Machine learning to identify precachexia and cachexia: a multicenter, retrospective cohort study

Background: Detection of precachexia is important for the prevention and treatment of cachexia. However, how to identify precachexia is still a challenge.

Objective: This study aimed to detect cancer precachexia using a simple method and distinguish the different characteristics of precachexia and cachexia.

Methods: We included 3896 participants in this study. We used all baseline characteristics as input variables and trained machine learning (ML) models to calculate the importance of the variables. After filtering the variables based on their importance, the models were retrained. The best model was selected based on the receiver operating characteristic value. Subsequently, we used the same method and process to identify patients with precachexia in a noncachexia population using the same method and process.

Results: Participants in this study included 2228 men (57.2%) and 1668 women (42.8%), of whom 471 were diagnosed with precachexia, 1178 with cachexia, and the remainder with noncachexia. The most important characteristics of cachexia were eating changes, arm circumference, high-density lipoprotein (HDL) level, and C-reactive protein albumin ratio (CAR). The most important features distinguishing precachexia were eating changes, serum creatinine, HDL, handgrip strength, and CAR. The two logistic regression models for screening for cachexia and diagnosing precachexia had the highest area under the curve values of 0.830 and 0.701, respectively. Calibration and decision curves showed that the models had good accuracy.

Conclusion: We developed two models for identifying precachexia and cachexia, which will help clinicians detect and diagnose precachexia.