Cholesterol-to-Lymphocyte Ratio as a Predictor of 1-Year Unplanned Readmission in Patients with Coronary Artery Disease and Type 2 Diabetes Mellitus: A Retrospective Cohort Study

Background

Despite notable progress in the prevention and treatment of cardiovascular disease (CVD) in recent decades, both its incidence and mortality have continued to rise, leaving CVD the leading cause of global disease burden.^1,2 Coronary artery disease (CAD), as one of the most common manifestations, poses substantial public health and socioeconomic challenges because of its high rates of morbidity and mortality.³ The growing prevalence of type 2 diabetes mellitus (T2DM) further intensifies the clinical burden associated with CAD.⁴ A growing body of evidence indicates that patients with both CAD and T2DM—owing to chronic metabolic disturbances, ongoing inflammation, and endothelial dysfunction—tend to present with greater clinical complexity and a significantly higher risk of unplanned hospital readmission compared with those with CAD alone.^5–8 Such recurrent readmissions often reflect inadequate disease control, elevate healthcare costs, and negatively impact patients’ quality of life. Identifying simple, accessible markers to help predict readmission risk in this patient population is therefore of considerable clinical value. Although conventional risk assessment models incorporate demographic characteristics and comorbidities, they may not fully capture the combined effects of metabolic imbalance and systemic inflammation that characterize patients with CAD and T2DM.

Dyslipidemia and inflammation play central roles in the development and progression of both CAD and T2DM.^9–11 Serum cholesterol reflects underlying lipid metabolism and is closely linked to the advancement of atherosclerosis, while lymphocyte counts serve as a marker of immune and inflammatory status; reduced lymphocyte levels often suggest systemic inflammation and have been associated with worse cardiovascular outcomes.^12–14 Based on these considerations, we hypothesized that the cholesterol-to-lymphocyte ratio (CLR), which integrates lipid and inflammatory information, may be associated with the risk of unplanned readmission within one year in patients with CAD and T2DM.

Compared with single inflammatory or lipid indicators, CLR combines routinely available laboratory parameters and may provide a more comprehensive reflection of cardiometabolic and immune status without additional testing burden. In addition, machine learning approaches have increasingly been applied to clinical prediction tasks because of their ability to model complex relationships among variables, while SHapley Additive exPlanations (SHAP) can help improve model interpretability.Accordingly, this study aimed to evaluate the predictive value of CLR for 1-year unplanned readmission among patients with CAD and T2DM.

Methods

This single-center retrospective cohort study included patients hospitalized at Minda Hospital of Hubei Minzu University between January 1 and December 31, 2023, who were diagnosed with both coronary artery disease (CAD) and type 2 diabetes mellitus (T2DM) based on coronary angiography. Patients were excluded if they had a known malignancy or other terminal illness with a life expectancy of less than one year, missing key clinical data or loss to follow-up, or if they died during the index hospitalization (Figure 1). CAD was defined as the presence of ≥50% stenosis in at least one major coronary artery confirmed by coronary angiography, and T2DM was diagnosed according to established clinical criteria documented in medical records. All enrolled patients were followed until December 31, 2024. The primary endpoint was the first unplanned rehospitalization for any cause. Unplanned readmission was defined as an unscheduled hospital admission occurring after discharge from the index hospitalization and excluded elective or planned admissions. Individuals without readmission were censored at the end of follow-up. The study protocol was approved by the Ethics Committee of Minda Hospital of Hubei Minzu University (Approval No. K2024005) and was conducted in accordance with the Declaration of Helsinki. Owing to the retrospective design of the study and the use of anonymized clinical data, individual written informed consent was not required.

Figure 1 Study flow diagram.

Clinical data were collected from electronic medical records, including demographic characteristics (age, sex, smoking history), comorbid conditions (hypertension, prior stroke, atrial fibrillation, hyperuricemia, dyslipidemia), medication use (diuretics, β-blockers, anticoagulants), and laboratory findings. Laboratory evaluations included hematologic indices (white blood cell count, neutrophils, lymphocytes, monocytes, hemoglobin, hematocrit, mean corpuscular volume, platelet count, mean platelet volume, plateletcrit), lipid profile (total cholesterol [TC], triglycerides [TG], high-density lipoprotein [HDL], low-density lipoprotein [LDL]), renal and hepatic function tests (potassium, alanine aminotransferase [ALT], aspartate aminotransferase [AST], total and direct bilirubin, creatinine, uric acid, urea), and thyroid-stimulating hormone (TSH). The cholesterol-to-lymphocyte ratio (CLR) was calculated as TC divided by lymphocyte count.¹⁵

To identify key predictors, we used least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation to determine the optimal penalty parameter (λ). All candidate variables collected at baseline were entered into the LASSO model without prior univariable screening to minimize selection bias. Based on these selected features, three predictive models were developed: logistic regression, a multilayer perceptron (MLP), and an extreme gradient boosting (XGBoost) model. Machine learning models were included to explore potential nonlinear relationships among predictors and to compare their performance with conventional regression approaches.

Model performance was assessed in both the training and test sets using accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), precision, recall, and average precision (AP). The dataset was randomly divided into training and test sets at a ratio of 7:3 to evaluate model generalizability. Calibration was evaluated using calibration plots and the Hosmer–Lemeshow goodness-of-fit test. The best-performing model was subsequently used for in-depth interpretation. Model explainability was explored using SHapley Additive exPlanations (SHAP), which quantified the relative contribution and direction of effect of each variable. SHAP analysis was applied to improve transparency and facilitate clinical interpretation of model predictions. Visualization included ranked importance plots and force plots to improve interpretability and clinical applicability.

All statistical analyses were conducted using R version 4.2.2 and Python version 3.9. Categorical variables were expressed as frequencies and percentages [n (%)] and compared using the χ²-test. Continuous variables with normal distribution were reported as mean ± standard deviation and compared using independent-samples t-tests. Non-normally distributed variables were presented as median [IQR] and compared with the Mann–Whitney U-test. A two-sided P < 0.05 was considered statistically significant.

Results

Baseline Characteristics

As shown in Table 1, a total of 973 patients were included in the study, of whom 769 were assigned to the non-readmission group (NRG) and 204 to the readmission group (RG). The age distribution was similar between the two groups, whereas the proportion of male patients was higher in the NRG. Atrial fibrillation was more common among patients in the RG, while the prevalence of hypertension, prior stroke, hyperuricemia, and dyslipidemia did not differ significantly. Regarding medication use, both β-blockers and oral anticoagulants were prescribed less frequently in the RG. In terms of laboratory findings, CLR was notably higher in the RG (3.63 vs 3.07, P < 0.001). Monocyte counts also differed significantly between groups. Differences were additionally observed in HDL, LDL, and serum creatinine levels, whereas other parameters—such as white blood cell and neutrophil counts, hemoglobin, platelet indices, potassium, liver function tests, uric acid, urea, and TSH—were largely comparable.

Table 1 Descriptive Characteristics of Overall Participants

Variable Selection

All collected variables were entered into a LASSO regression model to screen for potential predictors. As shown in Figure 2, six variables with non-zero coefficients were retained at the optimal λ value determined through cross-validation: sex, atrial fibrillation, β-blocker use, oral anticoagulant use, age, and CLR.

Figure 2 Feature selection using LASSO regression.

Model Performance

The performance of the three predictive models is summarized in Figure 3. In the training set, logistic regression achieved an AUC of 0.80, an accuracy of 0.77, a sensitivity of 0.63, and a specificity of 0.81. The MLP model demonstrated the highest sensitivity (0.82) with an AUC of 0.84, although at the cost of lower specificity (0.73). XGBoost delivered the strongest overall performance, yielding an AUC of 0.87, an accuracy of 0.78, a sensitivity of 0.81, and a specificity of 0.77. Model performance declined slightly in the test set, but the relative ranking remained unchanged, with XGBoost maintaining the best balance between discrimination and calibration (AUC 0.81, accuracy 0.72). As shown in Figure 4a, XGBoost consistently provided superior discrimination, while the calibration curves in Figure 4b indicate that all models performed well, with XGBoost demonstrating the closest agreement between predicted and observed probabilities.

Figure 3 Heatmap of predictive performance of different models in training and test sets.

Figure 4 ROC and calibration curves of the predictive models in the training and test sets. (a) ROC curves of the predictive models. (b) Calibration curves of the predictive models.

Model Interpretation

SHAP analysis results are shown in Figure 5a. CLR emerged as the most influential predictor in the model, followed by β-blocker use, oral anticoagulant use, and sex. Age and atrial fibrillation showed smaller SHAP value distributions centered near zero, indicating a more limited contribution to the prediction. As illustrated in Figure 5b, the mean absolute SHAP values further confirmed the prominence of CLR (0.42), underscoring its key role in explaining model output.

Figure 5 SHAP-based feature importance analysis of the optimal model (a): value distribution; (b): mean absolute SHAP values).

Discussion

In our study, CLR was significantly higher in patients who experienced unplanned readmission. Using LASSO regression, we identified six key variables—sex, atrial fibrillation, β-blocker use, oral anticoagulant use, age, and CLR—that were most closely associated with readmission risk. Among the predictive models constructed, the XGBoost algorithm demonstrated the strongest and most stable performance in both the training and test sets, showing superior discrimination and acceptable calibration. SHAP analysis further highlighted CLR as the most influential factor in model output, with a substantially higher mean absolute SHAP value than the other predictors, reinforcing its central role in readmission risk prediction. These findings suggest that CLR may provide additional supportive information for identifying patients at increased risk of rehospitalization.

CLR reflects a combined signal of inflammatory activity and lipid metabolism. An elevated CLR suggests the coexistence of increased cholesterol levels and reduced lymphocyte counts, pointing to a heightened inflammatory state with compromised immune regulation.^16,17 Inflammation plays a central role in the pathophysiology of cardiovascular disease: inflammatory cytokines damage the endothelium, upregulate adhesion molecules, and promote monocyte and lymphocyte recruitment, all of which accelerate atherosclerosis and contribute to plaque instability.^18–21 At the same time, lipid abnormalities exacerbate vascular injury, as cholesterol and LDL deposition induce oxidative stress and local inflammation, further promoting plaque progression.^22–24 An increased CLR may therefore capture the synergistic interplay between dyslipidemia and systemic inflammation, creating a cycle of endothelial damage and thrombogenicity that ultimately increases cardiovascular risk. Lymphopenia may also weaken anti-inflammatory defenses, allowing chronic low-grade inflammation to persist and facilitate vascular deterioration. Compared with single inflammatory or lipid indicators, CLR integrates routinely available laboratory parameters and may better reflect the cardiometabolic and immune burden in patients with concomitant CAD and T2DM.

Our results also shed light on the protective effect of β-blocker and oral anticoagulant use in the outpatient setting, both of which were associated with a reduced risk of 1-year readmission.^25–28 Conversely, atrial fibrillation was strongly linked to higher readmission rates, which may be explained by its contribution to hemodynamic instability, heart failure exacerbation, and elevated thromboembolic risk.^29–31 We also observed higher readmission risk among younger patients and women.^32–36 For younger individuals, this may relate to lower disease awareness or poorer adherence to treatment, while in women, differences in hormonal status, lipid profiles, and glucose metabolism may contribute to increased cardiovascular vulnerability. These findings align with previous studies and add further support to the robustness of our conclusions. However, these associations should be interpreted cautiously given the observational nature of the study.

Machine learning models were incorporated to explore potential nonlinear interactions among clinical variables and to compare their performance with conventional regression analysis. Among the models evaluated, XGBoost demonstrated superior overall performance, which may reflect the ability of ensemble tree-based algorithms to capture complex interactions among metabolic, inflammatory, and treatment-related factors. Furthermore, SHAP analysis enhanced model transparency by quantifying the relative contribution of each predictor, thereby facilitating clinical interpretation of model outputs.

Several limitations merit attention. First, the study was conducted at a single center, and the relatively homogeneous patient cohort may limit the generalizability of the findings. Second, the follow-up duration was restricted to one year, preventing evaluation of long-term outcomes. Third, potentially important factors such as lifestyle behaviors, psychosocial influences, socioeconomic conditions, and medication adherence were not captured and may have influenced readmission risk or model performance. Lastly, the observational design precludes any causal inference between CLR and unplanned readmission. In addition, external validation was not performed, and the relatively limited sample size may affect model robustness. Future multicenter prospective studies with longer follow-up periods are needed to validate these results and further clarify the clinical value of CLR.

Conclusion

The CLR was identified as a meaningful predictor of 1-year unplanned readmission in patients with CAD and T2DM. In the multivariable models evaluated, CLR contributed substantially to risk assessment, showing moderate discriminatory performance. These findings suggest that CLR may serve as a readily available biomarker to assist in early risk stratification in this high-risk population. However, given the retrospective single-center design and absence of external validation, further studies are needed to confirm its predictive value and to explore its potential role in guiding clinical decision-making.