M4-health: the basis for general AI in health care

Black background. You see several
M4Health. lev dolgachov

In the VR funded project “M4-health: the basis for general AI in health care” we make way for a more general and flexible type of AI in health care. The project is centered around the creation of digital twins; physiologically based computer models for patients. We want to use models to promote health, an aim which has given the project it’s name, M4-health, which is short for “models for health.” M4-health is also an abbreviation of the type of models we develop: mechanistic, multi-level, multi-timescale and multi-specie models.

This project contributes to the research field artificial intelligence (AI) in health care. AI can contribute with better advice and insights, which in turn provides more correct diagnoses, and help find the best treatment, because there is a better basis for making sound medical decisions. Potentially, AI can do this instead of physicians, which would reduce the cost of healthcare and make it more accessible. One reason for the great potential for AI in the field is the recent explosion of data, which makes it impossible to just look at the numbers and draw conclusions, like we used to do. Another reason is that a new method for AI, called deep learning, has been developed. In recent years, there have been many new and important examples of how deep learning can be used. For example, it is now possible for an AI to automatically diagnose the type of tumour in skin images with the same degree of certainty as a dermatologist.

However, every example that exists so far is quite specialised. Put another way, they take one particular type of data (for example an image) and can predict a certain type of output (for example if the image shows a tumour), and do nothing else. This is due to limitations in the AI methods that are currently used: they are all within the realm of so called narrow AIs. The key limitation with these kinds of methods is that they require large amounts of training data, and can not go beyond the scenarios they have been trained for. For this reason, most of the data that can be collected for different diseases are not useful in the AI system.

Furthermore, traditional AI will only give specific risk estimations and identifications, and not give a general overview of the state of the body, that can be used for different purposes. Such narrow AI will not build an understanding of the body.

The foundation for a new type of AI in healthcare

In this project, we lay the foundation for a new type of AI in healthcare: one that builds a general overview to help us understand the body. The basis for this understanding are so called mechanistic models. Such models describe the physiological processes that happen in the body, and use data to support or reject hypotheses of what these processes look like. This type of model does not require such large amounts of data, because they do not acquire all their information directly from the data.

Gunnar Cedersund has during the last 20 years shown this many times in collaboration with a large number of experimental and clinical research groups all over the world. Together we have uncovered physiological insights for many of the body’s organs. In this project, we are going to use these previously separate organ projects as a base and take a quantum leap towards a level no one has reached before, where all these models are connected and can be used by an AI.

The project consists of three steps:

  1. Connect all the previous models of the organs into one large model of the whole body. This will be done by connecting the organ models with an existing model of blood flow for all vessels in the body.
  2. Create individual versions of this model, which are specific for each person, by training and testing them on our unique data gathered in 10 large clinical studies. In this step, we will connect the mechanistic models to both statistical AI systems, and AI systems based on the structure of the brain. The latter will be done in two ways; by including the brain as an organ in the body, and by constructing so-called hybrid models. These hybrid models can both simulate scenarios of what would happen to a specific patient during different treatments, and calculate how the risk of, for example, cardiac disorders would change in these scenarios.
  3. Test the usefulness of the models on 4 different patient groups from 3 different countries.

The potential in this project is great: to create the basis for a new type of healthcare, one where AI does not only perform specific tasks, but instead contribute to a useful overview of each patient’s body showing what it looks like and their health status. All researchers can contribute to updating this overview, something that both patients, medical doctors, and pharmaceutical companies will benefit from.