Classification Tree-Based Machine Learning to Visualize and Validate a Decision Tool for Identifying Malnutrition in Cancer Patients

Background The newly proposed Global Leadership Initiative on Malnutrition (GLIM) framework is promising to gain global acceptance for diagnosing malnutrition. However, the role of machine learning in facilitating its application in clinical practice remains largely unknown. Methods We performed a multicenter, observational cohort study including 3998 patients with cancer. Baseline malnutrition was defined using the GLIM criteria, and the study population was randomly divided into a derivation group (n = 2998) and a validation group (n = 1000). A classification and regression trees (CART) algorithm was used to develop a decision tree for classifying the severity of malnutrition in the derivation group. Model performance was evaluated in the validation group. Results GLIM criteria diagnosed 588 patients (14.7%) with moderate malnutrition and 532 patients (13.3%) with severe malnutrition among the study population. The CART cross-validation identified 5 key predictors for the decision tree construction, including age, weight loss within 6 months, body mass index, calf circumference, and the Nutritional Risk Screening 2002 score. The decision tree showed high performance, with an area under the curve of 0.964 (kappa = 0.898, P < .001, accuracy = 0.955) in the validation group. Subgroup analysis showed that the model had apparently good performance in different cancers. Among the 5 predictors constituting the tree, age contributed the least to the classification power. Conclusion Using the machine learning, we visualized and validated a decision tool based on the GLIM criteria that can be conveniently used to accelerate the pretreatment identification of malnutrition in patients with cancer.