Machine learning models to predict outcomes at 30-days using Global Leadership Initiative on Malnutrition combinations with and without muscle mass in people with cancer

Background

Equipment to assess muscle mass is not available in all health services. Yet we have limited understanding of whether applying the Global Leadership Initiative on Malnutrition (GLIM) criteria without an assessment of muscle mass affects the ability to predict adverse outcomes. This study used machine learning to determine which combinations of GLIM phenotypic and etiologic criteria are most important for the prediction of 30-day mortality and unplanned admission using combinations including and excluding low muscle mass.

Methods

In a cohort of 2801 participants from two cancer malnutrition point prevalence studies, we applied the GLIM criteria with and without muscle mass. Phenotypic criteria were assessed using ≥5% unintentional weight loss, body mass index, subjective assessment of muscle stores from the PG-SGA. Aetiologic criteria included self-reported reduced food intake and inflammation (metastatic disease). Machine learning approaches were applied to predict 30-day mortality and unplanned admission using models with and without muscle mass.

Results

Participants with missing data were excluded, leaving 2494 for analysis [49.6% male, mean (SD) age: 62.3 (14.2) years]. Malnutrition prevalence was 19.5% and 17.5% when muscle mass was included and excluded, respectively. However, 48 (10%) of malnourished participants were missed if muscle mass was excluded. For the nine GLIM combinations that excluded low muscle mass the most important combinations to predict mortality were (1) weight loss and inflammation and (2) weight loss and reduced food intake. Machine learning metrics were similar in models excluding or including muscle mass to predict mortality (average accuracy: 84% vs. 88%; average sensitivity: 41% vs. 38%; average specificity: 85% vs. 89%). Weight loss and reduced food intake was the most important combination to predict unplanned hospital admission. Machine learning metrics were almost identical in models excluding or including muscle mass to predict unplanned hospital admission, with small differences observed only if reported to one decimal place (average accuracy: 77% vs. 77%; average sensitivity: 29% vs. 29%; average specificity: 84% vs. 84%).

Conclusions

Our results indicate predictive ability is maintained, although the ability to identify all malnourished patients is compromised, when muscle mass is excluded from the GLIM diagnosis. This has important implications for assessment in health services where equipment to assess muscle mass is not available. Our findings support the robustness of the GLIM approach and an ability to apply some flexibility in excluding certain phenotypic or aetiologic components if necessary, although some cases will be missed.