The TWIN-X model – A Multi-Layered System

This layer comprises modality-specific embeddings, i.e., heterogeneous data types structured according to their clinical context, including laboratory results, CT imaging data, and clinical examination reports.

The interaction layer acts as the interface between clinicians and the AI system. It is based on a so-called Mixture of Experts (MoE) model, a type of neural network architecture in which, rather than using the entire model for every input, only a small subset of specialised components ("experts") is activated to process a given query. The model interprets user queries, retrieves relevant information from the different data representations, coordinates the required reasoning processes, and generates structured responses. These responses are accompanied by transparent evidence links and calibrated confidence estimates to support trust, transparency, and interpretability.

The cross-attention fusion transformer is a model component designed to integrate information from multiple data sources into a unified representation. The cross-attention mechanism enables the AI system to identify which elements from different modalities, such as laboratory results, medical imaging, and clinical reports, are most relevant to a specific user query. Based on this information, it generates a query-specific joint embedding, i.e. a combined representation of the information that is most relevant to the clinical question being asked. Unlike static patient models, this representation is recomputed dynamically for each query. This enables the system to provide adaptive, context-aware support for clinical reasoning and decision-making.