SAN DIEGO – An artificial intelligence tool based on self-supervised learning appears superior to the standard-of-care scoring methods for predicting myelodysplastic syndrome (MDS) outcomes in a retrospective study.
Elisabetta Sauta, a data scientist at Humanitas Research Hospital in Rozzano, Italy, presented results from a study exploring an AI-driven framework for personalized prognostic assessment in MDS at the American Society of Hematology's annual meeting in San Diego on Saturday. In the study, she and her collaborators used a deep learning model combining two layers of processing: a customized version of Prov-GigaPath that predicts patients' outcomes from whole slide images and a type of neural network that can integrate omics and clinical data into those predictions.
MDS, a disorder of clonal hematopoietic stem cells associated with older age, progresses to acute myeloid leukemia in about a quarter to a third of cases. Patients present with a variety of clinical symptoms and outcomes, even when they have the same MDS subtype. The cause is known in about 15 percent of cases and can be due to heritable variants in genes such as DDX41, GATA2, RUNX1, ANKRD26, and ETV6 and in genes within the telomerase complex.
MDS is diagnosed via blood counts, cytomorphology of blood and bone marrow, and cytogenetics of bone marrow cells. In some cases, bone marrow biopsy and cytogenetic analysis by next-generation sequencing can also be conducted.
Treatment for patients with MDS is heavily dependent on their risk of progression. Higher-risk patients whose MDS could evolve to AML receive aggressive treatment such as allogeneic stem cell therapy, hypomethylating agents, or chemotherapy. However, the aim for lower-risk patients is on managing their symptoms such as anemia and cytopenia and optimizing their quality of life.
The International Prognostic Scoring System-Revised (IPSS-R) and the newer IPSS-Molecular (IPSS-M), which integrates somatic gene mutations into prognostic determinations, are standard tools for classifying MDS patients into higher- or lower-risk categories. Although incorporating genomic screening results into patients' risk assessments have improved the IPSS as a tool for predicting clinical outcomes compared to IPSS-R and IPSS-M, Sauta and her colleagues hypothesized that adding transcriptomic and immunomic information could improve it even further.
At the 2023 ASH meeting, Sauta's team presented data showing that a computationally generated risk score incorporating transcriptomic data in addition to genomic, clinical, and cytogenic information had a C-index of 0.83. A C-index with a value of 1 indicates that the model perfectly predicted the survival of patients in the group and 0.5 means the model is no better at predicting the outcome than random chance. In comparison, the IPSS-R score and the IPSS-M had C-index values of 0.68 and 0.76, respectively.
Building on that work in the current study, Sauta and her colleagues incorporated transcriptomic data from bulk RNA-seq of CD34-positive bone marrow cells and deep flow cytometry analysis of T lymphocytes as well as of natural killer and myeloid cells into a new model, dubbed MEGAERA, on the theory that the added data would enhance the accuracy of the risk scores generated by the algorithm.
In the study, the researchers retrospectively determined risk scores for 605 patients from Humanitas Research Hospital using MEGAERA with the added transcriptomic and immunomic data and factoring in clinical characteristics, cytogenetics, somatic mutation screening on 31 genes, and whole slide images of bone marrow biopsies prepared with hematoxylin and eosin staining or a staining technique called May-Grunwald Giemsa. MEGAERA predicted overall survival in the patient group with a C-index of 0.85 compared to 0.68 for IPSS-R and 0.76 for IPSS-M. MEGAERA also predicted leukemia-free survival with a C-index of 0.83 and an individual probability of response to a hypomethylating agent with a C-index of 0.84.
Sauta and her team also applied the SHapley Additive exPlanations (SHAP) model to enhance the interpretability of the MEGAERA to identify a set of features most associated with treatment failure. Those features included variations in ASXL1, TP53, RUNX1, and TET2 as well as immune features, cellular response to stress genes, platelets, and loss of chromosome 3.
Because transcriptomic and immunomic data are not routinely collected when patients are diagnosed with MDS, Sauta and her team explored how they could reduce the complexity of the model as a way to ease MEGAERA's integration into clinical practice. Using only data that is routinely collected in clinical practice such as cytogenetic and genomic information, they analyzed an independent subset of 31 MDS patients and were able to predict the risk class of the patients with 94 percent accuracy.
Sauta clarified that while all of the data modalities, including transcriptomic and immunomic data, were necessary to define the risk classes within the model, with the complexity reduction analysis, the model was able to "approximately" predict the correct risk class for each patient but without the granularity available from the full analysis.
Based on the overall analysis, "the MEGAERA model … effectively enhanced clinical outcome predictions in MDS patients, including the probability of response to hypomethylating agents," Sauta concluded at the meeting. "A framework that is completely explainable represents a novelty because we can extract multimodal insights to improve decision-making, and we have provided preliminary evidence that we can predict the multilayered landscape of MDS starting from information available for routine workup of the patient."
Christopher Park, a professor of pathology at New York University Grossman School of Medicine, said the study's findings are "pretty exciting" since the researchers were able to beat the current best prognostic scoring system. "Anything that improves our ability to prognose patients' outcomes is always going to be good," he said.
However, he added that ensuring the model can be integrated into clinical practice using routinely collected data will be important. "If we did literally what she just said [was necessary for the model], it would cost us four times more to work up a patient," Park said, referring to the MEGAERA inputs before the complexity reduction. "In this day and age, the way things are reimbursed, I don't think anybody here thinks we can do [transcriptomic and immunomic analysis on every patient] for real."
Park also expressed some skepticism about the model's predictions and explainability. He noted that unsupervised models have limitations in how accurately they generalize to real-world patient populations and that models incorporating both supervised and unsupervised features tend to perform better.
He also challenged Sauta's characterization of MAGAERA as a completely explainable framework. Sauta and colleagues' ability to identify a set of features that most contributed to the model's predictions does not meet the definition of explainability, in Park's view, because it does not provide any information about the underlying biological mechanisms contributing to the predicted outcomes. "You train [the model] to match the outcome, but whether or not that reflects true biological differences, we don't know," he said.
In the future, Sauta and her collaborators hope to enlarge the independent validation cohort for MAGAERA and improve their understanding of the features that are more or less important in its predictions.