The growing diversity of multimodal machine learning approaches makes it increasingly challenging to systematically analyse how methodological choices—such as data modalities, fusion strategies, validation protocols, and evaluation metrics—are reported across the literature. In this work, we present a proof-of-concept framework that combines topic modelling and Large Language Model (LLM)–assisted information extraction to support a structured analysis of multimodal machine learning studies. Starting from publica- tions retrieved through a structured Web of Science query, the proposed pipeline transforms unstructured textual descriptions of machine learning studies into a structured representation of methodological configurations derived from explicit textual evidence. As a representative case study, the framework is applied to the literature on multimodal learning in breast cancer. The analysis highlights heterogeneous combinations of clinical, imaging, and multi-omic data sources, a predominance of early fusion strategies, and limited reporting of reliability- related evaluation practices.