Cachexia is associated with multiple adverse outcomes in cancer. However, clinical decision-making for oncology patients at the cachexia stage presents significant challenges. This study aims to develop a machine learning (ML) model to identify potentially reversible cancer cachexia (PRCC). This was a multicenter cohort study. Cachexia was retrospectively diagnosed using Fearon's framework. PRCC was defined as a diagnosis of cancer cachexia at baseline that turned negative one month later. Body weight dynamics accessible upon patient admission were screened and modeled to predict PRCC. Multiple ML models were trained and cross-validated using 70% of the data to predict PRCC, with the remaining 30% reserved for model evaluation. The interpretability and clinical usefulness of the optimal model were assessed, and external validation was performed in an independent cohort of 238 patients. The study enrolled 1983 men and 1784 women (median age=58 years). PRCC was identified in 1983 patients (52.6%). Breast cancer exhibited the highest rate of PRCC (72.1%), while cachexia associated with various gastrointestinal cancers was less likely to be reversed. Weight change (WC) from six months ago to one month ago, WC from one month ago to baseline (-1 to 0) and baseline body mass index were selected for modelling. A multilayer perceptron model showed good performance to predict PRCC in the holdout test set (AUC [95%CI] = 0.887 [0.866, 0.907], accuracy=0.836, sensitivity=0.859, specificity=0.812) and the external validation set (AUC [95%CI] = 0.863 [0.778, 0.948]). The WC -1 to 0 showed the highest impact on model output. The model was demonstrated to be clinically useful and statistically relevant. This study presents an explainable ML model for the early identification of PRCC that utilizes simple body weight dynamics. The findings showcase the potential of this approach in improving the management of cancer cachexia to optimize patient outcomes. Data described in the manuscript was derived from a registered research project (URL: http://www.chictr.org.cn/showproj.aspx?proj=31813; ID: ChiCTR1800020329).
Read full abstract