“If you build it, they will come.”
WP Kinsella, Canadian author of Shoeless Joe
This timely viewpoint from JAMA Ophthalmology asks the proverbial question “if the artificial intelligence tool is made available, will clinicians use it?”
While this editorial appeared in the journal for ophthalmologists, it provides practical insight for every clinician and healthcare stakeholder. The authors correctly point out that, even for FDA-approved AI-based technologies that have reached the market, there is no guarantee of adoption by clinicians. In addition, the authors remind all of us that it is precisely the loss of faith in AI-based technologies that produced the previous “AI winters” a few decades ago that resulted in withdrawal of funding for research in this domain.
The deep learning methodology in ophthalmology is trained to map inputs of images to outputs of diagnoses such as diabetic retinopathy. The authors briefly describe the issue of algorithm bias that originates from how data are captured and algorithms are developed and evaluated. To minimize this bias, the FDA used a framework called Software as Medical Device (SaMD) to evaluate the software throughout the lifecycle. This term, SaMD, as well as “locked algorithm”, is not my favorite, as they imply that agile software can be perceived in the same way as much less agile medical devices with much longer development cycles. Another challenge for these AI algorithms is that unanticipated modifications beyond the original intended use of the SaMD will require a new application.
Finally, the authors bring about the tenet that analytic and clinical validity of the AI algorithm does not ensure fairness and equity or even generalizability. The inequities of an algorithm can be perpetuated or even exaggerated. The literature is replete with examples of how algorithms can compound disparities in healthcare utilization and outcomes. One potential solution is the FAIR Data Principles: findability, accessibility, interoperability, and reusability of digital assets. The authors conclude this commentary by stating that it is essential for clinicians (not only ophthalmologists) to develop robust data sets that capture real-world use of algorithms, and concomitantly contribute to improving these algorithms by identifying failure modes in order to stave off yet another AI winter.
Read the full paper here
The post AI in clinical practice is here – now what? appeared first on AIMed.
Author: