AI Tool May Better Predict Transplant Risk in Myelofibrosis

An AI model may help identify patients with myelofibrosis at high risk of early death after transplant, according to an expert: © stock.adobe.com.

A machine learning model may help improve transplant risk assessment for patients with myelofibrosis by more accurately identifying those at high-risk of early death, according to an expert.

Dr. Adrián Mosquera, a hematologist at the University Hospital of Santiago de Compostela in Spain, shared valuable insights during an interview with CURE in which he highlighted the development of the machine learning model. According to Mosquera, this model is designed to enhance transplant risk assessment for patients with myelofibrosis.

CURE: How does the machine learning model outperform the current models and predict and transplant outcomes?

Mosquera: There were some previous efforts. The conventional statistical method used in medicine for a long time was more or less like performing traditional Cox regression, and based on that strategy, patients were classified as high risk, low risk, and intermediate risk, or similar categories.

The issue is that despite this being a valid approach that could somewhat stratify patients, we still had too many patients in the intermediate risk group who behaved as high risk. So, if 30% of patients die early due to toxicity, and you can only identify 8% of them, you are clearly underperforming.

The approach then was to use artificial intelligence. Let's generate not a rule-of-thumb-based score, but a quantitative score, and then try to find the optimal point where we can cut and clearly identify these high-risk patients. To our surprise, that threshold was somewhere between the 25% and 30% higher-risk patients, and these patients did actually have very poor survival. So, you have to be very careful when considering an allogeneic stem cell transplantation in this patient profile, because the likelihood of the one-year death rate is 40%, and the two-year death rate goes over 50%, and that's significant.

What inspired the development of the machine learning model for transplant risk assessment in patients with myelofibrosis?

We've been working on several artificial intelligence modeling projects in hematology for a long time, and I always encounter the same kind of question. This is interesting because when you go to the bank to get credit or something you need to pay off over time, you pay a lot of attention to all the payments you need to make and the interest rates. You want to know a lot about how you will face any difficulties with payment you may have.

However, in medicine, we are not as accustomed to using data-driven approaches to balance decision-making in something far more important than the economy — human health. Our commitment here is bringing advancements in big data and artificial intelligence to address important unmet medical needs, which can sometimes be improved with the implementation of data-driven approaches. This is particularly focused on areas where we have to make difficult decisions that can lead either to disease duration, prolonged disease remission, or, in the worst case, toxic death.

How do key patient characteristics, such as age and comorbidities, factor into the model's predictions?

Well, these characteristics actually perform well. We initially analyzed the whole dataset, which included more than 5,000 patients with long follow-up. This is a significant advantage of the EBMT registry, which is probably the best in the world for conducting this kind of research. We had over 52 potentially prognostic variables available for analysis. We initially fed all those variables into the model, and we obtained an initial model that performed quite well.

However, the problem is that in clinical routine, at least at this stage of technological development, you cannot input 52 variables into a calculator. You have to simplify this somehow to make it easy for clinicians to use and adopt this technology in their clinical routine. So, we then set out to identify the variables that contained most of the independent prognostic value, and we derived a 10-variable base model which contained the same prognostic accuracy but with fewer variables—specifically, 42 fewer. This model is actually pretty much aligned with what we know about prognostication in allogeneic stem cell transplantation.

So, there are a few variables related to demographics, particularly age. Then we have a few variables related to comorbidity and prior treatment. Following that, we have a few variables related to myelofibrosis stage—how aggressive is the myelofibrosis at the moment allogeneic stem cell transplantation is prescribed? And finally, we have a few variables related to transplant type, donor type, and conditioning regimen for allogeneic stem cell transplantation. This makes a lot of sense because it takes into consideration variables that we know are involved in different layers of complexity that our patients have.

What impact could this tool have on the decision-making process for clinicians and patients regarding allergenic hematopoietic cell transplant?

I think it has a big impact, and this is not only my belief but that of our collaborators as well. For myelofibrosis, allogeneic stem cell transplantation is the only potentially curative therapy we have. However, thanks to advancements in pharmacology, there is a growing number of drugs being developed for this disease. We know that although allogeneic stem cell transplantation might be curative, many patients do not fare well.

It's particularly regrettable when you recommend an allogeneic stem cell transplantation and the patient starts to experience significant complications, like graft-versus-host disease (GVHD). This can lead not only to early mortality but also to a poor quality of life. So, it doesn't mean that you cannot perform allogeneic stem cell transplantations in these patients, but they have the opportunity to make a better-informed decision, to discuss it more thoroughly with their healthcare team and the patient, and perhaps even to develop potential risk-mitigating strategies.

So, it will have an impact, and we expect this impact will be in the best interest of all patients. Furthermore, there are several similar projects being planned now that actually follow the same guidelines, but for other conditions like myelodysplastic syndromes (MDS) and chronic myelomonocytic leukemia (CMML).

What are potential next steps in refining the model and incorporating additional disease factors into future versions?

I would argue that the next step is to apply the same technology to other existing data, even if it comes from different disease subgroups, like MDS and CMML, as I mentioned previously. Our approach aims to provide efficiency to the health system, offering a fast way to proceed. However, I agree with you. When you analyze the results of our paper, although the strategy is sound and we have attractive results, we still see that a significant portion of the predictability of the mortality variable is not explained by baseline variables, at least the clinical ones.

So, we need to take a step forward and incorporate other technologies. We are discussing including molecular testing of tumors, as well as imaging from bone marrow, and definitely also some generative AI tools that we are developing, which aim to better classify or estimate frailty in patients.

This is a very important aspect because we tend to overlook it. Frailty estimation is often very subjective, and there is a significant need to objectify it. While there are scores being used, they are not always well-applied, and there is a need to facilitate the implementation and reproducibility of these scores. This is an area where large language models are being developed, at least within our team, with that specific purpose.

How do you envision this open access tool being integrated into daily clinical practice,

It seems it will be integrated. You know, we've been using risk models throughout our professional lives, and they are often suboptimal. In my experience, you only tend to use these risk models when you actually have actionable information that can guide you to make a better decision for a patient.

In this case, I think it's going to be used quite frequently because in myelofibrosis, we are never entirely sure about how much the disease risk is balanced with the toxicity of the transplant. We often make decisions based on our experience and some indirect data from the literature.

However, this data-driven approach will probably enhance or even reinforce the decisions that some of our colleagues might have already considered. So, it is valuable, and I think it's going to be used quite frequently. Furthermore, it's a kind of breakthrough in the way it is actually moving or changing paradigms in data management for a large international society like the EBMT, shifting towards AI-driven agreement approaches that are very much needed in the transplantation process.

For more news on cancer updates, research and education, don’t forget to subscribe to CURE®’s newsletters here.