Episode 47: Anemia Management with AI and Computational Modeling with Luca Neri



About This Episode

Luca Neri, senior director of clinical advanced analytics at Fresenius Medical Care, discusses Anemia Management with AI and computational modeling.

Featured Guest:

Luca Neri, MD, PhD
Senior Director, Clinical Advanced Analytics 
Global Medical Office
Fresenius Medical Care

Luca Neri joined Fresenius Medical Care as a medical data analyst in 2016.  He has experience in epidemiology and outcomes research and has acquired a broad range of analytical and methodological skills in the field of outcomes research and advanced analytics.

Before joining Fresenius Medical Care, he served as scientific consultant and advisory board member for several commercial, science, and industry clients and academic institutions. His research interest focused on the epidemiology and outcomes of chronic kidney disease and other chronic medical conditions. He earned his medical degree at the University of Milan School of Medicine, where he also earned a specialty degree and a PhD in environmental and occupational medicine.


Listen to This Episode


Episode Transcript:

Dr. Maddux: Approximately 2 million people worldwide are living with kidney failure, and a majority are treated with hemodialysis three times per week in a clinic. Unfortunately, most kidney patients also develop anemia at some point during their disease progression. Anemia of chronic kidney disease is often severe, difficult to treat, and there is wide variation in the patient's response to therapy. Dr. Luca Neri, Senior Director of Clinical Advanced Analytics at Fresenius Medical Care, joins us to discuss the impact of anemia management models using artificial intelligence and computational modeling. 

Welcome Luca.

Luca Neri: Thank you. Thank you, Frank.

Dr. Maddux: Luca, tell us a little bit about the treatment of anemia and what are the levers that we have that go into the development of our treatment of anemia for people with chronic kidney disease.

Luca: For people of chronic disease, there are two main treatments, which are iron and erythropoietin stimulating agents. However, finding the right dosage of the two for each patient in each moment of their life is very complex because not only do we have differential patient's response ability to the drugs, depending on their general characteristics, but also this response change over time. And so it's very difficult for the doctor to understand how to manipulate the dosage of the drugs to keep the patient on target and minimize the side effects that come with high doses of the drug. So we need to minimize the dosage while keeping the patient on target.

Dr. Maddux: You've spent a good bit of your career working on digital solutions that begin to address how do we optimize some of these therapies. Tell me about some of that work and the kind of things that we can do that begin to modify the course and the prescribing patterns that patients need.

Luca Neri: We have several, let's say, options. The easiest one is using protocols that help physicians understand how to adjust the treatment based on patient responses, observe the patient response. And this is the original way of managing anemia. But when we look at the outcomes, targeted treatment grades reported in different national registries, UK, US, China and Japan all report that target achievement for patients range between 45% and 65% of the whole dialysis population.

And then one can ask whether we can do better or it is really the maximum that doctors can do for the treatment of renal anemia. And this is why we started working on more advanced digital solution. A few years ago. In Europe, we started working with the Artificial Neural network to understand the response patterns of patients, and we trained an artificial neural network over 900,000 treatment records, and we were able to map the characteristic of the patient in a specific moment in time, including complication, inflammation, or inter-current events like hospitalization and the following outcome.

And based on this predictive model, we built a digital model that helped the physician find the right dosage for each moment, for each patient. And also, helped adjust the dosage of erythropoietin stimulating agent in Diuron to optimize the treatment of patient. The second has been developed in the United States by the Renal Research Institute. And it is a completely different approach to the problem. It is an example of partial differential equations that map the processes that govern the physiology of erythropoietin and the red blood cell physiology in the body. And by that it should predict how the drug would respond out of the patient, would respond to the drug in each moment of the patient's life and clinical course. I think that both may have a great impact on how we treat renal anemia in our patients.

Dr. Maddux: As we look forward, we know that there will be some additional drugs that have impact on the production of red blood cells, the hypoxia inducible factor drugs have impact. There are drugs in phase two trials that are inhibitors of IL6 and other cytokines and they have impact on anemia. How do these models perform and how do you adjust the models when you have new agents that will come either to replace or in addition to some of the existing medications that are needed?

Luca Neri: Well these two models will adapt very differently. To be able to adapt to a new agent, you need first to collect a lot of data that would enable to map the relationship between patient characteristics with a specific time point and the following response to the drug. So we first need the drugs to enter the market and then we observe what happens as the physician use it. 

Dr. Maddux: How do you think the physicians will interact and adopt these particular models? What do you think the drivers are for them? Outcomes? Logistics?

Luca Neri: The study of AI ergonomics is very fascinating. It's a new field. So, there are studies that show how doctors interact with artificial intelligence on computational models and show that after some times of usage, doctors learn how the model behave. In their brain, doctors create a model of the model and somewhat understand how to use the best of the model without falling into the mistakes of the models.

This has been shown very elegantly in a recent study that evaluated how doctors interact with the image recognition software for the diagnosis of colon cancer. That was shown very clearly that the doctors, after a while had developed the capacity, the ability to detect whether the model was wrong and then decided not to follow the suggestion. But the doctor and the model combined, is certainly a better performance than the doctor alone. 

Dr. Maddux: Internally, as you know, we've been talking quite a bit about generative AI and the large language models that have begun to become available in our society and certainly in health care. Do you envision in the future any role for generative AI to participate in these kinds of modeling efforts? 

Luca Neri: I don't know whether large language model exactly, but for sure, the ability to generate cases that are exemplary and are instructive for the doctor to create a training course for using artificial intelligence would be a great advancement because that can increase the efficiency through which we push the uptake of such technologies in clinical practice.

Dr. Maddux: We know that these techniques have gone through various phases of regulatory requirements and cycles. What do you envision the future looking like from the regulatory environment for using AI in clinical decision making, like drug dosing?

Luca Neri: For drug dosing there is worldwide a general agreement that these products should be qualified as medical devices because we need to ensure a high level of safety.  And the regulation is meant to induce manufacturers to follow a disciplined production process that documents safety from different perspective, not only clinical safety, but also cybersecurity and data privacy.

Other models may not be required to be qualified as medical device in some regions and in other regions we may need to follow that path, but I think for drug adjustment models, decision support models that deal with adjusting the dosage of drugs there is general agreement that they should be qualified as a medical device.

Dr. Maddux: Thinking about anemia as a model where you've been able to incorporate these digital methods, where do you think some of the other issues related to kidney disease care offer opportunity for us to build similar models that may influence either the diagnostic or the therapeutic approach to that co-morbidity?

Luca Neri: In our company we are looking at different therapeutic areas within chronic kidney disease. Metabolic disorders, for example, we have been working for a while on a similar model to optimize the dosage of vitamin D or sinical set or phosphate binders. I know that also RRI is working toward this space, but also when we think about the management of the vascular access, we are doing great progresses within our company in developing a suite of digital solution, I would say an ecosystem of digital solution to help doctors manage in a more efficient and accurate way, vascular access, for example.

We are developing an artificial intelligence model based on deep neural network that uses signals collected with a transtado scope to the detect stenosis. We have an AFV failure score that is based on machine data that is very good predictive ability. And we are already piloting 15 clinics around the world.

And also there is one version in the United States. But we are also working on IDH prevention. Both by preventing and changing the dialysis parameters during the dialysis by using a feedback method based on Crit-line data. And this is developed by RRI, but also by predicting the likelihood, the risk of hypotension before the onset of the dialysis session so that the physician can already change the dialysis prescription in advance.

So we believe that by combining these two methods, we can reduce substantially the rate of intradialytic hypotension. There are also other potential areas of development. For example, in prevention of CKD progression, we developed models for predicting the progression of patients towards renal replacement therapy within two years, but also on the short term so that we can suggest to doctors when to start transition management for example. We have a very good model that predicts renal replacement therapy onset within six months that that can be used for that purpose.

Dr. Maddux: One of the things that I have always wondered whether there's a potential application for is, we are constantly worrying about patients on peritoneal dialysis that develop membrane failure and the ability to predict which patients either because of some innate predisposition they may have or because of the kinds of prescriptions they've required of osmotic agents, the ability to predict membrane failure in those patients would be quite valuable in trying to figure out how long a therapy may last for somebody. Any thoughts on that?

Luca Neri: Yes, it would be really, really a very good application. The difficulties that we had while developing a model for PD patients is that we have very few data and we certainly need to improve our ability to capture data momentarily because now we are very good in capturing data during medical encounters. But PD patients have very few medical encounters.

So, we need to capture data while they're at home, and we need to improve in that sense. So we need to explore how to use sensors and wearables to help patients collect the data effortlessly so that we have the information that we need to develop further models.

Dr. Maddux: Luca, anything else around the field of how you're utilizing AI to develop clinically relevant tools. Before we finish.

Luca Neri: There are maybe two additional developments that we are working on. First, we are trying to use a larger language models to help physicians collect information from patients in a more efficient way. In some therapeutic areas, the time doctor spend in administrative work is greater than the time they spend with the patient, and we want to reverse that.

To do this, we need to digitalize the way we conduct our medical work. And so we are collecting data in a small for now, prospective study in Czech Republic to see whether the new large language model can help us doing this exercise of collecting information during a normal and resting conversation between patient and doctor. Collect the data structure in a database and then input in Euclid and then use the same system to create referral notes that can be used by the doctor to refer the patient to other specialists.

And another topic that is quite interesting is the way we use artificial intelligence to advance clinical governance. We have developed an advanced benchmarking tool that is based on 21 models that work together to provide our medical directors a way to compare their clinic’s performance to the performance of other clinics in the network so that we can pick up the best performance and teach other doctors how to do best.

But also we can spot the lowest performance centers where we can improve. And also the model assesses the expected improvement that a center can make based on the case mix. And this is done using,  , some modeling techniques that can project the expected performance of a center given a certain case mix in that center. So, I think that this is a tool that is quite helpful to advance our clinical governance in the network.

Dr. Maddux: Thank you. I've been here with Dr. Luca Neri, who leads many of our projects around not only anemia, but other advanced analytic artificial intelligence projects. Luca, thanks for being here today on Dialogues.

Luca Neri: Thank you for having me.