Liver transplantation (LT) has become the optimal treatment for end-stage liver disease. Although liver is considered an immunologically privileged organ among solid organ transplants (1), the incidence of acute rejection following LT can still reach 10–30% (2, 3). Theoretically, T-cell-mediated rejection (TCMR) typically occurs within 4 to 6 weeks post-transplantation, representing the period of most intense immune rejection in LT recipients (4). If not promptly intervened, it can lead to graft dysfunction and even graft failure (5). Although Banff classification criteria remain the gold standard for diagnosing TCMR (6), percutaneous liver biopsy, has limitations in the prompt identification. Concurrently, pneumonia is a common infectious complication during the perioperative period. It exacerbates hepatic ischemia-reperfusion injury, contributing to delayed graft function recovery (7), adversely impacting clinical outcomes of LT and posing a significant threat to recipient survival (8). Hence, it is critically urgent to refine prognostic tools. Precise risk stratification for TCMR and pneumonia occurrence is essential for optimizing perioperative care. This may provide a framework for elucidating the pathophysiological mechanisms of perioperative TCMR and pneumonia in LT and developing individualized dynamic surveillance protocols. Such advancements hold significant implications for enhancing patient quality of life and prolonging survival outcomes, thereby optimizing long-term prognosis in this high-risk population. Postoperative pneumonia risk stratification has been enhanced by recently developed scoring systems, including the ISAN score (Intracerebral Hemorrhage, Sex, Age, NIH Stroke Scale) (9), Pneumonia Risk Index (non-cardiac surgery) (10), and Systemic Inflammation Score (post-gastrectomy) (11). However, these scoring systems are not applicable for assessing pneumonia after LT. They fail to account for the multifactorial synergy driving the co-occurrence of TCMR and pneumonia and struggle to capture nonlinear relationships or higher-order interactions among variables. Furthermore, conventional models exhibit low sensitivity in predicting TCMR post-LT (12), making it difficult to optimize immunosuppressive regimens for individual patients. For instance, while pulse steroid therapy can reduce rejection risk, it increases infection probability; traditional models are inadequate for quantifying such dynamic trade-offs.

In contrast, machine learning (ML) algorithms—through automated feature extraction and nonlinear modeling—demonstrate superior prediction accuracy. Techniques such as gradient-boosted ensembles outperform conventional scoring systems in predicting perioperative TCMR and pneumonia occurrence by decoding complex clinical variable interactions. Compared to traditional scoring systems, ML models have shown promising performance in predicting outcomes across various solid organ transplantation (13, 14). Currently, ML is widely applied to predict outcomes after LT (15, 16). Chen et al. (17) developed an ML model to predict pneumonia occurrence after LT. Maryam et al. (18) developed and validated an ML model demonstrating good performance in predicting plasma cell-rich rejection after LT. To mitigate the incidence of rejection following LT, Yoon et al. (19) employed ML algorithms to predict the optimal therapeutic range of tacrolimus, thereby advancing the clinical implementation of personalized immunosuppressive regimens. However, ML models specifically predicting the co-occurrence of perioperative TCMR and pneumonia after LT are scarce.

This study aims to develop and validate an ML-based predictive model for the co-occurrence of perioperative TCMR and pneumonia after LT. Ensemble algorithms and SHapley Additive exPlanations interpretability (SHAP) will be employed to identify nonlinear interactions and latent risk patterns that conventional methods may overlook, thereby addressing critical gaps in current prognostic frameworks. The clinical significance of this study lies in its potential to transform perioperative management strategies. With precise risk stratification, the established model could guide personalized therapeutic interventions, optimize surveillance protocols, inform immunosuppression regimen adjustments, and optimize the rational use of antimicrobial agents. These advances hold promise for reducing serious complications arising from perioperative TCMR and pneumonia after LT, thereby providing a basis for clinical decision-making.

Accurate prediction of risk factors for perioperative co-occurrence TCMR and pneumonia following LT is critical for timely intervention and improved outcomes. This study focused on recipients undergoing LT and validated a predictive model for assessing the simultaneous occurrence of TCMR and pneumonia during the perioperative period. The model integrated preoperative clinical characteristics, immunological interventions, dynamic changes in surgical parameters, preoperative concurrent symptoms, postoperative clinical characteristics, and donor information utilizing multiple ML algorithms. Comparative analysis of ML models revealed that the SVM algorithm achieved the optimal predictive accuracy, with AUC values of 0.881 (95% CI: 0.83–0.93) in the training set and 0.786 (95% CI: 0.69–0.88) in the test set. Calibration demonstrated high predicted-observed risk concordance. Furthermore, the SVM model demonstrated a high specificity, sensitivity, and F1 score. These findings indicate that ML models incorporating recipient-specific multidimensional data can effectively stratify perioperative co-occurrence TCMR and pneumonia, offering actionable insights for optimizing clinical decision-making and postoperative management.

The comparative analysis demonstrated the substantial advantage of ML models in prediction tasks (15, 20, 21). Torres et al. developed a machine learning-based model (K-prototype clustering algorithm) to predict post-liver transplant complications including acute rejection and infection, demonstrating favorable performance (22). This advantage can be attributable to ML’s ability to cope with complex data through advanced regularization techniques and ensemble learning mechanisms (23). Particularly, ML exhibits higher accuracy in capturing non-linear relationships, suggesting that statistical models may simplify complex biomedical interactions. Previous studies predominantly rely on clinical experience for variable selection, have confirmed the utility of ML in predicting TCMR or pneumonia prediction (17, 18). In contrast, this study integrated preoperative and postoperative clinical characteristics, immunological interventions, preoperative concurrent symptoms, donor information, and dynamic changes in surgical procedures to objectively identify key predictors (postoperative DBIL, postoperative INR, postoperative ALT, HDL, NK, FK506, Na⁺, operative time, anhepatic phase, induction regimen, and ICU stay) through LASSO regression. These predictors originated from objective data. A ML-based predictive model for perioperative co-occurrence TCMR and pneumonia following LT was subsequently developed.

As the most critical predictor in this model, the blood concentration of FK506 is paramount for preventing graft rejection and reducing the risk of infection, particularly pneumonia (24). A delicate therapeutic balance exists between these competing outcomes. FK506 suppresses the cellular immune response against the hepatic allograft primarily by inhibiting T-lymphocyte activation and proliferation, achieved through blockade of key cytokine transcription, including interleukin-2 (25). Subtherapeutic concentrations result in insufficient immunosuppression, failing to adequately inhibit recipient T-cell recognition and attack of the donor liver. This predisposes patients to acute cellular rejection, manifested by abnormal liver function tests (elevated transaminases and bilirubin). Severe or recurrent rejection episodes can lead to graft dysfunction or loss (5). Conversely, supratherapeutic concentrations, while theoretically offering enhanced rejection prophylaxis, incur significant costs: a marked increase in drug toxicity (26) and the risk of severe infections (27). Consequently, indiscriminately maintaining excessively high concentrations solely to achieve “absolute rejection avoidance” is clinically contraindicated, as the associated risks substantially outweigh potential benefits. FK506 exhibits a narrow therapeutic index between effective immunosuppression (preventing rejection) and toxic concentrations (predisposing to infection/toxicity) (28). Furthermore, significant interindividual variability in blood concentrations arises due to factors influencing drug metabolism, including genetics (e.g., CYP3A5 polymorphisms), age, hepatic and renal function, diet, and concomitant medications (29). Regular monitoring of FK506 trough levels is therefore fundamental. Based on these results, alongside individual rejection risk, infection susceptibility, and manifestations of drug toxicity, transplant clinicians must dynamically titrate the dosage through an individualized approach. The therapeutic goal is to maintain concentrations within a target range that effectively prevents rejection while minimizing the risks of infection and drug-related toxicity (27). ICU stay is also a significant predictor variable in our study. The duration of ICU hospitalization following LT serves as a critical indicator reflecting surgical complexity, graft functional recovery, severity of early complications, and the patient’s overall clinical status. While prolonged ICU stay itself is not a direct cause of rejection or pneumonia, it is strongly associated with and significantly increases the risk of developing both severe complications, with complex interactions existing between them. Prolonged ICU hospitalization is a well-established, highly significant independent risk factor for hospital-acquired pneumonia (30, 31). It is generally not the initiating cause of rejection; rather, severe rejection is often one of the primary factors leading to extended ICU stay. Furthermore, the complexity of the ICU environment and the patient’s critical condition significantly increase the difficulty of maintaining effective and stable immunosuppression, thereby indirectly elevating the risk of TCMR both during the ICU stay and shortly after transfer out of the ICU. Immunosuppression management poses substantial challenges in patients requiring protracted ICU care. In summary, the interplay between ICU length of stay, rejection, and pneumonia frequently establishes a vicious cycle. Severe rejection episodes often necessitate ICU admission and prolong hospitalization. The ICU environment and associated risk factors markedly increase the susceptibility to pneumonia (32). Severe pneumonia, in turn, necessitates reduction or discontinuation of immunosuppressive therapy to control the infection. Inadequate immunosuppression subsequently triggers rejection episodes or exacerbates existing rejection, thus creating a self-perpetuating cycle of adverse events. Operative time is also recognized as a significant predictive factor. Prolonged surgical time typically correlates with an extended anhepatic phase and increased total ischemic time. Sustained ischemia-reperfusion injury (IRI) leads to hepatic sinusoidal endothelial cell damage and microcirculatory disturbances, triggering the substantial release of damage-associated molecular patterns (33, 34). These damage-associated molecular patterns robustly activate the innate immune system, characterized by macrophage and neutrophil infiltration, and complement activation, resulting in the release of large quantities of pro-inflammatory cytokines (35). Damaged hepatocytes and endothelial cells exhibit upregulated expression of MHC molecules and co-stimulatory molecules (36), rendering them more recognizable as “non-self” by the recipient’s immune system. This facilitates the accelerated presentation of donor antigens to recipient T cells (37). Although the liver possesses inherent immunotolerance properties, severe IRI disrupts this microenvironment (38), significantly increasing the incidence and severity of early acute cellular rejection (39). Furthermore, prolonged operative duration is often associated with greater blood loss and substantial transfusion requirements. Allogeneic blood transfusion can induce complex immunomodulatory effects. It may also increase the risk of alloimmunization, including rejection directed against the graft, potentially through the introduction of allogeneic leukocyte antigens or by activating the recipient’s immune system (40). Extended operative time is a strong independent risk factor for postoperative pneumonia. A direct and well-established mechanistic link involves its association with longer durations of mechanical ventilation and an increased risk of pulmonary atelectasis (41). NK cells emerged as the fourth most important predictive variable in this study, exhibiting a negative regulatory role in the co-occurrence of TCMR and pneumonia. NK cells possess the ability to recognize and lyse virus-infected cells without requiring presensitization. They induce target cell apoptosis either through the release of perforin and granzymes or via the Fas ligand (FasL)/Fas pathway (42). Activated NK cells robustly secrete interferon-gamma and tumor necrosis factor-alpha. These cytokines not only exert direct antiviral effects (43) but also recruit neutrophils, monocytes/macrophages, and T cells to the site of infection, thereby amplifying the anti-infection immune response. Furthermore, NK cells can mediate antibody-dependent cellular cytotoxicity via FcγRIIIa (CD16a), enabling more effective killing of infected cells (44). In summary, insufficient NK cell counts can directly contribute to the development of perioperative pneumonia following LT. Viral infections themselves, particularly cytomegalovirus infection, constitute a significant risk factor for rejection (45). This creates a complex interplay: on the one hand, viral infection activates NK cells to combat the virus; on the other hand, cytokines like IFN-γ secreted by activated NK cells may simultaneously promote rejection. Conversely, the inflammation and tissue damage caused by rejection reactions also heighten the risk of infection.

This paper demonstrated that ML models could predict perioperative co-occurrence TCMR and pneumonia following LT. ML models exhibit superior abilities to capture complex, non-linear relationships and intricate interactions among clinical variables compared to traditional methods, thus offering valuable insights into individual patient risk profiles. This information can potentially inform personalized treatment strategies, optimize post-LT surveillance protocols, and ultimately improve patient outcomes.

This study developed a robust predictive model for perioperative co-occurrence TCMR and pneumonia in LT, and the SVM model achieved superior discriminative performance. Key predictors, including postoperative DBIL, postoperative INR, HDL, postoperative ALT, NK, FK506, Na⁺, operative time, anhepatic phase, induction regimen, and ICU stay, were identified as critical determinants of model. These results advance our understanding of the multifactorial pathogenesis of perioperative co-occurrence TCMR and pneumonia after LT, offering actionable insights for optimizing clinical decision-making and postoperative management.