Hierarchical clustering, a common 'unsupervised' machine-learning algorithm, is advantageous for exploring potential underlying aetiology in particularly heterogeneous diseases. We investigated potential embolic sources in embolic stroke of undetermined source (ESUS) using a data-driven machine-learning method, and explored variation in stroke recurrence between clusters.
We used a hierarchical k-means clustering algorithm on patients' baseline data, which assigned each individual into a unique clustering group, using a minimum-variance method to calculate the similarity between ESUS patients based on all baseline features. Potential embolic sources were categorised into atrial cardiopathy, atrial fibrillation, arterial disease, left ventricular disease, cardiac valvulopathy, patent foramen ovale (PFO) and cancer.
Among 800 consecutive ESUS patients (43.3% women, median age 67 years), the optimal number of clusters was four. Left ventricular disease was most prevalent in cluster 1 (present in all patients) and perfectly associated with cluster 1. PFO was most prevalent in cluster 2 (38.9% of patients) and associated significantly with increased likelihood of cluster 2 [adjusted odds ratio: 2.69, 95% confidence interval (CI): 1.64-4.41]. Arterial disease was most prevalent in cluster 3 (57.7%) and associated with increased likelihood of cluster 3 (adjusted odds ratio: 2.21, 95% CI: 1.43-3.13). Atrial cardiopathy was most prevalent in cluster 4 (100%) and perfectly associated with cluster 4. Cluster 3 was the largest cluster involving 53.7% of patients. Atrial fibrillation was not significantly associated with any cluster.
This data-driven machine-learning analysis identified four clusters of ESUS that were strongly associated with arterial disease, atrial cardiopathy, PFO and left ventricular disease, respectively. More than half of the patients were assigned to the cluster associated with arterial disease.