EP201: Decoding Decisions: Eye-Tracking Insights in Diagnostics

The Impact of Gaze and Fatigue on Medical Decision-Making with Dr. Bulat Ibragimov

In this episode of The Girl Doc Survival Guide, Christine interviews Dr. Bulat Ibragimov, an Associate Professor of Machine Learning and Medical Imaging at the University of Copenhagen. Dr. Ibragimov shares personal anecdotes and discusses his research on the role of artificial intelligence and eye tracking in medical decision-making. Key topics include the impact of gaze patterns and fatigue on diagnostic accuracy, the potential for AI to recognize when doctors may make errors, and how individualized gaze patterns can indicate the level of expertise and certainty in medical professionals. The conversation explores the implications of this research for improving the integration of AI in medical practices and enhancing training and decision-making processes for healthcare professionals.

00:00 Introduction and Guest Introduction

00:49 Personal Anecdote and Background

01:46 Eye Tracking and Medical Decision Making

03:18 Patterns in Gaze and Error Prediction

11:00 Fatigue and Its Impact on Accuracy

16:09 AI and Gaze Analysis in Medical Training

20:07 Conclusion and Final Thoughts

Transcript

Christine Ko: [00:00:00] Welcome back to the Girl Doc Survival Guide. Today I'm very happy to be with Dr. Bulat Ibragimov. Dr. Bulat Ibragimov PhD is an Associate Professor of Machine Learning and Medical imaging at the University of Copenhagen. He studies the role of artificial intelligence as well as gaze in medical decision making. He has worked at Johnson and Johnson as well as Stanford and Johns Hopkins Universities in the United States, and currently, as I said, he's in Europe.

Thank you so much for joining me today. 

Bulat Ibragimov: Hi, Christine. Thank you for inviting me. It's a pleasure to be here. 

Christine Ko: I contacted you because you've written articles on perception, on gaze. What we are looking at influences our perceptions and our medical decision making. We'll get into that, but can you first share a personal anecdote?

Bulat Ibragimov: Yes. I came from a family of of doctors. My mom is a radiologist. My aunt, she's a dentist. My grandma, she's a Professor of Medicine. And my uncles, cousins: all [00:01:00] doctors. So I thought that I will be a doctor eventually, but then my life turned a different point. I become a software developer, but as a child I went a lot to my mom's work, and I remember one episode very clearly. There was a patient coming to her and starting to argue about her decision about the reading of the image, and the patient was saying, No, I think the decision should be this. And I become unsure if my mom is right until the second doctor came and said, No, it's all fine. Now my mom participates in my experiments of eye tracking of doctors and I can see that even expert doctors can be wrong when they make decisions. 

Christine Ko: I love that story. Okay, so your mom and other people in your family are doctors, but your mom participates in your studies right now. That's really cool. When you said, based on eye tracking, you know for sure now that experts even make mistakes. Can you just expand on that a little bit? 

Bulat Ibragimov: Yes. We conducted several experiments to see how we [00:02:00] perform under high workload. We were starting to investigate if fatigue or some other factors may manifest in errors. So we asked a group of radiologists to read images during the whole day. The first day they read 400 images. Which was like already a lot. But we saw that they didn't get maybe tired enough and next day was to read 600 images. We were recording how we are reading changes during this day. And we saw some clear patterns of what happens when the users make errors. There are some predictive facts. We saw that all of them, very consistently, they start to cover 4% of image less with every one hundred images. And the more tired you become, less you cover, and the more you tend to overlook things. When the user goes too much, wander around over the image jumps from one organ to another, back, forward, this is a very predictive indicator that the user is uncertain of what's going on. Or, for example, if he returns to same area. He looks somewhere, but then moves away. [00:03:00] Goes back and stays at the same area, but there is little information gained with every step. So this is again the indication that the error might have been made. And that was very clearly visible on the patterns. These models or analysis of gaze paths was very predictive of errors. 

Christine Ko: So when someone's about to make an error, for one thing, you just overlook what you need to see. And that's maybe a separate issue, but sometimes you see what you should see, what you wanna see to make the right diagnosis, the right decision. You see it, but you are instead jumping from organ to organ or focusing in on something else rather than that more important thing. Is that correct? 

Bulat Ibragimov: Not only this. There are different patterns we can observe. Sometimes the reader leaves the abnormality area too quickly. This may indicate that the reader needs to spend enough attention [00:04:00] into it, but sometimes, which is much more common case, the reader actually returns back to the same abnormality area and wanders around there, returns back, leaves it, returns back again, and so on. So, although this does not indicate an error by itself, this suggest that the reader might be uncertain of what is going on here. And the probability of error is increasing if this pattern is happening. Okay. The similar type of abnormality, if a reader scan it through fast or efficiently and then move on, this is the indication that he's much more certain about what's going on here. 

Christine Ko: Got it. So just to contrast, can you talk about then the gaze pattern when a mistake is not being made, or when the probability of error is very low, like what does that gaze pattern look like? 

Bulat Ibragimov: It's very unique for every reader, at least in our experience. What we saw, one most surprising things which we observed, that from a gaze pattern, we can predict who exactly is reading the [00:05:00] image with very high accuracy, like 80% accuracy or so, something like this. Because every participant we had, they develop their own pattern of reading the image, like how the eye moves. And they, I'm just now giving an example. They look at one area, then they move to another area, and they move the third area and then they can make a decision and they do it again and again.

But sometimes this pattern breaks and this is an indication that something is unusual. There's uncertainty going on here. And that was actually one of the most surprising observation that we can, without even knowing who reads the image, we can guess from how the eye moves.

Christine Ko: That's fascinating. I'm just thinking about what you just said. So each individual is unique in how they come to the right diagnosis, but people are all similar, the way their gaze goes is all similar, when the probability of error is increasing? 

Bulat Ibragimov: Yes. So basically when they make an error, they're similar in the [00:06:00] way that there is a very chaotic behavior. The gaze start going in a very unpredictable manner all over the image. It doesn't happen every single time, but statistically we can see this. 

Christine Ko: Yes. Okay. That's interesting. So like for me, for example, if I were a radiologist, which I'm not, typically I would have a certain normal way of looking. If I'm going to be right. It's generally the same way over and over again. I look at it the same way, but my probability of error statistically is increased if I'm jumping around or I'm going back to the right spot that I should be looking at to make the right diagnosis, but then I go away and then come back and go away and come back. As someone who looks at microscopic slides, not radiologic images, but microscopic slides, I think that makes sense because most people do tend to go through the microscopic side, a certain way. There's a certain way of looking at it. When I am putting the slide myself on the microscope and looking at it, [00:07:00] versus if I'm looking through a multi-headed microscope and someone else, we call it driving the slide, someone else is putting the slide up and moving it around, it's harder for me to make a diagnosis when someone else is moving the slide around and showing the findings. Sometimes they're looking in a certain place and I'm like, Oh no, I wanted to go back and look somewhere else. It does change things when someone else versus me is directing where the gaze goes, I guess it relates to where you're able to look. 

Bulat Ibragimov: Yeah, absolutely. And I think in the case of microscopic images, we may need to rely on other types of statistical patterns, not the exact path, how the user travels, but more statistics if he goes back to the same areas, if he makes more fixations, the distance he travels, if his eye is shorter, it is also one of a predictive indicator that if a distance between fixation is short or if it's long, we can make some decision about whether the user is more certain about [00:08:00] what's going on or lesser.

Christine Ko: For me, looking at a microscopic image, if I feel like my gaze is jumping around or I'm looking here and then there, I agree, it would suggest to me a certain amount of uncertainty. Like I don't just immediately know what I wanna call this thing that I'm looking at. But is there a recommendation of what I should do when I feel like my gaze is jumping around like that?

Bulat Ibragimov: Yes, we're investigating this and we're thinking of not what you should do, because usually the users, they actually don't even sometimes understand that they made a mistake or we might have made a mistake. What we're trying to do, we're trying to think how to help and pinpoint this mistake. The easy situation is if some abnormality areas overlooked completely; if user just scan through it and didn't see it. In that case, you can highlight this area. We can say it to the user saying, Okay, look here. You didn't see something. Maybe there is nothing but take a look. This is one way. Another easy [00:09:00] scenario if we just believe that the user didn't pay enough attention or didn't grasp what is going on. His reading was too chaotic. Now it's an open question, How can we help? What can we do in this scenario? And there are a number of things we tried, for example, the second opinion. If AI thinks that the user made a mistake, he can try to send it to somebody else to take a look. And in that case, if the AI triggers your second opinion statistically more often in the cases where the user is wrong, we can deem some benefit from a second person looking at it. In that case, the requirements for AI prediction quality are very high. So you have to be very certain that the user made a mistake in order to , such a request.

But the question is how can AI directly tell the reader itself that he or she made a mistake? That's a big challenge in AI because if you show something to the reader, you can really guide the reader in the wrong direction in some [00:10:00] situations, or if AI is wrong, the readers tend to lose confidence in AI very quickly and just do things their own way. 

Christine Ko: I agree with what you're saying. When I make a mistake, I don't know I'm making it because otherwise I wouldn't do that. This is the struggle that I have because, as a doctor, when I'm making a diagnosis on a microscopic slide, I would love to be 100% correct. All the time. I don't want to ever make an error. The flip side of that is I am struggling with the fact that of course, I am human, and I cannot be 100% right all the time, unfortunately. Not just me; there's a certain percentage of medical error in diagnosis that's happening all the time. That's why I find your work fascinating because, maybe at the very least, I am trying to get at predicting when I am going to be more likely to make a [00:11:00] mistake. Going back to your initial response about having the radiologist look at 400 images and get tired, and then the next day you had them look at 600 images to make them tired, and every a hundred images there would be a 4% decrease in what the gaze was looking at. So automatically for one thing, it seems that one way to prevent error, at least an error of overlooking something, not even really seeing it, would be to try to make sure you're not fatigued. 

Bulat Ibragimov: Exactly. And this was one of the aims of this study. Can you predict if a user is fatigued just from image reading without knowing any prior information? For example, somebody can have a very difficult night shift. And we can predict this for individual users. From several readings we can see that it seems like the reader is fresh or average or more tired. But again, the next question is, what do we do now? We see that somebody is tired, but [00:12:00] can we now completely interfere with the process, saying, Okay let's replace the doctor. Let's find another one who's not tired. That's a challenging part. We can do it, we can find these moments, but the question is, How can we act upon this knowledge? 

Christine Ko: Yes. That's hard because like, even today, my son, who's 15 is, I am sick. So I was like, Okay, maybe you shouldn't go to school, if you're sick. And he's, No I'm okay to go to school. I can go to school. I'm not that sick. And yet I know he's tired. If he really had the ability to just wind the clock back 12 more hours, and just sleep for 12 hours and not miss school, he would probably do that. And I think probably for me too, on a given day, if I'm tired, but have a bunch of work to do. If I could like just take four hours and rest before I do my work, that would probably be ideal. But the reality of life is that we can't manipulate time in that way. 

Bulat Ibragimov: No, exactly. Exactly. So this fatigue [00:13:00] analysis was to see if it can predict it. If it's visible, and it is visible. 

Christine Ko: It is visible. 

Bulat Ibragimov: The action points are now to think about if something can be done. 

Christine Ko: If I'm honest, I know when I'm tired. I have sensed that when I'm more tired, it takes me longer actually to come to a decision, to come to a diagnosis on each slide. I don't know if you've found that in your studies?

Bulat Ibragimov: We surprisingly found that they made decisions faster. Faster maybe not in the way of time, but they needed less information to make a conclusion, like a satisfaction of search. You just see something. Okay. Okay. I'm already good. I see this abnormality. I'm good to go. Let's move on. 

Christine Ko: Yes. Like meaning that you question it less. You're more dogmatic. You're like, This is the answer and that's it. I don't need to question it. I don't need to think about what else could it be.

Bulat Ibragimov: And also when you find the first abnormality, you already say, okay, now everything's clear. I found the [00:14:00] problem. And we are good to go. 

Christine Ko: Yes. Okay. Yeah. 

Bulat Ibragimov: This manifests more when you're tired. Of course people suffer even when they're fresh. But when they're tired, they make decision much faster, sometimes in less than a second.

Christine Ko: You mentioned search satisficing, that's a cognitive bias. I have read that the more tired we are or, after having made a hundred decisions, then we do, we use heuristics. We use them because of decision fatigue.

Bulat Ibragimov: Absolutely.

Christine Ko: Based on your work, I think one way we can decrease decision errors is to try to make sure that we're fresh and not tired. But do you have other ideas of what we can do?

Bulat Ibragimov: From the gaze, we can predict so many things. There is so much information hidden in the gaze. It sounds a little bit like maybe futuristic or maybe too unreal. We're trying to use the gaze as a way to predict what the user thinks when he looks at the image, what he wanted to say.

Christine Ko: Do you think that your [00:15:00] work applies to decision making just in life? 

Bulat Ibragimov: That's a good question. We have been trying mathematical questions in some different ways to see if there is some kind of a good way so people can make less errors. We didn't yet find any kind of strong pattern saying, Okay here is how we can help, like, how we can reduce the errors by showing the objects in a more beautiful, more appealing or maybe some other more crisp way to the user. But I think eye movement analysis has a really great potential.

Christine Ko: You've collaborated with Claudia Mello-Thoms. She said that perception is subconscious. She said that when I'm looking at something, that's all subconscious, so then when I put words to what I've seen, I'm jumping up to a different level, to the conscious level because language is conscious, and so then I'm trying to explain what I've seen. If AI could predict at the stage before I jump [00:16:00] to my consciousness, or even at the same time, it might tell me something different than what I tell myself in my own words. I don't know if you agree or disagree with any of that, if you have thoughts.

Bulat Ibragimov: Oh, this is one of the aims which we investigate. We are trying to predict what the user thought or what the user wanted to say only from gaze. We're still looking how to investigate it better, how to see these patterns better. There is a difference in the way we look if we agree with what we see or if we disagree with what we see, and we're trying to capture this. Can we understand what he wanted to correct, or what he wanted to do, what he thinks in his head?

Christine Ko: What does that mean exactly? If we agree with what we see versus if we disagree with what we see, what does that mean? 

Bulat Ibragimov: For example, let's say we have an AI algorithm and we show some solution, some annotation, let's say in the pathology on the language of histopathology, let's say AI labeled them [00:17:00] this, and let's say this way is correct, and let's say this way is wrong. There is a mistake made. Can we see from the user's gaze if he agrees or if he doesn't agree? Because why is it important? Because if we think he doesn't agree, we can instantly propose some other solution saying, maybe this one you would like better, instead of asking him to manually go and correct things.

Christine Ko: Okay. Have your ideas on perception changed based on this work you've done with eye tracking? 

Bulat Ibragimov: One of the big things which I observe is that we can recognize the reader, as I said, we can recognize the reader of surprisingly high probability. This brings us to the one of our great areas of application, recognizing the level of expertise. Experts are very different from amateurs, from novices. AI analysis of gaze shows this very clearly. We can separate them very clearly. The area where it can be used practically is training. Radiologist training, sonography training, and so on. We can do it as a [00:18:00] part of evaluation. We can look at how we look at the images, and then using this information we can understand if they're ready, if they're ready to go further.

From gaze we can even understand what the people think. So I look at you, and I can see everything which goes around. But then I need 10 minutes to explain what I saw. If we can use this means of communication better there might be a great potential for integrating doctors and AI better. 

Christine Ko: Yes. Visually, everyone's face is a little different. But for me to describe exactly all the reasons, like, why your face is different from my face, is hard. It's much harder than I can just look and say, Look, they're different. But for me to explain like, how much longer is my nose than yours or wider, or like, how far apart are my eyes? These are the differences that we perceive fairly quickly but it's hard to really describe objectively.

Bulat Ibragimov: Yeah, exactly. 

Christine Ko: This has really been [00:19:00] fascinating. Do you have any final thoughts? 

Bulat Ibragimov: We have been investigating a number of areas where AI can be used to automate some work of doctors, of radiologists and so on. But the area of communication between the computer and the doctor, this area is less investigated in the field of radiology, and I think a lot can be done. This work has lot of potential in unraveling maybe how we think and how we look.

Christine Ko: How we look. I think that's cool. I think it'd be really cool if AI could just tell me everything that I just looked at. Very quickly. Before I've even been able to explain it myself. That would be amazing. Thank you so much for your time.

Bulat Ibragimov: It was really great pleasure for me talking to you.