What Evidence Will Cut it for AI Solutions in Healthcare?

My discussions with many of the experts in AI in Healthcare has highlighted the fact that well-designed, large-scale, multi-center trials have not been done so far. These types of trials would establish the efficacy and safety of these algorithms in the real-world settings where there are different types of patients. Also, the algorithm gets tested on whether it can generate its output during the normal clinical workflows so it can be timely and based on realistically available inputs (data.) Some of the reasons cited for this lack of quality studies include that the cost and the length of time that it would take to do such studies are beyond the means of research centers or small companies. For larger companies to invest the capital, they would need a clear path to their return on investment and that is not as clear for algorithms as it would be for a drug or medical device. Intellectual property protections for algorithms are not very strong and the barriers to entry are low.

Experts emphasized the importance of doing well-designed trials to establish the benefits of AI models.  They stated that the current state of research is chaotic with a wide range of quality and publications are scattered in clinical and Computer Science (CS) journals and they indicated that high-quality studies will need to be designed to examine these algorithms based on their impact on one or more of these areas:

Accepted practices in medicine are prospective trials involving several centers with diversity of patient populations and datasets would provide far more convincing data about the quality of these models.Of course, doing real-world studies is not easy and questionable return on investment may not be the only reason they are not being done. Current state of healthcare data may serve as a barrier in short- and medium-term to adoption of AI technologies. In the real-world setting, the data needed to run the algorithm will need to be available at the right moment to generate its output and if not, it will not help the clinician or the user when its help is needed. So, if you design a trial, you need to be sure that the flow of data is air-tight. If it is not, the study may not show the benefits of your algorithm. In that situation, you have invested time and money but you fail to show any benefits!

These types of trials are not just needed for adoption of these models by clinicians. Insurance companies, public or private, demand solid evidence to pay for the use of these technologies. In healthcare, it is very difficult to succeed in drive adoption of  innovation if it is not paid for.  As long as the promised impact on patient outcomes is not shown in well-designed trials, reimbursement will be difficult to secure and thus medical centers will be slower to purchase.