Risk modeling has become a crucial component in the effective delivery of health care, as it allows health systems to better understand which patients are at risk of adverse events. A key challenge in building effective risk models is accounting for patient heterogeneity among the large and diverse populations often present in health systems. Incorporating such heterogeneity into risk models is crucial for aiding the development of tailored care strategies, as it can provide more descriptive information about patients with different comorbidity profiles and can result in more accurate risk prediction. We propose a flexible and interpretable risk modeling approach based on sufficient dimension reduction that simultaneously accounts for patient heterogeneity, results in models which provide information specific to patients with different comorbidity profiles, borrows strength in estimation across related subpopulations of patients, and can act as a useful visualization tool. We demonstrate that our approach improves estimation performance in the presence of heterogeneity in both simulated examples and in a risk prediction study of hospital admission for a large health system.