Beyond language: generative artificial intelligence as a general computing model for medicine

Area of Science:

• Artificial Intelligence in Medicine
• Health Informatics
• Computational Health

Background:

• Traditional AI models often rely on textual translation of medical data, limiting their ability to capture complex temporal health patterns.
• Lack of direct processing of structured medical data hinders the development of accurate and personalized patient care models.

Purpose of the Study:

• To advocate for direct tokenisation of medical data for transformer-based AI models.
• To explore the potential of tokenised representations for learning patient health timelines and improving clinical decision-making.
• To propose a framework for privacy-preserving collaborative development of AI models in healthcare.

Main Methods:

• Direct tokenisation of discrete medical data units (e.g., lab results, medications, vital signs).
• Utilisation of transformer-based models, such as Enhanced Transformer for Health Outcome Simulation, for forecasting health timelines.
• Development of a privacy-preserving model-sharing framework for local training and sharing of trained models.

Main Results:

• Tokenisation enables transformer models to learn directly from the temporal structure of patient health timelines.
• The proposed approach facilitates more accurate and personalised patient care.
• Privacy-preserving model sharing allows for collaborative development across institutions, enhancing data diversity.

Conclusions:

• Embracing tokenised representations is crucial for developing scalable, multimodal, and equitable artificial intelligence in medicine.
• Direct tokenisation overcomes limitations of textual translation, improving AI's understanding of patient health trajectories.
• Collaborative efforts and diverse datasets are essential for advancing fairness, generalisability, and equity in healthcare AI.