Prevayl’s vision is to enable the understanding of the human body through the creation of a multimodal foundation model. Using artificial intelligence we aim to facilitate the understanding of relationships between human biology, behaviour, disease biomarkers and health outcomes.
Key to the vision are large-scale multimodal datasets, like the UK Biobank. The goal is to develop a pre-trained foundation model that can be adapted or fine-tuned for specific tasks and applications in health care. Health data is inherently multimodal and therefore the project aims to solve 2 key challenges:
1. the ability to process multiple types of data, including structured and unstructured text, image, video, and physiological sensor data such as ECG and IMU; and
2. the ability to integrate and analyse diverse and multi-dimensional human data to reveal patterns and relationships.
We will approach the project in 4 stages:
1. data acquisition – we will collect a vast array of medical data from sources like the UK Biobank and other health databases;
2. data processing – we will develop a multimodal encoder to align all data modalities into an embedding space;
3. model training – we will train the model on the integrated dataset to learn general physiological patterns and relationships;
4. model validation – we will validate the model output via the encoder to ensure accuracy and also compare encoder output to final LLM output to evaluate hallucination risk.
Solving these data challenges will enhance the healthcare profession’s ability to predict, diagnose, and treat a wide range of disease conditions, as well as promote strategies for disease avoidance. It also offers the potential for groundbreaking insights into disease development, paving the way for future medical breakthroughs.
Project duration: 3 Years