12-05-2025
AI in Liver Care Needs Vigilance and Tailoring to Population
AMSTERDAM — As artificial intelligence (AI) becomes increasingly embedded within healthcare, including liver care, it will be essential to tailor AI to the local population and ensure regular model monitoring to ensure both effective and safe outcomes for patients.
Ashley Spann, MD, is a transplant hepatologist interested in developing informatics and AI to optimize outcomes in liver disease, including transplantation, at Vanderbilt University Medical Center, Nashville, Tennessee. At European Association for the Study of the Liver (EASL) Congress 2025, she shared advice on ethics and how to implement AI into the liver clinic in a session on the impact of AI on the hepatology practice.
Ashley Spann, MD
'We need to include patients and providers from the very beginning, not build in silos. The data must be representative of the population of concern, the technical solution must fit the clinical problem, and the model must not cause harm,' Spann told Medscape Medical News in an interview after her talk.
She stressed that principles of clinical care, particularly non-maleficence, also have a place in the use of AI. 'AI is already around us. The question is: Should we use it? And if so, how do we do it responsibly?'
To this end, Spann discussed best practices for model development, clinical implementation, and a key safeguard she termed algorithmovigilance : The ongoing monitoring of AI models after deployment to detect performance drift and prevent harm. 'We can minimize harm by setting parameters for the model so we know when the performance starts to lag in real time and patients might be affected. If this happens, we turn the model off, reassess, retrain, and redeploy.'
'Each step of the way, from inception to deployment, we must track what the model is doing and ensure it isn't making care worse for patients,' she said.
Buy or Build — Key Questions for Adoption
Spann stressed the importance of starting with the clinical problem and then layering in appropriate AI technology while always considering the protection of patients.
Whether building or buying a model, ensuring that it reflects the population of concern is paramount. Most AI models are trained on historical healthcare data. This means that it could reflect systemic inequalities, such as risk factor prevalences among a specific population, underdiagnosis, undertreatment, or lack of access to care among marginalized populations. In this case, the model learns and replicates those patterns.
'We must make sure biases and disparities don't worsen,' she said. 'If a model begins to underperform, we need to know when and how to intervene.'
Spann urged clinicians and institutions to interrogate the data available when deciding which model is appropriate. For example, when building a model, she suggested asking whether some patient groups in your dataset are more affected than others. If buying a model, she suggested asking whether the model addressed the clinical problem in need of solving.
For instance, AI might be a solution for identifying people with undetected cirrhosis within a population-level approach to the problem. 'It's crucial to ask what data are available that could be useful to make that prediction and are there patients who are disproportionately affected? There might be certain patients without available data even though they may have a disease, and what are the implications of that?'
She cited an example from her institution, where the Fibrosis-4 (FIB-4) Index was integrated into the electronic health records to automate liver fibrosis risk stratification. More than half the patients lacked the key lab values needed to generate a FIB-4 score. 'They might still have disease, but without those labs, we can't know the risk severity. The question becomes, what are the data that we need and how do we get it? That's a data gap with real implications,' Spann explained.
Mismatched Populations Can Render a Model Useless
When buying an AI model, Spann cautioned against applying them to populations too different from those on which it was trained. She cited a model that was developed using data from the US Veterans Affairs system, which mostly contains patients who are older White men and, as such, may not generalize well to urban centers serving more diverse populations. 'That population is a very unique subset of patients. The only way to determine the suitability or not is to take that model and test it retrospectively and look at how the model might change and then locally track performance over time.'
She also underscored how sociodemographic and economic factors such as proximity to transplant centers or to a liver clinic can skew outcomes, and these are most likely not accounted for in a model's clinical inputs. 'We need to consider how well a model performs in those subgroups because it may be erroneous in them.'
AI's Role in Population Health
Session co-moderator, Tom Luedde, MD, director at Heinrich Heine University Düsseldorf in Düsseldorf, Germany, considered the impact promised by AI for liver care. 'Prevention, detection, risk prediction, and actually getting patients into the healthcare system are our biggest deficits in liver disease. AI could help fill those gaps,' he said. 'Right now, general practices are not implementing FIB-4 in daily practice, for example, but we might get there with an LLM [large language model] or an AI system that provides patients with access to the hepatology system. I believe, with these approaches, we will have a greater impact than with any single drug or complex intervention. In the future, I can envisage AI being implemented in a type of health kiosk. And with all the resource issues we have, this might help.'
