Prediction of the Total Liver Weight using anthropological clinical parameters: does complexity result in better accuracy?

Abstract

Background: The performance of linear models predicting Total Liver Weight (TLW) remains moderate. The use of more complex models such as Artificial Neural Network (ANN) and Generalized Additive Model (GAM) or including the variable "steatosis" may improve TLW prediction. This study aimed to assess the value of ANN and GAM and the influence of steatosis for predicting TLW.

Methods: Basic clinical and morphological variables of 1560 cadaveric donors for liver transplantation were randomly split into a training (2/3) and validation set (1/3). Linear models, ANN and GAM were built by using the training cohort and evaluated with the validation cohort.

Results: The TLW is subject to major variations among donors with similar morphological parameters. The performance of ANN and GAM were moderate and similar to that of linear models (concordance coefficient from 0.36 to 0.44). In 28-30% of cases, TLW cannot be predicted with a margin of error ≤20%. The addition of the variable "steatosis" to each model did not improve their performance.

Conclusion: TLW prediction based on anthropological parameters carry a significant risk of error despite the use of more complex models. Others determinants of TLW need to be identified and imaging-based volumetric measurements should be preferred when feasible.