Phenotypic clustering of allogeneic hematopoietic stem cell transplant recipients associated with cardiovascular events
T Goncalves, G Roncin, B Sibilia, M Robin, F Chevillon, A Xhaard, N Dhedin, M Jestin, R Peffault de Latour, F Sicre de Fontbrune, A Unger, J G Dillinger, P Henry, S Toupin, T PezelAbstract
Background
Allogeneic hematopoietic stem cell transplantation (alloHSCT) recipients face substantial cardiovascular morbidity and mortality. Traditional risk stratification tools may not adequately capture the heterogeneity of cardiovascular risk in this population. Phenotypic clustering using unsupervised machine learning (ML) approaches may identify distinct patient subgroups with differential cardiovascular risk profiles.
Purpose
To identify distinct phenotypic clusters among alloHSCT recipients and assess their association with cardiovascular events (CE).
Methods
We conducted a retrospective cohort study of 1,381 consecutive patients who underwent alloHSCT at Saint-Louis Hospital (Paris, France) between July 2011 and July 2023. We applied a Latent Class Model, a probabilistic unsupervised ML approach, to baseline demographic, clinical, biological, echocardiographic, and transplant-related variables. Final clustering was selected based on statistical robustness and clinical relevance. A decision tree classifier was developed to predict cluster membership. The primary outcome was a composite of CE including cardiovascular death, heart failure, atrial fibrillation/flutter, ventricular tachycardia, and conductive disorders resulting in pacemaker implantation, and acute arterial events (acute coronary syndrome, ischaemic stroke). Cox proportional hazards models were used to assess the prognostic significance of each phenotypic cluster to predict the primary outcome.
Results
Four distinct phenotypic clusters were identified: Cluster 1 (28%) comprised young patients with non-malignant hematologic diseases; Cluster 2 (38%) included young patients with malignancies and high cardiotoxic therapy exposure; Cluster 3 (35%) consisted of older patients with myeloproliferative/myelodysplastic syndromes; and Cluster 4 (6%) represented patients with the highest cardiovascular risk burden. During a median follow-up of 3 (1-6) years, 247 patients (18%) experienced CE. Kaplan-Meier survival curves demonstrated significant differences in CE-free survival across clusters (log-rank test p<0.001) (Figure 1). Compared to Cluster 1, Cluster 3 (adjusted HR: 1.80, 95% CI: 1.07-3.03, p=0.026) and Cluster 4 (adjusted HR: 2.00, 95% CI: 1.08-3.71, p=0.027) were independently associated with higher risk of CE. A decision tree classifier achieved 87% accuracy using five readily available variables (Figure 2). Phenotypic clustering significantly improved cardiovascular risk prediction beyond the hematopoietic stem cell transplantation-comorbidity index (C-statistic improvement: 0.08, integrative discrimination index: 0.050, net reclassification index: 0.640, all p<0.001).
Conclusion
ML-derived phenotypic clustering identifies distinct patient subgroups with differential cardiovascular risk among alloHSCT recipients. This approach may enhance risk stratification and guide personalized cardiovascular surveillance and prevention strategies.Figure 1.Kaplan Meier survival curvesInterpretable decision tree