In the last post, we discussed the myriad of issues that need to be analyzed from the point of view of builders or buyers of health AI solutions. Each one of those issues can make or break a promising solution. The business analysis for building or buying a health AI solution should include issues such as data availability, implementation complexity, and workflow issues. That’s because even if there’s an opportunity to solve a mission-critical problem with a high-impact solution, those issues can prevent adoption. Many systems that show great result in the development environment fail to show the same results in a real-world setting.
One of the key issues to explore are the workflow issues. This has proven to be the achilles heel for many of the solutions developed to date. Users at clinical settings or in life science companies have not been willing to change their workflows just because there is a new shiny digital tool provided to them. Unless it makes their lives easier and eliminates manual steps, the incentive to do things using a new approach is not enough for them to make the effort of changing their habits. One of the major hurdles we need to overcome is to ensure that the new AI solutions fit into existing workflows. These technologies need to have frictionless workflows or facilitate better workflows for the users. Fitting into existing workflows means that they won’t cause delays or result in extra steps for clinicians.
In radiology, which many see as the initial frontier for healthcare AI to improve patient care, key barriers include IT implementation and integrating into legacy systems. One example of this is AiDoc’s AI solution for the diagnosis of intracranial hemorrhage on head CT scans. My discussions with some of the pilot sites indicated that although the algorithm identified subtle hemorrhages that could potentially be missed by radiologists, the initial setup and implementation took a long time and was taxing on the IT staff. The multiple interfaces that had to be created between the scanner, the cloud and the PACS system weren’t simple. Also, the algorithm had a high false-positive rate, which created extra work and scrutiny for the clinicians and resulted in serious discussions about whether long-term use was worth the effort.
For decision support systems, the AI should deliver advice and information to physicians in real-time when they’re in front of their patients. If the data that the algorithm needs is fragmented, not yet entered into the system, or initially unstructured, the algorithm won’t be able to provide its output in a timely manner. These types of issues could make the algorithm obsolete.
Another workflow and implementation issue is taking an algorithm that was trained on external data and deploying it on local data, because we’ll need to ensure that it performs as expected on the local patient population. Selecting that data and further training the model will take work and create extra friction, which will mean a lower amount of short and medium-term adoption by health systems and clinicians.