Abstract
Background
Delirium is a common complication following coronary artery bypass grafting (CABG). This study aims to develop and validate a predictive model for postoperative delirium in CABG patients using a Bayesian Network (BN).
Methods
Data from the MIMIC-IV and eICU-CRD databases were analyzed, with the MIMIC-IV dataset used for model training and internal validation, and the eICU-CRD dataset for external validation. A directed acyclic graph was constructed using BN based on the Max-Min Hill-Climbing algorithm, followed by model inference. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC) and compared with logistic regression, LightGBM, and a BN model based on the Hill-Climbing algorithm.
Results
A total of 3,708 CABG patients from the MIMIC-IV database and 630 from the eICU-CRD database were included, with postoperative delirium incidence rates of 17% and 14.9%, respectively. The developed BN predictive model comprises 14 nodes and 22 directed edges, with Richmond Agitation-Sedation Scale and Sequential Organ Failure Assessment score appearing as parent nodes of delirium, indicating a probabilistic dependency within the network. The model achieved an AUROC of 0.79 in the internal validation cohort and 0.72 in the external validation cohort. Additionally, a Shiny platform application based on the BN model was developed.
Conclusions
This study successfully constructed a BN predictive model for postoperative delirium following CABG, demonstrating robust predictive performance and high interpretability.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12888-025-07299-w.
Keywords: Coronary artery bypass grafting, Bayesian network, Delirium, Predictive model
Background
Coronary artery bypass grafting (CABG) is a key therapeutic approach for coronary artery disease, particularly effective for patients with multi-vessel or complex lesions [1, 2]. While CABG significantly improves cardiovascular outcomes, its perioperative complications continue to pose challenges to patient recovery [3–5]. Postoperative delirium is one of the most common complications following CABG, with reported incidence rates ranging from 18–50% [6–8]. Postoperative delirium not only prolongs hospital stay but also significantly increases postoperative mortality and the risk of cognitive decline, adversely affecting patients’ long-term quality of life [9–12]. Therefore, early identification of high-risk patients for postoperative delirium is crucial for implementing targeted preventive strategies and improving clinical outcomes.
In recent years, predictive models based on big data and artificial intelligence have become increasingly prevalent in the medical field, serving as vital tools for disease risk prediction and treatment optimization [13, 14]. Traditional delirium risk prediction models predominantly rely on logistic regression and other linear models, which struggle to capture the complex interactions among multiple factors [15, 16]. In contrast, the Bayesian network (BN), as a probabilistic graphical model, excels at handling uncertainty and modeling conditional dependencies among variables. This capability grants the BN greater flexibility and interpretability in complex systems, providing a robust and adaptable framework for clinical risk prediction and decision making [17, 18]. With advancements in BN technology, its application in medical predictive modeling is expanding, highlighting its potential for addressing intricate clinical challenges [19, 20].
Numerous studies have reported on predictive models for delirium following cardiac surgery [16, 21]. However, specific predictive models for postoperative delirium after CABG remain relatively scarce, with most existing research focusing on risk factor analysis [6, 22]. In this study, we aim to develop and validate a BN-based predictive model for postoperative delirium in CABG patients, facilitating early identification of high-risk individuals.
Methods
Data sources
The test dataset for this study was derived from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, which included comprehensive intensive care unit (ICU) patient information from 2008 to 2019 at the Beth Israel Deaconess Medical Center in the United States [23]. External validation data were obtained from the eICU Collaborative Research Database (eICU-CRD), a multicenter ICU database comprising patient data from 335 ICUs across 205 hospitals in the United States between 2014 and 2015 [24]. Both databases provide comprehensive patient information, including demographics, vital signs, laboratory test results, treatments, and nursing records. Researchers in this study completed the necessary training required by the collaborating institutions to gain access and use these databases.
The data used in this study were de-identified and publicly available, ensuring that patient confidentiality and privacy were maintained. The Institutional Review Boards (IRB) of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center provided ethical approval for the use of these databases in research. The IRBs of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center granted a waiver of informed consent due to the retrospective nature of the study and the use of de-identified data.
Study population
The study included all patients who underwent CABG surgery recorded in the databases. Exclusion criteria were as follows: (1) patients under the age of 18 years; (2) patients who did not receive an ICU admission after surgery; (3) patients with multiple ICU admissions (only data from the first ICU admission were included); (4) patients without postoperative delirium assessment records; and (5) patients with delirium assessments marked as “unable to assess”. The study flowchart is shown in Additional file 1: Fig. S1.
Definition of delirium
Delirium was diagnosed based on the following criteria: (1) a positive Confusion Assessment Method for the ICU (CAM-ICU) result during the ICU stay, or (2) an Intensive Care Delirium Screening Checklist (ICDSC) score of ≥ 4 at any point during the ICU stay.
The CAM-ICU assessment consists of four features: (1) an acute change or fluctuation in mental status, (2) inattention, (3) disorganized thinking, and (4) an altered level of consciousness. A diagnosis of delirium (i.e., a positive CAM-ICU result) requires the presence of both the first and second features, along with either the third or fourth feature [25].
The ICDSC score is an alternative tool for assessing delirium in ICU patients, with a maximum score of 8. A score of ≥ 4 suggests the presence of delirium. The ICDSC assessment includes eight items: acute or fluctuating changes in consciousness, inattention, disorientation, hallucinations or illusions, psychomotor disturbances, disorganized speech or thinking, sleep-wake cycle disturbances, and abnormal mood or affect. Each item scores 1 point, with a cumulative score of ≥ 4 indicating delirium [26].
Other variables
Additional variables included: (1)demographic information: age, sex, body mass index; (2) vital signs and laboratory results within the first 24 h of ICU admission; (3) interventions within the first 24 h of ICU admission, including mechanical ventilation, use of vasopressors, continuous renal replacement therapy, use of propofol, midazolam, and dexmedetomidine; (4) disease severity scores within the first 24 h of ICU admission: Sequential Organ Failure Assessment (SOFA) score, Glasgow Coma Scale (GCS) score, Richmond Agitation-Sedation Scale (RASS) score, and pain score; (5) comorbidities: hypertension, diabetes, congestive heart failure, chronic kidney disease, and stroke.
Data processing
Structured Query Language was used to extract the above variables. Patient data with missing values exceeding 30% were excluded, while the remaining missing values were imputed using the median. Continuous variables were discretized for model development. Vital signs and BMI were categorized based on clinical thresholds, laboratory values were classified according to clinical reference ranges, and age and SOFA scores were grouped by quartiles. GCS, RASS, and pain scores were discretized based on clinical expertise, with details of the discretization shown in Additional file 1: Table S1.
Statistical analysis
Continuous variables were described as medians with interquartile ranges. Differences in continuous variables between groups were statistically compared using the Mann-Whitney U test. Categorical variables were expressed as frequencies and percentages, with group differences assessed via the Chi-square test or Fisher’s exact test. The MIMIC-IV data served as the test set, while the eICU-CRD data were used for external validation, the test set was randomly divided in a 7:3 ratio for the training and internal validation sets.
The primary modeling approach used a BN based on the Max-Min Hill-Climbing (MMHC) algorithm to construct a directed acyclic graph. Parameters within the BN were then estimated using Maximum Likelihood Estimation (MLE) to derive conditional probability tables for each node. The MLE approach identifies parameter combinations that maximize the likelihood function, thus assigning the highest probability to the observed data under the assumed model. Each local distribution was represented by a linear regression model. Through model simplification, balancing complexity and interpretability, we constructed a clinically reliable and practical predictive model.
The primary evaluation metric for model performance was the area under the receiver operating characteristic curve (AUROC), with additional metrics including accuracy, sensitivity, specificity, precision, and F1 score. For comparative purposes, models were also built using Logistic Regression (LR), Light Gradient Boosting Machine (LightGBM), and a BN model based on the Hill-Climbing (HC) algorithm, with AUROC values compared across models. Finally, a Shiny platform application was developed based on the BN model for clinical use.
To evaluate the robustness and generalisability of the model, we performed a series of sensitivity analyses. In addition to the primary approach of median imputation, alternative methods of handling missing data were employed for comparative analysis, including multiple imputation (MICE, R package), k-nearest neighbours (KNN) imputation and random forest imputation (missForest, R package) [27, 28]. Due to the presence of different delirium assessment tools (CAM-ICU and ICDSC) in the external validation cohort, the performance of the model was evaluated separately according to each assessment criterion. In model structure learning, we compared performance under various scoring functions, such as Akaike Information Criterion (AIC) and Bayesian Dirichlet equivalent uniform (BDeu). Furthermore, we trained and evaluated the model using different training–validation set partition ratios to examine the impact of partitioning strategies on model performance.
All statistical analyses were conducted using R 4.4.3 and STATA 15.1, with a P-value < 0.05 considered statistically significant.
Results
Baseline data
After applying exclusion criteria, a total of 3,708 post-CABG patients were included from the MIMIC-IV database, with 3,078 patients in the non-delirium group and 630 in the delirium group, resulting in a delirium incidence of 17%. Baseline characteristics of the two groups are presented in Table 1. In the eICU-CRD database, 529 post-CABG patients were included, with 450 in the non-delirium group and 79 in the delirium group, resulting in a delirium incidence of 14.9%. Baseline characteristics for both groups in the eICU-CRD database are also shown in Table 1.
Table 1.
Baseline characteristics of participants
| Variables | MIMIC-IV | P | eICU-CRD | P | ||
|---|---|---|---|---|---|---|
| Non-delirium patients (n = 3078) |
Delirium patients (n = 630) |
Non-delirium patients (n = 450) |
Delirium patients (n = 79) |
|||
| Age | 68.4 (61.4, 75.2) | 73.3 (64.9, 79.3) | < 0.01 | 66 (58, 73) | 70 (61, 76.5) | 0.05 |
| Sex Male, n (%) | 2419 (78.6) | 456 (72.4) | < 0.01 | 342 (76) | 56 (70.9) | 0.41 |
| BMI (kg/m2) | 28.3 (25.4, 32.2) | 28.1 (25.6, 32) | 0.55 | 29.2 (26.1, 33.5) | 29.3 (25.3, 34.6) | 0.91 |
| Vital signs | ||||||
| Temperature, (℃) | 36.7 (36.4, 36.9) | 36.7 (36.5, 36.9) | 0.92 | 36.3 (35.9, 36.8) | 36.3 (35.9, 36.8) | 0.92 |
| Heart rate, (min−1) | 80 (74, 89) | 81 (76, 88) | 0.39 | 84 (76.2, 90) | 89 (80, 96) | < 0.01 |
| Respiratory rate, (min−1) | 17 (15, 20.4) | 18 (15, 21) | 0.12 | 16 (12, 18) | 16 (14, 20.5) | 0.04 |
| Spo2, (%) | 98 (96, 100) | 99 (97, 100) | < 0.01 | 99 (96, 100) | 99 (96.5, 100) | 0.24 |
| Systolic BP, (mmHg) | 111 (101, 121) | 110 (100, 121) | 0.52 | 112 (103, 125) | 112 (106.5, 128) | 0.42 |
| Diastolic BP, (mmHg) | 56 (50, 63) | 55 (49, 61) | < 0.01 | 64 (57, 71) | 64 (60, 70) | 0.57 |
| Mean arterial BP, (mmHg) | 73 (67, 80) | 72 (66, 79) | < 0.01 | 79 (72, 84) | 79 (74, 87) | 0.13 |
| Laboratory tests | ||||||
| WBC, (K/uL) | 13.2 (10.3, 16.8) | 13.5 (10, 17.8) | 0.58 | 11.9 (9.1, 14.7) | 13.2 (9.8, 16.4) | 0.07 |
| Hemoglobin, (g/dL) | 10.2 (9.1, 11.3) | 9.7 (8.6, 10.8) | < 0.01 | 10.8 (9.5, 12.3) | 10.5 (8.9, 11.6) | 0.07 |
| Platelet, (K/uL) | 148.5 (122, 182) | 139.5 (110, 172) | < 0.01 | 147 (115, 183) | 134 (97.5, 158) | < 0.01 |
| Creatinine, (mg/dl) | 0.9 (0.7, 1.1) | 1 (0.8, 1.3) | < 0.01 | 0.9 (0.8, 1.1) | 1.1 (0.8, 1.4) | < 0.01 |
| BUN, (mg/dl) | 15 (13, 20) | 17 (14, 23) | < 0.01 | 15 (11, 20) | 16 (14, 22) | 0.01 |
| Glucose, (mg/dl) | 136 (117, 161) | 138 (116, 168) | 0.15 | 128 (112, 152.8) | 132 (118, 166.5) | 0.15 |
| Sodium, (mEq/L) | 136 (134, 138) | 136 (134, 138) | 0.49 | 139 (136, 141) | 139 (136, 141) | 0.99 |
| Chloride, (mEq/L) | 107 (104, 109) | 107 (104, 109) | 0.14 | 106 (104, 109) | 107 (104.5, 109) | 0.26 |
| Potassium, (mEq/L) | 4.4 (4, 4.8) | 4.3 (3.9, 4.8) | 0.15 | 4.1 (3.8, 4.4) | 4.2 (3.8, 4.6) | 0.15 |
| Magnesium, (mg/dl) | 2.4 (2.2, 2.6) | 2.5 (2.3, 2.9) | < 0.01 | 2.3 (2, 2.8) | 2.3 (2.1, 2.6) | 0.92 |
| Bicarbonate, (mEq/L) | 23 (22, 24) | 22 (21, 24) | < 0.01 | 23 (22, 25) | 23 (21, 24) | 0.13 |
| Anion gap, (mEq/L) | 12 (10, 13) | 12 (10, 14) | < 0.01 | 8 (7, 9) | 8 (7, 11) | 0.22 |
| pH | 7.4 (7.3, 7.4) | 7.4 (7.3, 7.4) | 0.68 | 7.4 (7.3, 7.4) | 7.3 (7.3, 7.4) | 0.01 |
| Treatment measures | ||||||
| MV, n (%) | 1791 (58.2) | 490 (77.8) | < 0.01 | 328 (72.9) | 66 (83.5) | 0.06 |
| Vasopressor, n (%) | 2437 (79.2) | 560 (88.9) | < 0.01 | 202 (44.9) | 38 (48.1) | 0.68 |
| CRRT, n (%) | 21 (0.7) | 18 (2.9) | < 0.01 | 6 (1.3) | 1 (1.3) | 1.00 |
| Dexmedetomidine, n (%) | 662 (21.5) | 186 (29.5) | < 0.01 | 169 (37.6) | 33 (41.8) | 0.56 |
| Propofol, n (%) | 3001 (97.5) | 616 (97.8) | 0.79 | 19 (4.2) | 6 (7.6) | 0.24 |
| Midazolam, n (%) | 51 (1.7) | 34 (5.4) | < 0.01 | 106 (23.6) | 7 (8.9) | < 0.01 |
| Scores | ||||||
| GCS | 15 (14, 15) | 15 (14, 15) | 0.46 | 14 (10, 15) | 12 (9, 15) | < 0.01 |
| SOFA | 5 (4, 7) | 7 (5, 9) | < 0.01 | 6 (4, 9) | 8 (6, 10) | < 0.01 |
| RASS | −0.5 (−1, 0) | −2 (−4, −0.5) | < 0.01 | 0 (−1, 0) | −1 (−2, 0) | < 0.01 |
| Pain | 5 (4, 6.5) | 5 (3, 6.5) | < 0.01 | 5 (4, 6) | 5 (4, 5) | 0.42 |
| Comorbidities | ||||||
| CHF, n (%) | 577 (18.7) | 236 (37.5) | < 0.01 | 38 (8.4) | 7 (8.9) | 1.00 |
| CKD, n (%) | 486 (15.8) | 179 (28.4) | < 0.01 | 20 (4.4) | 6 (7.6) | 0.26 |
| Hypertension, n (%) | 1930 (62.7) | 318 (50.5) | < 0.01 | 109 (24.2) | 21 (26.6) | 0.76 |
| Diabetes, n (%) | 774 (25.1) | 164 (26) | 0.68 | 100 (22.2) | 10 (12.7) | 0.07 |
| Stroke, n (%) | 59 (1.9) | 18 (2.9) | 0.18 | 24 (5.3) | 5 (6.3) | 0.79 |
Abbreviations:BMI body mass index, BP blood pressure, WBC white blood cell count, BUN blood urea nitrogen, MV mechanical ventilation, CRRT continuous renal replacement therapy, GCS Glasgow Coma Scale, SOFA Sequential Organ Failure Assessment, RASS Richmond Agitation-Sedation Scale, CHF congestive heart failure, CKD chronic kidney disease
BN modeling
A predictive model for postoperative delirium after CABG was constructed using the BN. In this model, each predictor is represented as a node, and directed edges between nodes illustrate the conditional dependencies and direction of information flow between variables. This facilitates an understanding of the complex interplay between factors in delirium prediction. The BN model for postoperative delirium comprises 14 nodes and 22 directed edges. Fig. 1 illustrates the BN structure of the delirium model. Within this structure, RASS and SOFA scores are parent nodes of delirium, demonstrating significant conditional dependencies with its occurrence. Mechanical ventilation is structurally connected to the RASS score, suggesting that it may be indirectly associated with delirium through its impact on sedation levels. Furthermore, the interrelated pathways between the SOFA, GCS, and RASS scores suggest that these variables may contribute jointly to delirium risk prediction.
Fig. 1.
Predictive model for postoperative delirium following coronary artery bypass grafting based on Bayesian network. Abbreviations: BUN, blood urea nitrogen; GCS, Glasgow Coma Scale; SOFA, Sequential Organ Failure Assessment; RASS, Richmond Agitation-Sedation Scale; CHF, congestive heart failure; CKD, chronic kidney disease
Model inference and results
Using the available evidence variables, the BN model can estimate changes in the probability of delirium occurring. For instance, when adjusting the patient’s RASS score to 0, the probability of delirium decreases from 29 to 15%, as shown in Fig. 2. Further inference reveals that, holding other variables constant, if the patient is not on mechanical ventilation, the probability of dexmedetomidine use decreases from 25 to 17%, while the probability of a RASS score of 0 increases from 47 to 64%, ultimately reducing the probability of delirium from 29 to 24%, as shown in Fig. 3.
Fig. 2.
Bayesian inference with adjusted RASS score. Abbreviations: BUN, blood urea nitrogen; GCS, Glasgow Coma Scale; SOFA, Sequential Organ Failure Assessment; RASS, Richmond Agitation-Sedation Scale; CHF, congestive heart failure; CKD, chronic kidney disease
Fig. 3.
Bayesian inference with adjusted mechanical ventilation. Abbreviations: BUN, blood urea nitrogen; GCS, Glasgow Coma Scale; SOFA, Sequential Organ Failure Assessment; RASS, Richmond Agitation-Sedation Scale; CHF, congestive heart failure; CKD, chronic kidney disease
Model performance
Figure 4 shows the receiver operating characteristic (ROC) curves of the BN model based on the MMHC algorithm in both the internal and external validation sets, with an AUROC of 0.79 (95% confidence interval [CI]: 0.77–0.81) in the internal validation set and 0.72 (95% CI: 0.66–0.79) in the external validation set. In comparison, the LightGBM model achieved an AUROC of 0.97 (95% CI: 0.97–0.98) in the internal validation set and 0.73 (95% CI: 0.67–0.79) in the external validation set for predicting postoperative delirium following CABG. The LR model had an AUROC of 0.79 (95% CI: 0.77–0.81) in the internal validation set and 0.66 (95% CI: 0.60–0.72) in the external validation set. The BN model based on the HC algorithm achieved an AUROC of 0.78 (95% CI: 0.76–0.80) in the internal validation set and 0.66 (95% CI: 0.59–0.73) in the external validation set. Detailed performance metrics are provided in Additional file 1: Table S2, and statistical significance testing of AUROC differences among the models is presented in Additional file 1: Table S3. The variable importance ranking plots for the LightGBM and LR models are shown in Additional file 1: Fig. S2 and Fig. S3, respectively.
Fig. 4.
Receiver operating characteristic curves and area under the receiver operating characteristic curve values for each model. A Internal validation set; (B) External validation set
Application of the BN model
To facilitate clinical application and broader dissemination, we developed an application based on the BN model using the Shiny framework in R. This application allows users to input relevant variables from the BN model, such as age, dexmedetomidine use, and RASS score, to calculate the probability of delirium occurrence. In addition, the application provides a visual representation of the model inference process, clearly illustrating the conditional dependencies among predictive factors and displaying model performance metrics (such as the ROC curve and AUROC value). The application is accessible at: https://amu-secondhospital-ccmii.shinyapps.io/StaticBayes-CABG/. A screenshot of the application is shown in Additional file 1: Fig. S4.
Sensitivity analysis
To evaluate the impact of various methods of handling missing data on model performance, we conducted a comparative analysis using multiple imputation (MICE), KNN and random forest imputation (missForest), as well as median imputation. The results showed that median imputation produced the best performance in the external validation cohort (Additional file 1: Fig. S5). Due to the variety of delirium assessment tools in the external validation set, we divided the eICU-CRD data into subgroups based on the assessment instrument used (CAM-ICU or ICDSC) and assessed the predictive accuracy of the models within each subgroup. We found comparable predictive accuracy between the two subgroups (Additional file 1: Fig. S6).
During BN structure learning, we compared the default Bayesian Information Criterion (BIC) scoring function with the AIC and BDeu scoring functions. For BDeu, we further optimised the equivalent sample size using a grid search. The results showed that model performance was best under the default BIC criterion (Additional file 1: Fig. S7). Additionally, we trained and evaluated the model using different training–validation set partition ratios (60/40, 70/30 and 80/20), revealing minimal variation in performance across different split strategies (Additional file 1: Fig. S8).
Discussion
In this study, we developed a BN-based predictive model for postoperative delirium following CABG, achieving an AUROC of 0.79 in the internal validation set and 0.72 in the external validation set. Additionally, we created an application to facilitate the model’s clinical implementation. These results demonstrate the effectiveness and practicality of our model in predicting delirium, although there remains room for further enhancement.
Numerous studies have reported predictive models for delirium across various clinical scenarios and surgical types. The earliest delirium prediction model, the PRE-DELIRIC model from 2012, was based on 3,056 adult ICU patients and included ten predictive factors. It achieved an AUROC of 0.86 in internal validation and 0.84 in external validation, both of which were significantly higher than the predictive accuracy of healthcare providers based on clinical experience [15]. A meta-analysis examined 22 machine learning-based delirium prediction models, demonstrating promising predictive performance [29]. Postoperative delirium following cardiac surgery, with an incidence of approximately 26–52%, is a common complication that not only impacts short-term recovery but is also associated with long-term adverse outcomes, including prolonged hospital stay, increased mortality, and cognitive impairment [30, 31]. Some predictive models for cardiac postoperative delirium have been reported. For example, the DELIPRECAS model, which includes four clinical risk factors, predicts postoperative delirium in cardiac surgery patients with an AUROC of 0.83 in internal validation and 0.79 in external validation [16]. However, the majority of these models are predicated on traditional regression analyses or opaque, black-box machine learning techniques. They are also predominantly oriented towards the broader population of post-cardiac surgery patients and do not specifically model distinct cardiac surgical cohorts. As the most common cardiac surgical procedure, CABG is characterised by unique underlying pathophysiological and postoperative management features, which may result in distinct postoperative delirium mechanisms compared to other types of cardiac surgery. Therefore, developing a predictive instrument specifically for CABG patients is of considerable clinical importance.
One of the major advantages of BN is their ability to perform inference and prediction; by incorporating existing evidence and data, the model can calculate the probability of delirium under specific conditions [32]. This inference process, based on prior and posterior probabilities, enables the model to adapt its predictions to varying clinical contexts, providing clinicians with a flexible approach to managing complex patient scenarios. In addition, the model supports hypothetical analyses by adjusting certain clinical variables to observe their impact on delirium risk, providing strong support for personalized interventions [17, 18]. For example, in response to a modifiable clinical risk factor, BN can infer its dynamic influence on delirium risk, helping clinicians to evaluate and optimize intervention strategies. The graphical structure of BN not only provides clear visualization but also models the complex interactions among multiple variables, making it particularly suitable for multifactorial clinical issues like delirium. By elucidating these relationships, clinicians can better identify high-risk patients and refine perioperative management.
The association between mechanical ventilation and delirium has been demonstrated in several studies, particularly in ICU patients. A systematic review of risk factors for delirium in the ICU, which included 33 studies (70% of which were of high quality), provides robust evidence that mechanical ventilation is a significant risk factor for delirium [33]. Reports indicate that the incidence of delirium is 36% in ICU patients receiving non-invasive ventilation, while it reaches 40.6% among those undergoing mechanical ventilation [34, 35]. The association between mechanical ventilation and delirium may be attributed to various mechanisms, including prolonged sedation, disruption of circadian rhythms, inflammatory responses, and inadequate oxygen supply. In a mediation analysis by Bose et al., the use of sedative medications was found to mediate approximately 39% of the effect of mechanical ventilation on delirium [36]. In our study, the BN model revealed a distinct associative pathway whereby mechanical ventilation is linked to the occurrence of postoperative delirium through its impact on the depth of sedation. While the observational nature of this research precludes definitive causal inference, these pathways offer valuable insights for clinical interventions. For patients undergoing CABG, optimising sedation strategies and minimising the duration of mechanical ventilation could help reduce the incidence of delirium.
The relationship between sedation and delirium has long been a focus of clinical research. Certain sedatives, such as benzodiazepines, are known to increase the risk of delirium, while the effects of other drugs, like dexmedetomidine, remain controversial [37, 38]. Dexmedetomidine, a selective α2-adrenoceptor agonist, is thought to reduce the incidence of delirium in mechanically ventilated and non-cardiac postoperative patients [39, 40]. However, some randomized controlled trials have found no reduction in delirium incidence with dexmedetomidine use [41, 42]. Furthermore, the DECADE study reported that dexmedetomidine does not decrease postoperative delirium in cardiac surgery patients, and thus it is not recommended for delirium prevention in this group [43]. These discrepancies may be attributed to patient variability, drug dosage, and timing of administration. A meta-analysis indicated that dexmedetomidine could lower the incidence of postoperative delirium in non-cardiac surgery patients over 65 years old, but showed no such effect in patients under 65 [44]. In this study, we observed an association between dexmedetomidine use and the occurrence of delirium, suggesting that this agent may be linked to an increased risk of delirium in specific patient populations. However, as this finding is derived from retrospective data, it should not be interpreted as evidence of causality. It is more likely to reflect underlying clinical context, such as patient severity or sedation strategies, than a direct pharmacological effect. Therefore, this association should be regarded as a hypothesis-generating observation that requires further investigation in prospective studies. Moreover, existing studies on the relationship between sedation depth and delirium risk remain inconclusive [45]. However, our BN analysis indicates a potential link between deeper levels of sedation and an increased risk of delirium. These findings provide a theoretical rationale for optimising sedation strategies to reduce the incidence of delirium. Nevertheless, it must be emphasised that this study is based on retrospective data and that the inferences drawn differ from causal evidence derived from randomised controlled trials, thus necessitating further validation.
Mechanical ventilation and sedation have been identified as important variables in several previous delirium prediction models, including the PRE-DELIRIC model, the PRIDE model, and the delirium prediction model developed by Gong et al. [15, 46, 47]. Other influential factors in the model, such as age, GCS score, SOFA score, and comorbidities, have been shown to be associated with delirium in various risk factor and predictive model studies [47, 48]. Notably, many of these factors are modifiable risks, and Bayesian inference enables assessment of how adjusting these modifiable factors could potentially impact the risk of delirium. This provides a foundation for personalized interventions, assisting clinicians in optimizing patient management to reduce the likelihood of delirium.
This study focuses on patients undergoing CABG and presents a BN-based predictive model with several notable advantages. This is the first time that a BN model has been introduced for predicting postoperative delirium risk following CABG. The model offers robust predictive performance and structural interpretability, elucidating conditional dependencies and potential risk pathways among variables. This facilitates clinical understanding and the formulation of intervention strategies. The model also incorporates key variables that can be easily obtained in the ICU, making it suitable for the early identification of high-risk patients and particularly valuable for intensivists and surgeons. Finally, we have developed an interactive model visualisation tool (a Shiny application) to further enhance the model’s usability and scalability in frontline clinical settings.
This study has several limitations. First, the relatively small number of patients included in the external validation may affect the model’s generalizability to larger populations. Second, the study population was limited to patients in the United States, so further research is needed to determine the model’s applicability to other regions and different ethnic groups. Finally, due to limitations of public databases, intraoperative anesthesia-related information could not be included, which is a significant limitation of this study and may affect the comprehensive assessment of delirium risk.
Conclusion
This study successfully developed a predictive model for postoperative delirium following CABG using BN analysis. The model demonstrates not only robust predictive performance but also high interpretability, providing an effective tool for the early identification and intervention of postoperative delirium, with promising potential for clinical application.
Supplementary Information
Acknowledgements
Thanks to all the staff members associated with the MIMIC-IV and eICU-CRD databases.
Abbreviations
- CABG
Coronary artery bypass grafting
- BN
Bayesian network
- MIMIC-IV
Medical information mart for intensive care-IV
- ICU
Intensive care unit
- eICU-CRD
Electronic intensive care unit collaborative research database
- CAM-ICU
Confusion assessment method for the intensive care unit
- ICDSC
Intensive care delirium screening checklist
- SOFA
Sequential organ failure assessment
- GCS
Glasgow coma scale
- RASS
Richmond agitation-sedation scale
- MHHC
Max-min hill-climbing
- MLE
Maximum likelihood estimation
- AUROC
Area under the receiver operating characteristic curve
- LR
Logistic regression
- LightGBM
Light gradient boosting machine
- HC
Hill-climbing
- KNN
k-nearest neighbours
- AIC
Akaike information criterion
- BDeu
Bayesian dirichlet equivalent uniform
- ROC
Receiver operating characteristic
- CI
Confidence interval
- BIC
Bayesian information criterion
Authors’ contributions
LX and YZ conceived the idea, performed the analysis, and drafted the manuscript. MY and QL interpreted the results and helped to revise the manuscript. WX and JZ helped to formulate the idea of the study. YL contributed to the analysis of the data. All authors read and approved the final version of the manuscript.
Funding
This work is supported by National Natural Science Foundation of China (No. 82072134) and the Anhui Province Key Research and Development Plan High-tech Special Project (No. 202304a05020071).
Data availability
Publicly available datasets were analyzed in this study. This data can be found here: https://physionet.org/about/database/.
Declarations
Ethics approval and consent to participate
The study analyzed data obtained from the MIMIC-IV and eICU-CRD public databases. The data used in this study were de-identified and publicly available, ensuring that patient confidentiality and privacy were maintained. The Institutional Review Boards (IRB) of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center provided ethical approval for the use of these databases in research. The IRBs of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center granted a waiver of informed consent due to the retrospective nature of the study and the use of de-identified data. Details of the databases are available at the following links: MIMIC-IV: https://mimic.mit.edu, eICU-CRD: https://eicu-crd.mit.edu.
Consent for publication
All the authors agree to the publication of this work.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Lei Xu and Yang Zhang contributed equally to this work.
Contributor Information
Yu Liu, Email: 11072@ahu.edu.cn.
Qi Li, Email: qili_md@126.com.
Min Yang, Email: yangmin@ahmu.edu.cn.
References
- 1.Head SJ, Milojevic M, Taggart DP, Puskas JD. Current practice of state-of-the-art surgical coronary revascularization. Circulation. 2017;136:1331–45. [DOI] [PubMed] [Google Scholar]
- 2.Thakare VS, Sontakke NG, Wasnik P, Sr., Kanyal D. Recent advances in coronary artery bypass grafting techniques and outcomes: A narrative review. Cureus. 2023;15:e45511. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Sipahi I, Akay MH, Dagdelen S, Blitz A, Alhan C. Coronary artery bypass grafting vs percutaneous coronary intervention and long-term mortality and morbidity in multivessel disease: meta-analysis of randomized clinical trials of the arterial grafting and stenting era. JAMA Intern Med. 2014;174:223–30. [DOI] [PubMed] [Google Scholar]
- 4.Montrief T, Koyfman A, Long B. Coronary artery bypass graft surgery complications: a review for emergency clinicians. Am J Emerg Med. 2018;36:2289–97. [DOI] [PubMed] [Google Scholar]
- 5.Jawitz OK, Gulack BC, Brennan JM, Thibault DP, Wang A, O’Brien SM, et al. Association of postoperative complications and outcomes following coronary artery bypass grafting. Am Heart J. 2020;222:220–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Greaves D, Psaltis PJ, Davis DHJ, Ross TJ, Ghezzi ES, Lampit A, et al. Risk factors for delirium and cognitive decline following coronary artery bypass grafting surgery: A systematic review and Meta-Analysis. J Am Heart Assoc. 2020;9:e017275. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Smulter N, Lingehall HC, Gustafson Y, Olofsson B, Engstrom KG. Delirium after cardiac surgery: incidence and risk factors. Interact Cardiovasc Thorac Surg. 2013;17:790–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Greaves D, Psaltis PJ, Ross TJ, Davis D, Smith AE, Boord MS, et al. Cognitive outcomes following coronary artery bypass grafting: a systematic review and meta-analysis of 91,829 patients. Int J Cardiol. 2019;289:43–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Norkiene I, Ringaitiene D, Misiuriene I, Samalavicius R, Bubulis R, Baublys A, et al. Incidence and precipitating factors of delirium after coronary artery bypass grafting. Scand Cardiovasc J. 2007;41:180–5. [DOI] [PubMed] [Google Scholar]
- 10.Lechowicz K, Szylinska A, Listewnik M, Drozdzal S, Tomska N, Rotter I, et al. Cardiac delirium index for predicting the occurrence of postoperative delirium in adult patients after coronary artery bypass grafting. Clin Interv Aging. 2021;16:487–95. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Zhang S, Tao XJ, Ding S, Feng XW, Wu FQ, Wu Y. Associations between postoperative cognitive dysfunction, serum interleukin-6 and postoperative delirium among patients after coronary artery bypass grafting: a mediation analysis. Nurs Crit Care. 2024;29:1245–52. [DOI] [PubMed] [Google Scholar]
- 12.Loponen P, Luther M, Wistbacka JO, Nissinen J, Sintonen H, Huhtala H, et al. Postoperative delirium and health related quality of life after coronary artery bypass grafting. Scand Cardiovasc J. 2008;42:337–44. [DOI] [PubMed] [Google Scholar]
- 13.Wu WT, Li YJ, Feng AZ, Li L, Huang T, Xu AD, et al. Data mining in clinical big data: the frequently used databases, steps, and methodological models. Mil Med Res. 2021;8: 44. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.de Hond AAH, Leeuwenberg AM, Hooft L, Kant IMJ, Nijman SWJ, van Os HJA, et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review. NPJ Digit Med. 2022;5:2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.van den Boogaard M, Pickkers P, Slooter AJ, Kuiper MA, Spronk PE, van der Voort PH, et al. Development and validation of PRE-DELIRIC (prediction of delirium in ICU patients) delirium prediction model for intensive care patients: observational multicentre study. BMJ. 2012;344: e420. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.de la Varga-Martinez O, Gomez-Pesquera E, Munoz-Moreno MF, Marcos-Vidal JM, Lopez-Gomez A, Rodenas-Gomez F, et al. Development and validation of a delirium risk prediction preoperative model for cardiac surgery patients (DELIPRECAS): an observational multicentre study. J Clin Anesth. 2021;69: 110158. [DOI] [PubMed] [Google Scholar]
- 17.Nistal-Nuno B. Tutorial of the probabilistic methods Bayesian networks and influence diagrams applied to medicine. J Evid Based Med. 2018;11(2):112–24. [DOI] [PubMed] [Google Scholar]
- 18.Arora P, Boyne D, Slater JJ, Gupta A, Brenner DR, Druzdzel MJ. Bayesian networks for risk prediction using real-world data: a tool for precision medicine. Value Health. 2019;22:439–45. [DOI] [PubMed] [Google Scholar]
- 19.McLachlan S, Dube K, Hitman GA, Fenton NE, Kyrimi E. Bayesian networks in healthcare: distribution by medical condition. Artif Intell Med. 2020;107: 101912. [DOI] [PubMed] [Google Scholar]
- 20.Reijnen C, Gogou E, Visser NCM, Engerud H, Ramjith J, van der Putten LJM, et al. Preoperative risk stratification in endometrial cancer (ENDORISK) by a bayesian network model: A development and validation study. PLoS Med. 2020;17:e1003111. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Cai S, Cui H, Pan W, Li J, Lin X, Zhang Y. Two-stage prediction model for postoperative delirium in patients in the intensive care unit after cardiac surgery. Eur J Cardiothorac Surg. 2022;63:ezac573. [DOI] [PubMed] [Google Scholar]
- 22.Zhang WY, Wu WL, Gu JJ, Sun Y, Ye XF, Qiu WJ, et al. Risk factors for postoperative delirium in patients after coronary artery bypass grafting: a prospective cohort study. J Crit Care. 2015;30:606–12. [DOI] [PubMed] [Google Scholar]
- 23.Johnson A, Bulgarelli L, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV Clinical Database Demo (version 2.2). PhysioNet. 2023.
- 24.Johnson A, Pollard T, Badawi O. Raffa J. eICU Collaborative Research Database Demo (version 2.0.1). PhysioNet. 2021.
- 25.Ely EW, Margolin R, Francis J, May L, Truman B, Dittus R, et al. Evaluation of delirium in critically ill patients: validation of the confusion assessment method for the intensive care unit (CAM-ICU). Crit Care Med. 2001;29:1370–9. [DOI] [PubMed] [Google Scholar]
- 26.Bergeron N, Dubois MJ, Dumont M, Dial S, Skrobik Y. Intensive care delirium screening checklist: evaluation of a new screening tool. Intensive Care Med. 2001;27:859–64. [DOI] [PubMed] [Google Scholar]
- 27.Van Buuren S, Groothuis-Oudshoorn K. Mice: multivariate imputation by chained equations in R. J Stat Softw. 2011;45:1–67. [Google Scholar]
- 28.Stekhoven DJ, Buhlmann P. Missforest–non-parametric missing value imputation for mixed-type data. Bioinformatics. 2012;28:112–8. [DOI] [PubMed] [Google Scholar]
- 29.Xie Q, Wang X, Pei J, Wu Y, Guo Q, Su Y, et al. Machine Learning-Based prediction models for delirium: A systematic review and Meta-Analysis. J Am Med Dir Assoc. 2022;23:1655–68. e6. [DOI] [PubMed] [Google Scholar]
- 30.Brown CH. Delirium in the cardiac surgical ICU. Curr Opin Anaesthesiol. 2014;27:117–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Mattimore D, Fischl A, Christophides A, Cuenca J, Davidson S, Jin Z, et al. Delirium after cardiac surgery-a narrative review. Brain Sci. 2023;13: 1682. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Kyrimi E, Neves MR, McLachlan S, Neil M, Marsh W, Fenton N. Medical idioms for clinical bayesian network development. J Biomed Inform. 2020;108: 103495. [DOI] [PubMed] [Google Scholar]
- 33.Zaal IJ, Devlin JW, Peelen LM, Slooter AJ. A systematic review of risk factors for delirium in the ICU. Crit Care Med. 2015;43:40–7. [DOI] [PubMed] [Google Scholar]
- 34.Tabbi L, Tonelli R, Comellini V, Dongilli R, Sorgentone S, Spacone A, et al. Delirium incidence and risk factors in patients undergoing non-invasive ventilation for acute respiratory failure: a multicenter observational trial. Minerva Anestesiol. 2022;88:815–26. [DOI] [PubMed] [Google Scholar]
- 35.Jeon K, Jeong BH, Ko MG, Nam J, Yoo H, Chung CR, et al. Impact of delirium on weaning from mechanical ventilation in medical patients. Respirology. 2016;21:313–20. [DOI] [PubMed] [Google Scholar]
- 36.Bose S, Kelly L, Shahn Z, Novack L, Banner-Goodspeed V, Subramaniam B. Sedative polypharmacy mediates the effect of mechanical ventilation on delirium in critically ill COVID-19 patients: a retrospective cohort study. Acta Anaesthesiol Scand. 2022;66:1099–106. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Mart MF, Williams Roberson S, Salas B, Pandharipande PP, Ely EW. Prevention and management of delirium in the intensive care unit. Semin Respir Crit Care Med. 2021;42:112–26. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Wilson JE, Mart MF, Cunningham C, Shehabi Y, Girard TD, MacLullich AMJ, et al. Delirium. Nat Rev Dis Primers. 2020;6: 90. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Pandharipande PP, Pun BT, Herr DL, Maze M, Girard TD, Miller RR, et al. Effect of sedation with dexmedetomidine vs lorazepam on acute brain dysfunction in mechanically ventilated patients: the MENDS randomized controlled trial. JAMA. 2007;298:2644–53. [DOI] [PubMed] [Google Scholar]
- 40.Su X, Meng ZT, Wu XH, Cui F, Li HL, Wang DX, et al. Dexmedetomidine for prevention of delirium in elderly patients after non-cardiac surgery: a randomised, double-blind, placebo-controlled trial. Lancet. 2016;388:1893–902. [DOI] [PubMed] [Google Scholar]
- 41.Kawazoe Y, Miyamoto K, Morimoto T, Yamamoto T, Fuke A, Hashimoto A, et al. Effect of dexmedetomidine on mortality and ventilator-free days in patients requiring mechanical ventilation with sepsis: a randomized clinical trial. JAMA. 2017;317:1321–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Hughes CG, Mailloux PT, Devlin JW, Swan JT, Sanders RD, Anzueto A, et al. Dexmedetomidine or propofol for sedation in mechanically ventilated adults with sepsis. N Engl J Med. 2021;384:1424–36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 43.Turan A, Duncan A, Leung S, Karimi N, Fang J, Mao G, et al. Dexmedetomidine for reduction of atrial fibrillation and delirium after cardiac surgery (DECADE): a randomised placebo-controlled trial. Lancet. 2020;396:177–85. [DOI] [PubMed] [Google Scholar]
- 44.Qin C, Jiang Y, Lin C, Li A, Liu J. Perioperative dexmedetomidine administration to prevent delirium in adults after non-cardiac surgery: a systematic review and meta-analysis. J Clin Anesth. 2021;73: 110308. [DOI] [PubMed] [Google Scholar]
- 45.Long L, Ren S, Gong Y, Zhao H, He C, Shen L, et al. Different depths of sedation versus risk of delirium in adult mechanically ventilated patients: A systematic review and meta-analysis. PLoS ONE. 2020;15:e0236014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Gong KD, Lu R, Bergamaschi TS, Sanyal A, Guo J, Kim HB, et al. Predicting intensive care delirium with machine learning: model development and external validation. Anesthesiology. 2023;138:299–311. [DOI] [PubMed] [Google Scholar]
- 47.Hur S, Ko RE, Yoo J, Ha J, Cha WC, Chung CR. A machine learning-based algorithm for the prediction of intensive care unit delirium (PRIDE): retrospective study. JMIR Med Inform. 2021;9: e23401. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Kim Y, Jin Y, Jin T, Lee SM. Risk factors and outcomes of sepsis-associated delirium in intensive care unit patients: a secondary data analysis. Intensive Crit Care Nurs. 2020;59: 102844. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Publicly available datasets were analyzed in this study. This data can be found here: https://physionet.org/about/database/.