Liver regeneration-associated machine learning architecture integrating time-phased predictions for post-hepatectomy liver failure

Le résumé fourni par la source

Background: Post-hepatectomy liver failure (PHLF) is the leading cause of morbidity and mortality following major hepatectomy. Existing prediction models inadequately capture the dynamic liver regeneration and perioperative changes, limiting their predictive accuracy. We aimed to develop a machine learning (ML) modelling system (PILOT architecture) integrating liver regeneration-associated biomarkers with time-phased perioperative data for PHLF prediction. Methods: This retrospective multicentre study included 1071 patients undergoing major hepatectomy at three centres (2019-2024), divided into training (n = 623) and two external validation cohorts (n = 206 and 242). Fifty-five perioperative variables, including novel liver regeneration-associated biomarkers (GATA3, RAMP2, VEGFA, PEDF), were categorised into three time-phased datasets (preoperative, intraoperative, postoperative). Thirteen ML algorithms were evaluated across these datasets, with gradient-based feature reduction strategies applied to optimise the PILOT models. This study is registered with ClinicalTrials.gov (NCT05779098). Findings: < 0.050). A risk-stratification framework integrating PILOT-Pre and PILOT-Intra predictions achieved a class-specific precision of 94.4%-96.6% for PHLF events in the consensus high-risk group and 92.1%-95.5% for non-PHLF events in the consensus low-risk populations. SHAP analysis revealed that serum phosphorus levels >2.4 mg/dL on postoperative day 3, liver RAMP2-GATA3 ratios <10.1, and serum PEDF-VEGFA indices >4.9 were associated with an increased predicted PHLF risk. Interpretation: The PILOT architecture integrates liver regeneration-associated biomarkers with time-phased data to accurately predict PHLF within the first 6 h postoperatively. Based on consistency analysis of predictions of PILOT-Pre and PILOT-Intra models, this framework enables early risk stratification, thereby providing a practical tool for personalised perioperative management. Funding: This research was funded by the projects from National Natural Science Foundation of China (82403243), Program for National Postdoctoral Researchers Funding of China (GZC20231943), and Shanghai Municipal Commission of Science and Technology (23Y11905900).

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.