Introduction: The Challenge of Post-Hepatectomy Liver Failure
Post-hepatectomy liver failure (PHLF) remains the most formidable complication following major hepatic resection, serving as a primary driver of postoperative morbidity and mortality. Despite advancements in surgical technique and perioperative care, the incidence of PHLF in patients undergoing major hepatectomy for malignancy continues to pose a significant clinical challenge. The fundamental pathophysiology of PHLF lies in the imbalance between the volume/function of the future liver remnant and the metabolic demands of the body, often exacerbated by impaired liver regeneration.
Traditional predictive models, such as the Child-Pugh score, MELD score, and various preoperative liver function tests (e.g., indocyanine green clearance), provide a snapshot of hepatic reserve but frequently fail to capture the dynamic, time-sensitive nature of liver regeneration and the physiological shifts occurring during the intraoperative and early postoperative periods. To bridge this gap, the PILOT (Perioperative Integrated Liver-regeneration Optimized Time-phased) architecture was developed, utilizing machine learning to integrate novel biomarkers with longitudinal clinical data.
Highlights of the PILOT Architecture
The study introduces several pivotal advancements in the field of hepatobiliary surgery and predictive analytics:
1. Integration of Novel Biomarkers: The model incorporates specific liver regeneration-associated biomarkers, including GATA3, RAMP2, VEGFA, and PEDF, which reflect the molecular state of the liver remnant.
2. Time-Phased Predictive Modeling: By categorizing data into preoperative, intraoperative, and postoperative phases, the system allows for continuous risk reassessment.
3. Superior Predictive Accuracy: The PILOT models achieved Area Under the Curve (AUC) values up to 0.904, vastly outperforming traditional scoring systems.
4. Early Clinical Actionability: The framework enables high-precision risk stratification within the first six hours post-surgery, providing a critical window for intervention.
Study Design and Methodology
This retrospective multicenter study (ClinicalTrials.gov: NCT05779098) analyzed data from 1,071 patients who underwent major hepatectomy across three high-volume centers between 2019 and 2024. The cohort was divided into a training set (n = 623) and two independent external validation cohorts (n = 206 and 242) to ensure the generalizability of the findings.
The researchers evaluated 55 perioperative variables. A unique aspect of this study was the inclusion of four novel biomarkers identified through previous transcriptomic research as critical to liver regeneration: GATA3 (GATA binding protein 3), RAMP2 (Receptor activity modifying protein 2), VEGFA (Vascular endothelial growth factor A), and PEDF (Pigment epithelium-derived factor). These variables were organized into three distinct datasets:
1. PILOT-Pre: Preoperative clinical data and biomarkers.
2. PILOT-Intra: Integration of preoperative data with intraoperative factors (e.g., blood loss, operative time).
3. PILOT-Post: Inclusion of early postoperative data (within 6 hours).
Thirteen different machine learning algorithms were tested, including LightGBM and XGBoost, with SHAP (SHapley Additive exPlanations) analysis used to interpret the contribution of individual features to the model’s predictions.
Key Findings and Performance Metrics
The PILOT architecture demonstrated exceptional discriminative ability across all phases of the perioperative journey.
Model Discrimination and Validation
In the training cohort, the PILOT-Pre and PILOT-Intra models (utilizing LightGBM with 10 and 15 features, respectively) achieved AUCs of 0.754 and 0.787. The PILOT-Post model, which utilized XGBoost with 20 features, reached an AUC of 0.904. These results were successfully replicated in the external validation cohorts, with AUCs ranging from 0.740 to 0.895. In contrast, traditional models such as the MELD score or ALBI (Albumin-Bilirubin) grade showed significantly lower predictive power, with AUCs between 0.502 and 0.644 (P < 0.050).
Risk Stratification and Precision
A consensus risk-stratification framework was developed by integrating PILOT-Pre and PILOT-Intra predictions. This framework categorized patients into high-, intermediate-, and low-risk groups. The consensus high-risk group demonstrated a class-specific precision for PHLF events of 94.4% to 96.6%. Conversely, the consensus low-risk group showed a precision of 92.1% to 95.5% for non-PHLF events. This level of accuracy allows clinicians to identify with high confidence which patients require intensive monitoring and which can be fast-tracked in their recovery.
Identification of Critical Risk Thresholds
SHAP analysis identified several key physiological and molecular thresholds associated with an increased risk of PHLF:
1. Serum Phosphorus: Levels >2.4 mg/dL on postoperative day 3 were identified as a significant predictor. While higher phosphorus is often associated with regeneration, the study highlights the complexity of its timing and levels.
2. RAMP2-GATA3 Ratio: A ratio 4.9 was linked to increased PHLF risk, reflecting a pro-inflammatory or anti-angiogenic state that hinders liver recovery.
Expert Commentary: Mechanistic Insights and Clinical Utility
The success of the PILOT architecture lies in its ability to quantify the biological reserve of the liver through the lens of regeneration. The inclusion of RAMP2 and VEGFA is particularly significant; these proteins are central to the sinusoidal endothelial cell signaling required for hepatocyte proliferation. When the RAMP2-GATA3 axis is disrupted, the orchestrated process of vascular and parenchymal growth is compromised, leading to functional failure of the remnant liver.
Furthermore, the time-phased approach acknowledges that surgery is a dynamic stressor. An intraoperative event, such as a prolonged period of ischemia or excessive blood loss, can shift a patient from a low preoperative risk to a high postoperative risk. By providing a reliable prediction within six hours of surgery, PILOT allows for early pharmacological interventions, such as the use of growth factors or specialized nutritional support, and optimized fluid management to mitigate the severity of PHLF.
However, limitations must be noted. As a retrospective study, there is an inherent risk of selection bias. While the external validation was robust, the integration of tissue-based biomarkers (like RAMP2-GATA3) requires rapid pathological processing which may not be available in all clinical settings. Future prospective trials are necessary to determine if interventions based on PILOT predictions directly improve patient survival rates.
Conclusion
The PILOT architecture represents a significant leap forward in personalized surgical care. By integrating the molecular underpinnings of liver regeneration with advanced machine learning, it provides a practical and highly accurate tool for predicting PHLF. This framework moves the field away from static, one-size-fits-all assessments toward a dynamic, data-driven approach that can identify high-risk patients in the earliest stages of the postoperative period, potentially transforming the management of patients undergoing major liver surgery.
Funding and Registration
This research was supported by the National Natural Science Foundation of China (82403243), the Program for National Postdoctoral Researchers Funding of China (GZC20231943), and the Shanghai Municipal Commission of Science and Technology (23Y11905900). The study is registered with ClinicalTrials.gov under the identifier NCT05779098.
References
1. Shen H, Yuan T, Si A, et al. Liver regeneration-associated machine learning architecture integrating time-phased predictions for post-hepatectomy liver failure. EClinicalMedicine. 2025;90:103661. doi:10.1016/j.eclinm.2025.103661.
2. Reissfelder C, Brand K, Sobotka C, et al. Prediction of posthepatectomy liver failure: A systematic review of existing scoring systems. HPB (Oxford). 2021;23(1):1-12.
3. Rahbari NN, Garden OJ, Padbury R, et al. Posthepatectomy liver failure: a definition and grading by the International Study Group of Liver Surgery (ISGLS). Surgery. 2011;149(5):713-724.