Optimizing graft weight prediction in living donor liver transplantation: a machine learning approach integrating pathologic, radiological and biological factors

Abstract

Background

Graft weight (GW) prediction is essential in living donor liver transplantation (LDLT) to mitigate the risks associated with small-for-size grafts. This study aimed to develop and validate a machine learning (ML)-based model integrating clinical, biochemical, radiological, and histopathological features to enhance GW prediction accuracy.

Materials and methods

This study included 1001 LDLT donors between 2014 and 2023. Preoperative clinical data, graft volume (GV), radiological and histological steatosis grading were analysed. Multiple models were developed using an 80% training and 20% test split. Model performance was assessed using mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) and compared with conventional GW prediction methods.

Results

The ML model demonstrated superior performance, with an MAE of 48.51 g, an RMSE of 65.12 g, and a MAPE of 8.49%. Compared with traditional methods, the ML approach significantly reduced prediction errors ≥15% (10.94% vs. 37.31% for conventional models, p < 0.001). Feature analysis highlighted GV, donor anthropometrics, lipid profiles, and CT attenuation values as key predictors, while hepatic steatosis had minimal impact.

Conclusions

ML-based GW prediction outperformed conventional methods by incorporating radiological and biochemical parameters, achieving greater accuracy and reducing clinically significant errors.