Development and Internal Validation of a Web-Based Dynamic Nomogram for Dynamic Risk Stratification of Early Bacterial Infection in Liver Transplant Recipients

Introduction

Early bacterial infections remain a clinically important complication after liver transplantation and continue to contribute to early post-transplant morbidity and adverse outcomes in solid-organ transplant recipients [1,2]. In living donor liver transplantation (LDLT), early infectious events have been reported and linked to perioperative factors and postoperative complications, underscoring the ongoing need for reliable early risk identification in this population [3–5].

The early post-transplant phase is characterized by high susceptibility to infection driven by immunosuppression, surgical factors, and device exposure [2]. In contemporary cohorts of liver transplant recipients, infections during early follow-up commonly involve major sites such as the bloodstream, respiratory tract, and intra-abdominal sites, with gram-negative bacteria frequently predominating; these patterns have been described across adult and ICU-based transplant populations and supported by microbiologic epidemiology studies [6–8]. Moreover, multidrug-resistant gram-negative pathogens and bloodstream infections have been reported as clinically consequential entities after liver transplantation, reinforcing the importance of timely stratification and surveillance strategies [9,10]. Reported risk correlates for early bacterial infections after liver transplantation span a broad spectrum of recipient-related, perioperative, and postoperative factors, reflecting the complex and multifactorial nature of infection risk rather than the influence of a single dominant determinant [11–14]. However, existing prediction models have several limitations, including limited applicability to LDLT-specific populations, reliance on variables not readily usable at the bedside, and lack of tools for real-time individualized risk estimation [4,15,16]. Nomograms provide a practical approach by translating multivariable models into individualized risk estimates, and web-based dynamic nomograms further enable automated, real-time prediction in a clinically accessible format. However, such tools for early bacterial infection prediction after LDLT remain limited [17].

Therefore, this study aimed to develop and internally validate a dynamic risk stratification tool for early bacterial infections after LDLT using routinely available clinical variables and to translate it into static and web-based dynamic nomograms.

Material and Methods

STUDY DESIGN AND PATIENT SELECTION:

This single-center retrospective cohort study included adult patients (≥18 years) who underwent living donor liver transplantation between October 1, 2017, and October 31, 2025. The last eligible transplant was performed on October 31, 2025, and follow-up for the 30-day study outcome was completed on November 30, 2025. All included patients had complete 30-day follow-up for early bacterial infection ascertainment, and no patient was administratively censored due to incomplete follow-up. Patients were excluded if they had documented infections within 14 days prior to transplantation, received grafts from brain-dead deceased donors, or died within 48 hours after transplantation. Follow-up was continued until the occurrence of early bacterial infection, death, or 30 days after transplantation, whichever occurred first. The study protocol was approved by the Institutional Ethics Committee of the 108 Military Central Hospital, and the requirement for informed consent was waived due to the observational nature of the study.

DATA COLLECTION:

Demographic, pre-transplant, intra-transplant, and post-transplant data were collected from electronic medical records and a prospectively maintained transplant database. Pre-transplant clinical parameters included the Model for End-Stage Liver Disease score (MELD), white blood cell count (WBC), neutrophil-to-lymphocyte ratio, and the use of continuous renal replacement therapy or therapeutic plasma exchange (CRRT/TPE). Intra-transplant factors included operative, intra-transplant blood loss volume, and intra-transplant blood transfusion volume. Post-transplant variables included vasopressor use, post-transplant blood transfusion within the first week, bile leakage within 30 days after transplantation, early acute kidney injury (AKI), early allograft dysfunction (EAD), and duration of ICU stay. These post-transplant variables were collected during the early postoperative period and may reflect the evolving clinical course after transplantation, rather than strictly preceding the onset of infection.

The primary outcome of interest was the occurrence of early post-transplant bacterial infection. For patients who developed early bacterial infections, all infection sites and microbiologically confirmed pathogens identified during the early post-transplant period were recorded for descriptive analysis.

DEFINITIONS:

Early acute kidney injury was defined as acute kidney injury occurring within the first post-transplant week, according to the Kidney Disease: Improving Global Outcomes (KDIGO) clinical practice guidelines [18]. Early allograft dysfunction was defined according to the criteria proposed by Olthoff et al [19]. Bile leakage within 30 days after transplantation was defined according to the International Study Group of Liver Surgery criteria, based on biochemical evidence of biliary leakage and/or radiological or surgical confirmation [20]. Early post-transplant bacterial infection was defined as a microbiologically confirmed bacterial infection occurring within 30 days after transplantation [2]. For microbiological analysis, the unit of analysis was the individual isolate. In patients with polymicrobial infections, each pathogen was counted separately. A single patient could contribute multiple isolates and/or multiple infection sites if applicable. For the regression analysis, the outcome was defined at the patient level as the occurrence of any microbiologically confirmed bacterial infection within 30 days after transplantation, regardless of the number of isolates or infection sites.

IMMUNOSUPPRESSION PROTOCOL:

All recipients received a standardized immunosuppressive regimen consisting of basiliximab induction, perioperative corticosteroids, and maintenance therapy with tacrolimus and mycophenolate mofetil according to institutional protocols.

STATISTICAL ANALYSIS:

Statistical analyses were performed using SPSS version 27.0 (IBM Corp., Armonk, NY, USA) and R software version 4.3.3 (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables are presented as the mean±standard deviation for normally distributed data or as the median with interquartile range for non-normally distributed data, whereas categorical variables are presented as counts and percentages. Group comparisons were conducted using the t test or the Mann-Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. All statistical tests were two-sided, and a P value <0.05 was considered statistically significant.

Univariable logistic regression analysis was performed to identify potential factors associated with early post-transplant bacterial infections. Variables with clinical relevance or a P value <0.10 in univariable analysis were entered into a multivariable logistic regression model to identify independent predictors. In the multivariable model, post-transplant blood transfusion within the first week and bile leakage were coded as binary variables (0=no, 1=yes). Continuous predictors were entered using their original units without transformation and were assumed to have linear relationships multicollinearity among candidate predictors was assessed using variance inflation factors (VIF), and no significant multicollinearity was detected (all VIF <5). No imputation was performed, and analyses were conducted using complete-case data. No missing data were observed for the variables included in the analysis; therefore, the complete-case dataset was identical to the full cohort (n=205, including 45 events). Bootstrap validation was performed using the same dataset. Patients who died during the 30-day follow-up were retained in the dataset; death was not treated as missing data.

MODEL DEVELOPMENT, VALIDATION, AND CLINICAL UTILITY:

A multivariable logistic regression framework was used to develop a dynamic risk stratification tool for estimating early post-transplant bacterial infection risk during the postoperative course, in accordance with established principles for clinical prediction modeling [21]. Model discrimination was assessed using receiver operating characteristic (ROC) curve analysis and quantified by the area under the curve (AUC) [22]. The reported AUC represents the apparent model performance. Model calibration was evaluated using bootstrap-based calibration plots and quantitative error metrics, including mean absolute error (MAE), mean squared error (MSE), and the 90th percentile absolute error, which reflect the agreement between predicted probabilities and observed outcomes [21]. These error metrics were calculated from the residuals between observed outcomes and predicted probabilities. Internal validation was performed using bootstrap resampling with 1000 iterations to assess model stability and reduce optimism [23]. The clinical utility of the dynamic risk stratification tool was evaluated using decision curve analysis (DCA), which estimates net clinical benefit across a range of threshold probabilities and facilitates comparison with alternative strategies [24]. Because the tool was intended to reflect dynamic risk estimation rather than prediction at a fixed time point, no single classification threshold was prespecified.

DEVELOPMENT OF STATIC AND WEB-BASED DYNAMIC NOMOGRAMS:

Based on the final multivariable model, both static and web-based dynamic nomograms were constructed to facilitate individualized risk estimation within the dynamic risk stratification framework. Model development and visualization were implemented in R using the rms package for regression modeling and ggplot2 for graphical representation. The web-based dynamic nomogram was developed and deployed using the shiny framework, enabling automated real-time calculation of predicted probabilities and interactive access to model performance metrics.

Results

PATIENT ENROLLMENT AND GROUP CLASSIFICATION:

A total of 240 patients who underwent living donor liver transplantation between October 2017 and October 2025 were screened for eligibility. Of these, 35 patients were excluded based on predefined criteria: 9 were younger than 18 years old, 12 had documented infections within 2 weeks prior to transplantation, 11 received grafts from brain-dead deceased donors, and 3 died within 48 hours after transplant. The remaining 205 eligible patients were enrolled in the study. All included patients completed 30-day follow-up for infection assessment. These patients were subsequently categorized into 2 groups: the infection group (n=45, 21.9%) and the non-infection group (n=160, 78.1%) (Figure 1).

COMPARISON OF PATIENT CHARACTERISTICS BETWEEN 2 GROUPS:

A comparison of patient characteristics between the 2 groups (Table 1) showed that recipients who developed early post-transplant bacterial infections were more often female, had a higher prevalence of pre-transplant CRRT/TPE, higher MELD score, higher baseline WBC counts, higher neutrophil-to-lymphocyte ratios, and were more frequently transplanted for acute liver failure or acute-on-chronic liver failure (ALF/ACLF) compared with those without early post-transplant bacterial infections. Intra-transplant, the volume of blood transfused was significantly greater in the infection group. these patients exhibited higher post-transplant rates of vasopressor requirement, blood transfusion within the first week, bile leakage, early AKI, and EAD, along with a longer ICU length of stay. In contrast, no significant differences between groups were observed in age, history of diabetes mellitus, duration of liver transplantation surgery, post-transplant mechanical ventilation, or intra- and post-transplant WBC levels.

DISTRIBUTION OF BACTERIAL ISOLATES ACROSS INFECTION SITES:

The distribution of bacterial pathogens across infection sites after LDLT is summarized in Table 2. The microbiological analysis was performed at the isolate level; therefore, the total number of pathogens exceeds the number of infected patients because some patients had polymicrobial infections and/or infections involving multiple sites. Overall, gram-negative bacteria predominated, accounting for 75.8% of all post-transplant bacterial infections, whereas gram-positive bacteria comprised 24.2%. Among gram-negative pathogens, Klebsiella pneumoniae was the most frequently isolated organism (33.3%), followed by Escherichia coli (12.1%) and Acinetobacter baumannii (10.6%). Gram-negative infections were most commonly identified in the bloodstream, respiratory tract, and intra-abdominal sites. Gram-positive bacteria were less frequently isolated, with Enterococcus faecium being the most common gram-positive pathogen (9.1%), followed by Enterococcus faecalis and Staphylococcus haemolyticus (each 4.5%). Gram-positive infections were predominantly observed in the respiratory tract and intra-abdominal sites, while urinary tract and surgical site infections were relatively uncommon across all pathogens.

FACTORS ASSOCIATED WITH INFECTION IDENTIFIED BY LOGISTIC REGRESSION:

Table 3 shows that univariable logistic regression analysis identified several factors associated with early post-transplant bacterial infections, including pre-transplant CRRT/TPE use, higher MELD score, elevated pre-transplant WBC count, higher neutrophil-to-lymphocyte ratio, and ALF/ACLF as the indication for transplantation, as well as post-transplant vasopressor use, post-transplant blood transfusion within the first week, bile leakage, early AKI, EAD, and longer ICU length of stay. In the multivariable logistic regression model, higher MELD score (OR 1.04, 95% CI 1.01–1.07; P=0.035), post-transplant blood transfusion (OR 2.71, 95% CI 1.15–6.40; P=0.023), bile leakage (OR 4.12, 95% CI 1.62–10.50; P=0.003), and longer ICU length of stay (OR 1.11 per day, 95% CI 1.01–1.23; P=0.040) remained independently associated with the development of early post-transplant bacterial infections and were included as components of the dynamic risk stratification tool.

Multivariable model equation:

The regression coefficients of the final multivariable model are presented in Figure 2, corresponding to the equation used for construction of the dynamic risk stratification tool and its implementation in the web-based nomogram.

MODEL DISCRIMINATION AND CLINICAL UTILITY:

ROC analysis demonstrated that the predicted probability derived from the multivariable model showed the best discrimination, with an apparent AUC of 0.792 and 95% confidence interval (CI) 0.716–0.867, outperforming all individual predictors (Figure 3A). Among these, ICU length of stay (AUC=0.709) and pre-transplant MELD score (AUC=0.699) showed moderate discrimination, whereas bile leakage (AUC=0.629) and post-transplant blood transfusion (AUC=0.609) exhibited lower discriminative ability. Consistently, decision curve analysis (Figure 3B) indicated that the full model provided greater net clinical benefit across a broad range of threshold probabilities compared with individual predictors, supporting its potential clinical utility within the dynamic risk stratification framework of early bacterial infection in liver transplant recipients.

INTERNAL VALIDATION OF THE DYNAMIC RISK STRATIFICATION TOOL:

Internal validation using 1000 bootstrap resamples in 205 patients demonstrated good calibration and predictive accuracy of the dynamic risk stratification tool (Figure 4). The prediction error, derived from residuals between observed outcomes and predicted probabilities, was low (MAE=0.02; MSE=0.00103). Furthermore, the 90th percentile of the absolute error was 0.059, indicating that most predicted probabilities deviated minimally from the observed outcomes. The bias-corrected calibration curve closely approximated the ideal reference line across a wide range of predicted risks, reflecting good agreement between predicted and actual probabilities.

STATIC NOMOGRAM FOR RISK STRATIFICATION:

A static nomogram derived from the multivariable logistic regression model was developed to estimate individualized risk within the dynamic risk stratification tool of early post-transplant bacterial infections (Figure 5). Each predictor in the model was assigned a corresponding number of points on the upper scale, reflecting its relative contribution. The total score, obtained by summing the individual points, can be projected onto the probability axis to generate an individualized risk estimate. In addition, the density curves displayed above continuous variables illustrate the distribution of predictor values within the study cohort, indicating the range and frequency of observations used to construct the model.

WEB-BASED DYNAMIC NOMOGRAM FOR DYNAMIC RISK STRATIFICATION:

A web-based dynamic nomogram was developed to enable automated and individualized risk estimation for early post–liver transplant bacterial infections during the early postoperative period, as illustrated in Figure 6. By entering patient-specific clinical variables, including pre-transplant MELD score, bile leakage, ICU length of stay, and post-transplant blood transfusion within the first week, the application automatically calculates the corresponding total points and provides an individualized estimated probability of infection in real time. In addition to the risk estimation interface, the web application incorporates interactive navigation tabs that allow users to access supplementary information on the dynamic risk stratification tool, including model summary, internal validation using bootstrap resampling, decision curve analysis, and ROC curve performance, thereby facilitating transparent interpretation of the estimated risk. Accordingly, the dynamic risk stratification tool is intended to support dynamic, real-time risk stratification during the early postoperative period and should be interpreted as an association model rather than a prediction tool at a fixed postoperative time point.

The dynamic nomogram is available at: https://dynnomogramsicu.shinyapps.io/infectionpostldlt/. The web-based application was deployed on December 13, 2025, and accessed on March 21, 2026.

Discussion

Early post-transplant bacterial infections remain a frequent and clinically relevant complication after LDLT, with a predominance of gram-negative pathogens and infections mainly involving the bloodstream, respiratory tract, and intra-abdominal sites. In this context, we developed and internally validated a dynamic risk stratification tool to estimate individualized infection risk, which was subsequently implemented as both static and web-based dynamic nomograms.

The observed infection rate of 21.9% in our cohort aligns with the broad range reported in the literature, which varies from approximately 10% to 60% depending on study population, transplant type, and infection definitions [3–5,9,11,25,26]. Park et al reported a 9.7% incidence of bacteremia at 1 month after adult LDLT [4]. Mukhtar et al identified infectious complications in 52% of 246 LDLT patients within 3 months after transplant [5]. The predominanceof bloodstream, respiratory, and intra-abdominal infections in our cohort is consistent with previously reported patterns following liver transplantation. Respiratory tract infections are frequently described as the most common early post-transplant infections, with gram-negative pathogens such as Acinetobacter baumannii and Klebsiella pneumoniae commonly isolated [6,7]. Bloodstream infections represent another major infectious complication after liver transplantation and are often associated with substantial morbidity, with gram-negative bacteria – particularly Klebsiella pneumoniae – being frequently implicated [8,10,27]. Intra-abdominal infections, including biliary tract infections, are also commonly reported in the post-transplant setting and are typically related to surgical factors, with Enterococcus faecium and various gram-negative organisms among the predominant pathogens [27].

Four independent predictors emerged from our multivariable analysis: higher pre-transplant MELD score, post-transplant blood transfusion, bile leakage, and longer ICU length of stay. The association between elevated MELD score and increased infection risk has been well documented in the literature and reflects the multifaceted relationship between pre-transplant disease severity, intra-transplant vulnerability, and post-transplant complications [12–14]. Avkan-Oguz et al found that MELD score >20 was significantly associated with early bacterial infections in liver transplant recipients, while Park et al included MELD among univariable correlates of bacteremia in their LDLT cohort [4,12]. Post-transplant blood transfusion, identified as an independent predictor in our model, has similarly been associated with increased infection risk across multiple studies [9,11,12]. Gotoh et al identified surgical blood loss ≥5100 mL as an independent risk factor for bacteremia in LDLT recipients [11], while Lee et al demonstrated that intra-transplant blood loss >3000 mL and greater transfusion volumes correlated with higher bloodstream infection risk and mortality [9]. The immunomodulatory effects of blood transfusion, combined with markers of operative complexity and physiologic insult, likely contribute to this association. Bile leakage emerged as the strongest independent predictor in our model (OR 4.12), consistent with prior reports demonstrating OR ranging from 3.5 to 6.2 for bloodstream infections following biliary complications [5,9,25]. The pathophysiologic basis for this association is straightforward: bile leakage creates a nidus for bacterial contamination, triggers persistent local inflammation, and often necessitates interventional procedures or reoperation, all of which increase exposure to nosocomial pathogens and facilitate bacterial translocation into the bloodstream. Finally, prolonged ICU length of stay was independently associated with early bacterial infection risk, likely reflecting both increased exposure to invasive devices, antibiotics, and multidrug-resistant organisms, and greater post-transplant severity requiring prolonged critical care. Because some predictors, such as ICU length of stay and bile leak, are time-dependent and can occur after infection onset, reverse causality cannot be excluded. Therefore, the dynamic risk stratification tool should be interpreted as supporting dynamic, early postoperative risk stratification and as reflecting associations during the postoperative course rather than strictly pre-infection predictors [11,12,26].

The discriminative performance of our dynamic risk stratification tool, with an apparent AUC of 0.792, falls within the range typically reported for post-transplant infection prediction models. Published models have demonstrated AUCs ranging from 0.707 to 0.91, depending on outcome definitions, predictor selection, and cohort characteristics [4,15,16,28]. Park et al developed a bacteremia risk model with an AUC of 0.707, while Elkholy et al achieved an AUC of 0.91 for a composite risk index predicting early postoperative infections [4,15]. However, this performance represents apparent model performance based on internal validation and may be subject to optimism, and therefore should be interpreted with caution in the absence of external validation. Our model’s calibration metrics (MAE=0.02; MSE=0.00103) indicate minimal deviation between predicted probabilities and observed outcomes across the risk spectrum. While calibration metrics such as MAE and MSE are infrequently reported in prediction models for post-transplant infections, very low MAE and MSE values have been consistently used in other clinical prediction models as indicators of accurate calibration and minimal deviation between predicted probabilities and observed outcomes, supporting their validity as measures of model performance [29,30]. The decision curve analysis in our study demonstrated that the full model provided greater net clinical benefit compared to individual predictors across a broad range of threshold probabilities, supporting its potential utility for clinical decision-making. This finding aligns with the principle that multivariable models integrating complementary predictors typically outperform single-variable approaches by capturing the multifactorial nature of infection risk [16,28].

The translation of our dynamic risk stratification tool into both static and dynamic nomograms addresses a critical gap between statistical modeling and clinical implementation. Nomograms have been increasingly recognized as valuable tools for risk stratification, enabling clinicians to identify high-risk patients who may benefit from intensified surveillance, prophylactic interventions, or tailored antimicrobial strategies [15,16,28]. Elkholy et al framed their composite risk index as a tool to identify high-risk LDLT recipients for earlier intervention and monitoring, while Jie Yu et al demonstrated that their nomogram yielded a higher net clinical benefit on decision curve analysis across a clinically relevant range of threshold probabilities compared with the extreme strategies of treating all or no patients, supporting its potential utility for early risk stratification and post-transplant management [15,16]. The web-based dynamic nomogram developed in our study offers several advantages over traditional static nomograms: it automates calculation of predicted probabilities in real time, eliminates manual scoring errors, and incorporates interactive navigation tabs that provide transparent access to model performance metrics, including ROC curves, calibration plots, and decision curve analysis. This design facilitates bedside use by clinicians who may lack statistical expertise while maintaining transparency regarding model assumptions and limitations. Similar dynamic prediction tools have been advocated in other transplant and critical care settings to enhance clinical decision support and enable personalized risk assessment [17,31–33]. Beyond risk stratification, nomograms can inform antimicrobial stewardship by identifying patients who may benefit from targeted surveillance cultures, empiric coverage adjustments based on local resistance patterns, or early de-escalation strategies in low-risk individuals [15,28]. However, it is important to emphasize that risk stratification tools should complement, not replace, clinical judgment, and that their implementation must be accompanied by ongoing evaluation of impact on patient outcomes and resource utilization.

Several limitations should be acknowledged. This was a retrospective, single-center study, which may limit the generalizability of the findings, and the dynamic risk stratification tool has not yet been externally validated. The analysis focused exclusively on bacterial infections and did not include fungal or viral pathogens that can also contribute to early post-transplant morbidity. In addition, the sample size and number of events were relatively modest (n=205, number of events=45), potentially constraining model complexity. The reported AUC represents apparent performance derived from internal validation and may be optimistic; therefore, the true performance and generalizability of the tool remain uncertain without external validation. Importantly, because the tool incorporates time-dependent and postoperative variables such as ICU length of stay, bile leakage, and post-transplant transfusion, it reflects a dynamic risk profile that evolves during the clinical course rather than predicting future risk at a fixed time point.

Future studies should prioritize external validation of the dynamic risk stratification tool in independent patient cohorts to assess its generalizability, transportability, and true performance. Such validation would enable progression from a TRIPOD Type 1b study to Type 3 and is essential before clinical application. Prospective evaluation of the clinical impact of the dynamic risk stratification tool is also warranted. Incorporation of additional preoperative and donor-related factors, as well as dynamic post-transplant variables, may further improve performance and clinical usefulness.

Conclusions

We developed and internally validated a dynamic risk stratification tool for early bacterial infections after LDLT, implemented as both static and web-based dynamic nomograms. By incorporating postoperative and time-dependent variables, the tool reflects evolving risk during the early clinical course rather than predicting outcomes at a fixed time point. It demonstrated good apparent performance in this single-center cohort; however, its true performance and generalizability remain uncertain. Therefore, external validation in independent cohorts is required before any clinical application.