Why healthcare needs Causal AI?
Imagine a future where we could study huge populations and not only conclude about mass scale correlations (e.g. eating salt increases one’s chances of having high blood pressure), but actually discover the cause & effect relationships for each individual. In this future, medicine is personalized, therapy is optimized, and quality of life & safety are maximized.
Science is about transforming observations into timeless laws of cause & effect. I have always been impressed by the progress science made in the last few centuries without having the mathematics to study cause & effect. It is truly remarkable that we derived so many conclusions by trial and error and by the ingenuity of the scientists who tinkered with nature.
Most of us know that correlation is not causation, but finding causation is difficult. Think about an incredibly complex system like the human body with roughly 200 different types of cells, 25,000 genes, 30 trillion cells, and close to 40 trillion microbiome cells all interacting with one another. Our limitation in understanding the cause & effect relationships between these components, our lifestyle, and the environment inhibits us to help people live healthier.
In the last decade we’ve seen an exciting revolution unfolding in front of our eyes: the AI revolution. We’ve seen AI, and especially deep learning, doing magical things: performing complex surgeries, driving cars, classifying images better than humans, creating new art forms, beating humans in computer games, composing music, creating disturbingly realistic deep fake videos, and mimicking the writing of famous authors like Shakespeare. The list goes on and records are still being broken.
This technology has far-reaching potential and bringing it to healthcare will undoubtedly improve human lives in ways we are only just now beginning to imagine. It truly is the new and unexplored frontier of human improvement and advancement. However, while (traditional) AI is amazing at approximating data, learning from it, and even generalizing it to make predictions, it has two big weaknesses. First, the more complex the data and the AI we are using the harder it is to know what the AI model is actually learning. In that sense, the best conventional AI models we have are black boxes. Secondly, conventional AI excels at finding correlations in the data but correlation is not causation.
In healthcare, especially with therapies, it is essential to find cause & effect relationships in the data. We must be able to answer causal questions. We need the therapy we give patients to be effective and safe, and both safety and efficacy are about cause and effect. If you know what would positively affect a health problem (and how much), you have found a therapy. If you know what health issues the therapy you found may cause and under what circumstances, you hereby discover its side effects and safety profile.
No matter how you slice it, healthcare without the mathematics for studying cause & effect is like doing physics without knowing calculus.
In short, (conventional) AI has an amazing capacity to handle big complex data and learn from it. Causal AI has the ability to turn black boxes into white boxes and to convert population data to N=1 data. Causal AI is the missing key to unlocking the bright future of personalized medicine.