Construction and validation of a predictive model for the risk of malnutrition in hospitalized patients over 65 years of age with malignant tumours: a single-centre retrospective cross-sectional study

Study design and participants

This retrospective cross-sectional study was conducted at the Anning First People’s Hospital Affiliated to Kunming University of Science and Technology. This study retrospectively analyzed clinical data from 3,387 older inpatients aged 65 years and older with malignant tumours who were admitted between July 1, 2021, and December 31, 2023.The data were divided into training and validation sets based on patient admission timeframes, maintaining an 8:2 ratio, with the training set comprising 2,715 patients admitted between July 1, 2021, and July 31, 2023, and the validation set consisting of 672 patients admitted from August 1, 2023, to December 31, 2023.This study was conducted following the Declaration of Helsinki by the World Medical Association. It was approved by the Ethical Review Committee of the First People’s Hospital of Anning, affiliated with Kunming University of Science and Technology (Ethical Approval No. 2024-020 (Self-horizontal)-01). Given its retrospective nature, the requirement for informed consent was waived. All data were treated confidentially and analyzed anonymously to protect patient privacy, no information identifying individual participants was available to the authors during or after data collection.

The inclusion criteria for this study were as follows: (I) patients aged 65 years or older; (II) patients with histopathologically confirmed malignant tumours; and (III) patients with complete data available during their hospitalization, retrievable from the Hospital Information System (HIS), Laboratory Information Management System (LIS), or Electronic Medical Record System (EMRS), encompassing comprehensive information. The exclusion criteria encompassed patients with severe cardiac, hepatic, renal, or pulmonary dysfunction, and those with bleeding from vital organs such as the digestive, respiratory tract, or liver. Patients who were admitted to the Intensive Care Unit (ICU) or Coronary Care Unit (CCU), those anticipated to have a life expectancy of <3 months and those with >20% missing data were also excluded from the study.

Sample size estimation

The calculation of an adequate sample size for predictive studies, encompassing both modelling and validation phases, depends on the number of outcome events. Following guidelines from predictive modelling research, including those described by Hosmer, Lemeshow & Sturdivant (2013), the number of positive malnutrition cases should ideally exceed the number of examined risk factors by 5–10 times to mitigate the likelihood of overfitting and prediction errors. In this study, 26 risk factors were considered. A minimum of 130 malnutrition cases was deemed necessary to maintain prediction accuracy. The modelling sample size was 3,387, with 1,369 positive instances, which exceeds the requisite minimum to ensure a robust assessment.

Data extraction and nutritional status evaluation

The analysis incorporated a range of variables, including demographic details, medical history, clinical observations, and laboratory indicators from patients admitted to our hospital with malignant tumours between July 2021 and December 2023. The selection of risk factor variables for initial screening was informed by potential causes of malnutrition in older patients with malignant tumours, existing literature, and insights derived from clinical experience and consultations with nutrition experts. The collected demographic information encompassed gender, age, marital status, Hukou (household registration) status, and BMI. Moreover, medical history was documented, encompassing habits such as smoking and alcohol consumption, and the presence of conditions such as hypertension, type 2 diabetes mellitus, coronary artery disease, and hyperlipidaemia. Details of any surgical interventions within the last year, and the incidences of co-infections were recorded. The clinical information gathered encompassed the clinical diagnosis, the TNM stage of the tumour, treatment plans, the presence of thoracic and abdominal effusions, the NRS-2002 score, and the KPS (Karnofsky Performance Status) score.

Laboratory findings considered in this analysis encompassed red blood cell (RBC) count, white blood cell (WBC) count, haemoglobin (HGB) levels, platelet (PLT) count, neutrophil-to-lymphocyte ratio (NLR), serum albumin (ALB), aspartate aminotransferase (AST),alanine aminotransferase (ALT), total bilirubin (TBIL), creatinine (Cr), and blood urea nitrogen (BUN) levels.

The GLIM criteria were adopted for malnutrition diagnosis. Nutritional screening constituted the first phase, using the NRS-2002 as the initial tool. A score of ≥3 on the NRS-2002, verified through medical records, indicated a malnutrition risk. Subsequently, the second stage involved a detailed malnutrition assessment based on five criteria: involuntary weight loss (>5% within 6 months or >10% at 6 months), low BMI (Asian standard (Kanazawa et al., 2005): BMI <18.5 kg/m² (<70 years), <20 kg/m² (≥70 years)), and decreased muscle mass (below the standard values for different body composition measures; Decreased muscle mass is defined as a reduction in muscle mass below accepted reference values, determined through physical measurements conducted by clinical staff using standardized anthropometric techniques. These measurements, including skinfold thickness and circumferences, were used to assess and compare muscle mass to reference values, identifying significant reductions where applicable as part of the phenotypic criteria. The remaining two etiologic criteria involve reduced food intake or absorption (food intake <50% for >1 week, or reduced intake for >2 weeks, or chronic gastrointestinal dysfunction impairing food digestion and absorption), and the disease/inflammation burden. Inflammation burden was assessed using clinical biomarkers with established laboratory reference ranges. Specifically, inflammation was quantified using C-reactive protein (CRP) levels and erythrocyte sedimentation rate (ESR). For this study, Significant inflammation was defined according to the measurement reference ranges provided by our testing instruments, with CRP levels >10 mg/L and/or ESR >30 mm/hr as indicators of elevated inflammation. To diagnose malnutrition, at least one phenotypic criterion and one etiologic criterion, including these inflammation cut-off values, must be met.

Model development and validation process

In this study, 2,715 patients admitted consecutively between July 1, 2021, and July 31, 2023, were included in the training set. Meanwhile, 672 patients admitted from August 1, 2021, to December 31, 2023, comprised the validation set, maintaining an 8:2 ratio. The modeling process classified malnutrition cases in the training set as the dependent variable, thereby differentiating participants into malnutrition and non-malnutrition groups. A forward selection method was employed to identify and select statistically significant variables (P < 0.05) for inclusion in the multivariate logistic regression analysis. This analysis was conducted to evaluate the influence of these factors on the likelihood of malnutrition, with significance assessed using P-values and the impact quantified using odds ratios. Column plots were generated using the “rms” package in R software (version 4.2.0) (R Core Team, 2023; Harrell, 2023). The model’s validity was confirmed by employing Statistics software (version 26.0) (IBM, Armonk, NY, USA) to generate receiver operating characteristic (ROC) curves, calculate the area under the curve (AUC) for both sets and compare the model’s predictive accuracy against the NRS-2002 scale, thereby evaluating its discriminative capabilities. The calibration curves for the training and validation sets were developed using the “rms” package in R programming language (version R 4.2.0). Model calibration was assessed through visual inspection and the Hosmer-Lemeshow test, a statistical test used to evaluate the goodness-of-fit for logistic regression models by comparing the observed and predicted event rates across different subgroups (Hosmer, Lemeshow & Sturdivant, 2013). The test helps to determine whether the model adequately fits the data or if there are systematic discrepancies between the observed and predicted values. Clinical applicability was evaluated through decision curve analysis (DCA) using the “rmda” package. The model underwent internal validation with bootstrap resampling conducted 1,000 times, considering P < 0.05 statistically significant.

Statistical analysis

Data were inputted using Excel version 2022, with a double-entry method applied to minimize data entry errors. Data summarization and analyses were performed using SPSS 26.0 software (IBM SPSS 26.0; SPSS Inc., Chicago, Illinois, USA). Measurement data following a normal distribution were presented as $\overline{\mathrm{x}}$ ± s, and t-tests were applied for comparisons between groups. Count or categorical data were expressed as frequencies (n) and percentages (%), with chi-square tests or Fisher’s exact test used for group comparisons. Medians and interquartile ranges represented measurement data not adhering to a normal distribution (M (P25, P75)). Regarding missing data, particularly for muscle mass measurements, we utilized multiple imputation methods. This approach involves generating several complete datasets to replace missing values, performing analyses on each dataset, and then combining the results to provide a comprehensive estimate. This method reduces bias and enhances the robustness of our findings despite the missing data. The nomogram’s performance was determined using the concordance index (C-index), which evaluates the agreement between predicted probabilities and actual observations of malnutrition. Bootstrapping with 500 resamples was performed. A higher C-index signifies more accurate prognostic differentiation. The total points for each patient in the validation cohort were calculated utilizing the constructed nomogram to assess the risk of malnutrition. The C-index, calibration curve, and DCA were derived from this analysis using the R package (version 4.2.0) for all nomogram-related computations.

Characteristics and predictive factors of malnutrition in development and validation cohorts

After applying the criteria, 3,387 patients were included in the study, as shown in Fig. 1.The characteristics of patients aged ≥65 years in the development and validation cohorts are listed in Table 1. These cohorts displayed similarities in gender, Hukou status, TNM staging, and Karnofsky Performance Status (KPS) scores. However, differences were observed in alcohol consumption history, hypertension, type 2 diabetes mellitus, hyperlipidaemia, and coronary heart disease (CHD). Table S1 presents an exhaustive analysis of malnutrition indicators in patients aged 65 and older diagnosed with malignancies. The table encompasses detailed data on phenotypic factors, including the number of patients experiencing significant weight loss, the prevalence of low body mass index (BMI), and the degree of muscle mass reduction. Specifically, Table S1 delineates the proportion of patients experiencing weight loss, stratified by severity, alongside the count of patients exhibiting low BMI values below predefined thresholds. Additionally, it elucidates the methodologies employed by clinical staff to assess muscle mass, encompassing the anthropometric techniques utilized, and details the approach to handling missing muscle mass data through multiple imputation methods. Table 2 summarizes the association between clinical characteristics and malnutrition in the development cohort of older cancer inpatients. No significant differences were noted in gender, smoking history, hyperlipidaemia, and alanine aminotransferase (ALT) between the negative and positive malnutrition groups (P > 0.05). Conversely, factors such as age, marital status, Hukou status, tumour type, TNM staging, surgical history within the past 12 months, KPS score, alcohol consumption history, co-infection, hypertension, type 2 diabetes mellitus, CHD, presence of ascites or pleural effusion, counts of WBC, RBC, PLT, HGB, total bilirubin, aspartate aminotransferase (AST), Cr, urea, and the NLR were significantly associated with the development of malnutrition. Table S2 provides the methods used for assigning variables in the logistic regression analysis during model construction. This includes strategies such as setting threshold values, integrating laboratory reference ranges, applying clinical expertise, and categorizing variables to identify the most effective assignment methods. Table S3 presents the Univariate Logistic Regression Analysis Identifying Risk Factors for Malnutrition in the Training Cohort, showcasing the initial screening of potential predictors. The multivariate analyses identified 10 variables as independent risk factors for malnutrition (Table 3). These included the system of the tumour, patient age, KPS score, TNM staging, history of alcohol consumption, co-infection presence, ascites or pleural effusion, HGB and Cr levels, and the NLR, all of which were significant predictors of malnutrition in older patients with cancer.

Prevalence of malnutrition in older patients with malignant tumours

Analysis of the case data from the development and validation cohorts revealed 1,369 instances of malnutrition, resulting in a prevalence rate of 40.42%. Analysis of malnutrition rates by age, gender, Hukou status, and tumour type revealed that older patients were most susceptible to malnutrition, with men experiencing a higher prevalence than women. Patients from rural areas had a higher malnutrition rate than those from urban areas; digestive system tumours were most frequently associated with malnutrition. Stage IV tumours were most frequently associated with malnutrition, with the distribution of malnutrition rates across various subgroups depicted in Fig. 2.

Development of the malnutrition prediction nomogram

The prediction model’s 10 independent variables (age, presence of a digestive system tumour, TNM staging, KPS score, history of alcohol consumption, history of infections, presence of combined thoracic and abdominal fluid, HGB levels, Cr levels, and the NLR) were employed to construct a nomogram. The logistic regression analysis identified these variables as significant predictors of malnutrition. Weights for each predictor in the nomogram were derived from their respective β-coefficients. The instructions for using the nomogram are outlined in the caption of Fig. 3A, which includes a comprehensive guide. Figure 3B displays an interactive version of the nomogram, enhancing user interaction and interpretation.

To use the nomogram, a vertical line was drawn from the point on the horizontal axis corresponding to each independent variable’s value for a patient to the “Point” value representing a specific score. The scores for the five independent variables were summed to calculate the total score, after which a vertical line was drawn downwards. The predicted risk of malnutrition for the patient was determined by locating the corresponding point on the horizontal axis labelled “Risk”.

Evaluation of the predictive performance, calibration, and clinical applicability of the malnutrition prediction models

The Hosmer–Lemeshow test results, with a P-value of 0.688, suggest no significant difference between predicted and observed malnutrition rates across subgroups, supporting the model’s calibration. Internal bootstrap validation revealed a C-index of 0.793, indicating the model’s effective predictive capability. The predictive accuracy of the malnutrition nomogram was further assessed through ROC analysis. For the training set (Fig. 4A) and validation set (Fig. 4B), the AUC values of the nomograms were 0.793 (95% confidence interval (CI) [0.776–0.810]) and 0.832 (95% CI [0.801–0.863]), respectively. The P-value of 0.564 indicates no significant difference in predictive performance between the training and validation cohorts. In contrast, the predictive model for the training set demonstrated a significantly superior AUC compared to the NRS-2002 scale, which had an AUC of 0.737 (P = 0.012), as illustrated in Fig. 4C. These results indicate the model’s enhanced capability to predict malnutrition risk in older patients with malignant tumours. The nomogram demonstrated robust calibration within the validation cohort (Figs. 5A and 5B), indicating its excellent discriminative ability. Furthermore, bar charts were generated to assess the discriminatory power of the nomogram in identifying malnutrition based on calculated risk scores. DCAs were conducted for the nomograms predicting malnutrition (Figs. 6A and 6B) in both the training and testing sets. Threshold probabilities ranging from 0 to 0.3 for malnutrition were identified as most advantageous for predicting malnutrition using our nomograms. The DCAs revealed that employing the nomogram for malnutrition prediction would benefit more than universally classifying patients as either malnourished or well-nourished.