Ischemic stroke recurrence remains a major clinical challenge driven by heterogeneous and interacting risk factors. In this study, we analyzed a prospective cohort of ischemic stroke patients from the Mind-the-Heart study, integrating clinical characteristics and left atrial (LA) structural information to identify patient phenotypes using an unsupervised multiple kernel learning (MKL) framework. Clustering in the MKL space (first two dimensions) with K-means++ identified three distinct phenogroups with differing clinical profiles and outcomes, including early (0-90 days) and late (90 days-2 years) stroke recurrence, mortality, and functional status. Phenogroups 1 and 2 exhibited higher stroke recurrence rates compared with Phenogroup 0, with distinct temporal patterns. Phenogroup 2 was characterized by older age, higher prevalence of atrial fibrillation, and cardioembolic stroke etiology, and showed the highest risk of early stroke recurrence, mortality, and worse functional outcomes. This group presented increased prevalence of left atrial appendage (LAA) slow flow, supporting a cardioembolic risk profile. In contrast, Phenogroup 1 exhibited the highest rates of late stroke recurrence and was characterized by a vascular vulnerability, including high prevalence of coronary, aortic, and carotid plaques (with mild stenosis), along with cardiovascular comorbidities and smoking history. In multivariable logistic regression, Phenogroup 1 was independently associated with an increased risk of late ischemic stroke recurrence after adjustment for age and antithrombotic therapy (adjusted odds ratio 7.97; 95% CI 1.72–77.7; p = 0.006). Phenogroup 0 showed a comparatively lower risk profile across outcomes. The presence of an atrial diverticulum was more prevalent in Phenogroups 0 and 1, suggesting a potential role in longer-term recurrence, although its clinical relevance requires further investigation. Overall, this phenogroup-based approach captures the multidimensional heterogeneity of stroke and reveals distinct temporal recurrence patterns without relying on outcome-driven models. These findings highlight the value of multi-domain patient characterization for personalized secondary prevention strategies.