Artificial Intelligence in Medical Diagnostics
When you think of artificial intelligence, do any of these come to mind?
Enabling Autopilot and letting your Tesla drive you through rush hour traffic on the 405
Watching IBM’s Watson defeating Ken Jennings on Jeopardy
The cell phone surveillance system that Morgan Freeman created for Batman in The Dark Knight
Regardless of what you know or think about artificial intelligence and the associated hype, rest assured that it is real and has already made fundamental changes in the world. One of the areas where AI is making a major impact is in healthcare, and specifically in medical diagnostics.
Medical diagnostics work by measuring some aspect of your physiology, comparing it to what “normal” should be, and determining whether or not your result is healthy.
In the current paradigm of product development for medical diagnostics, developers have to “lock in” key characteristics of the diagnostic test early in the development process. This includes deciding
How much material (blood, tissue, etc.) is needed to perform a test?
How accurate will the results be?
What physiological conditions or diseases can be determined based on the results?
How frequently does the diagnostic device need to be checked or calibrated?
Once these parameters are determined, the developer has to reach a few major milestones before the product can improve patient lives
Complete extensive testing to prove the device can meet the requirements
Analyze all the results, summarize the data and submit the information to the appropriate regulatory authority
Receive approval to market the diagnostic device as described in the submission
This process typically takes about two years.
As the device is used in the real world, data is generated that can pinpoint ways to improve the performance of the test. This could include how to perform the test with less tissue, how to improve the test accuracy, or expanding the types of diseases that can be measured. This data is generated by customers who are using the existing diagnostic test, internal testing performed by the developer, and sometimes through scientific or technical advancements made by others in the field. Once the developer has gathered enough information to justify an improvement, the product development process is started once more. This improvement cycle occurs once a year or so.
If you add up the time it takes to develop and launch a medical diagnostic, with the time it takes to gather enough data to start an improvement initiative, you’ll realize that if you decided to start a medical diagnostic company in September 2019, it would be late 2024 before you’d released a second version!
This time frame makes sense for “traditional diagnostics”, where the majority of the analytical work is performed by consuming some amount of human tissue, purifying it, and observing how it reacts with chemicals which were precisely dispensed within a large test system.
In this paradigm
The tissue being tested has a limited shelf life
It is consumed during the analysis and must therefore be replaced
Performing one analysis requires a non-trivial amount of consumable material such as specimen containers or test chemicals
The diagnostic value of the test is fixed and pre-determined based on the understanding of the chemical interactions at the time that the test was created
Compare this paradigm to that of next generation of diagnostic devices, where the majority of the analytical work is performed by computer code which analyzes information about the tissue and observes how it compares to a given dataset.
Once information is extracted, it can be stored forever without degradation
The information can be re-used for an infinite number of tests
Performing a test has negligible cost or required material
As new information improves any aspect of the computer code, the diagnostic value can increase
This last distinction is where artificial intelligence can drive exponential improvements. The developer of a next generation diagnostic device (or a traditional diagnostic with some next generation capabilities) can imbue the product with the ability to improve itself - independent of input from the developer. Examples of such improvements include
Improved accuracy in detection of the test
The ability to detect new disease states
The ability to analyze patients in a new demographic group
Needless to say, this self-improvement ability can drastically reduce the amount of time it takes for an improved diagnostic product to be developed. Instead of taking two years to develop a version, one year to gather data to improve the performance, then another two years to develop the improved version, artificial intelligence can allow a diagnostic test to immediately improve in real time - as soon as new information is available and validated!
However, determining performance parameters and proving the diagnostic test’s performance is only half the battle when it comes to launching a medical diagnostic. The other half is creating the regulatory submission package and achieving approval to market the device. If the regulatory steps in the product development process don’t accelerate, the overall pace of medical improvements won’t change. Fortunately, the FDA has recognized the looming imbalance between the pace of technological innovation and the pace of the regulatory process and is actively developing a new approach to support the use of artificial intelligence in medical diagnostics.
In the next article in this series, I'll explain the FDA’s proposed approach and the major roadblocks that must be addressed so the regulatory process can keep pace with the exciting potential of Artificial Intelligence in medical diagnostic innovation.