I’m speaking today on a panel at the Atlantic Live healthcare forum.
Below are the themes I’m going to hit in my talk - but it felt like it
was worth posting them here as well.
We have engines to generate correlations out of massive data that are
unreasonably
effective. We’re not using them in health, because the machines don’t
work very well on health data. It isn’t just that the data be massive,
but that it be available (i.e. open) and at least somewhat standardized.
Health data’s almost never either one.
Here’s a key sentence from the Google paper linked above:
“With a corpus of thousands of photos, the results were poor. But once
they accumulated millions of photos, the same algorithm performed quite
well.”
Have we ever had millions of data records about health for algorithm
training? Nope.
That’s a tragedy. It’s not that these engines - machine learning - are
the magic answer. But they’re an incredibly powerful toolkit for finding
correlations, which helps us decide what experiments to run to test for
causation. It’s like we’re carpenters and we’re sitting around building
a house and saying, nah, we don’t want to use power tools. Hand saws for
everyone!
The Amish approach won’t scale. We have no idea how many people it’s
going to take, or how much data about each person it’s going to take, to
discover exactly how valuable machine learning is going to be in the
health space. But that’s a terrible excuse to do nothing. In fact, it
should be the spur that gets us started on the experiment itself.
That’s what i am doing with the Portable
Legal Consent project. It’s an experiment to test three things. 1. can
we use a standardized approach to informed consent that disintermediates
the traditional study systems, which are crusty and have insanely high
transaction costs? 2. will people enroll and upload data, and of what
type and quality? 3. what kind of results will emerge from the
computational research?
Until we know these three things we’re basically arguing about
philosophy and not reality. We know that these systems are tremendously
powerful and predictive in many areas. We know that we aren’t using them
in health in a meaningful way. And we know that there’s nothing stopping
us from trying except ourselves.