Dr Orgueira on Refining the Use of Machine Learning to Predict Survival Outcomes After Transplant in Myelofibrosis

"It is still surprising that a large part of [patient] mortality is still not explained by these machine learning models, even with such big cohorts and multinational data. This means that we need to explore other areas of improvement in the future by taking probably other layers of complexity into the model, like imaging or molecular biology."

Adrián Mosquera Orgueira, MD, PhD, a specialist in Hematology and Hemotherapy and lead researcher of the Computational and Genomic Hematology Group at the Health Research Institute of Santiago de Compostela, discussed the limitations of machine learning–based prognostic models for overall survival into clinical practice for patients with myelofibrosis undergoing allogeneic hematopoietic stem cell transplant (allo-HSCT). He also talked through avenues for future improvement of one such model generated by the EBMT.

The model, which used a random survival forest (RSF) approach, demonstrated modest improvements in predictive accuracy over traditional statistical models,such as Cox regression, Orgueira began. One key strength of the RSF model was its ability to avoid overfitting through cross-validation, allowing for more robust generalizability across diverse patient cohorts, he added. The study drew on data from a large multinational population of patients with myelofibrosis who received allo-HSCT, Orgueira noted.

Despite these advantages, a substantial proportion of post-transplant mortality remains unexplained by the model, likely reflecting unmeasured variables or biological complexities not captured by existing inputs, Orgueira acknowledged. To improve future predictive accuracy, integration of additional data types—including imaging biomarkers, molecular profiling, and dynamic clinical parameters—should be considered, he suggested.

The study also explored whether the model could inform decision-making regarding transplant conditioning regimens or graft-vs-host disease (GVHD) prophylaxis but found no consistent patterns to guide clinical strategy, Orgueira stated. He emphasized that although the tool may help identify patients at highest risk of poor post-transplant outcomes, it should not be used to determine specific therapeutic interventions, such as conditioning intensity or GVHD management.

Rather, the model’s utility lies in facilitating early discussions among hematologists and patients about risks and goals of care before transplantation, Orgueira explained. The model is best applied as a risk stratification tool rather than a decision-making algorithm, he concluded.