Multimodal Deep Learning and Data Fusion in Precision Breast Oncology: Clinical Applications, Fusion Strategies, and Future Directions

Multimodal deep learning and data fusion represent a transformative approach in precision breast oncology, addressing the disease's profound biological and clinical heterogeneity. This review synthesizes recent advances in integrating diverse data streams—including radiology, histopathology, multi-omics, and clinical records—to enhance prediction models for key clinical tasks. We detail the taxonomy of fusion strategies, from early feature-level to intermediate representation-level and late decision-level fusion, highlighting the increasing adoption of attention-based and hybrid architectures. The application of these models consistently demonstrates superior performance over unimodal alternatives in predicting neoadjuvant therapy response (pathological complete response), stratifying prognosis and recurrence risk, improving diagnostic classification, and inferring molecular subtypes directly from imaging. Despite these promising results, challenges in data harmonization, interpretability, and robust clinical validation persist. Future directions must prioritize the development of standardized benchmarks, the incorporation of dynamic biomarkers like circulating tumor DNA, and the adoption of privacy-preserving federated learning frameworks. Ultimately, for multimodal artificial intelligence to realize its full potential, it must evolve into clinically deployed, pathway-aligned decision-support systems that are rigorously validated in prospective trials to demonstrate tangible improvements in patient care and outcomes.