Construction and Validation of a Predictive Model for the Risk of Ventilator-Associated Pneumonia in Elderly ICU Patients

Background Ventilator-associated pneumonia (VAP) is among the most important hospital-acquired infections in an intensive-care unit setting. However, clinical practice lacks effective theoretical tools for preventing VAP in the elderly. Aim To describe the independent factors associated with VAP in elderly intensive-care unit (ICU) patients on mechanical ventilation (MV) and to construct a risk prediction model. Methods A total of 1851 elderly patients with MV in ICUs from January 2015 to September 2019 were selected from 12 tertiary hospitals. Study subjects were divided into a model group (n = 1219) and a validation group (n = 632). Two groups of patients were divided into a VAP group and a non-VAP group and compared. Univariate and logistic regression analyses were used to explore influencing factors for VAP in elderly ICU patients with MV, establish a risk prediction model, and draw a nomogram. We used the area under the receiver operating characteristic curve (AUROC) and the Hosmer–Lemeshow goodness-of-fit test to evaluate the predictive effect of the model. Findings regarding the length of ICU stay, surgery, C-reactive protein (CRP), and the number of reintubations were independent risk factors for VAP in elderly ICU patients with MV. Predictive-model verification results showed that the area under the curve (AUC) of VAP risk after MV in the modeling and verification groups was 0.859 and 0.813 (P < 0.001), respectively, while P values for the Hosmer–Lemeshow test in these two groups were 0.365 and 0.485, respectively. Conclusion The model could effectively predict the occurrence of VAP in elderly patients with MV in ICUs. This study is a retrospective study, so it has not been registered as a clinical study.


Introduction
Ventilator-associated pneumonia (VAP) is defned as pneumonia that occurs 48 hours after intubation or tracheotomy. Pneumonia that occurs within 48 hours after decanting and extubation from mechanical ventilation (MV) is also included in VAP [1]. VAP prolongs MV time and hospital stay and increases medical costs. VAP also results in atelectasis, MV-related lung injury, respiratory tract obstruction, and other complications. Te morbidity and mortality of individuals with VAP vary by country, intensive-care unit (ICU) type, and diagnostic criteria. Te incidence of VAP reported in European centers is 18.3 cases per 1000 MV days (MVDs) [2], 18.5 cases per 1000 MVDs in low-and middle-income countries, and 9.0 cases per 1000 MVDs in high-income countries such as the United States (U.S.) [3]. A multicenter investigation in China showed that the incidence of VAP in patients with MV was 9.7-48.4%, or 1.3-28.9 per 1000 MVDs, and the fatality rate was 21.2-43.2% [4]. In addition, delayed diagnosis of VAP can impede treatment or promote the overuse of broadspectrum antimicrobials [5]. A recent cost assessment in the U.S. estimated the attribution cost of VAP to be $40,144 (95% confdence interval [CI], $36,286-44,220) [6].
Te number of elderly patients requiring MV is increasing every year. Due to a decline in immune system function, elderly individuals are more prone to opportunistic infections, which increase the possibility of VAP [7]. At present, several risk factors, including age, history of severe chronic obstructive pulmonary disease, admission type, and gender, have been identifed as being associated with VAP [8,9]. However, few studies exist that assess the risk of VAP in elderly ICU patients with MV. Furthermore, models for predicting VAP in these patients are limited. Clinical practice lacks efective theoretical tools for preventing VAP in the elderly.
Te VAP prediction nomogram is recognized as a userfriendly prognostic tool that helps clinicians evaluate the risk of VAP and make better clinical decisions [8]. In this study, we used the classic logistic regression method to develop a VAP prediction model and draw a nomogram based on the baseline clinical characteristics of elderly ICU patients with MV to explore the factors infuencing the occurrence of VAP and to provide an efective predictive tool for early identifcation of high-risk patients.

Study Population.
In this retrospective analysis, the study cohort included elderly patients who were in ICUs with MV from January 2015 to September 2019 in 12 tertiary hospitals in Shanxi Province, China. Te inclusion criteria were as follows: (1) ventilator use ≥48 h; (2) no VAP infection prior to MV; and (3) age ≥65 years. Te sole exclusion criteria were as follows: (1) pneumonia diagnosed before MV; (2) death ceased treatment, discharged, or transferred within 48 h; and (3) patients with incomplete case data. Patients were diagnosed in accordance with the criteria of the Guidelines for the Diagnosis and Treatment of Chinese Adult Hospital-Acquired Pneumonia and VAP (2018 edition) [1]. We diagnosed VAP when the chest X-ray or computed tomography (CT) between 48 hours after the initiation of MV and discharge showed new or progressive infltrating shadows, consolidation shadows, or ground-glass shadows with two or more of the following symptoms: (1) fever (body temperature >38°C); (2) purulent discharge from the airway; or (3) peripheral-blood white blood cell (WBC) count >10 × 10 9 /L or <4 × 10 9 /L. Based on the study inclusion and exclusion criteria, 1219 patients were identifed from 9 tertiary hospitals from January 2015 to December 2017 and divided into the VAP and non-VAP groups (137 patients in the VAP group and 1082 in the non-VAP group). 632 elderly ICU patients receiving mechanical ventilation (MV) in 3 other tertiary hospitals from January 2018 to September 2019 were selected as the model validation group using the convenience sampling method. Te selection criteria for research subjects and VAP diagnostic criteria were the same. 67 individuals were in the VAP group, and 565 individuals were in the non-VAP group.

Research Tools.
A VAP clinical data questionnaire for elderly ICU patients with MV was compiled by researching the literature and convening an expert group meeting. Te questionnaire had four parts, covering 27 indicators: (1) clinical demographics of patients, including gender, length of ICU stay, fever, days with fever, smoking, alcohol consumption, admission department, underlying diseases, the reason for admission, VAP prevention, and the composition of the tracheal tube cuf; (2) surgery and antibiotics, including whether or not surgery is performed, the number of surgeries performed, the duration of antibiotics used after surgery, and the use of antibiotics in combination; (3) laboratory indicators such as CRP level, urinary WBC count, and procalcitonin (PCT) level; (4) invasive monitoring conditions, including duration days of CVC, indwelling CVC, number of indwelling CVC, duration days of a urinary catheter, indwelling urinary catheter, number of an indwelling urinary catheter, duration of MV, and number of reintubations.

Data Collection Methods.
Trough a computer terminal, we used the hospital infection real-time monitoring system to collect clinical data according to the questionnaire we designed. Data were collected by two researchers who worked in the Department of Respiratory Medicine and the Department of Nosocomiology. Qualifed respiratory physicians provided the diagnosis of VAP. Data were input and statistically analyzed using two-person unifed coding, and then a third party checked all data to ensure accuracy. Tis study was approved by the Ethics Committee of the First Hospital of Shanxi Medical University (2020KK072). Informed consent was not required because of the study's retrospective nature.

Statistical
Methods. SPSS version 20.0 (IBM Corp., Armonk, NY, USA) was used for data analysis. Te count data are statistically described by frequency and percentage, and the VAP and non-VAP groups were compared using the chi-square test or Fisher's exact probability method. Te Wilcoxon rank-sum test was used for comparisons between more than two groups. We included variables with P < 0.05 in the univariate analysis in our logistic regression analysis and used the forward stepwise method to determine which variables to include in the predictive model. P < 0.05 indicated that the diferences were statistically signifcant. Multicollinearity between independent variables was determined by calculating the variance infation factor (VIF). Finally, according to the corresponding partial regression coefcient of each variable, we constructed an equation to establish the predictive model of VAP in elderly ICU patients with MV and visualized patient risk using a line graph. Te area under the receiver operating characteristic curve (AUROC) and the Hosmer-Lemeshow goodness-of-ft test were used to evaluate the efect of the predictive model.

Incidence of VAP.
From the initial modeling cohort of 1307 elderly patients on MV in the ICU, 88 had incomplete data and were excluded for not meeting the inclusion criteria. Tus, the fnal modeling cohort consisted of 1219 patients, of whom 137 (11.24%) developed VAP after MV.

Univariate-Regression Analysis.
To facilitate statistical analysis and clinical application, we grouped continuous variables, such as length of ICU stay, days with fever, the number of surgeries, duration of antibiotics used after surgery, days and number of CVCs, days and number of urinary catheters and duration of MV days, and number of reintubations by the median. Single-factor analysis revealed 15 statistically signifcant factors: length of ICU stay, surgery, number of surgeries, fever, days with fever, COPD, CRP, PCT, urinary WBC count, indwelling central-venous catheter, duration of CVC days, number of indwelling CVC, number of indwelling urinary catheter, indwelling urinary catheter, and number of reintubations (all P < 0.05; Table 1).

Construction of VAP Risk Prediction
Model. Based on the above six independent infuencing factors and their corresponding regression coefcients, a predictive model of VAP occurrence in elderly ICU patients with MV was established: A nomogram of VAP in these patients was constructed (Figure 1). Scores corresponding to each indicator can be obtained from the classifcation of variables in the nomogram, and these scores are added together to calculate the total score. In this study, the predicted probability corresponding to the total score was the probability of VAP occurrence in elderly ICU patients on MV.

Validation of Predictive Models.
In this study, the differentiation and calibration of the predictive model were evaluated using ROC curve analysis and the Hosmer-Lemeshow test. Te results showed that the AUC of VAP risk after intubation in the modeling group was 0.859 (95% CI: 0.828-0.890; P < 0.001; Figure 2). Sensitivity, specifcity, and the Youden index were 0.723, 0.848, and 0.570, respectively. Te AUC for the risk of pulmonary infection after intubation in the validation group was 0.813 (95% CI: 0.700-0.850; P < 0.001; Figure 3). Sensitivity, specifcity, and the Youden index were 0.714, 0.806, and 0.504, respectively. Te Hosmer-Lemeshow test showed P values for the modeling and validation groups of 0.365 (χ 2 � 8.734) and 0.485 (χ 2 � 7.484), respectively.

Discussion
VAP is one of the most common infections in patients with MV [10]. Te present study found that the incidence of VAP in elderly MV patients in the ICU setting was 11.24%, which was close to that in elderly patients in other geographic locations (10%) [11] but lower than that found in Europe, where the rates were 16.6% for 65-74 years old and 13% for ≥75 years old [12]. Te reasons for the diference from the present study might be that the European study did not reclassify patients >65 years old and that the diagnostic criteria and defnitions of VAP difer between countries. Te present study also determined that length of stay, surgery, CRP, and several uses of a ventilator were independent risk factors for VAP in elderly ICU patients with MV, while PCT and COPD were protective factors against VAP in this cohort.
Previous studies showed that VAP occurred in 27% of ICU patients receiving MV [11,13]. Pneumonia is associated with a prolonged hospital stay. Te total hospital stay of patients with postoperative pneumonia is 4.7 days longer than average; that of patients with tumors is 5.1 days longer; and that of patients with cerebrovascular disease is 4.0 days longer [9]. We found that prolonged ICU hospitalization was a risk factor for VAP (odds ratio (OR) � 1.542), possibly because VAP was associated with longer MV duration, thereby increasing the length of ICU stay and medical costs [14]. Tese new and published data suggest that to reduce costs, it is imperative to identify the risk factors associated with VAP and initiate preventive measures promptly.
Te incidence of VAP in trauma patients is 17.8% [15], which, in part, may be explained by changes in immune function after major trauma [2]. Elderly patients generally have low immune function and complications from a variety of chronic diseases, making them susceptible to VAP. Surgery may lead to decreased lung capacity, failure to clear secretions, and a reduced cough [16], all of which increase the risk of VAP. We found that surgery was highly positively correlated with the risk of VAP in elderly patients with MV in ICUs (odds ratio (OR) � 6.937) and that the OR value of surgery was signifcantly higher than those of other infuencing factors. Together, this suggests that surgery was the  Canadian Respiratory Journal most important infuencing factor in the occurrence of VAP. Considering these fndings, it seems reasonable for clinical staf to provide additional attention to patients undergoing surgery and administer personalized VAP prevention and control measures to these individuals. In addition, the specifc type of surgery may afect the incidence of VAP. Te  types of surgery included in this study were mainly traumatic brain injury and cardiac surgery. Te results showed that the incidence of VAP was 36% in patients with traumatic brain injuries [17]. In patients who had mechanical ventilation after cardiac surgery for more than 48 hours, the average incidence of VAP was 35.2% (range, 17.9%-53%) [18]. Subsequent studies will further analyze the incidence and infuencing factors of VAP in elderly patients with diferent types of surgery. PCT and CRP are commonly used clinical markers of infection. Interestingly, CRP increased by 86.22% in patients with COVID-19 [19]. Predictive models for severe COVID-19 showed that CRP was associated with a higher chance of severe disease [20,21]. Recently, the combination of PCT and CRP was employed in the diagnosis of VAP. Te combination was found efective in the diagnosis of VAP and in tracking the treatment efects. Additionally, dynamic monitoring of PCT and CRP predicted the prognosis of VAP patients [22] and is consistent with the results of the present study. Diferences in VAP condition, pathogen type, and detection time lead to changes in the levels of infammatory markers such as PCT and CRP. Yet, such changes may not be synchronous with the disease. Unexpectedly, we found CRP to be a risk factor and PCT to be a protective one for VAP. Due to the limitations of retrospective data analysis, this study could not determine the correlation between the time and timing of the blood specimen examination and the incidence of VAP. In future work, the relationship between the occurrence of VAP and the time and value of blood specimen collection should be further studied.
Reintubation is an important predictor of VAP [23]. Te incidence of VAP in patients who were reintubated more than twice (28.4%) was higher than that in patients reintubated ≤2 times (16.2%), [7]. Tis is similar to the results of the present study. Tis may be secondary to the loss of upper respiratory tract fltration and humidifcation and the loss of the tracheal cough refex, all of which promote colonization of the trachea by pathogenic microorganisms. Colonized pathogens form bioflms and increase the likelihood of lower respiratory tract infection [24,25]. It is recommended that clinical staf employ the spontaneous breathing test and balloon leakage test prior to extubation to minimize reintubation rates.
Despite recent advances in microbial technology, the epidemiological data and diagnostic criteria of VAP remain controversial, complicating the interpretation of treatment, prevention, and outcome studies [10]. Te increased incidence observed in COPD patients might be due to prolonged invasive MV (muscle weakness), a high incidence of microinhalation and bacterial colonization (defective mucosal ciliary clearance), and changes in local and general host defense mechanisms [26]. COPD is a risk factor for pneumonia after craniotomy [9]. In contrast, we found that COPD was a protective factor against the occurrence of VAP. Te discrepancy might be secondary to the selection of study subjects and the classifcation of disease severity. Te low number of COPD patients with VAP included in this study might also be related to the low rate of pulmonary function examination and the high rate of missed diagnosis in China [27].
In this study, we established a risk prediction model for elderly ICU patients with MV. Te results showed that the AUC of the prediction model in the modeling group was 0.859 and that in the verifcation group was 0.813, a decrease of only 0.046. Tis indicated that the risk prediction model had a strong ability and high accuracy in predicting the risk of VAP in elderly patients with MV in the ICU. In addition, the calibration degree of the prediction model was good. More specifcally, in the calibration degree test, the P values for the modeling and verifcation groups were 0.565 and 0.370, respectively. Tis indicated that the probability of VAP risk predicted by the model was close to the actual incidence. For simplicity, we constructed a nomogram based on the model. Te nomogram is easy to understand and can be used quickly. Medical personnel can use this model after the MV of patients to perform dynamic risk assessment and screen the high-risk population, allowing for efective preventive measures to achieve early prediction, early prevention, and early intervention.
Similar studies have been carried out in the past but in older patients (2011-2015) [7]. We added in newer data, and expanding it to ≥65 does add new data to the literature. Tere are several limitations to this study. First, this is a multicenter, retrospective study, and thus information on specifc types of parameter settings and sedation while individuals were on the MV is lacking. VAP is known to lead to severe lung injury, and accumulating evidence has suggested that driving pressure and mechanical power could refect lung injury and alveolar damage [28,29]. Being limited by the data collected from the information system, our study could not account for the impact of diferent markers of respiratory mechanics on VAP. Furthermore, the nature of the endotracheal tube, and especially the tube cuf, might have played a role in the results [30]. Although we collected data on tube materials, most hospitals use lower-priced polyvinylchloride tubes due to the infuence of national policy to reduce the consumable ratio. Terefore, the infuence of tube material on VAP needs further research. In addition, the endpoint of this study was discharge from the hospital. Some patients can develop lung infections after discharge; the incidence of VAP reported in this study might be lower than the true incidence. In terms of model verifcation, although we adopted a rigorous external verifcation method, data were collected from only three centers. In the future, the developed model will be applied clinically at multiple centers to refne its accuracy.

Conclusion
Analysis of over one thousand elderly ICU MV subjects found that >19 days in the ICU, surgery, CRP >8 mg/L, and >2 times of reintubations were associated with increased incidence of VAP. A VAP predictive nomogram, calculated AUC, and the Hosmer-Lemeshow goodness-of-ft test demonstrated acceptable model ft and relatively good performance. In clinical practice, physicians can use our nomogram to assess the risk of VAP development in elderly ICU patients with MV and initiate early preventive strategies. Efective preventive treatment might confer a better prognosis in these critically ill patients.

ICU:
Intensive care unit VAP: Ventilator-associated pneumonia MV: Mechanical ventilation COPD: Chronic obstructive pulmonary disease CRP: C-reactive protein PCT: Procalcitonin WBC: White blood cell CVC: Central vein catheterization CT: Computed tomography DVT: Deep venous thrombi AUROC: Te area under the receiver operating characteristic curve AUC: Te area under the curve.

Data Availability
Te data that support the fndings of this study are available from the corresponding author (shanglp2002@163.com), upon reasonable request.

Conflicts of Interest
Te authors declare that there are no conficts of interest.