Rheumatic Digital Twin: Proposed Machine Learning-Based Multimodal Framework to Inform Clinical Decision-Making

Area of Science:

• Computational medicine
• Rheumatology
• Precision medicine

Background:

• Rheumatic diseases are complex and heterogeneous, making snapshot assessments insufficient for understanding their longitudinal progression.
• Current clinical approaches lack the ability to capture the dynamic nature of chronic immune-mediated conditions.
• There is a need for advanced tools to support precision medicine in rheumatology.

Purpose of the Study:

• To present the design of the Rheumatic Digital Twin, a conceptual framework for dynamic patient modeling.
• To integrate heterogeneous multimodal data for a comprehensive representation of the patient journey.
• To enable in silico cohorting for improved clinical decision-making in rheumatic diseases.

Main Methods:

• Utilizing domain-specific foundation models to process distinct data modalities (EHR, clinical notes, imaging, omics).
• Employing Transformer architectures with self-attention mechanisms to model temporal disease progression.
• Fusing unimodal representations via joint embedding techniques to create a shared multimodal space.

Main Results:

• The framework maps patients into a latent space where proximity indicates clinical and biological similarity.
• Identification of "nearest neighbors" (patients with similar trajectories) enables in silico cohorting.
• The system theoretically allows forecasting of clinical events and treatment responses.

Conclusions:

• The Rheumatic Digital Twin offers a novel approach to dynamically represent patient journeys in rheumatic diseases.
• This framework addresses data integration challenges and supports precision medicine implementation.
• It has the potential to enhance clinical forecasting, treatment prediction, and understanding of disease trajectories.