Table 4.

Perspectives of machine learning implementation in pediatric medicine from qualitative interviews.

Themes and subthemes					Example quotations
Benefits of machine learning implementation
	Facilitates decision making
		Complex scenario	To me was very disturbing scenario where a very complex child with a number of issues, [...] Having some kind of system which alerts physicians who are directly involved as to not any in their own domains, but in other domains’ risk would be helpful
		Support less experienced clinicians	Well, you know where I see potential strength is not so much for the highly experienced physician, but more for the person who’s starting out [...] and just doesn't have that experience base yet.
		Reduce cognitive load	It can offload some of the cognitive load. So yeah, absolutely. I mean there's many times you find yourself in the middle of the night very tired, half groggy and trying to make a decision and kind of going back and forth in your brain. You know, for like half an hour - should I do this or that?
		Reduce cognitive bias	[...]  it's not that it replaces your judgment, it supplements another sense. ... your decisions informed no matter by your experience but it's informed by thousands of experiences, computed even more times to see all the possibilities and then come up with a best sort of path forward. The most likely scenario. And understanding that it is not a perfect prediction but it's a much more ... It's where that big data come in, right? It's really powered by real knowledge. It's not personal perceptions or personal experience, which is very biased and skewed.
	Improve quality of care
		Standardize care	There probably is some significant interpersonal variability in terms of interpreting the guidelines and then decision making around management, and so if we could use machine learning so that there’s less of that, all the while providing I guess more accurate or better care. I think that would be very helpful.
		More effective triage	I feel like if we were able to use machine learning to risk stratify so that kids who are at higher risk could get more timely access to a referral. Recognizing that in this particular situation, certainly early diagnosis and management can really impact the trajectory of a child’s outcome. I think that would be helpful.
		Facilitate precision medicine	And what I mean by that if you look at it, look at a population of babies who were all born, say at 25 weeks. There will be individual differences that should [...] be detectable by machine learning or artificial intelligence. So instead of treating every baby as simply a member of the population, I can sort of drill down onto specific physiological and clinical factors for that baby, [...] get closer to the idea of personalized medicine.
	Reduce physician workload
		Freeing up time for physicians	If it was really useful, then maybe it would free me up to do things that only I can do.
Challenges with machine learning implementation
	Hinders decision making
		Algorithmic bias	It's all about the biases like built into the system and how it's learned the data that you're putting in, and then how you get that out and how it would either pick up on our own biases, or like pre-existing, whether those are like systemic like sort of racial, ethnic or gendered biases [...] And so then that's not really helping us.
		Lack of transparency and trust	Understanding what it is doing: like if it's doing things that I can't follow or don't understand, I'm going to be less to trust its opinion [...] I want to understand how it came to that decision so I can ask myself if I agree.
		Not incorporating clinical expertise into decisions	I think it's like all the tools we have in medicine that if you use it appropriately, it can be incredibly powerful. But if it's used as a, you know, let me abandon all my other skills and I'll just follow this kind of direction, it potentially could be harmful, so I think a lot of thought will be needed. I mean in some ways it helps to predict, but I think I've always been a little skeptical about machine learning because biology and people do not follow an algorithm, they don’t follow a formula.
	Negative impact on quality of care
		Need for outcome evaluation	[...] looking at what the outcomes are and that we're actually improving patient care. So if we're admitting more but the outcomes are the same and the return visits are the same, then did it really matter and are we improving patient care or we just increasing cost to the system? And so, I think it needs constant evaluation, just like anything else that we do...
		Data quality	Of course, you know your outcome or the recommendation, or how machine learning is used is always only as good as the input, right?
	Practical concerns
		Challenges in workflow implementation	I guess there’s going to be some learning curve. How do we use it? Is it feasible? Is it on my iPhone? Do I have to go into certain area, how fast will it take me to get the response and along with the interface, how friendly is the interface? You know things that are related to stuff that we have not seen yet.
		Accountability	The challenge with machine learning over clinical decision rules is right now with the accountability piece and it's just getting to what that's going to be like. We don't blame, you know, the lab test or the lab. You know, if we don't pick it up. But right now, I think people feeling if they go against it, what does that mean and do we have to add like admit everybody or treat everybody based on that, knowing that like you alluded on the first question that it is a probability [...] So what does that mean for the provider thing choose to ignore it versus if they choose to follow it in harm happens
	Physician role
		Uncertainty in physician role	On the other hand, you know, maybe it also kind of takes away a little bit from like, I guess there's a fear of what exactly is the doctor's role. If the computer can do a better job at diagnosing then I can