Background: The subtyping of breast cancer based on features of tumour biology such as hormonal receptor and HER2 status has led to increasingly patient-specific treatment and thus improved outcomes. However, such subgroups may not be sufficiently informed to best predict outcome and/or treatment response. The incorporation of multi-modal data may identify unexpected and actionable subgroups to enhance disease understanding and improve outcomes.
Methods: This retrospective cross-sectional study used the cancer registry Surveillance, Epidemiology and End Results (SEER), which represents 28% of the U.S. population. We included adult female patients diagnosed with breast cancer in 2010. Latent class analysis (LCA), a data-driven technique, was used to identify clinically homogeneous subgroups ("endophenotypes") of breast cancer from receptor status (hormonal receptor and HER2), clinical, and demographic data and each subgroup was explored using Bayesian networks.
Results: Included were 44,346 patients, 1257 (3%) of whom had distant organ metastases at diagnosis. Four endophenotypes were identified with LCA: 1) "Favourable biology" had entirely local disease with favourable biology, 2) "HGHR-" had the highest incidence of HR- receptor status and highest grade but few metastases and relatively good outcomes, 3) "HR+ bone" had isolated bone metastases and uniform receptor status (HR+/HER2-), and 4) "Distant organ spread" had high metastatic burden and poor survival. Bayesian networks revealed clinically intuitive interactions between patient and disease features.
Conclusions: We have identified four distinct subgroups of breast cancer using LCA, including one unexpected group with good outcomes despite having the highest average histologic grade and rate of HR- tumours. Deeper understanding of subgroup characteristics can allow us to 1) identify actionable group properties relating to disease biology and patient features and 2) develop group-specific diagnostics and treatments.