MoCaPS: a machine learning model for stratification of cancer-associated cachexia based on blood biomarkers

Identification of minimally invasive biomarkers of cancer-associated cachexia may help to recognize high risk patients for progression to more severe cachectic stages. We developed a machine learning-based Model for Cachectic Patients Stratification (MoCaPS) to determine the sets of blood biomarkers that differentiate between noncachectic (NCa), precachectic (PCa), or cachectic (Ca) patients. The model was applied to data collected from treatment-naïve patients with pancreatic ductal adenocarcinoma through the Florida Pancreas Collaborative multi-institutional cohort study and biobanking initiative. Cachexia status of all participants was classified according to modified criteria by Vigano and colleagues. The MoCaPS model pipeline was designed to work effectively with datasets of moderate size to robustly select predictive data features, and to efficiently handle data imbalance. MoCaPS identified between 4 and 5 biomarkers out of 37 candidates that distinguished precachectic and cachectic stages, and demonstrated accuracies near or greater than 75% for predictors of NCa, PCa, and Ca.