Breast cancer is a heterogeneous disease comprising multiple molecular subtypes, each with distinct clinical outcomes and therapeutic challenges. MicroRNAs (miRNAs), as key regulators of gene expression, hold great promise as biomarkers for cancer subtyping and developing personalized treatments. In this paper, we propose a machine-learning discriminative modeling framework to uncover subtype-specific miRNA biomarkers for breast cancer. Our approach jointly integrates miRNA expression data with patient clinical data, to identify miRNA signatures that differentiate between luminal A, luminal B, HER2-enriched, and Basal-like subtypes. We conduct extensive validation across multiple discriminative models and provide evidence that microRNAs are strong discriminators within breast cancer subtypes, achieving up to \(90\%\) F1. Furthermore, we contrast our findings against state-of-the-art multi-omics integration and biomarker discovery, and provide a reassessment of its predicted miRNA signatures. To this end, a model explainability approach is employed to analyze and pinpoint subtype-specific miRNA profiles. These potentially highlight subtype-specific biologically meaningful and functionally relevant miRNAs, that can now be therapeutically validated through in vitro and in vivo experiments.