Machine learning and artificial intelligence application in medical care systems is a hot topic being discussed actively in both academia, with countless papers focusing on different stages of the process: from theoretical research to thorough exploration into healthcare systems. While researchers are still actively examining every step, not many real applications have been done. Regarding this issue, some researchers from the University of Pennsylvania, the University of California at Berkeley, and VA Medical Center proposed five criteria for developing similar framework in terms of evaluating and regulating predictive algorithms.

 

(1) Meaningful Endpoints

In most theoretical work, the most popular metrics are area under the curve, which is may not be readily understandable by clinicians or patients, or even useless without providing a cutoff suggestion, making it not very clinically meaningful. The proposed endpoints should be based on some established standards of clinical benefit, even though this may involve some downstream outcomes such as overall survival and change in cost of clinical management. 

 

(2) Appropriate Benchmarks

Since the focus points of these algorithms may not be exactly the same as clinical endpoints, there is usually not a universal standard to test that. As standards for comparison are not well-defined in the FDA’s premarket clearance program, studies can be performed without clear comparison with meaningful clinical practice to get approved. So far, a best bet is still the “clinician + algorithms” framework, where algorithms are compared with clinician’s expert review. But this is very limited in terms of recruiting enough experts to do this stuff.

 

 

(3) Interoperable, Generalizable

FDA 510(k) is usually for such algorithms in general use. The WAVE system, the first predictive surveillance system that received FDA 510(k) clearance for general clinical use, is largely based on five common key factors: heart rate, respiration rate, oxygen saturation, temperature, and blood pressure. With researchers more and more “greedy”, many other data forms are required in building their algorithms. This may lead to the problem that some of the factors used are very specific to their health systems, so that interoperation becomes a huge obstacle when transporting their methods into another system. As the other side of the coin, their algorithms work pretty well in their system, usually. Some common data models, such ODHSI, may be a good solution for tackling this problem.

 

 

(4) Specify Interventions

This is a general problem for these algorithms, which even though provide good accuracy in predicting upcoming events, but are not accompanied with solutions. This may be severe (or useless) in critical settings such as ICU, where a decision should be made momentarily once an error came up. It is still a hard problem for the system to make predictions and give solutions at the same time. So is this a huge limitation to machine learnings?

 

 

(5) Audit Mechanisms

Algorithms are buggy, and way from perfect. Just as what FDA did for post market surveillance for drugs and medical equipments, similar strategies should also be applied to algorithms. How the system works, how much it helps with improving healthcare, how many error it prevents, how many error itself makes, are some easy yet very important questions to answer when monitoring a predictive system. Making guidelines for auditing algorithms to be implemented should never be ignored as part of the original plan.