This study was led by the Geriatric Nutrition Study Group of the Chinese Society for Parenteral and Enteral Nutrition, Chinese Medical Association. From June to September 2014, a clinical nutrition survey was conducted to assess the changes in nutritional status during hospitalization in 7122 Chinese patients from 34 large hospitals in 18 cities. The patients enrolled in this study were from a variety of clinical departments, including gastroenterology, respiratory medicine, neurology, oncology, general surgery, thoracic surgery, orthopedics, and geriatrics. The inclusion criteria were as follows: aged≥18 years; hospital stay of 7-30 days; no urgent surgery performed before 8 AM the day following admission; patients had to be conscious, willing to accept the multiple nutrition assessment, and have signed written informed consent. The patients belonging to the following categories were excluded from the examination: those who were younger than 18 years; with an expected hospital stay of <7 days or >30 days; unconscious; refused participation in the nutritional assessment or did not sign the written informed consent.

The study protocol was approved by the Ethics Committee of Beijing Hospital (PIC approval number: 2014BJYYEC-022-02) and registered in the Chinese Clinical Trial Registry (No. ChiCTR-EPC-14005253). This study included only older adults (>65 years of age; n=2734 people). The flowchart of patients is shown in Figure 1.

For participants at malnutrition risk (NRS2002≥3 in this study), at least one phenotypic criterion and one etiologic criterion should be met to establish the GLIM diagnosis: only if NRS2002 score is ≥3, GLIM score would be calculated. At least 1 phenotypic criterion and 1 etiologic criterion were assigned GLIM (+); the participants could be diagnosed with malnutrition; all phenotypic criteria and etiologic criteria were GLIM (-), and the diagnosis was negative. In addition to the above, the diagnosis can be “Not Known.” Participants who could not develop GLIM diagnosis were deleted. Please find the original scale of NRS2002 and GLIM in Multimedia Appendix 1.

In this study, the GLIM diagnosis was retrospectively determined according to the information collected in the study. A low BMI was diagnosed according to the Asian standards described in the GLIM criteria, and reference values for severe malnutrition were defined according to a previous study conducted among an Asian population [8,16]. Calf circumference was used to evaluate whether there was a reduced muscle mass. A value of less than 15% (percentile) according to a previous study (male<30 cm; female<29.5 cm) was defined as positive for moderate muscle loss [17]. Other criteria were derived from the results of NRS2002 (Table 1).

In this study, feature selection was carried out in 2 steps. Firstly, based on recommendations and a review of literature, we chose 25 variables for the ML models, which included demographics (ie, gender, age, education, and type of Medicare), medical history (ie, weight loss, food consumption, gastrointestinal symptoms, and metabolic demand), physical examination (ie, subcutaneous fat, muscle consumption, edema, hand grip strength, BMI, and calf circumference), and laboratory test (ie, albumin, hemoglobin, white blood cells, lymphocyte count, total protein, total bilirubin, alanine transaminase, blood urea nitrogen, creatinine, platelet count, and glucose). According to the statistical differences between groups and expert opinions, 23 variables, except blood urea nitrogen (P=.65) and Creatinine (P=.66), were used to construct the prediction model [18,19].

After checking for outliers and errors of data, we applied the imputation method for completing missing values using k-nearest neighbors. Each missing feature was imputed using values from the n nearest neighbors that have a value for the feature. The feature of the neighbors was averaged uniformly or weighted by distance to each neighbor. The development data set was split into training and test sets, with a ratio of 8:2. A stratified strategy was adopted for consistent outcome class proportions (normal nutrition vs abnormal nutrition) in the sub–data sets and data sets. Preprocessing was based on the features of the training set to avoid data leakage. Ordinal values for categorical variables were assigned to represent duration and converted into categorical values; categorical values were one-hot encoded. All other quantitative variables were standardized. After preprocessing, of 2660 patients diagnosed with malnutrition, 855 (32%) were divided as 684 (80%) into the training set, and 171 (20%) into the test set, respectively.

For training models, we applied 5 widely used ML algorithms in assisted diagnosis prediction , including logistic regression [20], tree-based model (light gradient boosting machine [LightGBM] [21]), random forest (RF) [22], and extreme gradient boosting (XGBoost) [23], multilayer perceptron [24], naive Bayes [25], and support vector machine–based methods [26]. The logistic regression model represents the linear model, which is a state-of-the-art classification model for a baseline construction. RF, LightGBM, and XGBoost were representative ensemble learning models, which do well in dealing with multiple types of features. RF is a bagging ensemble algorithm containing multiple decision trees, which uses the voting method to classify samples and integrate the final voting results produced by multiple decision trees. XGBoost is an optimized distributed gradient boosting library designed to be highly efficient, flexible, and portable. It implements ML algorithms under the gradient boosting framework. LightGBM is a gradient boosting framework that uses tree-based learning algorithms. It uses a gradient-based one-sided sampling algorithm to reduce the sample dimension and a mutually exclusive feature bundling algorithm to reduce the feature dimension.

After development, prediction performance was assessed using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and accuracy. To enhance the clinical applicability, we combined the diagnostic prediction results with clinical feature importance explanations to identify a patient with malnutrition. In this study, we used Shapley additive explanation (SHAP) values, to interpret feature contributions and assess the clinical significance of predictive models [27,28].

The SHAP value is the measurement of the marginal contribution of each feature in different combinations (Equation 1).

Where ϕ₀ is the average predicted value of all the samples, known as the base value, ϕ_j is the SHAP value of the feature, and M is the total number of features. When ϕ_j is greater than zero, this feature improves the predicted value and has a positive effect. According to the SHAP value, the effect direction and intensity of each feature can be obtained. Figure 2 shows the workflow for the establishment and interpretation of the malnutrition diagnosis tool.

For the ML model interpretation, we used both globally explainable and locally explainable methods [29]. The global explanation determines the importance of features by comparing the magnitude of the model prediction error change before and after replacing a feature. If the prediction error changes more, it indicates that the feature is more important. Meanwhile, we used SHAP values to explore individual-based decision-making processes in the view of local explanation.

Data preprocessing, model development, and external validation were performed in Python (version 3.8; Python Software Foundation). Model algorithms and hyperparameter tuning tools were based on Python scikit-learn (v0.24.2).

Continuous data were expressed as means (SD) and were compared using a t test. The normality of continuous data was tested using a Shapiro-Wilk test. Categorical data were expressed as n (%) and were compared using a chi-squared test. All tests were 2-tailed, and P<.05 was considered statistically significant. Statistical analyses were performed using SPSS Statistics (version 16; IBM Corp).

The GLIM criteria diagnosed 855 patients with malnutrition. GLIM diagnosed that malnutrition was significantly associated with age, BMI, hand grip strength, albumin, total protein, hemoglobin, lymphocyte count, total bilirubin, alanine transaminase, and platelet count (Table 2).

The performance of predictive models was evaluated in terms of sensitivity, specificity, accuracy, and AUROC. In a 10-fold cross-validation performed on the training set, the XGBoost performed best in AUROC (92.5%), and logistic regression had the worst AUROC and second-worst sensitivity (Table 3). In the external validation cohort, the top 3 models by AUROC were the RF model (95.1%), LightGBM (94.7%), and XGBoost (94.0%), as shown in Table 4 and Figure 3.

To gain a thorough understanding of the general impact of features, the importance of features included in the XGBoost was quantified. Features and their subtypes were sorted by the magnitude of impact.

Categorical predictor variables were presented separately. The importance of features was sorted by the sum of the mean absolute value of the SHAP value magnitude over all samples. SHAP summary plots succinctly display the magnitude, prevalence, and direction of a feature's effect. As each dot corresponds to a single sample, the swarm plot avoids conflating the magnitude and prevalence of the feature effect into a single number. Therefore, apart from the magnitude of the features' global importance, it is informative concerning the exact impact on specific individuals.

Figure 4(a) presents the feature importance as the strongest predictor to effect GLIM. Of the top 10 important features, there are 2 body measurement features, 4 laboratory features, and 4 nutrition-related histories. The results revealed that BMI, weight loss, and calf circumference were associated with a higher risk of GLIM events. Moreover, Platelet count and total protein also were important features. Additionally, we found that Lymphocyte count and Alanine transaminase seemed to be important features in predicting malnutrition.

Figure 4(b) summarizes the SHAP value plot by combining feature importance with feature effects. The y-axis is defined by the feature, and the x-axis is defined by the SHAP value. The plot describes the features’ overall influence on the model prediction. Each point in each feature represents an individual case, with colors ranging from blue (low feature value) to red (high feature value). The data points further to the right represent the features that contribute to the higher risk of malnutrition using GLIM for a given individual case. The data points to the left represent the features that contribute to the lower risk of GLIM. The vertical line in the middle represents no change in risk. We found that the data points (individual cases) with lower BMI values had a higher risk of GLIM.

Furthermore, the lab tests and BMI were used as the continuous variable in our prediction model. This study examined the marginal effect of those features on the predicted outcome of an ML model using a SHAP dependence plot. As shown in Figure 5(a), the results showed that the value of BMI over approximately 21 kg/m² was associated with a higher risk of GLIM. Patients had a high risk of GLIM when their platelet count value was below approximately 280×10⁹/L. As shown in Figure 5(b), when calf circumference was below 30, it increases the risk of GLIM. Moreover, hemoglobin below 100 increases the risk of GLIM.

At the same time, we use the explainable method of ML to explain the constructed ML model. As shown in Figure 6, compared with previous studies, our method can not only propose diagnostic cutoff values for characteristic populations, but also interpret individual individuals, enabling medical workers to know in detail what each characteristic of a patient represents. As shown in Figure 6a, this is an individual diagnosed as malnourished by the machine model, and through the machine learning explainable method, we can clearly see that although his BMI is within the normal range, the characteristics of weight loss, calf circumference and age reflect that her nutritional status is not in a normal state, especially weight loss has a huge negative impact. At the same time, we can also provide health warnings for individuals identified as nutritionally healthy, as shown in Figure 6b, although the patient was diagnosed as nutritionally healthy by the machine learning model, her gastrointestinal symptoms and changes in eating habits have begun to increase her risk of malnutrition.

In this study, GLIM was retrospectively used to evaluate the nutritional status of older inpatients, and an assisted diagnosis model of malnutrition with a quite high performance was established by the ML algorithm. Using ML, we validated and visualized a tool that incorporated medical history and laboratory tests for identifying malnutrition in older inpatients based on the GLIM criteria. Through the SHAP value, we identified the most important features and their recommended cutoff values in evaluating malnutrition in older people. We also interpreted and visualized the model in individuals. Our model provided insights to the diagnosis of malnutrition and possesses the characteristic of high interpretability.

To the best of our knowledge, this is the first study using ML to validate nutritional status in older patients. The etiology of malnutrition in older individuals is multifactorial and consists of physiological, social, and economic parameters, often different from the young. Thus, it is important to develop a practical malnutrition tool and optimal cutoff values targeted to older people. The AUROC of the ML model achieved 0.921, which is comparable to findings by previous studies performed in different populations [17,18,30]. This could help clinicians acquire fast malnutrition diagnosis and implement interventions. According to the visualization of the model, the most important features are among phenotypic criteria in GLIM, that is, weight change, low BMI, and muscle mass. Furthermore, through the interpretation of individuals, we could also provide the extent to which individual risk factors influence diagnostic outcomes. This would allow doctors to know the focus of treatment more precisely. Compared with previous tools, our model could provide more individualized information to direct clinical intervention [17].

According to the analysis of the dependence plot, we could observe the cutoff values of the features with a higher risk of malnutrition. The cutoff point of BMI was approximately 21, which suggests that when BMI of older Chinese people is below 21 kg/m², they are exposed to a higher risk for malnutrition. This coincides with the result from the current national BMI data set [19]. The results also provided helpful laboratory test information for identifying the risk of malnutrition. Based on the visualization for feature importance, albumin, platelet count, lymphocyte count, and hemoglobin were the top 4 laboratory tests.

This study has some limitations. First, the GLIM diagnosis was retrospectively determined. Although the original study collected a wide range of nutritional factors, more information was needed for accurate diagnosis and grading. In this study, GLIM-recommended muscle measurements such as dual-energy absorptiometry or bioelectrical impedance are not available; calf circumference was used as an alternative measure. In addition, we established the GLIM diagnosis based on Asia BMI thresholds, so the applicability of the model in a non-Asian population should be reevaluated. Furthermore, the training and test sets were derived from the same study, which means the model needs to be validated by prospective research. Lastly, the data were slightly imbalanced, which means the precision of the model training was relatively low because of the lower number of malnutrition cases.

In conclusion, using ML, we could get an accurate and dynamic diagnosis of malnutrition based on the GLIM criteria. This model could be used as a decision tool to assist clinicians. The cutoff values of laboratory tests also provide references for the identification of malnutrition in Chinese older inpatients. Further studies are needed to simplify and validate the model in clinical conditions.