BMI-based prediction models for short-term and long-term mortality following coronary artery bypass grafting using the MIMIC-IV database: a retrospective cohort study

Abstract

Objective

This study aims to evaluate the relationship between obesity (measured by Body Mass Index (BMI)) and postoperative mortality in patients undergoing coronary artery bypass grafting (CABG) and to use machine learning algorithms to assess key factors in order to explore the ‘obesity paradox’ phenomenon.

Method

Data were obtained from Medical Information Mart for Intensive Care IV (MIMIC-IV) V.3.0. We included adult patients who underwent CABG, excluding those with an Intensive Care Unit (ICU) stay of <24 hours or missing BMI data. Primary outcomes were 7-day, 14-day,28-day and 365-day all-cause mortality. Patients were categorised by BMI into six groups. Logistic regression, Kaplan-Meier and restricted cubic spline analyses were performed with subgroup analyses. The random forest and Boruta algorithm were used for key factor identification. Multiple machine learning models were built and assessed using area under the curve (AUC) and decision curve analysis.

Result

Among 5790 patients who underwent CABG, being overweight (BMI 25–30) predicted the lowest 365-day mortality (adjusted OR<1). BMI showed a U-shaped association with mortality, with the nadir of risk observed between 25–35 kg/m². The protective effect persisted in patients aged ≥65 years. Key mortality drivers differed by BMI: acute physiology, comorbidity burden, metabolic stability and metabolic liver dysfunction. Extreme gradient boosting achieved the highest 365-day mortality prediction (AUC=0.70) with favourable clinical utility.

Conclusions

The obesity paradox is observed among patients following CABG, with distinct BMI-specific predictive factors of mortality risk identified across BMI categories. Consequently, risk-stratified monitoring strategies—tailored to BMI-defined subgroups—are warranted.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Obesity is a well-known risk factor for coronary artery disease.

• The ‘obesity paradox’ has been observed in some cardiac surgery studies, where overweight or patients with mild obesity have better postoperative outcomes than normal-weight individuals.

• However, evidence in patients who underwent coronary artery bypass grafting (CABG) remains inconclusive, and few studies have used machine learning to explore Body Mass Index (BMI)-specific mortality risk factors.

WHAT THIS STUDY ADDS

• This study confirms a U-shaped association between BMI and 365-day mortality after CABG, with the lowest risk in BMI 25–35 kg/m². Key mortality drivers differ by BMI category, and an extreme gradient boosting model achieved moderate predictive performance (area under the curve=0.70).

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• These findings provide empirical support for the moderate relaxation of weight-related intervention restrictions in postoperative management among patients with overweight or class I obesity (BMI 25–35 kg/m²). Furthermore, postoperative management strategies should be individualised according to the BMI category to optimise clinical outcomes.

Introduction

Obesity is a kind of metabolic disease that prevails globally, and its incidence has been on the increase continuously in recent years. According to the statistics of the WHO, the global population of overweight and obese individuals accounts for over 40% of the total population.¹ Obesity not only markedly elevates the risk of multiple chronic diseases (such as dyslipidaemia, type 2 diabetes, hypertension and sleep disorders) but is also considered a crucial risk factor for the onset and progression of coronary artery disease (CAD).²

CAD remains the leading cause of mortality and morbidity on a global scale. The underlying pathology involves atherosclerotic lesions within the coronary arteries that result in inadequate myocardial perfusion and may ultimately lead to myocardial infarction. Coronary artery bypass grafting (CABG) is an effective surgical intervention for patients with CAD, especially those who do not respond to drug therapy or who have multi-vessel disease, which can significantly improve myocardial perfusion and improve clinical prognosis and survival.³

Research in recent years has revealed a paradoxical phenomenon known as the ‘Obesity Paradox’. Although obesity is an important risk factor for CAD, among patients who undergo CABG, postoperative mortality and the incidence of adverse events may be lower in patients with overweight or moderate obesity than in patients with normal weight or underweight.^{4 5} There are also studies that show that obesity confers no protective effect on either early or late mortality in patients undergoing CABG.^{6 7} Therefore, the aim of this study was to evaluate the relationship between obesity and mortality following CABG, as well as to forecast adverse outcomes using machine learning algorithms for improved clinical postoperative management.

Methods

Data source

This was a retrospective cohort study using data from MIMIC-IV (Medical Information Mart for Intensive Care IV) V.3.0 (https://mimic.mit.edu), a large deidentified dataset of patients admitted to the emergency department or an intensive care unit at the Beth Israel Deaconess Medical Center between the years 2008 and 2019 in Boston, Massachusetts, USA.⁸ MIMIC-IV contains data for over 65 000 patients admitted to an ICU and over 200 000 patients admitted to the emergency department. Data collected and presented in the present study were extracted by the author H-JJ, who completed and passed the required training course (certificate number: 63966213).

Inclusion and exclusion criteria

Inclusion criteria:

1. Patients aged ≥18 years.

2. Patients who received surgical treatment, and the name of the surgery included ‘Bypass Coronary Artery’.

Exclusion criteria:

1. Patients with an ICU stay of less than 24 hours.

2. Missing height or weight data.

3. For patients with multiple hospitalisations or ICU admissions, only data from the first hospitalisation were included.

Outcome

The outcomes were all-cause mortality within 7, 14, 28 and 365 days after CABG.

Data extraction

Data extraction was performed using Navicat Premium V.17 software. Patient characteristics, including age, sex, temperature, height and weight, were collected. Information on comorbidities, such as hypertension, diabetes, cerebrovascular disease, congestive heart failure and chronic obstructive pulmonary disease (COPD), was extracted based on the International Classification of Diseases coding system and the Charlson table. Vital signs (heart rate, systolic blood pressure (SBP), respiratory rate and peripheral capillary oxygen saturation (SpO₂)) were all recorded as the average values within 24 hours after ICU admission. Laboratory tests (white blood cell (WBC) count, platelet (PLT) count, percentage of lymphocytes, neutrophil percentage, calcium, chloride, creatinine, glucose, potassium, sodium, haemoglobin, aspartate aminotransferase, alanine aminotransferase (ALT), total bilirubin, haemoglobin A1c, albumin, blood urea nitrogen (BUN), international normalised ratio, haematocrit, mean corpuscular haemoglobin (MCH), MCH concentration (MCHC), mean corpuscular volume (MCV), red blood cell (RBC) count, red cell distribution width (RDW), prothrombin time (PT) and partial thromboplastin time) and blood gas analysis (blood pH, base excess, total CO₂: Total Carbon Dioxide in Blood, anion gap and bicarbonate) were extracted. Furthermore, the Charlson Comorbidity Index and the Acute Physiology and Chronic Health Evaluation III (APACHE III) score were also extracted. Body Mass Index (BMI) is used to assess the degree of obesity with the following formula: BMI=weight (in kg)/height² (in m²).

Statistical analysis

Statistical analyses were performed using R and PyCharm Community Edition V.2024.3, and p values<0.05 (two-tailed) were considered statistically significant. Variables with a missing data rate exceeding 30% were excluded (see online supplemental figure S1 for the missing data rates of each variable). For variables with missing data rates below 30%, multiple imputation was performed using the IterativeImputer algorithm (a total of 7487 data points were interpolated). Extreme values where BMI was ≤14.5 or ≥42.7 were excluded (based on the stem-and-leaf diagram of BMI). Based on the BMI of each patient, they were divided into six groups as follows: Group 1 (underweight, BMI<18.5), Group 2 (normal, 18.5≤BMI<25), Group 3 (overweight, 25≤BMI<30), Group 4 (class I obesity, 30≤BMI<35), Group 5 (class II obesity, 35≤BMI<40) and Group 6 (class III obesity, BMI≥40).

For continuous variables that followed a normal distribution, means and SDs were calculated and compared using analysis of variance. Non-normally distributed continuous variables were summarised as medians with IQRs and analysed using the Kruskal-Wallis test. Categorical variables were presented as frequencies and percentages and evaluated using the χ² test or Fisher’s exact test, as appropriate. Kaplan-Meier survival curves were constructed to visually assess the differences in survival rates at 7 days, 14 days, 28 days and 365 days, while the log-rank test was used to evaluate overall survival time. A logistic regression model was used to estimate the OR and corresponding 95% CI for event occurrence. Model I did not account for covariates, whereas Model II incorporated adjustments for age, sex, heart rate, respiratory rate, SBP, blood oxygen saturation, body temperature, the Charlson Comorbidity Index and the APACHE III score.

By employing a random forest (RF) methodology to evaluate the features within the dataset, we identified the top 10 factors most closely associated with the target variable and visualised their relative importances in a heatmap. Subsequently, these 10 factors were analysed across the different groups to assess intergroup variations and to further investigate their potential influence on the outcomes.

Subgroup analysis

Subgroup analyses were conducted stratified by age and sex. Univariate and multivariate analyses were performed, with multivariate models adjusted for age, sex, heart rate, respiratory rate, SBP, blood oxygen saturation, body temperature, the Charlson Comorbidity Index and the APACHE III score. Notably, in the multivariate analyses of male and female subgroups, adjustment for sex was excluded. Patients were categorised into two groups based on age (<65 years and ≥65 years). Multivariate logistic regression analyses were carried out for each subgroup, and the results were visually represented using forest plots, displaying ORs and their corresponding 95% CIs.

Restricted cubic splines

In this study, we used restricted cubic spline (RCS) to visualise the potential nonlinear relationship between BMI (continuous predictor) and mortality rates (outcome variable) at 7, 14, 28 and 365 days.

Establishment and validation of the prediction models

The Boruta algorithm identifies the most important features by comparing the z value of each feature with the z value of the ‘shadow feature’. This method copies all the real features in each iteration and randomises these copied features to generate shadow features. The RF model is then used to calculate the z value for each feature, both the original and the shadow features. If the z value of a real feature is consistently higher than the maximum z value of the shadow feature in many independent experiments, the feature is judged to be important.⁹

In this study, we incorporated the selected feature variables into a machine learning algorithm and divided the data set into training and validation sets in a 7:3 ratio. To predict the risk of death at 7, 14, 28 and 365 days after CABG, we used support vector machines (SVM), RF, extreme gradient boosting (XGBoost), K-nearest neighbour algorithm (KNN), gradient boosting decision tree (GBDT), light gradient boosting machine (LightGBM) and Naive Bayes to analyse the filtered variables separately. In order to improve the performance of the model, we used hyperparameter tuning and synthetic minority oversampling technique (SMOTE). The hyperparameters of a variety of machine learning models were optimised using a combination of grid search and 50-fold cross-validation, and the SMOTE method was used to solve the problem of class imbalance in the data set. The training set was used for model construction, and the validation set was used for model evaluation. The model performance evaluation indices included the receiver operating characteristic curve and area under the curve (AUC), accuracy rate, recall rate and F1 score. In addition, we evaluated the model’s clinical validity through decision curve analysis (DCA).

Results

Baseline characteristics

Data on 5790 patients who underwent CABG were extracted from the MIMIC-IV V3.0 database (figure 1). There were 4570 male patients (78.93%), 3984 (68.81%) patients with hypertension, 2527 (43.64%) patients with diabetes, 928 (16.03%) patients with cerebrovascular disease, 1831 (31.62%) patients with congestive heart failure, 1256 (21.69%) patients with COPD, 15 (0.26%) patients who died within 7 days, 41 (0.71%) patients who died within 14 days, 60 (1.04%) patients who died within 28 days and 233 (4.02%) patients who died within 365 days. Patients were divided into six groups: Group 1 (BMI<18.5) consisting of 32 individuals, Group 2 (18.5 ≤BMI<25) consisting of 1315 individuals, Group 3 (25≤BMI<30) consisting of 2299 individuals, Group 4 (30≤BMI<35) consisting of 1445 individuals, Group 5 (35≤BMI<40) consisting of 571 individuals and Group 6 (BMI≥40) consisting of 128 individuals. Patients in Group 1 exhibited higher values for age, SpO₂ level, Charlson Comorbidity Index, blood pH, base excess, anion gap, MCH, MCV and RDW. In contrast, they demonstrated lower values for respiratory rate, temperature, heart rate, WBC count, calcium, glucose, sodium, haemoglobin, ALT, total bilirubin, glycated haemoglobin, haematocrit, RBC count and albumin. Meanwhile, patients in Group 6 showed higher values for PLT count, APACHE III score, total CO₂ and bicarbonate, along with a lower MCHC (online supplemental table 1).

Figure 1. Flowchart of study inclusions and exclusions. BMI, Body Mass Index; MIMIC-IV, Medical Information Mart for Intensive Care IV; ICU, Intensive Care Unit.

Clinical outcomes

Among all the groups analysed, Group 2 exhibited the lowest mortality rates at 7-day, 14-day and 28-day intervals, while Group 3 demonstrated the lowest mortality rate at the 365-day interval (online supplemental table 1). In the logistic regression analysis, with Group 2 (18.5≤BMI<25) set as the reference group, the results of Model I demonstrated that the 365-day mortality risk for Group 3 (25≤BMI<30) and Group 4 (30≤BMI<35) was significantly lower. Additionally, the findings of Model II indicated that the 365-day mortality risk for Group 3 (25≤BMI<30) was also significantly reduced (table 1). The Kaplan-Meier curve demonstrated that patients in the overweight group exhibited the highest 365-day survival probability, with a statistically significant difference compared with other groups (figure 2).

Table 1. Logistic regression model (365-day all-cause mortality).

BMI	Unadjusted OR (95% CI) (Model I)	P value	Adjusted OR (95% CI) (Model II)	P value
BMI<18.5 (Group 1)	2.466 (0.842 to 7.221)	0.1	1.992 (0.632 to 6.274)	0.239
18.5≤BMI<25 (Group 2)	Reference		Reference
25≤BMI<30 (Group 3)	0.590 (0.424 to 0.821)	0.002	0.659 (0.467 to 0.931)	0.018
30≤BMI<35 (Group 4)	0.632 (0.438 to 0.911)	0.014	0.704 (0.478 to 1.038)	0.077
35≤BMI<40 (Group 5)	0.692 (0.425 to 1.127)	0.139	0.701 (0.417 to 1.177)	0.179
BMI≥40 (Group 6)	1.151 (0.541 to 2.447)	0.715	0.989 (0.430 to 2.275)	0.979

BMI, Body Mass Index.

Figure 2. 365-day Kaplan-Meier survival curves by Body Mass Index (BMI) categories.

Subgroup analysis

We focused on the group with all-cause mortality within 365 days. When covariates were not adjusted for, in the subgroups of patients aged ≥65 years and male patients, Group 3 (25≤BMI<30) and Group 4 (30≤BMI<35) showed a lower risk of death (figure 3A). When the covariates were adjusted for, among the subgroups of patients aged ≥65 years, Group 3 (25≤BMI<30) showed a lower mortality rate (OR<1 in each subgroup, p<0.05) (figure 3B).

Figure 3. Subgroup forest plots for 365-day all-cause mortality stratified by Body Mass Index (BMI) categories. (A) The unadjusted covariate model. (B) The model was adjusted for covariates, including age, sex, heart rate, respiratory rate, systolic blood pressure, blood oxygen saturation, body temperature, the Charlson Comorbidity Index and the Acute Physiology and Chronic Health Evaluation III score.

Restricted cubic spline

The RCS analyses of all-cause mortality within 7 days (figure 4A), 14 days (figure 4B), 28 days (figure 4C) and 365 days (figure 4D) all showed an approximately U-shaped correlation between BMI and the risk of death. The RCS curve for mortality within 365 days exhibited a more pronounced inflection point in the BMI range of 25–35 kg/m², compared with the curves for 7 days, 14 days and 28 days.

Figure 4. Restricted cubic spline (RCS) curves of Body Mass Index (BMI). (A) 7-day all-cause mortality; (B) 14-day all-cause mortality; (C) 28-day all-cause mortality and (D) 365-day all-cause mortality.

Analysis of factors contributing to intergroup differences

We employed the RF algorithm to construct a feature importance model for analysing intergroup differences across various BMI categories. The variables—including age, BUN and APACHE III score—recur across various BMI subgroups and survival periods, indicating their substantial importance in predicting overall mortality risk. Group 1 primarily focused on predictive factors characterised by acute physiological indicators, including PT, base excess, blood pH and haematocrit. Within Group 2, WBC count, MCHC, age, RDW, APACHE III score and BUN were of relatively high significance. The key factors in Group 3 are mean SBP (sbp_mean), calcium, ALT, APACHE III score, RBC count and age. In Group 4, BMI, haematocrit, haemoglobin, RBC count, APACHE III score, age, WBC count and BUN exhibited relatively high weights. In Groups 5 and 6, glucose, age, BMI, BUN, total bilirubin, ALT and potassium exhibited high importance. When analysed from different survival time perspectives, in terms of short-term mortality risk, the factors age, BUN, APACHE III, glucose and haematocrit showed a relatively high weight. However, in the 365-day long-term mortality risk groups, the Charlson Comorbidity Index, BUN, age, APACHE III and glucose demonstrated relatively high weights (online supplemental figure S2).

Establishment and validation of the prediction model

The feature selection results for deaths within 7 days (figure 5A), 14 days (figure 5B), 28 days (figure 5C) and 365 days (figure 5D) were displayed using the Boruta algorithm. The variables within the green area are classified as significant features, whereas the variables in the yellow area are categorised as insignificant features according to the Boruta algorithm. The model showed that among patients who died within 365 days, the AUC was 0.67 for SVM, 0.68 for RF, 0.70 for XGBoost, 0.61 for KNN, 0.66 for GBDT, 0.64 for LightGBM and 0.73 for Naive Bayes (online supplemental figure S3). The DCA curve demonstrated that within the lower threshold range, the model exhibits a certain degree of clinical validity. Additionally, XGBoost and GBDT show greater advantages compared with other models (online supplemental figure S4).

Figure 5. Feature selection results of the Boruta algorithm for all-cause mortality. (A) 7 day all-cause mortality; (B) 14-day all-cause mortality; (C) 28-day all-cause mortality and (D) 365-day all-cause mortality. ALT, alanine aminotransferase; AST, aspartate aminotransferase; BMI, Body Mass Index; BUN, blood urea nitrogen; COPD, chronic obstructive pulmonary disease; INR, international normalised ratio; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; MCV, mean corpuscular volume; PT, prothrombin time; PTT, partial thromboplastin time; RBC, red blood cell; RDW, red cell distribution width; sbp_mean, mean systolic blood pressure; SPO₂, peripheral capillary oxygen saturation; WBC, white blood cell; apsiii, Acute Physiology and Chronic Health Evaluation III; totalco2, Total Carbon Dioxide in Blood; resp_rate_mean, Average respiratory rate.

Discussion

This study employed data from the MIMIC-IV database to examine the association between varying BMI levels and mortality at multiple time points following CABG. The findings contribute to the continuing discussion on the ‘obesity paradox’ by demonstrating that individuals classified as overweight or with Class I obesity (BMI 25–35 kg/m²) exhibit a lower risk of mortality at 365 days post-surgery compared with those with normal BMI.¹⁰ These results suggest a potential protective effect associated with higher BMI in the context of CABG, although this effect varies across subgroups and time intervals. In the following sections, we contextualise these findings, discuss their clinical relevance and outline the limitations of the present study.

Key findings and the obesity paradox

Our research is consistent with existing studies that provide evidence in support of the obesity paradox,^{10 11} demonstrating that patients with overweight and class I obesity (BMI 25–35) exhibit a lower mortality risk following CABG, particularly within the first 365 days post-surgery. This seemingly paradoxical association may be attributed to several underlying mechanisms. First, the greater metabolic reserve in these individuals may enhance their capacity to withstand surgical stress and postoperative catabolic states.¹² Second, obesity itself is frequently associated with a state of mild chronic inflammation. A prevailing hypothesis suggests that this underlying low-grade inflammatory condition may induce a form of ‘pre-adaptation’ or ‘tolerance’ in obese individuals, potentially mitigating tissue damage during subsequent intense acute inflammatory responses triggered by conditions such as infections or surgeries.^{13 14} Third, clinicians may implement more intensive risk factor management strategies for patients with obesity, thereby reducing the likelihood of adverse outcomes. However, these protective effects diminish in cases of grade II and III obesity (BMI≥35), suggesting that extreme obesity counteracts any potential survival advantage.¹⁵ The observed U-shaped relationship, as illustrated by the RCS curve, highlights the complex role of BMI in influencing postoperative outcomes.

Subgroup heterogeneity

We primarily performed a subgroup analysis of all-cause mortality over a 365-day period. The results indicated that the protective association of elevated BMI was most pronounced in male patients and those aged ≥65 years. Notably, even after adjusting for potential confounders, the protective effect of overweight remained significant among individuals aged ≥65 years. In the elderly population, sarcopenia is commonly observed, and an increased BMI may reflect preserved muscle mass rather than mere adiposity, which partially explains this paradox.¹⁶ As proposed by Rubino et al in The Lancet Diabetes and Endocrinology, BMI should no longer be used as the sole criterion for diagnosing obesity; instead, obesity and its related diseases should be redefined.¹⁷ Gender differences in fat distribution, hormonal profiles or comorbidity patterns may further account for these variations.¹⁸ However, the underweight group (BMI<18.5) exhibited the highest mortality rate despite having fewer participants (n=32), likely due to frailty, advanced age and a greater comorbidity burden, including higher Charlson Index scores and indicators of malnutrition such as low serum albumin levels.

Intergroup differences

The RF algorithm was used to examine intergroup variations in the weighting of various clinical indicators across BMI categories and survival durations. The results demonstrate that the low BMI group is primarily driven by acute and physiological parameters. This pattern is likely attributable to compromised baseline nutritional status in these patients, which increases susceptibility to rapid clinical deterioration following surgery, particularly due to acute physiological disturbances such as acid-base imbalances and coagulation disorders.¹⁹ In the normal BMI group, beyond inflammatory and stress markers, the importance of basic physiological reserve and chronic disease burden has become increasingly evident. Age and BUN, as key indicators of patients’ physiological reserve and long-term comorbidity burden, reflect the impact of prolonged physiological ageing and chronic organ dysfunction—particularly renal impairment—in determining clinical outcomes among patients in this group who were critically ill.²⁰ These individuals may lack the metabolic buffering capacity associated with obesity and have limited nutritional reserves, making their prognosis more heavily dependent on baseline health status and the ability to withstand acute inflammatory challenges.²¹ The increased weights of sbp_mean, calcium and ALT suggest that the mortality risk model for the overweight group places greater emphasis on the stability of the metabolism and haemodynamics. Patients in this group may exhibit inherent cardiovascular vulnerability due to prolonged weight burden.²² Following coronary artery bypass surgery, maintaining adequate blood pressure and perfusion pressure, along with careful management of subsequent electrolyte imbalances and potential hepatic injury, represents critical intervention targets for improving prognosis. In Group 4, the increased weight of BMI as a predictor indicates that obesity-related physiological changes have begun to directly influence prognosis. Concurrently, RBC-related parameters have gained greater significance. This may reflect chronic inflammation²³; alternatively, under the physiological stress following coronary artery bypass surgery, maintaining adequate haematocrit and oxygen-carrying capacity may become critical, suggesting impaired blood homeostasis in these patients.^{24 25} This group retains certain features observed in the normal and overweight groups (such as WBC and BUN), which may represent a transitional phase from overweight status towards a more severe obesity-associated risk profile. In Groups 5 and 6, glucose emerged as a key predictive factor, indicating that the severity of insulin resistance and stress-induced hyperglycaemia represents a critical mechanism underlying mortality risk in patients with extreme obesity.²⁶ Elevated serum levels of total bilirubin and ALT indicate clinically relevant hepatic impairment in this population, potentially reflecting underlying metabolic dysfunction-associated steatotic liver disease (MASLD).²⁷ The prominence of potassium as a predictor underscores the complexity of electrolyte management in these individuals. For patients with grade II and III obesity, the drivers of mortality risk shifted fundamentally to metabolic disorders and liver dysfunction. These findings raise the hypothesis that obesity-related hepatic dysfunction, potentially including MASLD, may contribute to mortality risk; however, this requires further validation.

From the perspective of different survival periods, our study demonstrates that, in comparison to short-term mortality risk, the influence of the Charlson Comorbidity Index weight substantially increases in predicting long-term mortality within 365 days, becoming a dominant predictor alongside BUN, age, APACHE III score and glucose levels. This finding indicates that after the surgical trauma and the acute phase of the disease, a patient’s long-term prognosis is primarily governed by their underlying chronic disease burden. Additionally, the common major predictive factors for both short-term and long-term mortality suggest that initial disease severity, renal function, glucose metabolism and age-related physiological reserve are key determinants of overall mortality risk after CABG.²⁸

Prediction model

The model indicates that among the characteristic variables associated with mortality at 7, 14 and 28 days following CABG, the APACHE III score is the most influential factor.²⁹ In contrast, for mortality at 365 days, the Charlson Comorbidity Index emerges as the predominant predictor.³⁰ This finding aligns with our prior conclusion that short-term mortality is primarily determined by the severity of the acute postoperative phase, whereas long-term mortality is more strongly influenced by the burden of underlying chronic conditions.

Our study focused on the long-term mortality rate over a 365-day period. The results show that, in addition to the Naive Bayes model—which remains suitable for small sample sizes—the XGBoost model achieved the highest predictive performance, with an AUC of 0.70. The DCA indicated that within a lower threshold probability range, the XGBoost model exhibits meaningful clinical utility, suggesting its potential to guide postoperative monitoring.

Limitations of the study

This study has several limitations. First, the retrospective design precludes causal inference and may be susceptible to information bias. Second, the XGBoost model was developed and internally validated using a single-institution dataset stratified into training and validation subsets at a 7:3 ratio; however, external validation in independent cohorts—particularly those representing distinct clinical settings, patient demographics and data acquisition protocols—was not undertaken. As a result, the robustness and transportability of the model across diverse healthcare systems remain unverified. Third, BMI may not adequately differentiate between lean and obese individuals and could misclassify muscular individuals as obese due to its inability to distinguish fat mass from muscle mass.¹⁷ Fourth, the relatively small sample size of the underweight group may have introduced statistical bias and reduced the power to detect significant associations. Finally, unmeasured confounding factors—such as perioperative medications and surgical techniques—were not accounted for in the analysis.

Impact on clinical practice

The findings of this study indicate that the influence of BMI should be comprehensively evaluated in risk assessment following CABG. For patients with normal weight, despite having a BMI within the normal range, clinicians should remain vigilant regarding potential metabolic reserve insufficiency. Elevated BMI alone should not be interpreted as an indicator of poor prognosis in patients with overweight or class I obesity after CABG. Weight management should be individualised rather than uniformly intensified based solely on BMI. Furthermore, the key clinical parameters requiring close monitoring during the perioperative period differ across BMI categories, warranting individualised monitoring strategies. For the XGBoost model, it may serve as a risk-stratification tool that can be embedded within electronic health record systems to estimate patients’ long-term prognostic risk and assign them to distinct categories such as ‘low-risk’ and ‘high-risk’. For patients classified as high-risk, protocol-driven enhanced monitoring is recommended to prioritise resource allocation towards those most likely to benefit from intensified clinical surveillance.

Conclusions

In conclusion, the obesity paradox is evident among patients following CABG, with a particularly pronounced U-shaped relationship between 365-day mortality and BMI. Patients who are overweight or have grade I obesity exhibit a lower long-term risk of mortality. Individualised key monitoring strategies are required for patients across different BMI categories.

Data availability statement

Data are available in a public, open access repository.