Letter to the Editor: Splitting the risk: Interpretable machine learning enters the era of split liver transplantation

Abstract

Split liver transplantation (SLT) has always existed in a delicate tension between expanding access and accepting complexity. By converting one deceased donor liver into two grafts, SLT directly addresses the persistent organ shortage, but at the cost of extended transection planes, altered hemodynamics, and a consistently higher rate of early postoperative complications (EPC) compared with whole liver transplantation, especially biliary and vasculobiliary events. In this context, the work by Wang et al, published in the recent issue of World Journal of Gastroenterology, on right tri-segment SLT and EPC represents an important conceptual step forward. The authors do not simply refine the list of risk factors, but deliberately place interpretable machine learning at the center of perioperative risk stratification in this highly vulnerable population.

Key Words: Artificial intelligence; Cirrhosis; Surgery; Split liver; Transplantation

Core Tip: In a single-center retrospective cohort of 109 adult recipients of right tri-segment split liver transplantation, Wang et al report an early postoperative complications (EPC) rate of 33.9%, including hemorrhage, vascular complications, biliary events, ascites and infections. They evaluated four modeling strategies, random forest, support vector machine, extreme gradient boosting and logistic regression, identifying random forest as the best-performing algorithm. Critically, they did not stop at prediction performance. Using SHapley Additive exPlanations, they quantified the contribution of each variable to the model and highlighted six major drivers of EPC risk.

TO THE EDITOR

Split liver transplantation (SLT) remains a valuable strategy to expand the donor pool, yet it is consistently associated with a higher rate of early postoperative complications (EPC), particularly biliary and vasculobiliary events, compared with whole liver transplantation[1]. In this context, we read with interest the study by Wang et al[1] published in the recent issue of the World Journal of Gastroenterology, which represents an important conceptual advance by integrating interpretable machine learning (ML) into perioperative risk assessment for adult right tri-segment SLT.

In a retrospective cohort of 109 recipients, the authors reported an EPC rate of 33.9% and identified random forest as the best-performing predictive model. The application of Shapley Additive exPlanations allowed transparent interpretation of model outputs, highlighting model for end-stage liver disease (MELD) score, log-transformed systemic immune-inflammation index (LnSII), intraoperative blood loss, operative time and segment IV partial lobectomy as key contributors to EPC risk. Multivariable analysis confirmed MELD score, LnSII, blood loss and absence of segment IV resection as independent predictors, forming a nomogram with good discrimination. The finding that resection of ischemic segment IV tissue may reduce EPC is particularly noteworthy. Segment IV has long been recognized as a major source of ischemia-related complications in extended right grafts[2-4]. When graft volume is adequate, its deliberate removal may represent a targeted risk-reduction strategy rather than unnecessary surgical trauma. Beyond technical aspects, the study reinforces the emerging role of systemic inflammation as a determinant of post-transplant outcomes. Elevated SII has been associated with early graft dysfunction and reduced survival after liver transplantation, supporting its inclusion in perioperative risk stratification alongside established severity scores such as MELD. Overall, Wang et al[1] demonstrate how interpretable ML can move SLT risk assessment beyond descriptive factors toward a more individualized and potentially modifiable framework, bridging advanced analytics with clinical and surgical decision-making.

WHAT DID WANG ET AL ACTUALLY SHOW?

In a single-center retrospective cohort of 109 adult recipients of right tri-segment SLT, Wang et al[1] report an EPC rate of 33.9%, including hemorrhage, vascular complications, biliary events, ascites and infections. They evaluated four modeling strategies, random forest, support vector machine, extreme gradient boosting and logistic regression, identifying random forest as the best-performing algorithm.

Critically, they did not stop at prediction performance. Using SHapley Additive exPlanations, they quantified the contribution of each variable to the model and highlighted six major drivers of EPC risk: LnSII, albumin-to-fibrinogen ratio, MELD score, partial lobectomy of segment IV (IV PL), intraoperative blood loss and operative time[5].

Multivariable logistic regression then distilled four independent predictors of EPC: Higher MELD score, higher LnSII, greater intraoperative blood loss, and absence of IV PL. From these variables, the authors built a nomogram with an area under the curve (AUC) of 0.788 and good calibration for EPC prediction, while MELD and LnSII were also associated with 5-year overall survival[5].

Perhaps the most provocative message is that resection of IV PL appears protective, reducing EPC and favoring early liver function recovery, provided that graft volume remains adequate[5].

SEGMENT IV: FROM “COLLATERAL DAMAGE” TO MODIFIABLE TARGET

The Achilles’ heel of extended right SLT grafts has long been segment IV. Earlier series repeatedly underlined that the split plane and compromised inflow/outflow can create ischemic parenchyma in segment IV, a well-recognized substrate for bile leaks, local necrosis, intra-abdominal infection and hemorrhage[1,4]. These concerns have driven technical innovation, including specific strategies for segment IV vascular reconstruction, aimed at preserving functional parenchyma and reducing ischemic complications without precipitating small-for-size syndrome[1,4].

Wang et al[1] extend this narrative in an original way: When segment IV is irreversibly ischemic, leaving it in situ may convert it into a necrotic, non-functional focus that fuels EPC under postoperative immunosuppression. Their finding that IV PL independently reduces EPC risk supports the view that, once graft-to-recipient weight ratio is secure, removing clearly ischemic segment IV tissue may function as a targeted de-risking maneuver rather than unnecessary surgical trauma[4,5].

This interpretation is biologically plausible and aligns with previous observations on segment IV-related complications in extended right grafts[1,4], but it remains hypothesis-generating: A single-center cohort with 37 EPC events cannot alone redefine standard practice. Prospective and multicenter validation will be essential before IV PL can be considered a default strategy in right tri-segment SLT.

INFLAMMATION, MELD AND BLOOD LOSS: CONVERGING SIGNALS OF VULNERABILITY

The study also sits firmly within a growing body of evidence linking systemic inflammatory indices and disease severity scores to post-transplant outcomes. The SII combining platelets, neutrophils and lymphocytes has been repeatedly associated with survival after liver transplantation for hepatocellular carcinoma, early allograft dysfunction and short-term mortality in acute-on-chronic liver failure[6-8]. Elevated SII reflects a shift towards neutrophil-driven inflammation and impaired lymphocyte-mediated immune surveillance, an immunological context that plausibly contributes to infection, graft dysfunction and poorer survival[6-8]. By identifying LnSII as an independent predictor of EPC and long-term overall survival, Wang et al[1] reinforce the notion that systemic inflammation is not just a bystander, but an active driver of postoperative trajectories after liver transplantation[5,6-8].

Similarly, the MELD score remains one of the most robust and validated predictors of perioperative mortality, complication rates and graft survival in liver transplantation[9,10]. High-MELD recipients consistently experience greater perioperative morbidity and more complex postoperative courses, even when long-term outcomes are acceptable in carefully selected candidates[9,10]. In the study by Wang et al[1], the association between higher MELD score, EPC and delayed early liver function recovery is entirely consistent with this broader literature[5,9,10].

The role of intraoperative blood loss is equally important. Long recognized as a surrogate for surgical complexity and intraoperative instability[11], major blood loss is associated with hemodynamic swings, transfusion-related immunomodulation, dilutional coagulopathy and amplified ischemia-reperfusion injury, all of which can magnify the risk of EPC. Wang et al[1] confirm intraoperative blood loss as an independent determinant of EPC in right tri-segment SLT[11], reinforcing the message that blood loss is not a neutral by-product, but a genuine, modifiable risk factor.

Taken together, these findings support a unifying view: EPC after SLT emerges from the interaction between a stressed graft and a vulnerable host, characterized by high inflammatory burden, limited functional reserve and exposure to prolonged, hemodynamically demanding surgery.

ML IN LIVER TRANSPLANTATION: BEYOND BLACK BOXES

Perhaps the most forward-looking aspect of the work by Wang et al[1] is how they build their risk model. ML has already entered liver transplantation in several domains: Prediction of acute kidney injury after liver transplantation using explainable ML models[12], pre-transplant and early post-transplant survival prediction using ensemble learners[13], ML-based tools for biliary complication prediction[14], and broader frameworks for integrating ML into clinical decision-making in hepatology and transplantation[15].

However, the adoption of ML has been hampered by concerns about opacity, overfitting and limited generalizability[15]. The key innovation of the study by Wang et al[1] is its explicit response to these concerns by comparing multiple algorithms, rather than presenting a single “best” model[5]. The authors employed SHapley Additive exPlanations values to enhance model transparency and used patient-level clinical characteristics as well as cohort-level features to interpret model predictions[5,12]. Importantly, this approach enabled cross-validation of ML-derived signals using conventional multivariable logistic regression before incorporating predictors into a nomogram[5].

This hybrid strategy, leveraging ML for feature discovery and ranking, followed by consolidation within more familiar statistical frameworks, may represent a pragmatic template for future studies using small- and medium-sized transplant datasets[12,13,15].

Nevertheless, important caveats remain. The dataset is single-center and modest in size; No external validation is presented; And the model uses essentially static perioperative variables without dynamic updating from early postoperative data. These limitations do not detract from the originality of the approach, but they do underscore that clinical implementation will require broader, prospective, multicenter efforts.

HOW SHOULD THIS CHANGE PRACTICE TODAY?

For teams engaged in SLT programs, several pragmatic and important messages emerge from this study, which could be revolutionary. The first is the concept to treat segment IV strategy as an explicit decision point and not an emergency choice. In right tri-segment SLT, the fate of segment IV should not be an afterthought but a calculated risk that could permit less risk and more security for patients. Where graft volume is secure, resection of clearly ischemic segment IV should be considered part of a deliberate risk-reduction strategy, grounded in accumulating evidence that segment IV ischemia contributes disproportionately to EPC[1,4,5]. Including SII or similar inflammatory indices in multidisciplinary discussions and patient counseling could help refine expectations and tailor surveillance, particularly in borderline candidates for SLT. This approach could allow earlier identification and targeted intervention to prevent catastrophic events and major complications[1,6-8]. In particular, intraoperative blood loss, often considered an inevitable but frequent occurrence in liver transplantation, could be used to stratify postoperative risk in real time during surgery, thereby transforming it into a potentially modifiable risk factor rather than a passive by-product of the procedure. The independent association between blood loss and EPC strengthens the rationale for low central venous pressure anesthesia, meticulous hemostasis and careful hemodynamic management in SLT[1,11]. Small technical and anesthetic optimizations may have a meaningful impact on complication rates.

The nomogram proposed by Wang et al[1], with an AUC of 0.788, should be used as a tool for structured clinical reasoning rather than as a definitive oracle. Although promising, it remains exploratory and is best regarded at present as a framework to guide perioperative risk assessment rather than a ready-to-use clinical decision support system.

WHERE SHOULD THE FIELD GO NEXT?

The study by Wang et al[1] also highlights priorities for future research: External and multicenter validation of the nomogram, ideally across programs with different SLT volumes, technical preferences and case mixes; Integration of dynamic postoperative data (early graft function tests, Doppler ultrasound, perfusion imaging) into time-updated models that refine EPC risk during the first postoperative days; Risk stratification enables the evaluation of artificial intelligence-guided intervention strategies and may inform clinicians and surgeons about optimal postoperative management, such as intensified monitoring, tailored immunosuppression, or early imaging in high-risk patients; Harmonization of inflammatory biomarkers in liver transplantation research to enable comparison and meta-analysis of the SII, neutrophil-to-lymphocyte ratio, platelet-lymphocyte ratio and related indices[6-8,15].

Ultimately, the central message is conceptual. It is no longer sufficient to state that SLT carries a higher risk of EPC. By combining interpretable ML with surgical insight and pathophysiological reasoning, Wang et al[1] show that this excess risk can be deconstructed into concrete, partially modifiable components: Inflammatory tone, disease severity, hemodynamic insult and the management of segment IV. For a procedure born of necessity to “split” a scarce resource, SLT may now become a proving ground for data-driven, personalized surgery in liver transplantation.