Supervised Machine Learning Predicts Transformation Risk and Key Features throughout the Patient Journey in Myelodysplastic Neoplasms

Myelodysplastic neoplasms (MDS) are hematologic malignancies defined by cytopenia, dysplasia, cytogenetic aberrations, and molecular mutations. MDS progression to acute myeloid leukemia (AML) significantly reduces patient survival, yet predicting which MDS patients will undergo transformation remains challenging. Current prognostic models, such as the Revised International Prognostic Scoring System (IPSS-R) and the IPSS-Molecular (IPSS-M), have improved risk stratification by incorporating cytogenetic and molecular data using conventional statistical methods, yet they do not provide a dedicated transformation prediction at different time points. Advanced artificial intelligence-based methods, such as the AIPSS-MDS, have recently shown superior prognostic performance compared to conventional prognostic models while focusing on clinical and cytogenetic data. We developed a machine learning (ML) pipeline to enhance the prediction of MDS transformation to AML. Prognostic features are selected autonomously by ML models and explainability is provided using Shapley additive explanation values (SHAP) that quantify the impact on model decision-making, aiming to provide a time-dependent analysis of feature importance.

Our pipeline was trained and tested on the original IPSS-M cohort, including clinical, cytogenetic, and molecular data from treatment-naïve patients diagnosed with MDS according to the WHO 2016 criteria. We excluded patients lost to follow-up or whose Leukemia-Free Survival (LFS) was shorter than the time to AML transformation from the original cohort, leaving 2,699 patients at the start of the study and 1,034 at 48 months. Median age was 72 years and 39.6% of patients were female. 20.3% of patients had MDS with multilineage dysplasia (MDS-MLD), followed by 15.4% of patients with excess of blasts-1 (EB-1), and 15.1% with EB-2 while 4.7% of patients had deletion 5q (del5q). The median number of molecular alterations per patient was four. TP53 mutations were classified as single hit (variant allele frequency [VAF] < 50%) or multi-hit (VAF > 50%, two mutations, or loss of heterozygosity). According to IPSS-R, 51% of the patients belonged to very low and low cytogenetic risk groups. Supervised ML models, specifically Extreme Gradient Boosting (XGB) and Random Forest (RF), were trained to predict transformation to AML at 6, 12, 24, 36, and 48 months after diagnosis using an 80:20 train-test-split. Both models were trained and tested over 1000 runs to improve robustness. The proportion of AML transformation was assessed at each time point, showing a steady increase from 6.2% at 6 months to 47.9% of uncensored patients at 48 months. Hyperparameter tuning was performed separately for each time point and model.

Prediction of transformation became more accurate as time from initial diagnosis progressed, showing an area-under-the-receiver-operating-characteristic (AUROC) of 0.84 and 0.81 for XGB and RF at 6 months, respectively, and reaching 0.90 for both at 48 months. Top features impacting model decisions across all points in time included bone marrow blast count and cytogenetic risk group per IPSS-R. Interestingly, hemoglobin levels were particularly important for the XGB model, ranking 4^th at 6 and 2^nd at 24 months. Consistent with these results, absolute neutrophil count (ANC) and platelet count had a high impact in the early stages, with ANC ranking 3^rd for XGB and platelets ranking 3^rd for RF at the 6-month mark, indicating that altered hematopoiesis is a robust predictor of early progression. Alterations of RUNX1 demonstrated high predictive values at 24, 36, and 48 months, ranking among the top predictors in both models. Alterations of ASXL1 had a substantial impact on the RF model at 48 months. Conversely, TP53-multihit was particularly impactful from 6 to 24 months post-diagnosis but dropped in model impact beyond that time point. Additionally, mutation count had an increasing impact on disease transformation over time. Differences between models were noted, with RF focusing on molecular data while XGB emphasized clinical information.

Patient features, both genetic and clinical, displayed divergent effects on MDS progression to AML over time, alluding to biological heterogeneity. Hence, ML models may enable personalized monitoring strategies to predict disease progression throughout the patient journey.