL treatment and the importance of avoiding overtreatment. However, accurately identifying EOL patients with cancer to provide appropriate care and improve their quality of life remains an unresolved issue.MethodsThis study was based on investigation on Nutrition Status and Clinical Outcomes of Common Cancer (INSCOC) cohort. Machine-learning methods analyzed the characteristics of EOL patients with cancer and identified the determinants associated with EOL risk. Population-attributable fractions, least absolute shrinkage and selection operator regression analysis, random forest, and logistic regression (LR) analysis were used to screen predictive indicators for EOL risk. LR, support vector machines, generalized linear models, gradient boosting machine, random forests, and artificial neural networks were then used to construct models based on the identified risk factors.ResultsIn total, 17,013 patients from the INSCOC cohort (including 1109 EOL patients) were analyzed. LR was the best-performing machine-learning model, and factors such as advanced stage, neutrophil-to-lymphocyte ratio, malnutrition, hypoalbuminemia, poor self-health assessment, limited mobility, prognostic nutritional index, lack of appetite, and cancer location were the key determinants of EOL risk.ConclusionThe findings of this study provide valuable insights for clinical practice, indicating potential pathways for the more precise management of patients with EOL cancer. Through timely identification and intervention of the identified risk factors, the risk of EOL in these patients may be reduced and their prognosis can be improved.