Interpretable SVM Model for Predicting CMV Infection in Seropositive Kidney Transplant Recipients: A Single-Center Retrospective Study

Introduction

Kidney transplantation (KTx) is the preferred treatment for end-stage renal diseases. Although immunosuppressants reduce rejection rates, they increase the risk of opportunistic infections.^1,2 Cytomegalovirus (CMV), a herpesvirus, usually remains latent in immunocompetent hosts but can cause severe complications in transplant recipients,³ CMV may manifest as asymptomatic viraemia, a syndrome, or tissue-invasive disease,^4,5 and has also been linked to rejection, cardiovascular events, and post-transplant diabetes.^6–8

CMV infection is influenced by multiple factors, with donor and recipient CMV serostatus serving as key determinants for risk stratification.^9,10 CMV-seropositive recipients (R+) were considered to have a moderate infection risk, regardless of donor serostatus (D+ or D−).¹¹ Moreover, the majority of kidney transplant recipients are CMV-seropositive,¹² yet the optimal preventive strategy for this group has not been clearly established.^13,14 Universal prophylaxis and preemptive therapy represent the two main strategies for CMV infection, both of which are effective when patients are closely monitored.^15,16 If a preemptive therapy strategy is adopted, a reduced frequency of CMV monitoring may increase the risk of infection and adversely affect graft outcomes. Despite reducing CMV infections, prophylaxis is associated with high costs, drug resistance, and leukopenia.^17,18 In addition, dose reduction of prophylactic agents appears to be a feasible approach, and several studies have shown that low-dose prophylaxis can be effective in R+ recipients at relatively low risk of infection. Therefore, accurate assessment of individual CMV risk is crucial for selecting the optimal preventive strategy in R+ patients.^19,20

Conventional CMV serostatus-based stratification is insufficient for accurately assessing infection risk in R+ recipient, especially for guiding individualized prophylaxis.²¹ Assessing whether a recipient is over- or mildly immunosuppressed may therefore provide additional value.²² This immune status of transplant recipients is closely associated with CMV infection, and peripheral blood lymphocyte subsets reflect the overall immune condition.²³ Non-specific lymphocyte subsets have limited accuracy in infection risk prediction, and research using CMV-specific lymphocyte subsets has shown improved predictive performance.²⁴ However, its clinical application remains restricted owing to high costs, technical complexity, and limited availability in some centers. Therefore, there is an urgent need to develop a risk assessment tool based on simple and readily available indicators with high accuracy.

In recent years, machine learning (ML) has shown great potential in the field of transplantation, with successful applications in predicting postoperative infections, assessing graft function, and other clinical aspects.^25,26 Machine learning algorithms can process high-dimensional data and capture complex nonlinear relationships, which can enhance the predictive accuracy of clinical models.^27,28 Moreover, recent studies have increasingly emphasized supervised and explainable machine learning frameworks, demonstrating their value in improving predictive accuracy and model interpretability in complex prediction tasks.^29,30 Combining interpretable machine learning with non-specific immune indicators holds the potential to further enhance the predictive capability and interpretability of the model.

In this study, we applied six machine-learning algorithms to develop a risk model for CMV infection in R+ recipients to support individualized prevention and management. Because post-transplant CMV infection occurs mainly during the early post-transplant period and the prophylactic treatment window, we focused on predicting CMV DNAemia within 6 months after transplantation. Clinical variables and non-specific lymphocyte subsets were included in the model. SHapley Additive exPlanations (SHAP) was used to improve model transparency. Ultimately, we will obtain an interpretable tool for assessing the risk of R+ receptor infection.

Materials and Methods

Data Source and Study Population

This study applied a machine-learning model to predict CMV infection within 6 months after kidney transplantation (KTx). Data were collected from multiple sources to improve completeness and transparency. Donor characteristics, recipient demographics, and transplant-related variables were obtained from the China Organ Transplant Response System (COTRS). Laboratory test results and treatment-related information were extracted from the electronic medical record system of Zhujiang Hospital. Follow-up and outcome data were supplemented through telephone follow-up when necessary. The final dataset included donor and recipient demographics, transplant-related characteristics, laboratory findings, medication regimens, and clinical outcomes.

A total of 162 adult patients who underwent KTx between June 2023 and December 2024 were included in this study. Patients were included (1) aged ≥18 years, (2) CMV-seropositive status before transplantation, and (3) negative CMV-DNA test before transplantation. Patients were excluded (1) death or allograft nephrectomy within a month post-transplant, (2) prior history of other solid organ transplantation, (3) loss to follow-up.

Study Variables and Data Preprocessing

Feature selection plays a crucial role before model training because it effectively reduces noise and alleviates overfitting by eliminating irrelevant and redundant features. According to clinical experience and previous reports,^10,31 common clinical indicators and immune parameters measured approximately a month post-transplantation were included in the analysis, while variables with more than 30% missing data were excluded to ensure model robustness. Finally, we included the remaining 36 features in further analysis. Additionally, we employed multiple imputation to reduce accidental bias by excluding participants with missing data, addressing missing data. Subsequently, we applied the Z-score algorithm to normalize the input values, achieving consistent evaluation of attributes and preventing model overfitting. The final variables included the following:

Recipient: sex, age, height, weight, body mass index (BMI), anti-CMV IgG titer kidney transplant history, CMV prevention strategy, panel reactive antibody (PRA), and dialysis modality.

Donor: sex, age, height, weight, BMI, serum creatinine level, and donor type.

Transplant-related: cold ischemia time, warm ischemia time, total perioperative intravenous steroid dose, HLA mismatch, and induction therapy.

Laboratory indicators (around a month post-op): CD4+, CD8+, CD4+CD27-, CD4+CD28- CD8+CD27-, CD8+CD28- counts, CD4+/CD8- ratio, and recipient’s scr.

Clinical events included: delayed graft function (DGF), BK virus infection, and rejection before infection.

Dataset Splitting and Variable Selection

Patients were randomly divided into training and validation sets in a 7:3 ratio. The Boruta algorithm was used for feature selection in the training set, which compared the Z-score of each variable against that of randomly permuted “shadow features.”³² After 500 iterations with a significance threshold of 0.05, the features were classified as: important (green area, consistently higher Z-scores than shadow features), unimportant (red area, consistently lower), or tentative (yellow area, intermediate). Both important and tentative features were retained for subsequent modeling. To ensure stability and avoid multicollinearity, the variance inflation factor (VIF) was calculated,³³ with a VIF <10 considered acceptable.

Definitions

Microbiological and Serology Tests

Quantitative surveillance of serum CMV DNA was performed using a commercial PCR kit (DaAn Gene Co., Ltd., China) on an ABI 7500 system (Thermo Fisher Scientific, USA). The assay had a lower limit of quantification (LLOQ) of 500 copies/mL. Samples with a detectable signal below the LLOQ were reported as “positive, <500 copies/mL”, whereas samples ≥500 copies/mL were quantified and reported as an integer value. Pretransplant CMV serostatus of donors and recipients was assessed using an electrochemiluminescence immunoassay (ECLIA) on a Cobas e801 analyzer (Roche Diagnostics, Germany).

CMV Infection and Monitoring Strategy

In this study, the primary outcome (“CMV infection”) was defined as CMV DNAemia, ie, any detectable serum CMV DNA by PCR, including results reported as “positive, <500 copies/mL” CMV syndrome and CMV disease were diagnosed according to established guideline definitions.³⁴ Following a preoperative baseline assessment, serum CMV DNA was monitored weekly during the first postoperative month and subsequently at months 1, 3, and 6 after transplantation. All included patients had at least 6 months of follow-up.

Prophylaxis and Preemptive Therapy

CMV management at our center follows a stepwise dose-escalation approach. For prophylaxis, valganciclovir is initiated at a low dose (112.5 mg daily) after renal function recovery—a dose lower than that reported in previous studies,^35,36 and maintained for at least three months; if CMV DNA replication occurs, the dose is escalated until clearance and then reduced back. For preemptive therapy, low-dose valganciclovir (225–450 mg daily) is started upon confirmed replication, escalated until control, and discontinued after two consecutive negative tests. Some patients do not receive prophylaxis for economic or other reasons.

Rejection

Rejection was defined based on the clinical criteria and/or confirmed by pathological findings from a graft kidney biopsy.³⁷

Delayed Graft Function

DGF was defined as the need for at least one dialysis treatment within one week after kidney transplantation.³⁸

Model Construction

After feature selection using the Boruta algorithm only in the training set, six machine learning models were developed: logistic regression (LR),³⁹ support vector machine (SVM),⁴⁰ XGBoost (XGB),⁴¹ random forest (RF),⁴² LightGBM (LGBM),⁴³ and multilayer perceptron (MLP).⁴⁴ CMV infection within 6 months of KTx was defined as a binary outcome. Hyperparameters for all models were optimized by grid search with 10-fold cross-validation within the training set to maximize the area under the receiver operating characteristic curve (AUC).⁴⁵ For the SVM model, the tuning process specifically included kernel type, regularization parameter C, and gamma. The final optimal hyperparameter settings for all models, including the SVM parameters, are provided in Table S1. The optimal parameter configuration obtained through grid search effectively reduces the risk of overfitting and improves the predictive performance of each machine learning model on the dataset.⁴⁶

Model Explainability

The SHAP (Shapley Additive Explanations) framework, based on the principles of cooperative game theory,⁴⁷ was used to interpret the best-performing model. The SHAP values quantify the contribution of each feature, allowing for the ranking of feature importance and overall model interpretation. Mean absolute SHAP values and Beeswarm plots were used to illustrate the relative importance of the features and their directional effects on CMV infection risk, respectively.

Statistical Analysis

All analyses were performed using the R software (version 4.2.7). Continuous variables were expressed as mean ± SD or median (IQR), and compared using Student’s t test or Mann–Whitney U-test, respectively. Categorical variables were summarized as frequencies (%) and analyzed using the Chi-square or Fisher’s exact test. A two-tailed p <0.05 was considered statistically significant.

Results

Patient Characteristics

After applying the exclusion criteria, 162 patients were included in this study. A total of 78 patients were excluded: 70 patients were under 18 years of age, 1 patient was CMV-seronegative, 1 patient had a history of liver transplantation, 2 patients died within a month post-transplant, 2 patients underwent graft nephrectomy within one month, and 4 patients were lost to follow-up. The flow of the study is shown in Figure 1.

Figure 1 Flow chart of the study. A total of 243 kidney transplant recipients were screened (June 2023–December 2024). After applying the exclusion criteria, 162 recipients were included and split into a training set (n=115) and a testing set (n=47).

Among the 162 CMV-seropositive recipients, 83 (51.2%) developed CMV DNAemia. Comparisons between the infected and non-infected groups in terms of donor and recipient characteristics, transplantation-related variables and laboratory variables around day 30 are presented in Table 1. Within the infected group, the median recipient age was significantly higher than that in the non-infected group (37.0% vs.46.0%, P = 0.007), suggesting that increasing age elevates the risk of infection. Donor serum creatinine levels were significantly higher in the infected group (92.0 vs 131.0, P = 0.005), and the proportion of living donor kidney transplants was significantly lower, suggesting that donor-related and perioperative factors may contribute to CMV risk, possibly through differences in graft condition, ischemia exposure, and early post-transplant inflammatory burden. No significant differences were observed in other recipient or donor variables between the two groups. Among transplantation-related variables, cold ischemia time was associated with infection, with prolonged cold ischemia time being a risk factor for infection, whereas no significant differences were found in other variables such as warm ischemia time or induction therapy. In post-transplantation laboratory examinations, CD4+ and CD8+ T-lymphocyte counts were significantly elevated in the non-infected group, suggesting that lymphocyte counts may reflect the risk of infection. The training and validation sets were well-balanced, are documented in Table S2.

Table 1 Characteristics of Non-Infected and Infected Recipients

To further assess the impact of different preventive strategies, patients were divided into a universal prophylaxis group and a preemptive therapy group (Table 2). The median time to infection was 33 days post-transplantation, and 48.2% of infected patients exhibited a viral load exceeding 500 copies/mL. Fourteen patients developed CMV syndrome, while three were diagnosed with tissue-invasive CMV disease confirmed by biopsy, including two cases of gastrointestinal involvement and one case of CMV hepatitis. In the preemptive therapy group, 65.3% of the patients developed CMV infection, and approximately 37.9% required immediate antiviral treatment. In contrast, the incidence of CMV infection was significantly lower in the universal prophylaxis group (31.3%), with only four patients (6.0%) showing viral loads exceeding 500 copies/mL. The incidence of CMV syndrome and tissue-invasive CMV disease was also lower in the prophylaxis group, although the difference was not statistically significant. Overall, in our cohort, low-dose valganciclovir prophylaxis was associated with a lower incidence of CMV DNAemia and fewer cases with viral load >500 copies/mL, while severe CMV manifestations were numerically less frequent but not statistically significantly different. These findings suggest that, when combined with risk stratification, prophylactic strategies may be further individualized in moderate-risk recipients.

Table 2 Comparison of CMV Infection Outcomes Between the Prophylaxis and Preemptive Therapy Groups

Model Development and Model Performance

Boruta selected seven predictive features in the training set: CD8⁺ T cell count, CD4⁺ T cell count, CD4⁺CD27⁻ cell count, recipient age, cold ischemia time, donor type, and CMV prevent strategy (Figure 2). No multicollinearity was detected (VIF < 10, Table S3). Six machine learning models were constructed and optimized.

Figure 2 Feature selection based on the Boruta algorithm. The horizontal axis is the name of each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. The green boxes represent important variables, yellow boxes represent tentative variables, and red boxes represent unimportant variables. Blue boxes represent shadow features variables, which serve as reference importance.

Model performance was assessed using receiver operating characteristic (ROC) curve analysis. Figure 3A and B present the ROC curves for the training and validation cohorts, respectively. Among six models, the SVM model achieved an AUC of 0.9463 (0.898–0.983) in the training set and the highest AUC of 0.821 (0.692–0.932) in the validation set. Although the SVM model has a slight overfiting, it has the highest AUC value in the validation sets of all models, and the degree of overfitting is less than that of other models. In addition, this model showed the highest specificity (85.00%) and PPV (86.96%), while also maintaining good clinical applicability, with a sensitivity of 68.97% and an F1 score of 76.92%. The performance metrics of all models are presented in Table 3, including AUC, Accuracy, Sensitivity, Specificity, PPV, NPV, and F1 score. Furthermore, the SVM model showed good calibration, with a Brier score of 0.135 in the training set (Figure 4A) and 0.160 in the validation set (Figure 4B), and provided greater net benefit across a clinically relevant threshold range of approximately 0.10–0.70 in the validation cohort (Figure 5). Therefore, the SVM model was identified as the optimal model for predicting the post-transplant infection risk. This may be because SVM is well suited to relatively small datasets with mixed clinical and immunological variables and can model nonlinear class boundaries while maintaining better generalization.⁴⁸ In contrast, several tree-based models showed near-perfect training performance but weaker validation performance, suggesting a greater tendency toward overfitting in the present dataset.

Table 3 Detailed Performance Metrics of Various Machine Learning Models for Predicting CMV Infection Risk in CMV-Seropositive Patients Across Training and Test Sets

Figure 3 Receiver operating characteristic (ROC) curves of the six prediction models. (A) Training cohort. (B) Validation cohort. The area under the ROC curve (AUC) with 95% confidence intervals for each model is shown in the legend.

Figure 4 Calibration curves of the six prediction models. (A) Training cohort. (B) Validation cohort. The dashed diagonal line indicates perfect calibration. Brier scores (95% CI) for each model are shown in the legend.

Figure 5 Decision curve analysis of the six prediction models. (A) Training cohort. (B) Validation cohort. Net benefit is plotted against threshold probability.

Model Interpretability Based on SHAP

The SHapley Additive exPlanations (SHAP) method was employed to interpret the output of the SVM model. This game theory–based approach quantifies the contribution of each feature to the model’s predictions. The SHAP summary plots (Figure 6). Feature importance was ranked in descending order based on mean absolute SHAP values, as illustrated in Figure 6A. Among all features, CD8+ T cell count, CD4+ T cell count, and CD4+CD27- T cell count contributed the most substantially to the prediction. This pattern is biologically plausible because post-transplant CMV control depends predominantly on cellular immunity, and these T-cell subsets may more directly reflect the host’s antiviral immune competence than broader demographic or perioperative variables.⁴⁹ Figure 6B shows the SHAP beeswarm plot, which visualizes the impact of individual feature values on the model predictions. Positive SHAP values indicated an increased risk of post-transplant infection in our model. Specifically, lower levels of CD8+, CD4+, and CD4+CD27- T cells were associated with positive SHAP values, thus pushing the prediction of infection. The remaining features made relatively small contributions. To further evaluate the structure of the model, we developed an exploratory model using only three immunological features. This resulted in a marked decline in the predictive performance, with AUC values decreasing in both the training and validation cohorts (Figure 7). This finding suggests that the inclusion of clinical variables improves predictive performance in internal validation. Clinical variables provide complementary prognostic information beyond immune-cell subsets alone, because they capture non-immunological contributors to CMV risk such as recipient vulnerability, graft-related injury, donor characteristics, and preventive management strategy. Therefore, the seven-feature SVM model may serve as a promising internally validated tool for predicting the risk of post-transplant CMV infection.

Figure 6 SHAP-based interpretation of the final model. (A) Importance ranking plot of features of the logistic regression model. (B) Characteristic attributes in SHAP. The abscissa is the SHAP value, and each line denotes a feature. Higher eigenvalues are indicated by red dots, and lower eigenvalues are indicated by blue dots.

Figure 7 Receiver operating characteristic (ROC) curves of prediction models built using the top three predictors. (A) Training cohort. (B) Validation cohort. The area under the ROC curve (AUC) for each model is shown in the legend.

Discussion

In this single-center study, we developed and validated a predictive SVM model for cytomegalovirus infection after kidney transplantation. The model incorporated routine clinical variables and nonspecific immune parameters, which were selected using the Boruta algorithm, and constructed using machine learning techniques. Compared with previously reported approaches based on CMV-specific immune assays,^50,51 our model relies on routinely available clinical and nonspecific immune variables, which may improve feasibility and accessibility in routine transplant care. Although direct performance comparisons across studies should be interpreted cautiously because of differences in study populations, endpoints, monitoring strategies, and validation methods, our findings suggest that a model based on readily available variables can still achieve encouraging predictive performance while offering greater practical applicability.

Boruta is a random-forest–based feature selection algorithm capable of handling both linear and nonlinear variables.⁵² By applying this method, we were able to accurately identify the true risk factors associated with post-transplant CMV infection. The use of the Boruta algorithm effectively reduced the number of input features, thereby simplifying the model while maintaining high predictive performance. This streamlined model can easily be implemented to assist clinicians in optimizing antiviral prophylaxis and post-transplant management strategies.

Machine learning offers a powerful computational approach capable of handling complex datasets and uncovering the nonlinear relationships among variables. In the present study, among the six tested machine learning algorithms, the support vector machine (SVM) model achieved the highest AUC and demonstrated superior predictive accuracy and clinical net benefit. The utility of SVM has been supported by several studies^53,54 In this study, we developed an SVM-based model incorporating seven easily accessible variables to predict the risk of post-transplant CMV infection.

This study aimed to predict early CMV infection after kidney transplantation. In the SVM model, seven variables (CD8+ T cell count, CD4+ T cell count, CD4+CD27- T cell count, recipient age, cold ischemia time, donor type, and CMV prophylaxis strategy) were identified as key predictors of infection risk. Immune induction therapy was completed before one month after surgery, but lymphocytes remained stable for a longer period of time under the action of cytodepleting agents,⁵⁵ thus cellular immune responses play a central role in controlling CMV infection.⁵⁶ CD8+ T cells are the major effector cells responsible for viral clearance.⁵⁷ Therefore, higher CD8+ T cell counts were associated with a lower risk of infection in our model. CD4+ T cells are also essential for regulating antiviral immunity and guiding immunosuppressive therapy. While high CD4+ T cell levels have been linked to an increased risk of rejection, low levels are associated with a greater susceptibility to infection.⁵⁸ By secreting cytokines such as IFN-γ and TNF-α, to modulate CD8+ T cells, NK cells, and B cells, CD4+ T cells contribute to viral control. This association is biologically plausible because CMV control after transplantation relies heavily on cellular immunity; lower CD4+ and CD8+ T-cell counts may indicate impaired antiviral immune competence under post-transplant immunosuppression. Loss of CD27 indicates terminal differentiation, which enhances IFN-γ secretion and antiviral activity. CMV infection drives the expansion of CD27-CD45RA+ terminally differentiated subsets within Vδ2-γδ T cells, particularly in seropositive recipients.

Recipient age has consistently been identified as a major risk factor for CMV reactivation. Older recipients showed a significantly higher incidence of infection,⁵⁹ likely due to immunosenescence and frailty-related immune dysfunction. Prolonged cold ischemia time has also been recognized as a high-risk factor for CMV infection reactivation,⁶⁰ as ischemia-reperfusion injury can trigger inflammatory responses that promote viral activation.³ CMV prophylaxis strategy plays a crucial role in infection prevention,⁶¹ this is clinically plausible because prophylaxis modifies the early post-transplant viral replication environment and may affect both the timing and magnitude of CMV DNAemia during the period of highest immunosuppressive intensity.⁶² The inclusion of this variable in our model allowed for individualized risk assessment across diverse clinical settings. For instance, among patients not receiving universal prophylaxis, those identified as low risk by the model may be safe with a reduced monitoring frequency. Conversely, for patients under universal prophylaxis, those classified as high risk may benefit from enhanced antiviral dosing or prolonged prophylaxis duration. In addition, induction therapy is an important determinant of CMV infection risk.⁶³ Although the induction regimen was not retained in the final model and did not differ significantly between groups, we observed a slightly higher proportion of anti-thymocyte globulin (ATG) use in the infected group, while the number of patients without induction was small. This may reflect limited statistical power due to the modest sample size. Future studies with larger cohorts are needed to better evaluate the contribution of induction therapy to CMV risk prediction.

From a clinical perspective, the main problem addressed in this study is the insufficient risk stratification of CMV-seropositive kidney transplant recipients. Although these recipients are generally classified as having intermediate CMV risk, they remain clinically heterogeneous, and serostatus-based classification alone is often insufficient to guide individualized prevention and monitoring strategies.⁶⁴ Our model was therefore developed to address a real-world management problem: how to identify, early after transplantation, which R+ recipients may require closer virologic surveillance or a more intensive preventive approach, and which recipients may be suitable for a less intensive management strategy. In real-world practice, the model may serve as a clinical decision-support tool rather than a purely hypothetical prediction model. Because all seven predictors are routinely available in standard post-transplant care, the model could be applied during the early post-transplant period without requiring specialized CMV-specific immune assays. Patients classified as higher risk may benefit from closer CMV DNA surveillance, earlier intervention, or a more individualized prophylactic strategy, whereas those classified as relatively lower risk may be considered for less intensive monitoring under appropriate clinical supervision. In this way, the model may help bridge the gap between uniform serostatus-based management and personalized CMV prevention.

The relatively high incidence observed in our study may be explained by several factors. First, the study endpoint was detectable CMV DNAemia rather than CMV disease alone, and therefore included both asymptomatic CMV replication and CMV disease. Second, the prophylactic dose used in our center was relatively low in a subset of patients, which is consistent with the current tendency to reduce prophylactic intensity in intermediate-risk recipients in order to minimize treatment-related complications and drug exposure; notably, not all patients received universal prophylaxis. Third, patients underwent relatively close virologic monitoring during follow-up, partly because of concern that lower-dose prophylaxis might miss clinically relevant infection, and even low-level DNAemia was captured.

In this study, pediatric recipients, as well as high-risk (D+/R−) and low-risk (D−/R−) populations, were not included due to differences in immune maturity, disease spectrum, and management strategies, as well as a high proportion of missing clinical data.^65,66 Nevertheless, the model may be applicable to these populations in future studies, as it relies on routinely available and clinically interpretable variables. It should be noted that this is an exploratory study with a relatively small sample size, and issues such as overfitting may exist; therefore, the model requires further validation in large, multicenter cohorts. Predictive models based on artificial intelligence can support clinical decision-making, but should be used as a reference tool and interpreted in conjunction with comprehensive clinical judgment.

In summary, we successfully established an interpretable machine learning model to predict CMV infection in CMV-seropositive kidney transplant recipients based on routinely available clinical and non-CMV-specific immune variables. The final SVM model demonstrated promising predictive performance in this single-center cohort. When integrated with a precision risk-prediction model, low-dose valganciclovir prophylaxis may offer a more individualized preventive approach for moderate-risk recipients.

Limitation

Although this study developed a machine learning–based model with promising predictive performance, several important limitations should be acknowledged. First, the relatively small sample size is a major limitation of the present study and may reduce the stability and robustness of model estimation and cause overfitting. Therefore, the current findings should be validated in prospective and multicenter cohorts. Second, owing to the retrospective design, several variables are frequently unavailable, such as donor CMV serostatus and recipient diabetes history. Including these variables in future analyses may further enhance predictive accuracy and support more precise risk stratification. Third, this study focused on early post-transplant CMV infection within the 6 months, and late-onset CMV infection beyond this period was not assessed. Studies with larger cohorts and longer follow-up designs are needed to further validate the model and assess its generalizability. Finally, differences in CMV prophylactic regimens across centers may affect model applicability, external validation in diverse clinical settings is essential before clinical implementation.

Conclusion

We developed an interpretable SVM model based on routinely available clinical variables and immune parameters to predict early CMV infection in CMV-seropositive kidney transplant recipients. In this single-center cohort, the model showed promising predictive performance and may support clinically meaningful risk stratification. In practice, such risk stratification may help identify recipients who require closer CMV DNA surveillance or a more individualized prophylactic strategy during the early post-transplant period. However, the model should be regarded as a decision-support tool rather than a substitute for clinical judgment. Prospective external validation in multicenter cohorts and direct comparison with CMV‑specific immune diagnostics are required before clinical implementation.