Artificial Intelligence
Artificial Intelligence and Liver Transplantation: Improving Waiting List Prioritization
AI and Equity in Liver Transplant Prioritization
Liver transplantation is built around a major challenge: allocating available grafts to the patients who need them most, while keeping prioritization as fair and accurate as possible. For several years, allocation models have been used to estimate the risk of death or clinical deterioration among patients on the liver transplant waiting list. However, the authors point out that current prioritization models mainly rely on linear statistical approaches. This may limit their accuracy in the sickest patients, especially when biological values fall outside conventional ranges. These are precisely the patients for whom accurate prioritization is especially important. In this context, a study published in Clinical Gastroenterology and Hepatology presents GEMA-AI, the Gender-Equity Model for Liver Allocation using Artificial Intelligence. This explainable artificial intelligence model was developed to predict outcomes for patients on the liver transplant waiting list, using the same input variables included in existing models.
Why liver transplant prioritization needs to evolve
The imbalance between available donors and patients waiting for liver transplantation requires allocation strategies that can reduce the risk of death on the waiting list. The MELD-Na score is currently considered a reference model for ranking patients according to mortality risk. It combines several objective laboratory parameters, including bilirubin, international normalized ratio, creatinine, and sodium. Despite the use of MELD-Na, the authors report that the risk of death or delisting due to clinical deterioration remains significant depending on geographical area, and is higher among women. More recent models, such as MELD 3.0 and GEMA-Na, have been developed to reduce waiting list mortality and address gender disparities in access to liver transplantation. GEMA-Na notably replaces serum creatinine with the Royal Free Hospital cirrhosis glomerular filtration rate, known as RFH-GFR.
What GEMA-AI changes
GEMA-AI is based on an explainable artificial neural network. It uses the same variables as GEMA-Na: international normalized ratio, bilirubin, sodium, and RFH-GFR. The main difference lies in the method. Conventional models assume a linear relationship between biological variables and waiting list risk. However, the authors explain that this does not always reflect clinical reality, especially when values are very low or very high. Sodium is an important example, as its relationship with risk may follow a “U-shaped” pattern. A nonlinear model such as GEMA-AI may better capture these situations. The model was also designed to be explainable, making its results easier to understand and potentially more acceptable for clinical implementation.
A study based on two international cohorts
The study included 9,320 adults listed for liver transplantation. Data from the United Kingdom were used for model training and internal validation. This cohort covered the period from 2010 to 2020 and included 7,682 patients, divided into a training cohort of 5,762 patients and an internal validation cohort of 1,920 patients. An Australian cohort of 1,638 patients listed between 1998 and 2020 was used for external validation. The primary outcome was a composite endpoint: death on the waiting list or delisting due to clinical deterioration within 90 days after listing.
More accurate predictions than comparator models
The authors compared GEMA-AI with GEMA-Na, MELD-Na, and MELD 3.0. In the internal and external validation cohorts, GEMA-AI showed better discrimination than the comparator models. This improvement was particularly pronounced in women and in patients with at least one extreme analytical value. In the training cohort, GEMA-AI showed better discrimination than MELD-Na and MELD 3.0. The difference with GEMA-Na was not significant in this cohort, but GEMA-AI performed better than GEMA-Na in both the internal and external validation cohorts. The study also reports that GEMA-AI showed better calibration in specific subgroups, particularly in patients with at least one extreme analytical value. In this subgroup, GEMA-AI was the only model with adequate calibration.
Potential impact on patient prioritization
The study did not only assess the statistical performance of the model. It also evaluated its potential impact on waiting list composition. Compared with existing models, GEMA-AI changed the position of a significant proportion of patients on the waiting list. A change of at least 2 score points, considered clinically meaningful by the authors, occurred in 27.8% of patients compared with GEMA-Na, 67.7% compared with MELD 3.0, and 61.5% compared with MELD-Na. Within the first 90 days, differential prioritization occurred in 6.4% of patients compared with GEMA-Na. This comparison is particularly important because GEMA-AI and GEMA-Na use the same input variables. It therefore reflects the impact of moving from a linear model to a nonlinear model. According to the study estimates, implementing GEMA-AI could potentially avoid 1 in 59 deaths overall compared with GEMA-Na, and 1 in 13 deaths among women.
Particular relevance for women and the sickest patients
One of the important findings of the study relates to gender disparities. The authors recall that there is a historical gender imbalance in access to liver transplantation, with women tending to wait longer and showing an increased risk of delisting due to clinical deterioration. GEMA-AI showed a particularly pronounced discrimination benefit in women. The study also indicates that the model may improve prioritization for the sickest patients, especially those with extreme analytical values. The authors also report that GEMA-AI gave more weight to serum sodium and relatively less weight to bilirubin and international normalized ratio. The model identified specific high-risk combinations, such as severe sodium abnormalities associated with reduced renal function, particularly in the presence of moderate or severe ascites.
A promising innovation with important limitations
The authors do not present GEMA-AI as ready for universal implementation without further validation. The study has several limitations. First, the model uses a limited number of objective variables to preserve explainability. Other clinical or laboratory parameters might have improved performance, but would have made the model more complex. Second, GEMA-AI was designed to predict 90-day waiting list outcomes. The authors indicate that the score should be reassessed at least every three months, or earlier if the patient’s clinical condition changes significantly. Additional validation would also be required before using GEMA-AI in allocation systems other than those studied. Patients with hepatocellular carcinoma or other MELD exceptions may need specific score adjustments. Finally, the study did not evaluate the impact of GEMA-AI on post-transplant outcomes.
What this study brings to liver transplantation
This publication shows that explainable artificial intelligence models may improve the prediction of liver transplant waiting list outcomes, particularly among the sickest patients and among women. The value of GEMA-AI lies in two main aspects: its ability to account for nonlinear relationships between biological variables and clinical risk, and its explainable design, which may facilitate acceptance in a medical context. The authors conclude that GEMA-AI provided more accurate predictions of waiting list outcomes than currently available models and could help reduce gender disparities in access to liver transplantation. They also consider its implementation feasible because of the model’s interpretability, while emphasizing the need for further validation.
This article is intended for scientific information purposes only and summarizes the findings reported by the authors. GEMA-AI remains a research model and requires further validation before routine clinical use.
Source
This article is based on the scientific publication:
Gender-Equity Model for Liver Allocation Using Artificial Intelligence (GEMA-AI) for Waiting List Liver Transplant Prioritization, published in Clinical Gastroenterology and Hepatology. DOI: 10.1016/j.cgh.2024.12.010.
Full article here.
This article is brought to you by IGL as part of its scientific news sharing.