I am a pediatric cardiologist and have cared for children with heart disease for the past three decades. In addition, I have an educational background in business and finance as well as healthcare administration and global health – I gained a Masters Degree in Public Health from UCLA and taught Global Health there after I completed the program.
“Artificial intelligence is creating a mind, hopefully as pure a mind as possible, for a computer.”
Frederick Lenz, American Buddhist
This Viewpoint from the Journal of American Medical Association reflects upon the diagnostic AI development as it relates to clinicians expectations and real world experience. Diagnostic AI, as the authors explain, refers to “a broad range of applications that use learning strategies that mimic human approaches to learning.” If the clinician can better understand the underlying mechanisms of diagnostic AI, they can perhaps appreciate both their advantages and limitations. The authors then go on to present three diagnostic AI learning methods (experts, examples, and experience), and draw from these methods parallels to how clinicians learn.
First, learning from experts is based on expert systems of the Go Old Fashioned AI (GOFAI) era. This methodology is severely limited by the narrow capabilities as well as labor intensive encodings. Next, learning from examples is executed recently by supervised type of machine learning. This methodology involves the algorithms learning in order to have rules for new incoming data, and has the potential to have superior performance compared to human experts. The limitations of this AI tool is its “black box” nature as well as its demand for human-derived labels that can be inaccurate or subjective. Lastly, learning from experience is possible with reinforcement learning (what I call “smart” AI), which was used in complex strategy games such as Go and Starcraft. One significant limitation is the lack of available simulated cases and the associated outcomes of those examples. In addition, many scenarios in biomedicine do not lead to a clear dichotomous outcome so this adds to the complexity of the situation. The authors state that the clinician example for this learning from experience is exploring variations around pathways, but the scale of this learning is not the same as reinforcement learning in which many more cases are studied.
In conclusion, the authors summarize by stating that diagnostic AI systems (not unlike clinicians) learn by mimicking experts, acquiring examples, and conducting experiments, and that these systems are best used in synergy with humans. The AI diagnostic tools can focus on computational and data intensive work while clinicians can focus on what humans do better: creative problem solving, complex decisions that involve uncertainty, and understanding the patient’s context.
We at AIMed believe in changing healthcare one connection at a time. If you are interested in discussing the contents of this article or connecting, please drop me a line – [email protected]