Bayesian network model for identification of factors associated with ventilator-associated pneumonia in mechanically ventilated patients in the ICU: retrospective cohort study

Abstract

Background

Ventilator-associated pneumonia (VAP) represents a complication occurring in patients undergoing mechanical ventilation. This study aimed to develop a Bayesian network model to identify factors associated with the occurrence of VAP.

Methods

A retrospective cohort analysis was conducted using data from patients aged ≥ 60 years who underwent mechanical ventilation in the Department of Intensive care unit at the Second Hospital of Shanxi Medical University between June 2018 and June 2022. Collected variables included demographic characteristics, clinical conditions, medication use, catheterization-related data, laboratory findings, nursing-related and mechanical ventilation-related data. The chi-square test and logistic regression analysis were applied to determine factors associated with VAP. A Bayesian network model was subsequently constructed using the max-min hill-climbing algorithm.

Results

A total of 502 patients were included, comprising 332 males and 170 females, with a median age of 72 years (range: 60–100 years). The prevalence of VAP was 9.6% (48/502 patients). The constructed Bayesian network included 10 nodes and 17 directed edges. Direct associations with VAP were identified for antibiotic use exceeding three agents, reintubation, mechanical ventilation methods, and an Acute Physiology and Chronic Health Evaluation II (APACHE II) score ≥ 15. Indirect associations were observed for disease category, diabetes mellitus, presence of an indwelling central venous catheter, transfusion, and corticosteroid administration. The area under the receiver operating characteristic curve for the Bayesian network model was 0.84 (95% CI: 0.79–0.90).

Conclusion

The Bayesian network model elucidates the interrelationships among multiple factors associated with VAP. The model may provide ancillary information to help clinicians identify patient profiles associated with higher VAP risk and facilitate the implementation of early during mechanical ventilation.

Contribution to the literature

• □ A Bayesian network model for VAP was developed, identifying direct associations with the use of more than three antibiotics, reintubation, mechanical ventilation methods, and an APACHE II score (≥ 15). Indirect associations were observed for disease category, diabetes, indwelling central venous catheter, transfusion, and corticosteroid administration.

• □ Based on a large dataset of patients aged ≥ 60 years who underwent mechanical ventilation in the ICU, this research provides a solid foundation for the development of targeted prevention strategies, supported by a model with strong discriminative performance (AUC = 0.84).

• □ The findings support the prioritization of interventions addressing directly associated VAP risk factors while also facilitating early identification of high-risk groups. This may enable timely implementation of preventive measures targeting both direct and indirect contributors to VAP.

Introduction

Mechanical ventilation (MV) constitutes a fundamental supportive intervention for patients in the intensive care unit (ICU) and remains critical for sustaining life in patients with severe illness. Despite its clinical benefits, MV is associated with a range of complications, among which ventilator-associated pneumonia (VAP) is particularly significant. VAP is defined as pneumonia developing more than 48 h after tracheal intubation or tracheotomy in patients receiving mechanical ventilation and may also occur within 48 h following ventilator withdrawal or extubation [1].

The reported incidence of VAP varies widely, ranging from 5% to 40% [2]. This condition is associated with a substantial increase in healthcare costs and contributes to adverse clinical outcomes. In the United States, VAP has been estimated to add approximately $47,238 to the cost of each ICU hospitalization [3]. Mortality among patients with VAP in the ICU has been reported to range from 16.2% to 74.1% [4]. Early identification of factors associated with VAP, recognition of high-risk patient groups, and timely implementation of preventive interventions are therefore of considerable clinical importance in reducing its occurrence.

In recent years, logistic regression methods [5–7] and a range of machine learning algorithms [8–10] have been widely applied to examine risk factors for VAP and to develop predictive models. Despite their utility, these approaches present certain limitations, including an inability to clearly differentiate between direct and indirect risk factors [11]. Bayesian network models offer a distinct advantage in this context, as they allow for the separation of direct and indirect associations while providing a graphical representation of the complex interrelationships between VAP and its contributing variables [12]. The Max-Min Hill-Climbing (MMHC) algorithm represents an efficient approach for Bayesian network structure learning [13], facilitating the reduction of the search space and identification of an optimal network configuration.

In the present study, data from patients aged ≥ 60 years who underwent mechanical ventilation in the Intensive Care Unit of the Second Hospital of Shanxi Medical University between June 2018 and June 2022 were retrospectively analyzed. Variable selection was conducted using the chi-square test in combination with logistic regression analysis. A Bayesian network model, constructed using the MMHC algorithm, was applied to examine factors associated with VAP, with the aim of supporting improved identification of patients at elevated risk. We further summarized published traditional VAP risk scoring tools and newly emerging CT-based artificial intelligence predictive models from recent literature [14, 15] to sort out the developmental context of existing predictive models for VAP.

Materials and methods

Participants

A retrospective cohort design was employed, with retrospective analysis of data from critically ill patients aged ≥ 60 years who underwent mechanical ventilation in the ICU of the Second Hospital of Shanxi Medical University between June 2018 and June 2022. The study cohort comprised 502 patients. Inclusion criteria were age ≥ 60 years and receipt of invasive mechanical ventilation for at least 48 h. Patients with a diagnosis of pneumonia prior to the initiation of mechanical ventilation were excluded. This study was conducted in accordance with the declaration of Helsinki. Ethical approval was obtained from the Ethics Committee of the Second Hospital of Shanxi Medical University [(2022) YX No. (161)].

Data collection

Candidate variables were determined through an extensive review of the literature and encompassed. Data for 21 variables were extracted from the electronic medical record system by two independent reviewers. Then, 21 available possible variables including demographic characteristics (Gender, Smoking), clinical conditions (Type of Disease, Diabetes, Hypertension, Tumor, Surgery, Transfusion, APACHE Ⅱ score), medication use (Antibiotic use, Acid inhibitor, Corticosteroids use), catheterization-related data (Nasal feeding, Indwelling CVC, Indwelling arterial catheter, Indwelling drainage tube), laboratory findings (Albumin), nursing-related characteristics (Number of oral care, Oral care methods), MV-related characteristics (Reintubation, MV methods). The time of variables extraction are described in Fig. 1.

Fig. 1.

Timeline of Variable Collection. Demographic characteristics: Gender, Smoking; Clinical conditions: Type of Disease, Diabetes, Hypertension, Tumor, Surgery, Transfusion, APACHE Ⅱ score; Medication use: Antibiotic use, Acid inhibitor, Corticosteroids use; Catheterization-related: Nasal feeding, Indwelling CVC, Indwelling arterial catheter, Indwelling drainage tube; Laboratory findings: Albumin; Nursing-related: Number of oral care, Oral care methods; Mechanical ventilation-related: Reintubation, MV methods

VAP diagnosis

Diagnosis was established jointly by the attending ICU physician and a physician specializing in infectious diseases. Clinical criteria for VAP [16] included radiographic evidence on chest X-ray or computed tomography demonstrating new or progressive infiltrates, consolidation, or ground-glass opacities, together with at least two of the following: (1) fever (≥ 38 °C); (2) purulent tracheal secretions; and (3) leukocyte count (≥ 10 × 10⁹/L or ≤ 4 × 10⁹/L).

Bayesian network

A Bayesian network comprises a directed acyclic graph (DAG) and an associated conditional probability table (CPT). The DAG represents potential relationships among variables, where nodes correspond to variables and directed edges indicate dependencies between them [17]. For instance, a directed edge from node A to node B denotes that A is the parent node and B is the child node. The CPT characterizes the uncertainty in these relationships and quantifies the strength of associations between variables [18]. For each child node, a conditional probability expresses the influence of its parent nodes, denoted as P[Xi| pa(Xi)], where P represents conditional probability, Xi represents the node, and pa(Xi) denotes the set of parent nodes associated with Xi [19].

Statistical analyses

Data analysis was performed using SPSS version 27.0 and R version 4.2.0. Missing values were addressed using the MissForest algorithm. Key laboratory variable missing proportions were albumin (8.4%), C-reactive protein (27.2%), and procalcitonin (18.1%). Missing data were processed via chained equation multiple imputation with random forest algorithm with five imputed datasets generated. We compared central tendency and frequency distribution of all variables before and after imputation, and final statistical outcomes from five imputed datasets were pooled following Rubin’s combination rules. Categorical variables are expressed as frequencies and percentages. Univariate analysis was conducted using the chi-square test, and variables with p < 0.1 were subsequently entered into multivariate logistic regression and incorporated into the Bayesian network model. Given the limited number of VAP-positive events (n = 48) in the present cohort, multiple strategies were adopted to minimize potential overfitting during model construction. First, univariate screening with a threshold of p < 0.1 was performed to filter low-correlation variables and reduce input dimensionality, avoiding over-parameterization. Second, stratified 5-fold cross-validation was applied to maintain a consistent case–control ratio across training and validation subsets, which improved the generalizability of the model. Third, the topology of the Bayesian network was moderately constrained by a predefined variable association whitelist to simplify network structure and reduce redundant parameter estimation, further enhancing model stability under a small-event sample size. Visualization of the Bayesian network was carried out using Netica software. Model performance was evaluated by calculating the area under the receiver operating characteristic (ROC) curve (AUC). All statistical tests were two-sided, with a significance threshold set at α = 0.05.

Results

Baseline characteristics

The study included 502 participants, of whom 332 (66.1%) were male and 170 (33.9%) were female. The median age was 72 years, with a range of 60 to 100 years. Disease categories comprised 235 cases (46.8%) classified as medical, 79 (15.7%) as non-traumatic surgical, and 188 (37.5%) as traumatic surgical. Among the total cohort, 454 patients (90.4%) did not develop VAP, whereas 48 (9.6%) were diagnosed with VAP, as shown in Table 1.

Table 1.

Baseline characteristics of VAP and non-VAP groups

Variables		non-VAP (N = 454)	VAP (N = 48)	χ2	p
Gender	Female	157 (34.6%)	13 (27.1%)	1.09	0.297
	Male	297 (65.4%)	35 (72.9%)
Smoking	No	316 (69.6%)	36 (75%)	0.603	0.437
	Yes	138 (30.4%)	12 (25%)
Type of Disease	Internal medicine	221 (48.7%)	14 (29.2%)	8.584	0.014
	Non-traumatic surgery	72 (15.9%)	7 (14.6%)
	Traumatic surgery	161 (35.5%)	27 (56.2%)
Diabetes	No	365 (80.4%)	44 (91.7%)	3.653	0.056
	Yes	89 (19.6%)	4 (8.3%)
Hypertension	No	272 (59.9%)	32 (66.7%)	0.829	0.362
	Yes	182 (40.1%)	16 (33.3%)
Tumor	No	416 (91.6%)	47 (97.9%)	2.394	0.122
	Yes	38 (8.4%)	1 (2.1%)
Surgery	No	223 (49.1%)	19 (39.6%)	1.581	0.209
	Yes	231 (50.9%)	29 (60.4%)
Transfusion	No	198 (43.6%)	10 (20.8%)	9.282	0.002
	Yes	256 (56.4%)	38 (79.2%)
APACHE Ⅱ score	<15	259 (57%)	3 (6.2%)	14.037	< 0.001
	≥ 15	195 (43%)	45 (93.8%)
Nasal feeding	No	33 (7.3%)	1 (2.1%)	1.849	0.174
	Yes	421 (92.7%)	47 (97.9%)
Antibiotic use	≤ 3	402 (88.5%)	28 (58.3%)	32.254	< 0.001
	>3	52 (11.5%)	20 (41.7%)
Acid inhibitor	No	69 (15.2%)	7 (14.6%)	0.013	0.91
	Yes	385 (84.8%)	41 (85.4%)
Corticosteroids use	No	320 (70.5%)	28 (58.3%)	3.014	0.083
	Yes	134 (29.5%)	20 (41.7%)
Indwelling CVC	No	148 (32.6%)	6 (12.5%)	8.246	0.004
	Yes	306 (67.4%)	42 (87.5%)
Indwelling arterial catheter	No	274 (60.4%)	25 (52.1%)	1.232	0.267
	Yes	180 (39.6%)	23 (47.9%)
Indwelling drainage tube	No	197 (43.4%)	17 (35.4%)	1.129	0.288
	Yes	257 (56.6%)	31 (64.6%)
Reintubation	No	437 (96.3%)	38 (79.2%)	24.91	< 0.001
	Yes	17 (3.7%)	10 (20.8%)
Albumin	<28	206 (45.4%)	25 (52.1%)	0.787	0.375
	≥ 28	248 (54.6%)	23 (47.9%)
Number of oral care	<3	25 (5.5%)	2 (4.2%)	0.153	0.696
	≥ 3	429 (94.5%)	46 (95.8%)
Oral care methods	Wipe	387 (85.2%)	44 (91.7%)	1.488	0.475
	Toothbrush	48 (10.6%)	3 (6.2%)
	Wipe + rinse	19 (4.2%)	1 (2.1%)
MV methods	Tracheal intubation	353 (77.8%)	18 (37.5%)	40.571	< 0.001
	Tracheotomy	30 (6.6%)	5 (10.4%)
	Combined	71 (15.6%)	25 (52.1%)

VAP Ventilator-associated pneumonia, CVC Central venous catheter, APACHE Ⅱ Acute Physiology and Chronic Health Evaluation Ⅱ, ICU Intensive Care Unit

Multivariate logistic regression analysis of factors associated with VAP

Nine variables identified in the univariate analysis (p < 0.1) were included in the logistic regression model. The analysis indicated that the use of three or more antibiotics, reintubation, and MV methods were independently associated with VAP, as presented in Table 2.

Table 2.

Multivariate Logistic regression analysis

Variables	B	S.E.	Wald	P	OR (95% CI)
Type of Disease
Internal medicine					1.000
Non-traumatic surgery	0.255	0.578	0.441	0.659	1.29(0.40,3.93)
Traumatic surgery	0.653	0.441	1.482	0.138	1.92(0.82,4.66)
Diabetes	-0.684	0.606	-1.129	0.259	0.50(0.13,1.51)
Corticosteroids use	0.373	0.406	0.918	0.358	1.45(0.65,3.21)
Transfusion	0.574	0.444	1.293	0.196	1.78(0.76,4.41)
APACHE Ⅱ score	17.472	928.529	0.019	0.985
Antibiotic use	1.548	0.419	3.695	< 0.001	4.70(2.08, 10.86)
Indwelling CVC	0.542	0.515	1.051	0.293	1.72(0.66,5.11)
Reintubation	1.867	0.597	3.126	0.002	6.47(2.00, 21.16)
MV methods
Tracheal intubation					1.000
Tracheotomy	1.419	0.611	2.324	0.020	4.13(1.16, 13.17)
Combined	1.732	0.401	4.319	< 0.001	5.65(2.60, 12.63)

APACHE Ⅱ Acute Physiology and Chronic Health Evaluation Ⅱ, ICU Intensive Care Unit

Bayesian network model of VAP

A Bayesian network and corresponding conditional probability table were constructed to examine factors associated with VAP, based on the nine variables identified in the univariate analysis (p < 0.1). As illustrated in Fig. 2, the resulting network comprised 10 nodes and 17 directed edges. These edges represent probabilistic dependencies between variables, while the percentages presented in the conditional probability table correspond to the prior probabilities of each node.

Fig. 2.

Network structure of VAP with its related factors

The prior probability of VAP was 9.6%, [P(VAP) = 9.6%]. The use of more than three antibiotics, reintubation, MV methods, and an APACHE II score ≥ 15 were identified as parent nodes, indicating direct associations with VAP. In contrast, disease category, diabetes mellitus, presence of an indwelling CVC, transfusion, and corticosteroid administration were linked indirectly to VAP, as depicted in Fig. 2.

Bayesian inference of VAP

The Bayesian network enables inference from observed variables to unobserved outcomes, facilitating estimation of the probability of disease occurrence. Through the conditional probability framework, the model quantifies the influence of associated factors on VAP and allows assessment of the contribution of individual nodes to overall risk.

Analysis of the network indicated that the use of more than three antibiotics, reintubation, MV methods, and an APACHE II score ≥ 15 were directly associated with VAP. For example, in patients undergoing reintubation, the probability of VAP increased to P(VAP|reintubation) = 40.3% (Fig. 3). When reintubation occurred in conjunction with a combined MV approach (initial tracheal intubation followed by tracheotomy), the probability further increased to P(VAP|reintubation, combined) = 47.9% (Fig. 4).

Fig. 3.

VAP risk inference with known partial characteristics

Fig. 4.

VAP risk inference with known partial characteristics

Evaluation of the Bayesian network model

All reported area under the ROC curve (AUC) values were calculated from stratified five-fold cross-validation. Each fold preserved fixed VAP positive-to-negative proportion; models were trained on inner subsets and predicted reserved hold-out samples, and pooled AUC was aggregated from all out-of-fold predictions to avoid overoptimistic training-set performance. Model performance was assessed using the AUC, as presented in Fig. 5. Head-to-head ROC comparison further revealed the incremental predictive advantage of Bayesian network over conventional logistic regression. The Bayesian network model developed for VAP demonstrated an AUC of 0.84 (95% CI: 0.79–0.90), indicating good discriminative ability.

Fig. 5.

ROC curves of the logistic regression model (A) and Bayesian network model (B) for predicting VAP. All AUC values were derived from stratified five-fold cross-validation to prevent overfitting

Discussion

Previous investigations have primarily relied on logistic regression to examine factors associated with VAP and to develop risk prediction models, typically quantifying the strength of associations through estimated probabilities of VAP occurrence [5–7]. While informative, such approaches provide limited insight into the interrelationships among variables and do not adequately distinguish between direct and indirect associations with VAP. We systematically compared the applicable scenarios, data requirements and clinical practicability of abovementioned traditional scoring systems and novel CT-AI prediction tools with the Bayesian network framework used in the current research.

In the present analysis, logistic regression identified the use of more than three antibiotics, reintubation, and MV methods as independent factors associated with VAP. By contrast, the Bayesian network model offered a more comprehensive representation, delineating not only the associated factors but also the structure of their relationships, including both direct and indirect contributions to VAP occurrence.

In this study, findings derived from the Bayesian network model were examined alongside those obtained from logistic regression analysis. Logistic regression assumes independence among variables; under this framework, antibiotic use of more than three agents, reintubation, and MV methods were identified as independent factors associated with VAP. but it could not reflect the influence of these three factors on VAP.

In contrast, Bayesian network modeling does not require independence among variables and allows for representation of hierarchical and interdependent relationships. The Bayesian network analysis indicated that the use of more than three antibiotics, reintubation, MV methods, and an APACHE II score ≥ 15 showed direct statistical dependency with VAP. Disease category, diabetes mellitus, presence of an indwelling CVC, transfusion, and corticosteroid administration were identified as indirectly correlated indicators based on conditional probability.

Antibiotic exposure represents a key predictor of VAP [2]. Multiple biological and clinical confounding pathways may explain the observed positive correlation between excessive antibiotic use and elevated VAP risk [20]. From a clinical perspective, anti-infective management should emphasize targeted antimicrobial therapy directed at identified pathogens while limiting unnecessary use of broad-spectrum agents to reduce the emergence of antimicrobial resistance [21]. Evidence from a meta-analysis [22] suggests that antibiotic prophylaxis may reduce the incidence of VAP, although no corresponding reduction in mortality has been demonstrated; further well-designed randomized studies are required to clarify its clinical role.

Reintubation has also been recognized as a factor associated with VAP in mechanically ventilated patients aged over 60 years [23], consistent with the present findings. Reintubation was statistically linked to higher VAP risk in our cohort, which may be affected by multiple underlying clinical conditions and airway-related factors [24]. In clinical practice, daily awakening protocols combined with spontaneous breathing trials may assist in accurately determining readiness for extubation, with the potential to reduce the incidence of reintubation [25].

The present findings suggest that patients managed with tracheostomy or combined MV approaches exhibit a higher likelihood of developing VAP compared with those receiving tracheal intubation alone, consistent with the observations reported by Li et al. [26]. Statistical discrepancies across existing studies and our results suggest that MV modality correlates with varying VAP risk, and multiple mixed clinical confounders could account for such differences [24]. In contrast, other evidence indicates a higher incidence of VAP among patients undergoing tracheal intubation compared to those with tracheostomy [27]. The influence of MV methods (tracheal intubation vs. tracheostomy) on VAP risk remains uncertain and warrants further comprehensive investigation.

An APACHE II score ≥ 15 was identified as a direct correlate of VAP in this study. The APACHE II scoring system is widely used to assess the severity of disease and predict prognosis in patients in the ICU. Previous research has demonstrated a close association between higher APACHE II scores and an increased risk of VAP, as well as with overall clinical outcomes [28]. Discrepant APACHE II results stem from distinct model mechanisms. Logistic regression does not require independent predictors but estimates each covariate’s adjusted individual effect. Correlated factors (reintubation, multiple antibiotics, ventilation modes) mask APACHE II’s predictive signal, rendering it non-significant in regression. The MMHC Bayesian network captures full variable correlations and nonlinear interactions instead of isolating single effects, retaining APACHE II ≥ 15 as a direct VAP node. Collectively, these findings indicate that factors associated with VAP do not operate in isolation but are statistically interrelated within the network, and such correlation structure is observed alongside VAP development rather than proven to causally drive disease progression.

The Bayesian network model for VAP not only facilitates identification of associated factors but also distinguishes variables with direct and indirect statistical dependency relationships from network probability distribution. Emphasizing interventions targeting directly associated factors may enhance the effectiveness of preventive strategies. Based on our BN-derived risk factors, clinical precautions include optimizing targeted narrow-spectrum antibiotics to reduce unnecessary multi-agent combination, standardizing daily sedation interruption plus spontaneous breathing trial protocols to cut avoidable reintubation, intensifying regular monitoring for patients with APACHEⅡ≥15 and personalizing tracheotomy timing balancing ventilation benefit and infectious risk. In future clinical practice, this prediction model could be integrated into the electronic medical record system of ICUs to dynamically calculate real-time individual VAP risk. Automatic risk alerts can be triggered once the predicted risk exceeds preset thresholds, helping clinicians implement early preventive interventions. In addition, the directed network structure enables clinicians to conduct personalized reverse probabilistic reasoning to identify individual high-risk pathways and formulate customized VAP prevention bundles.

It should be noted that this study has several limitations. Firstly, due to the retrospective design adopted, some variables may not have been included in the analysis, and some treatment information was not recorded, such as the duration of antibiotic use and the parameters of the ventilator. Certain laboratory indicators could not be obtained, and missing data were imputed using the random forest method. Secondly, the single-center data source and the high proportion of trauma patients have led to an uneven gender composition, which may limit the external generalizability of our research results to the general ICU population. Thirdly, patients receiving isolated thoracic surgical operations were not enrolled in our three general ICUs and the simplified three-part disease classification is listed as a research limitation. Future research may benefit from multicenter studies with larger sample sizes, along with external validation to further assess the performance and generalizability of the model.

Conclusions

Identification of risk factors remains central to the prevention of VAP. The present study indicates the use of more than three antibiotics, reintubation, MV methods, and an APACHE II score ≥ 15 are directly associated with VAP. In contrast, disease category, diabetes mellitus, the presence of an indwelling CVC, transfusion, and corticosteroid administration were identified as indirect correlates.

The Bayesian network model provides a structured representation of the interrelationships among these factors within a complex network. This approach may assist clinical staff in more effectively identifying VAP-related risk factors and offers a quantitative basis to inform preventive strategies.

Abbreviations

APACHE Ⅱ

Acute Physiology and Chronic Health Evaluation Ⅱ

AUC

Area under curve

CPT

Conditional probability table

CVC

Central venous catheter

DAG

Directed acyclic graph

ICU

Intensive Care Unit

MMHC

Max-Min Hill-Climbing

MV

Mechanical ventilation

VAP

Ventilator-associated pneumonia

ROC

Receiver operating characteristic

Authors’ contributions

Cailing Wang, Liyu Zhang and Yinghui Zhang conceived the idea for the review. Yuping Tang, Wenyan Wang, Xiuxiu Li and Jiao Lu designed, undertook the literature search, and coordinated the study. Jiaxin Wang, Xiaojing Qin and Huiyuan Wu gave crucial intellectual input and provided critical revision for the initial protocol. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Funding

No Funding.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the declaration of Helsinki.This study was approved by the Ethics Committee of the Second Hospital of Shanxi Medical University [(2022) YX No. (161)]. This study is retrospective, and does not involve informed consent of participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.