Identification of key factors for malnutrition diagnosis in chronic gastrointestinal diseases using machine learning underscores the importance of GLIM criteria as well as additional parameters

Abstract

Introduction: Disease-related malnutrition is common but often underdiagnosed in patients with chronic gastrointestinal diseases, such as liver cirrhosis, short bowel and intestinal insufficiency, and chronic pancreatitis. To improve malnutrition diagnosis in these patients, an evaluation of the current Global Leadership Initiative on Malnutrition (GLIM) diagnostic criteria, and possibly the implementation of additional criteria, is needed.

Aim: This study aimed to identify previously unknown and potentially specific features of malnutrition in patients with different chronic gastrointestinal diseases and to validate the relevance of the GLIM criteria for clinical practice using machine learning (ML).

Methods: Between 10/2018 and 09/2021, n = 314 patients and controls were prospectively enrolled in a cross-sectional study. A total of n = 230 features (anthropometric data, body composition, handgrip strength, gait speed, laboratory values, dietary habits, physical activity, mental health) were recorded. After data preprocessing (cleaning, feature exploration, imputation of missing data), n = 135 features were included in the ML analyses. Supervised ML models were used to classify malnutrition, and key features were identified using SHapley Additive exPlanations (SHAP).

Results: Supervised ML effectively classified malnourished versus non-malnourished patients and controls. Excluding the existing GLIM criteria and malnutrition risk reduced model performance (sensitivity -19%, specificity -8%, F1-score -10%), highlighting their significance. Besides some GLIM criteria (weight loss, reduced food intake, disease/inflammation), additional anthropometric (hip and upper arm circumference), body composition (phase angle, SMMI), and laboratory markers (albumin, pseudocholinesterase, prealbumin) were key features for malnutrition classification.

Conclusion: ML analysis confirmed the clinical applicability of the current GLIM criteria and identified additional features that may improve malnutrition diagnosis and understanding of the pathophysiology of malnutrition in chronic gastrointestinal diseases.

Keywords: GLIM criteria; decision trees; gastrointestinal diseases; liver cirrhosis; machine learning; malnutrition; supervised and unsupervised learning.

Figure 1

Overview of the data processing. LC, liver cirrhosis; CP, chronic pancreatitis; SB/II, short bowel/intestinal insufficiency; SHAP, SHapley Additive exPlanations; UMAP, uniform manifold approximation and projection.

Figure 2

SHAP feature importance summary plot from (A) the LGBM using all features including GLIM diagnostic criteria (marked in bold) or from (B) the Random Forest model omitting the GLIM malnutrition diagnosis criteria and the associated features malnutrition risk and C-reactive protein. Each dot represents the average SHAP feature importance value for one patient regarding the respective feature. The color of the dot represents the feature expression value (red - high, blue - low). Positive SHAP values indicate classification of malnourished patients, while negative SHAP values indicate classification of non-malnourished individuals. An alignment to the right therefore indicates higher importance of a feature for malnutrition classification. IL-6, interleukin 6; TNF, tumor necrosis factor.

Figure 3

Distribution of the clusters obtained by UMAP supported Feature Distributed Clustering (FDC). (A) Five clusters (0–4) were found; moreover 14 data points were denoted as noise (labeled as cluster -1). (B,C) Distribution of gastrointestinal diseases along with malnutrition in the clusters. It was observed that cluster 4 is mostly comprised of LC patients and also has the highest proportion of malnourished individuals. UMAP, uniform approximation and projection; HC, healthy controls; SB/II, short bowel/intestinal insufficiency; CP, chronic pancreatitis; LC, liver cirrhosis.

Figure 4

Distribution of the clusters obtained by UMAP supported Feature Distributed Clustering (FDC). (A) Five clusters (0–4) were found; moreover 14 data points were denoted as noise (labeled as cluster -1). (B–F) Distribution of the distinct GLIM criteria in the clusters. UMAP, uniform approximation and projection; LC, liver cirrhosis; BMI, body mass index; FFMI, fat-free mass index.

Figure 5

Exemplary decision tree using the top 10 features from LGBM using all features.