AI & productivity in healthcare

Everyone's looking for evidence of AI increasing productivity. There's some research indicating that it is happening in healthcare.

The AI doctor will see you now. Note the odd hallucination on display here: the AI doctor appears to be using a stethoscope to listen to a male patient’s heart, but an image of an ultrasound of a fetus is also included. Dall*E has no conception, of course, of gender, only that “stethoscope” and “fetus” and “ultrasound” are medical terms. The prompt used for this image was: generate an image of an AI doctor treating a human patient.

A few weeks ago, I wrote a post about AI & productivity. I noted that there were a lot of news articles speculating about an incipient boom in economic activity, given AI, but relatively little hard evidence that there was any actual increase in productivity. Well, someone commented on that post, noting that he had done research on the question of AI & productivity. Specifically, he had done research showing that opthamologists who use autonomous AI-enabled tools to perform diabetic retinopathy¹ exams can see more patients per hour than can opthamologists who don’t use such technology. This is interesting, in that it presents us with some empirical evidence for the claim that AI can increase productivity.

The researcher who commented on my post is Michael Abramoff, MD, PhD, the founder of Digital Diagnositcs. Digital Diagnostics makes software for opthamologists to use when treating patients with diabetic retinopathy². In any event, he and I spoke last week about his work, and this post is the result of our conversation. He’s reviewed this post before it was published, so, again, if that gives you pause, fair warning. I don’t have any financial stake in Digital Diagnostics, and have not been compensated in any way for this post.

The paper is available here. Its introduction provides some context for the hypothesis being tested (footnotes have been removed from this quote for readability):

We hypothesize that autonomous Artificial Intelligence (AI), where a computer rather than a human provider makes the medical decision, can improve clinic productivity as defined above. Such autonomous AI systems have recently been approved by the US Food and Drug Administration (FDA) as safe and effective for use in medical care and as reimbursable by Medicare, Medicaid and private insurance payors. However, the potential productivity impact of autonomous AI systems has received scant attention.

The study asserts that productivity is best measured in the number of completed care encounters per hour per specialist physician, and notes that using an autonomous AI-enabled digital retinopathy scanner yieleded an increase in productivity of 39.5%.

Another point which bears mentioning, from the study:

Autonomous AI systems have particular advantages in under-resourced settings, most obviously the benefit of improved productivity where trained personnel is scarce. While telemedicine platforms have been implemented in some cases, these do not allow instantaneous, point-of-care diagnostics, so that the care encounter cannot be completed in the same visit. The reason is that while the patient images can be taken in the clinic, the diagnostic result will only be available after the patient has already left the clinic, resulting in care completion rates of 30%, at lower diagnostic accuracy. Implementations of the AI system, including operator training, was delivered remotely. This suggests these AI systems are scalable and sustainable, especially in low- and middle-income countries….

What I take away from this research is that we finally have empirical evidence for the claim that autonomous AI will increase productivity. When opthalmologists can treat and manage 40% more patients, that is an immediate gain that in most industries took 50 years of relentless quality improvement to accomplish, healthcare is more scalable and productive. This seems like the kind of thing we ought to celebrate. It will be interesting to see whether these findings hold for other areas of healthcare. Further, if these findings hold for healthcare, what other parts of the economy might we see increased productivity from the widespread use of AI?

1. Diabetic retinopathy is the primary cause of blindness. You can find more information about the disease here.

2. Yes, this means that Dr. Abramoff has a vested interest in his research, and in publicizing his research. Nonetheless, I don’t think that this interest is reason to dismiss it.