Statistical methods versus machine learning techniques for donor-recipient matching in liver transplantation.

Area of Science:

• Organ transplantation research
• Medical informatics
• Statistical modeling in healthcare

Background:

• Donor-recipient matching is a critical challenge in liver transplantation due to organ scarcity.
• Efficient allocation methods are essential to optimize outcomes for the growing number of recipients.

Purpose of the Study:

• To compare the effectiveness of classical statistical methods versus machine learning techniques for liver donor-recipient matching.
• To evaluate model performance using multiple survival endpoints and established scoring systems.

Main Methods:

• Utilized the United Network for Organ Sharing database with 39,189 donor-recipient pairs and 28 variables.
• Compared Logistic Regression (LR) and Naïve Bayes (NB) against machine learning models like MLP, RF, GB, and SVM.
• Evaluated models against standard scores: MELD, SOFT, and BAR, across 3-month, 1, 2, and 5-year endpoints.

Main Results:

• Logistic Regression (LR) demonstrated superior performance (e.g., AUC=0.654 at 5 years) compared to complex machine learning models (e.g., MLP, GB, SVM, RF).
• LR outperformed standard scores (MELD, SOFT, BAR) across all evaluated time points.
• Advanced machine learning techniques did not yield improved predictive accuracy for liver allocation.

Conclusions:

• Classical statistical methods, specifically Logistic Regression, remain highly effective for liver donor-recipient matching.
• Complex machine learning models did not surpass the performance of LR, potentially due to database limitations.
• The findings underscore the enduring value of established statistical approaches in organ allocation strategies.