Diabetes mellitus (DM) is a major chronic disease worldwide, with prevalence rates steadily rising over recent decades (1). Diabetes and its associated complications pose significant threats to patient health and quality of life, greatly impacting patients’ daily activities and potentially leading to mortality in severe cases. Among these complications, hypoglycemia is particularly critical to prevent and manage due to its frequent occurrence and substantial health risks for diabetic patients (2). Hypoglycemic events in diabetic patients are multifactorial, influenced by medications, diet, lifestyle, and comorbidities (3–5). Recently, stringent glycemic control strategies have been associated with an increased risk of hypoglycemia. However, hypoglycemia can undermine the long-term benefits gained from good glycemic management, underscoring the importance of carefully balancing the benefits and risks of intensive glucose management.

Hospitals play a central role in glycemic control, adjustments of antidiabetic medications, and individualized care for diabetic patients, placing heightened responsibilities on healthcare professionals. Patients with diabetes commonly exhibit multiple comorbidities that contribute independently to hypoglycemia risk. Studies have identified older age, renal impairment, liver dysfunction, poor nutritional status, inappropriate medication use, and debilitating diseases as critical risk factors for hypoglycemia. Clinical practitioners thus must analyze these risk factors systematically to accurately assess the probability of hypoglycemia and implement effective preventive measures.

Predicting hypoglycemia accurately, however, remains challenging due to the complex interplay of various clinical and biological factors. Although multiple studies have confirmed insulin therapy (6), impaired renal function, and suboptimal glycemic control as significant predictors of hypoglycemia, reliably forecasting such events remains difficult. Traditional logistic regression models have been widely employed to identify risk factors but are limited by their assumption of linear relationships between predictors and outcomes.

Given the limitations of traditional statistical methods in capturing complex clinical relationships, this study was motivated to explore advanced machine learning algorithms for predicting hypoglycemia severity in hospitalized T2DM patients. Thus, the current study aimed to develop and systematically compare three predictive models—multinomial logistic regression, XGBoost, and RF—to identify independent risk factors and predict hypoglycemia severity. We further conducted a feature importance analysis to highlight clinically significant predictors of hypoglycemia risk. This comparative approach is expected to improve early clinical intervention and personalized management. The paper is structured as follows: the methods section describes the study population, data collection, and model development; the results section presents patient characteristics and model performances; the discussion elaborates clinical implications and comparative insights from the modeling results; finally, limitations and future research directions are provided.

Hypoglycemia remains a significant challenge in managing hospitalized patients with type 2 diabetes mellitus (T2DM), often resulting from complex interactions among clinical factors such as medication usage, renal function, and glycemic variability. Traditionally, logistic regression models have been widely employed to identify predictors of hypoglycemia risk due to their interpretability and simplicity. However, these models assume linear relationships and may not effectively capture intricate interactions among clinical variables, limiting their predictive capabilities in clinical practice.

With advancements in machine learning techniques, models such as Extreme Gradient Boosting (XGBoost) and Random Forest (RF) have shown remarkable performance in predicting complex clinical outcomes due to their ability to capture nonlinear interactions and handle high-dimensional data. Several previous studies have explored machine learning approaches to predict hypoglycemia in diabetic patients. For instance, Melih Agraz et al. used multi-view collaborative training of machine learning models on imbalanced datasets to improve hypoglycemia prediction (7). Harald Witte et al. used XGBoost alone to train models related to glycemic decompensation (8). Christopher Duckworth et al. also used XGBoost alone to predict glycemic decompensation (hyperglycemia and hypoglycemia) (9). Mai Shi et al. developed a multidimensional ML model based on electronic health records (EHR) to predict hypoglycemia in the elderly population (10). However, most of these studies have focused solely on hypoglycemia or hyperglycemia, lacking analysis of the severity of hypoglycemia. Additionally, few studies have used real hospital patient data to comprehensively compare multiple models. This study addresses this shortcoming by evaluating logistic regression, random forest, and XGBoost models to perform multi-class classification of hypoglycemia severity using clinical variables.

In the present study, we developed and systematically compared three predictive models—multinomial logistic regression, XGBoost, and random forest (RF)—to identify key predictors of hypoglycemia severity among hospitalized patients with diabetes. Our results demonstrated that both machine learning models (XGBoost and RF) exhibited superior discriminative performance compared to traditional logistic regression, especially for detecting moderate-to-severe hypoglycemia, this has some modeling studies with the same results (11). We utilized the One-vs-Rest strategy, a standard and effective approach for generating ROC curves in multiclass classification problems, allowing robust evaluation and comparison of each model’s predictive performance across hypoglycemia categories.

As detailed in Table 4 , the random forest model demonstrated the best overall predictive performance, with the highest accuracy (93.3%), Kappa coefficient (0.873), and macro-average AUC (0.960). These values reflect excellent classification agreement and discrimination for hypoglycemia severity prediction. The XGBoost model showed slightly lower but still robust performance (accuracy 92.6%, Kappa 0.860, AUC 0.955), confirming its strength in handling complex nonlinear data. In contrast, the logistic regression model yielded lower accuracy (83.8%) and Kappa (0.685), along with a substantially lower AUC (0.788), possibly due to its linear assumptions and inability to capture complex feature interactions. These findings highlight the superior performance of ensemble learning models in identifying subtle patterns and interactions in clinical data, and support their potential value in hypoglycemia risk stratification.

The dataset used to construct our predictive models comprised common clinical and laboratory variables routinely available for hospitalized patients with type 2 diabetes mellitus (T2DM). The logistic regression model identified seven statistically significant predictors (P < 0.05), while both XGBoost and RF models highlighted the top 10 most influential variables. Notably, six predictors—age, Charlson Comorbidity Index (CCI), serum creatinine, HbA1c, mean glucose levels, and insulin use—emerged consistently as significant factors across all three models. A number of previous modeling studies and factor analysis studies have similarly given the same conclusions for one or more of these factors (12–14). While previous studies have individually confirmed the significance of one or several of these factors, our research uniquely integrated multiple models for comprehensive comparison, offering a broader perspective than single-model studies previously reported.

A key advantage of our study lies in its reliance on routinely collected clinical data. This facilitates rapid assessment of hypoglycemia risk without additional specialized testing, thus granting medical staff valuable time to implement preventive strategies and appropriate treatments.

Regarding predictive variables, HbA1c has consistently been identified as a crucial factor in hypoglycemia research. Prior studies have reported that both excessively high HbA1c levels (>9%) and overly stringent glycemic control (HbA1c <7%) are associated with increased hypoglycemia risk (15). Additionally, insulin use within these HbA1c ranges further exacerbates hypoglycemia risk. Both our univariate and multivariate analyses supported these findings, reinforcing that elevated HbA1c levels and insulin therapy are critical risk factors for hypoglycemia in diabetic patients.

In recent years, growing attention has focused on the association between the Charlson Comorbidity Index (CCI) and hypoglycemia. Most prior studies investigating comorbidities typically examined only one or a few components of the CCI, such as cardiovascular disease, renal impairment, or malignancy (16). These comorbid conditions have been associated with glucose instability and heightened risk of adverse glycemic events. Patients with multiple chronic diseases often experience altered drug metabolism, polypharmacy, and malnutrition, potentially interfering with glucose regulation and insulin sensitivity. Incorporating the CCI into all three predictive models in our study enhances its utility as a comprehensive clinical indicator. Unlike individual diagnoses, the CCI provides an aggregated measure of overall disease burden, capturing complex interactions between comorbidities and thereby improving the generalizability and interpretability of the predictive models. Integrating the CCI into predictive assessments may assist clinicians in accurately stratifying risk, particularly among elderly hospitalized patients or those with multiple chronic diseases, who inherently have higher susceptibility to hypoglycemic episodes (17). Early identification of high-CCI patients could facilitate personalized monitoring, medication adjustments, and tailored nutritional plans, thereby mitigating hypoglycemia risk.

Lastly, we conducted feature importance analyses for the top ten variables in the XGBoost and RF models. Both models consistently identified glycemic control and glucose variability as key predictors. Additionally, XGBoost emphasized basal metabolic parameters (e.g., creatinine and C-peptide), whereas the RF model placed greater emphasis on medication use. These findings highlight nuanced differences between machine learning approaches, underscoring their ability to provide targeted clinical insights into managing and preventing hypoglycemic events in hospitalized diabetic patients.

In this study, we developed and compared three prediction models - multinomial logistic regression, XGBoost, and random forest - to identify factors associated with hypoglycemia severity in hospitalized diabetic patients. Our results showed that both machine learning models outperformed the traditional logistic regression model in terms of predictive performance, with higher accuracy and AUC values in all hypoglycemia categories. This time, the three models were designed with all commonly used clinical indicators, without the need for special tests, to prevent hypoglycemia as early as possible in type 2 diabetic patients and to avoid the adverse consequences caused by hypoglycemia. In a comparison of the three models and the focus of the models, the six variables of age, Charlson comorbidity index, creatinine, glycosylated hemoglobin, mean blood glucose level, and insulin use were consistently identified as core predictors in almost all models (7, 18, 19), This further enhances their clinical relevance. Future studies should focus on the prospective validation and practical application of these models to assess their clinical utility and impact on patient safety.

This study has several strengths. First, it comprehensively compared the performance of traditional and ensemble machine learning models in predicting hypoglycemia severity, highlighting the superior capability of tree-based algorithms. Second, it utilized a multiclass classification approach to stratify hypoglycemia into different severity levels, which has greater clinical relevance than binary classification. Third, model performance was rigorously evaluated using stratified 5-fold cross-validation, and visual comparisons were provided through ROC curves and boxplots, enhancing interpretability. Finally, the study was based on a large cohort of hospitalized patients with diabetes, offering real-world insights.

However, some limitations should be acknowledged. The data were derived from a single center, which may limit the generalizability of the findings. External validation using independent datasets from other institutions is needed. Additionally, the reliance on internal cross-validation without a separate test set may lead to optimistic performance estimates. Some potentially important predictors, such as postprandial glucose or nutritional interventions, were not available in the dataset. Moreover, the models have not yet been tested in real-time clinical settings, and their clinical utility remains to be validated.