Improving Healthcare with AI Diagnostics

(‘A Better World’ Series)

As someone who has done a lot of head scratching blood test results over recent years, I am surprised over the lack of high-quality analysis of those results.  A normal blood test may provide a dozen or more individual test results.  If they are all normal, nobody looks any further.  If there are abnormal results, it is up to the healthcare provider who ordered the tests to surmise a cause and next steps.

For many years, I have been obtaining the same abnormal blood test results.  Different doctors have focused on different individual results and unsuccessful therapies.  After those many years, I finally had a hematologist look at the results and he said that I was not anemic (past doctors have said I was due to low iron levels) and that I was probably suffering the effects of inflammation (which he doesn’t deal with).  But, to be safe, he ordered a couple of tests to make sure there wasn’t a blood problem.  One test result suggested that there was a problem with inflammation and the second showed definitively that I had an acquired genetic mutation which caused my bone marrow to overproduce blood components.  After some research on my own, I am lead to believe that the mutation may have been involved (at least in part) in at least two other serious medical issues in recent years.

Why am I telling you about my diagnostic story?  To demonstrate how difficult diagnostic results are to interpret even for very well-meaning and well-educated professionals.  For one thing, each doctor viewed my results through the lens of their personal experience and education.  And that lens caused them to emphasise some results over others.  In addition, the specialists that I visited did not interact with each other or seem to care about what was being found by the other specialists.  And, I feel, they all disregarded the importance of consistent abnormal test results year over year.

Again, I believe that these men and women truly cared about my health and were doing what they felt was best for me.  I never got any resistance from any of them in terms of running additional diagnostics or seeing other specialists.  The problem is that they are human and, like all humans, place greater importance on their own direct experience than the collective experience of all doctors.  Also, some patterns can be difficult to see and with a slightly large dataset, it can be hard to discern any patterns.

I believe that these doctors (and I) would have really benefited from the use of artificial intelligence (AI) in evaluating my test results and other physical metrics (body dimensions, physical injuries, complaints of pain, etc).  The AI analysis could provide the practitioner with suggested follow-on tests, the likelihood of certain causes, suggested treatments, etc.  The practitioner would then review the results and plan the next steps with the patient.  The outcome of those next steps would then be fed into the AI database to update the analysis.  I think there is a lot of fear of taking the human out of the process but I think this approach would let computers do what they do best and let humans provide QC of the analysis and the human-human interaction, which humans do best…so far.

Unfortunately, I can’t just take my test results, plug them into a computer, and have it evaluated by AI.  To be effective, an AI database must be populated with a large number of data points.  In this case, the database would need to collect things like the test results, physical attributes, patient complaints, medications taken, diagnosis, and outcome.  Fortunately, we have a lot of data floating around in computer networks in healthcare systems across the country.  There would be a lot of work to integrate the data from the systems into one AI database (all de-identified, of course), but eventually the machine learning will start seeing patterns and predicting outcomes.

The technical feasibility of AI has already been proven over and over again.  The key is to teach the system with a large quantity of high quality data. It seems like the best way to accomplish this task would be for a public organization, such as the National Institutes of Health (NIH) to collect all of the data from the myriad of sources, all formatted correctly, and generate a database.  Then private organizations (for-profit, non-profit, whatever) could use the data to run their machine learning to develop their AI and offer healthcare providers access to their AI and their unique interface. This approach would ensure that our data is publicly available (not owned by an individual company) in a consistent format and that it will contain the most data while encouraging innovation on how that data is analyzed and how it is presented to customers.

Those are my thoughts on it, what do you think?