Andrew Srisuwananukorn, MD, assistant professor at The Ohio State University Comprehensive Cancer Center, discussed the use of a novel artificial intelligence model that aids in the differentiating between prefibrotic primary myelofibrosis and essential thrombocythemia.
An artificial intelligence (AI) model is being investigated that will help clinicians distinguish between prefibrotic primary myelofibrosis (pre-PMF) and essential thrombocythemia (ET) using bone marrow biopsy images from 200 patients.
The AI tool had previously been trained with 32,000 pan-cancer biopsy images. It was also familiar with general pathologic features. From this, investigators tested if the AI could differentiate between the 2 types of myeloproliferative neoplasms (MPNs) in patients.
According to findings presented by Andrew Srisuwananukorn, MD, at the American Society of Hematology 2023 Meeting, the tool demonstrated a 92.3% rate of agreement with human experts, and the sensitivity and specificity for pre-PMF diagnosis was 66.6% and 100%, respectively.
Based on these promising findings, experts will continue to update the AI tool and test it in larger data sets.
“I view this type of tool as a companion diagnostic tool, but I do not believe [that] AI tools can replace the judgment of a human physician. I think it is really up to us as pathologists and clinicians to say when an AI algorithm tool is not working appropriately. What I hope is that this can be used for better information for the patient to understand their disease,” said Srisuwananukorn, assistant professor at The Ohio State University Comprehensive Cancer Center, in an interview with Targeted OncologyTM.
In the interview, Srisuwananukorn discussed the use of a novel AI model that aids in the differentiating between prefibrotic primary myelofibrosis and essential thrombocythemia.
Targeted Oncology: What can you tell us about this AI tool?
Srisuwananukorn: Our research is in developing an artificial intelligence tool to differentiate between rare myeloid malignancies, including prefibrotic myelofibrosis and essential thrombocythemia. As a brief overview, these diagnoses are challenging to differentiate because there’s similar criteria, including clinical and laboratory abnormalities, mutational profiling, and assessment of the bone marrow, which can be very subjective, particularly when looking at the megakaryocyte morphologies in the fibrosis. Our hope is to create a more objective or at least consistent tool using artificial intelligence to differentiate between the 2.
What is the motivation behind the development of this tool? What specific challenges or clinical needs there are in distinguishing between different disease types?
Our motivation is that this is a diagnostic dilemma for our patients. Essential thrombocythemia and prefibrotic primary myelofibrosis behave quite differently. The prefibrotic myelofibrosis cases have more symptoms and are at higher risk for progression to acute myeloid leukemia. To me, it behooves the physician to know exactly which disease they have. It’ll help guide their therapies in the future. Our hope is that a tool such as this can help guide that management and potentially to help enroll in clinical trials for more appropriate diagnosis and therapy creations.
In your study, how did AI demonstrate its efficacy when differentiating between the myelofibrosis and ET?
Our model had very high performance with [an] area under the receiver operator curve of 0.9, a sensitivity of 66.6% specificity of 100%, and an accuracy of 92.3% in diagnosing prefibrotic myelofibrosis. In addition, we did a qualitative analysis to try to understand what is being used in these AI algorithms to make those predictions. With this qualitative interpretation of quote unquote, opening the black box of our AI algorithms, we were able to see that preferentially, areas of bone marrow cellularity were chosen for the prediction of 1 vs the other disease. Reassuringly, the algorithm was not using nonsensical portions of the image, such as fat or cortical bone or even background artifacts. We believe this AI algorithm is using biological reasons.
What distinguishes this tool from others and how can it be interpreted in these settings moving forward?
I view this type of tool as a companion diagnostic tool, but I do not believe [that] AI tools can replace the judgment of a human physician. I think it is really up to us as pathologists and clinicians to say when an AI algorithm tool is not working appropriately. What I hope is that this can be used for better information for the patient to understand their disease.
Are there any particular patient subgroups that the AI tools showed notable effectiveness?
For right now, we’re only looking at 2 particular diseases: prefibrotic myelofibrosis and essential thrombocythemia. However, that does not mean that this is our only goal. These algorithms are agnostic of disease and outcome. Potentially, the same types of algorithms or workflows can be used for any type of disease that other clinicians might be able to implement in their clinical practice. Our particular goal is to do better for patients [with MPNs], and we have multiple ideas of how we can do that.
Moving forward, what are some of the next steps for this research?
Our next steps are in 2 domains. One, for this particular algorithm in differentiating pre-PMF and ET, we hope to validate it in further, larger retrospective cohorts at other academic centers and potentially within clinical trials that enroll patients [with ET]. Our thought is potentially, the ET trials did not have great performance because they were accidentally enrolling these pre-PMF cases. That’s our next step into rigorously seeing if this algorithm can be effective. Outside of that, we’re hoping that we can develop similar algorithms in other outcomes of interest, including risk stratification and therapy response.
What should oncologists know about the growing use of AI?
For physicians at large, what I want to reiterate is that it is important that we understand how it’s being used and when to use it. I do not believe that algorithms can supplant physician judgment, but I do think that they will be used in clinical practice and it’s up for us to know when it’s not working. I hope that it’s a support tool and will help guide your decision management but it’s always up to the physician.