Artificial intelligence driven malnutrition diagnostic model for patients with acute abdomen based on GLIM criteria: a cross-sectional research protocol

Abstract

Background

Patients with acute abdomen often experience reduced voluntary intake and a hypermetabolic process, leading to a high occurrence of malnutrition. The Global Leadership Initiative on Malnutrition (GLIM) criteria have rapidly developed into a principal methodological tool for nutritional diagnosis. Additionally, machine learning is emerging to establish artificial intelligent-enabled diagnostic models, but the accuracy and robustness need to be verified. We aimed to establish an intelligence-enabled malnutrition diagnosis model based on GLIM for patients with acute abdomen.

Method

This study is a single-centre, cross-sectional observational investigation into the prevalence of malnutrition in patients with acute abdomen using the GLIM criteria. Data collection occurs on the day of admission, at 3 and 7 days post-admission, including biochemical analysis, body composition indicators, disease severity scoring, nutritional risk screening, malnutrition diagnosis and nutritional support information. The occurrence rate of malnutrition in patients with acute abdomen is analysed with the GLIM criteria based on the Nutritional Risk Screening 2002 and the Mini Nutritional Assessment Short-Form to investigate the sensitivity and accuracy of the GLIM criteria. After data cleansing and preprocessing, a machine learning approach is employed to establish a predictive model for malnutrition diagnosis in patients with acute abdomen based on the GLIM criteria.

Ethics and dissemination

This study has obtained ethical approval from the Ethics Committee of the Sichuan Academy of Medical Sciences and Sichuan Provincial People’s Hospital on 28 November 2022 (Yan-2022–442). The results of this study will be disseminated in peer-reviewed journals, at scientific conferences and directly to study participants.

Trial registration number

ChiCTR2200067044.

Strengths and limitations of this study.

• This study is the first to investigate the prevalence of malnutrition in patients with acute abdomen using the Global Leadership Initiative on Malnutrition (GLIM) criteria.

• A machine learning approach is used to develop a prediction model for malnutrition in patients with acute abdomen based on GLIM criteria.

• Body mass index (BMI)<20 kg/m² is used as the new cut-off value for low BMI, and grip strength and calf circumference is used as the criteria for muscle mass loss, comparing the clinically used Nutritional Risk Screening 2002 and Mini Nutritional Assessment Short-Form, in order to explore the applicability and accuracy of the GLIM criteria in patients with acute abdomen.

• This is a single-centre study. The representativeness of the sample and the applicability of the established predictive model for the diagnosis of malnutrition in patients with acute abdomen in other regions need to be considered with caution.

Introduction

Background

Among common diseases in adult emergency patients, acute abdomen ranks third, accounting for 5%–10%.¹ The prevalence of malnutrition in patients with acute abdomen is 29.3% due to a decreased appetite for autonomous food and a high catabolic state caused by the disease.^{2 3} Furthermore, the elderly population (age >65) accounts for as high as 20% of patients with acute abdomen.^{4 5} Elderly patients with acute abdomen have a higher incidence of malnutrition due to multiple underlying diseases. Numerous studies both domestically and internationally have demonstrated the association between malnutrition and patient mortality, complications and hospitalisation.^{3 6} The guidelines published by the European Society for Clinical Nutrition and Metabolism (ESPEN) in 2017 explicitly state that perioperative nutritional status is an independent risk factor affecting patient prognosis, and severe malnutrition can even lead to patient death.⁷ Therefore, conducting a proper nutritional assessment of patients with acute abdomen during the initial stages of the disease is particularly important.

In recent years, the criteria for assessing malnutrition have been continuously revised, supplemented, and adjusted. It was not until 2015 that ESPEN published an expert consensus providing a clear definition of malnutrition as inadequate energy and macronutrient intake, specifically protein-energy malnutrition.⁸ This definition laid the foundation for the development of diagnostic criteria for malnutrition. In 2016, the Global Leadership Initiative on Malnutrition (GLIM) established unified diagnostic criteria for malnutrition.⁹ Furthermore, in 2017, the State Council of the People’s Republic of China released the ‘¹⁰Healthy China 2030’ blueprint, explicitly stating the need to improve the screening rate for malnutrition and the proportion of hospitalised malnourished patients receiving nutritional therapy.¹⁰ In recent years, various studies have emerged focusing on the validation of the GLIM criteria, demonstrating its strong feasibility in clinical practice.^11–15 However, the applicability of the GLIM criteria in patients with acute abdomen has not yet been validated.

In the 1990s, machine learning began to be applied in clinical settings to address simple clinical problems.¹⁶ With the advent of the big data era, medical information has experienced exponential growth, leading to increasingly extensive research and application of machine learning in healthcare data.¹⁷ A multicentre observational study conducted by Yin et al ¹⁸ found that a machine learning-based decision tool built on the GLIM criteria could rapidly and accurately identify malnutrition in patients with cancer, yielding a high area under the curve (AUC) of 0.964 (kappa=0.898, p<0.001). Additionally, studies have confirmed the feasibility of machine learning in clinical applications. The condition of patients with acute abdomen changes rapidly and the incidence of malnutrition is high. The use of machine learning methods to establish a diagnosis model of malnutrition in patients with acute abdomen can assist clinicians in quickly completing the diagnosis of the disease, which is of great significance for clinical practice.

Objectives

Primary objectives

• Investigate the nutritional status of patients with acute abdomen using the GLIM criteria and explore the prevalence of malnutrition in this population.

• Establish a predictive model for diagnosing malnutrition in patients with acute abdomen based on the GLIM criteria.

Secondary objectives

• Compare the commonly used clinical Nutritional Risk Screening 2002 (NRS-2002) and the Mini Nutritional Assessment Short-Form (MNA-SF) to explore the sensitivity and accuracy of the GLIM criteria.

• Explore the cut-off values for low body mass index (BMI) and reduced muscle mass in GLIM criteria appropriate for Asian populations.

• Explore the relationship between malnutrition and clinical outcomes in patients with acute abdominal disease.

Methods

Study design

This is a single-centre, cross-sectional study. The design scheme of the study is illustrated in figure 1. Clinical data from patients with acute abdomen are collected on multiple occasions to evaluate the incidence of malnutrition using the GLIM criteria. Subsequently, patients are categorised into two groups based on their nutritional status: the malnutrition group and the normal nutrition group. A predictive model for malnutrition in patients with acute abdomen will be developed. The study adheres to the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis reporting guidelines.

Figure 1.

Flow chart of the study design. GLIM, Global Leadership Initiative on Malnutrition.

Setting

Patients diagnosed with acute abdomen are selected in the emergency surgery department of a tertiary hospital in Sichuan province, China, from February 2023 to June 2024.

Eligibility criteria

Inclusion criteria

• Diagnosis of acute abdomen in accordance with the 11th revision of the International Classification of Diseases.

• Age ≥18 years.

Exclusion criteria

• Patients with significant cognitive or language impairments.

• Patients unable to provide complete clinical data.

• Pregnant or lactating women.

• Patients with severe ascites or oedema.

• Patients with repeated hospital admissions.

Study size

Based on a prospective cohort study published in 2017, reporting a malnutrition prevalence rate of 29.3% in patients with acute abdomen, denoted as ‘p’ for sample size calculation purposes in this study.³ The following formula is used to determine the minimum sample size required:

$${\displaystyle N=\frac{Z_{1-\alpha /2}^{2}p\left(1-p\right)}{{\delta }^2}}$$

In this study, the acceptable margin of error, denoted as δ, is determined to be 10% of the prevalence rate (p). With a confidence level of 95% for a two-sided test, the corresponding Z-score, Z_1-α/2 is calculated to be 1.96. Using the formula described earlier, the minimum sample size is determined to be 463 cases. Therefore, a total of 1000 cases will be collected for this study to ensure an adequate sample size.

Study procedures

Patients who meet the inclusion criteria are invited to participate in the study and provided with detailed research explanations and informed consent forms. Data collection is performed by trained researchers using a predesigned case questionnaire on the day of admission, at 3 and 7 days after admission. The collected data include demographic information, baseline characteristics, blood biochemistry analysis, body composition indicators, disease severity scores, nutritional risk screening, malnutrition diagnosis and nutritional support details. And a month follow-up collects outcome indicators (as shown in table 1). Clinical data and outcome indicators will be collected separately by two independent researchers.

Table 1.

Timeline of the various data collections

	Visit 1	Visit 2	Visit 3	Follow-up
Time point	Start point	3 days	7 days	1 month
Eligibility	×
Informed consent	×
Demographic information	×
Baseline characteristics	×
Blood biochemical analysis	×	×	×
Height	×	×	×
Weight	×	×	×
Calf circumference	×	×	×
Grip strength	×	×	×
APACHE II	×	×	×
SOFA	×	×	×
GCS	×	×	×
NRS-2002	×	×	×
MNA-SF	×	×	×
30-day mortality				×
Length of hospital stay				×
Admission to the ICU				×
Infections				×

APACHE II, the Acute Physiology and Chronic Health Evaluation II; GCS, the Glasgow Coma Scale; ICU, the intensive care unit; MNA-SF, the Mini Nutritional Assessment Short-Form; NRS-2002, the Nutritional Risk Screening 2002; SOFA, the Sequential Organ-Failure Assessment.

The primary outcome is 30-day mortality, while secondary outcomes include total length of hospital stay, admission to the intensive care unit and incidence of infection. Infection complication is defined as the absence of infection on admission but with subsequent evidence of infection based on pathogen culture results or clinical symptoms, signs, radiological findings or haematological evidence.

A standardised tool is used to conduct investigations and assessments, calibrated before use. In cases where there are doubts about the survey results, senior physicians are consulted for decision-making.

Blood biochemical analysis

On the day of admission or the following morning, fasting blood samples are collected from patients by the researchers and analysed in the hospital’s laboratory for blood biochemistry and blood cell analysis. The analysed parameters include total protein, albumin, prealbumin, transferrin, haemoglobin, creatinine, blood urea nitrogen and others.

Body composition measurement

Conventional scales are used to measure height and weight in a standing position. All measurements are taken with the patients fasting and without shoes, wearing lightweight clothing or patient gowns. The measurements are performed by trained researchers, and the average of two measurements is recorded. Height is recorded to the nearest 0.1 cm, and weight is recorded to the nearest 0.1 kg. Calf circumference is measured using a flexible tape measure at the thickest part of the calf following standard procedures. Measurements are taken for both legs, and the average value is recorded, with results accurate to 0.1 cm. Grip strength is measured separately for the left and right hands, with at least two measurements taken and the highest value is recorded, with results accurate to 0.1 kg. The measuring tool used is an electronic hand dynamometer produced by Zhongshan Camry Electronic.

Disease severity scoring

The researchers assess disease severity using the following scoring systems: The Acute Physiology and Chronic Health Evaluation II (APACHE II), the Sequential Organ-Failure Assessment (SOFA) and the Glasgow Coma Scale (GCS). The APACHE II score is used for critically ill patients, consisting of acute physiology score, age score and chronic health condition score, with the final score being the sum of these three components. The maximum theoretical score is 71, and a score greater than 15 indicates critical illness.¹⁹ The SOFA score is used to assess the degree of organ dysfunction, employing six criteria to reflect the functioning of organ systems, including respiratory, haematologic, hepatic, cardiovascular, neurological and renal systems, with each criterion scored from 0 to 4.²⁰ The GCS is used to evaluate the level of consciousness. It assigns specific scoring criteria for eye-opening, verbal response and motor response, with the total score representing the degree of consciousness impairment. The maximum score is 15, indicating normal consciousness, while scores below 8 indicate coma, with a minimum score of 3.²¹

Nutritional risk screening

Within 24 hours of admission, the researchers use the NRS-2002 and MNA-SF tools to conduct nutritional risk screening for the patients. The NRS-2002 consists of three indicators: disease status, nutritional impairment and age. A score of ≥3 indicates the presence of nutritional risk.²² The MNA-SF assesses appetite and weight changes in the past 3 months, whether the patient has experienced psychological trauma or acute illness in the past 3 months, psychological issues, BMI and activity level. The total score is 14, with the following scoring criteria: 12–14 points indicate good nutritional status, 7–11 points indicate nutritional risk and below 7 points indicate malnutrition.²³

Diagnosis of malnutrition

Following the positive nutritional risk screening, the researchers proceed to the second step of diagnosing malnutrition using the GLIM criteria (figure 2). To diagnose malnutrition in patients at nutritional risk, the presence of at least one phenotypic and one aetiological criterion is required.²⁴ Involuntary weight loss is defined as 5% weight loss within 6 months or 10% weight loss over 6 months. BMI is calculated as weight in kg divided by the quadratic of height in metres and <20 kg/m² was used as the low BMI cut-off value in this study. Muscle mass loss was used in this study as an indicator of body composition, referencing the Asian Working Group on Sarcopenia for calf circumference: male: <34 cm, female: <33 cm; grip strength: male: <28 kg, female: <18 kg. As acute abdominal disease is an acute condition, all patients with acute abdominal disease included in this study met the aetiological indicators.

Figure 2.

Global Leadership Initiative on Malnutrition criteria diagnostic procedure. MNA-SF, the Mini Nutritional Assessment Short-Form; NRS-2002, the Nutritional Risk Screening 2002; NUTRIC, Nutrition Risk in the Critically Ill.

Modelling

The collected data underwent necessary data cleaning and processing, including handling missing values and outlier data, to ensure data quality and accuracy. Indicators that accounted for more than 30% of missing data will be deleted and the mean or median value of each variable will be used to compensate for the missing. Feature normalisation is used to scale features to standard ranges, which are usually converted to a standard normal distribution with mean 0 and SD 1 to ensure the consistency of scale between different features. The feature engineering method is used to select relevant and representative features to reduce the complexity of the model and improve performance. Subsequently, the data set is divided into training and testing sets at a ratio of 80:20. Appropriate machine learning algorithms, including logistic regression (LR), elastic net, decision trees, random forests, extreme gradient boosting and support vector machines are selected to construct the malnutrition diagnostic model. The algorithms’ parameters, or hyperparameters, are initially defined by establishing a range of values. A parameter grid is constructed based on this range, enumerating all conceivable combinations. Techniques such as 10-fold cross-validation are employed to assess the stability and generalisability of the model. Each parameter combination undergoes evaluation through cross-validation. The objective is to discern the parameter combination that yields optimal model performance, and then applied to the final diagnostic model.

The performance of the model is evaluated using the receiver operator characteristic curve (ROC), AUC, accuracy, precision, recall and F1 score. The shape value is used to evaluate the contribution of features to the model. The diagnostic prediction result is combined with the interpretation of the importance of clinical features to enhance the interpretability of the model and improve the clinical applicability²⁵ (figure 3). Data preprocessing, model development, external validation and performance evaluation are performed in Python (V.3.8; Python Software Foundation).

Figure 3.

The process of malnutrition diagnostic modelling. AUC, area under the ROC curve; DT, decision trees; EN, elastic net; LR, logistic regression; RF, random forests; ROC, receiver operator characteristic curve; SVM, support vector machines; XGB, extreme gradient boosting.

Data analysis

Data analysis is performed using SPSS V.26 software. Normally distributed continuous variables are expressed as mean±SD. Homogeneity of variance is evaluated using analysis of variance, followed by standard t-test for group comparisons. Non-normally distributed continuous variables are presented as median with IQR, and group comparisons are analysed using the Wilcoxon rank-sum test. Categorical variables are reported as frequencies or percentages, and statistical analysis is conducted using the χ² test. Propensity score matching is performed to balance the two groups of patients based on age, sex, APACHE II score, etc, with a 1:1 matching ratio and a matching tolerance set at 0.02. A p value<0.05 is considered statistically significant, indicating a noteworthy difference in baseline characteristics between the two groups at enrolment, while a p value≥0.05 indicates non-significant differences.

Correlation analysis is conducted to examine the relationship between nutritional indicators and disease severity, length of hospital stays and other continuous variables. The correlation coefficients are calculated based on the distribution of the data. If all the feature data followed a normal distribution, the Pearson correlation coefficient is used to assess the correlation between features. If there are non-normally distributed features, the Spearman correlation coefficient is employed. A correlation coefficient greater than 0.7 with a corresponding p value>0.05 indicates a strong and significant correlation between the two features. Multiple linear regression and LR are performed to evaluate the association between malnutrition and disease severity, length of hospital stay, as well as clinical outcomes, with appropriate adjustments made during the analysis.

Patient and Public Involvement Statement

Patients and/or the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Study state

Patient enrolment started in February 2023. At the time of protocol preparation, 20% of the sample had been recruited (200/1000). The planned end date for the study is June 2024.

Discussion

Since its inception, the GLIM criteria have aimed to develop a standardised approach for diagnosing malnutrition, improving the accuracy and consistency of clinical nutritional diagnosis and promoting effective intervention measures. Unlike current nutritional screening tools, the GLIM criteria include muscle loss as a phenotype indicator for malnutrition, and studies have shown its good correlation with body composition analysis indicators.²⁶ However, the cut-off values for low BMI and reduced muscle mass in the GLIM criteria are still controversial, and there is limited research on diagnosing malnutrition in emergency patients. In 1995, the WHO advocated a BMI of <18.5 kg/m² as a general cut-off for underweight.²⁷ However, with the improvement in standards of living, the average BMI has been increasing, making 18.5 kg/m² inadequate as a cut-off for diagnosing malnutrition. The GLIM Consensus Group reached an agreement on using BMI<20 kg/m² for individuals under 70 years old and BMI<22 kg/m² for those over 70 years old, but the cut-off values specific to the Asian population remain uncertain. Wang et al developed a predictive model based on a large-scale, multicentre cross-sectional study to explore new BMI cut-off values that align with the characteristics of the Chinese population. The results indicated that adopting a BMI<20 kg/m² as the cut-off for diagnosing malnutrition in China is consistent with the BMI cut-off values in the GLIM criteria.²⁸ In this study, a low BMI cut-off of <20 kg/m² will be used, which is in line with the characteristics of the Chinese population and can improve the accuracy of malnutrition diagnosis.

ESPEN recommends referring to the diagnostic criteria for sarcopenia when assessing reduced muscle mass.²⁹ Kaegi-Braun et al ³⁰ conducted a secondary analysis of data from a multicentre randomised controlled trial and found that compared with NRS-2002, the GLIM criteria have strong predictive value for adverse clinical outcomes. However, in this study, the criterion for reduced muscle mass was grip strength. In a prospective clinical study by Sánchez-Torralvo et al,³¹ it was found that using grip strength as the criterion for reduced muscle mass in the GLIM criteria resulted in a significant difference in the prevalence of malnutrition compared with using calf circumference. In this study, we will refer to the Asian Working Group for Sarcopenia’s diagnostic process and use a combination of calf circumference and grip strength as the criteria for reduced muscle mass to improve the accuracy of malnutrition diagnosis based on GLIM.³²

Machine learning algorithms can efficiently analyse copious amounts of patient data to identify individuals at risk of or suffering from malnutrition, enabling early intervention. In a multicentre observational study by Wang et al,²⁵ a machine learning-based diagnostic model for malnutrition in elderly patients was developed using the GLIM criteria. The results demonstrated the applicability of machine learning in diagnosing malnutrition in elderly patients, with an accuracy of 80.1% achieved using the gradient boosting model. Similarly, Yin et al ¹⁷ validated a decision tool for identifying malnutrition in patients with cancer using a machine learning model based on classification trees, achieving an impressive accuracy of 95.5%. Therefore, applying machine learning methods to nutritional assessment in patients with acute abdominal conditions can effectively identify those at risk for malnutrition and aid in diagnosis.

Conclusion

This study investigates the nutritional status of patients with acute abdominal conditions and provides a theoretical basis for the application of the GLIM criteria in diagnosing malnutrition in this population. The developed predictive model for malnutrition based on the GLIM criteria can serve as an adjunct tool to assist healthcare professionals in promptly identifying and intervening in patients with nutritional issues, thereby improving their clinical outcomes and prognosis. However, further refinement and validation of the model are still warranted.

Ethics and dissemination

This study has obtained ethical approval from the Ethics Committee of the Sichuan Academy of Medical Sciences and Sichuan Provincial People’s Hospital on 28 November 2022 (Yan-2022–442). This study was registered with the Chinese Clinical Trial Registry. This study adheres to good clinical practice principles and strictly complies with relevant national laws, regulations and the Helsinki declaration. The results of this study will be disseminated in peer-reviewed journals, at scientific conferences and directly to study participants.

Ethics statements

Patient consent for publication

Not applicable.