Machine Learning-Based Prediction of Complications and Prolonged Hospitalization with the GLIM Criteria Combinations...

Malnutrition is widely present and influences the prognosis of elderly inpatients, so it is helpful to be able to identify it with a convenient method. However, in the widely accepted criteria for malnutrition, the Global Leadership Initiative on Malnutrition (GLIM), a lot of metrics can be used to define the phenotypic and etiological criteria. To identify muscle mass reduction, anthropometric parameters such as calf circumference (CC) and hand grip strength (HGS) are preferable to other expensive methods in many situations because they are easy and inexpensive to measure, but their applicability needs to be verified in specific clinical scenarios. This study aims to verify the value of CC- and HGS-identified muscle loss in diagnosing malnutrition and predicting in-hospital complications (IHC) and prolonged length of hospital stay (PLOS) in elderly inpatients using machine learning methods.

A sample of 7122 elderly inpatients who were enrolled in a previous multicenter cohort study in China were screened for eligibility for the current study and were then retrospectively diagnosed for malnutrition using 33 GLIM criteria that differ in their combinations of phenotypic and etiological criteria, in which CC or CC+HGS were used to identify muscle mass reduction. The diagnostic consistency with the subjective global assessment (SGA) criteria at admission was evaluated according to Kappa coefficients. The association and the predictive value of the GLIM-defined malnutrition with 30-day IHC and PLOS were evaluated with logistic regression and randomized forest models.

In total, 2526 inpatients (average age 74.63 ± 7.12 years) were enrolled in the current study. The prevalence of malnutrition identified by the 33 criteria combinations ranged from 3.3% to 27.2%. The main IHCs was infectious complications (2.5%). The Kappa coefficients ranged from 0.130 to 0.866. Logistic regression revealed that malnutrition was identified by 31 GLIM criteria combinations that were significantly associated with 30-day IHC, and 22 were significantly associated with PLOS. Random forest prediction revealed that GLIM 15 (unconscious weight loss + muscle mass reduction, combined with disease burden/inflammation) performs best in predicting IHC; GLIM 30 (unconscious weight loss + muscle mass reduction + BMI reduction, combined with disease burden/inflammation) performs best in predicting PLOS. Importantly, CC alone performs better than CC+HGS in the criteria combinations for predicting adverse clinical outcomes.

Muscle mass reduction defined by a reduced CC performs well in the GLIM criteria combinations for diagnosing malnutrition and predicting IHC and PLOS in elderly Asian inpatients. The applicability of other anthropometric parameters in these applications needs to be further explored.