AUTHORS
Wei Xiao, PhD, Patricia Oneal, MD2, Menglun Wang, PhD3, Qi Liu, MSc, PhD4, Tanya Wroblewski, MD5, Qin Ryan, MD, PhD6
AFFILIATIONS
1 University of Texas MD Anderson Cancer Center, Houston, United States
2 Food and Drug Administration, Silver Spring, Maryland, United States
3 Food and Drug Administration, Silver Spring, Maryland, United States
4 Food and Drug Administration, Silver Spring, Maryland, United States
5 Food and Drug Administration, Silver Spring, Maryland, United States
6 Food and Drug Administration, Silver Spring, Maryland, United States
ABSTRACT
Introduction
Sickle cell disease (SCD) is clinically heterogeneous, and this variability shapes individual patients’ disease manifestations, therapeutic responses, and outcomes. Subphenotype classification of SCD has the potential to guide treatment strategies but remains undefined. Our previous studies identified hemolysis-dominant (HD) and vaso-occlusive-primary (VP) subphenotypes using machine learning-based clustering of baseline data from over 3,000 SCD clinical trial patients; however, their relevance in patients receiving hydroxyurea (HU) has not been well characterized. Here, we define pre-treatment baseline subphenotypes in HU-treated patients and evaluate differences in post-HU laboratory parameters and disease manifestations.
Methods
We analyzed 2,205 patients with SCD treated with HU alone across multiple trials supporting FDA approval. Baseline clinical features (VOC, stroke, leg ulcers, priapism, baseline HU use, transfusion status), laboratory variables (hemoglobin, HbF, indirect bilirubin), demographic variables (age, BMI), and trial era (before or after the HU approval) were binarized and used as inputs for K-modes clustering. Patients with missing data for any baseline feature were excluded from clustering analysis (complete-case analysis). Multiple binarization cutoff combinations were evaluated to ensure robustness, and patients were assigned to HD or VP subphenotypes based on consensus clustering across iterations.
Within-patient changes during treatment were evaluated for laboratory markers (hemoglobin, HbF, LDH, indirect bilirubin, reticulocytes, leukocytes, and neutrophils) and VOC events, using Wilcoxon signed-rank tests. Differences in magnitudes of change between subphenotypes (HD vs. VP) were compared using Mann-Whitney U tests. Binary clinical parameters from follow-up data were screened using chi-square tests, and significant features (FDR<0.05) were used to train an XGBoost classifier to predict subphenotype assignments. Feature contributions were interpreted using SHapley Additive exPlanations (SHAP).
Results
Among 1,046 patients with completed baseline data, the cohort had a median age of 30 years (range: 1–68); 54% were female. Genotypes included SS (82%), S-β-thalassemia (15%), SC (2%), and other variants. Unsupervised clustering yielded three clusters, labeled based on dominant baseline features: 598 patients as VP (higher vaso-occlusive burden), 190 as HD (increased hemolysis), and 258 as indeterminate.
HU increased HbF across the entire cohort, resulting in overall improvements in hemoglobin, leukocyte, and hemolysis markers (reticulocytes, LDH and indirect bilirubin). Subphenotypes comparison revealed hemoglobin improvement from baseline was greater in the HD group (p<0.001), while VOC reduction from baseline was more pronounced in the VP group (p<0.001), suggesting differential response to HU across subphenotypes.
Outcome-based modeling demonstrated robust discrimination between HD and VP subphenotypes. Among 1,383 SCD clinical variables during HU treatment, SHAP analysis identified key distinguishing features: skin ulcers, priapism, and acute chest syndrome were more strongly associated with the HD subphenotype, consistent with hemolysis-driven endothelial dysfunction and vasculopathy. In contrast, VOCs were associated with the VP subphenotype, reflecting microvascular obstruction and inflammatory processes. Infection-related complications, also ranked top by SHAP analysis, suggesting differences in systemic inflammatory burden between subphenotypes.
Conclusions
Using large clinical trial datasets, we show that SCD subphenotypes derived from unsupervised learning reflect underlying biological heterogeneity with differing outcome patterns under HU therapy. By integrating downstream disease manifestations with machine learning, we provide independent support for these subphenotypes and identify patterns consistent with hemolysis-driven vasculopathy and vaso-occlusive mechanisms. These findings suggest that subphenotype-based stratification may serve as a practical framework for informing more individualized treatment and management strategies in SCD.
DISCLOSURES
Wei Xiao, PhD: No relevant conflicts of interest to declare.