C. Michael Gibson, M.S., M.D.’s Post

Transcript

Wrote on Kira Mike Gibson coming to you live from H 2024. We're talking about the Pan Echo study, AI, machine learning, deep learning, really helping facilitate the interpretation of imaging. Talk to us a little bit about what you did. Thanks, Mike, for having me here. It's a great honor for our group to be presenting this study at AHA. You know, there has been a, you know, echocardiography is the cornerstone of cardiovascular diagnostics and there has been a lot of interest in using it effectively. But the workflow for AI, for echocardiography requires, you know, expert acquisition, expert interpretation, lots of clicks. Yeah. Okay, expert acquisition and expert interpretation and increasingly, even though we have access to cheaper handheld echocardiographic devices that that presents a bottleneck for us to actually get the diagnostics at scale. So our our and people have suggested AI as a strategy to interpret echocardiographic videos, but it's always been done for single task detect the F detectorist Nosis including work from our group. It doesn't really translate to an automated solution that people we can deploy in lower. Resource settings or even in hospitals in here where it can be triaging and treating a full report and we can just overhead like we would do for a trainee. And so this is Span Echo where we are actually built. We built a lightweight model that is deployable world over and we extract 39 diagnostic labels that are you know routinely captured in echocardiography, taking a full echocardiographic video, all views, all dopplers and interpreting them using deep learning to generate these results so. And we're excited with the great performance we get both in our data and also external data where we tested the model. Can I ask you something? Was this Rohan? Was this really true deep learning? A lot of algorithms out there use automated edge detection. They're adapting technologies from the 90s. Is this true Black box AI? We used an image encoder using convolutional neural network to turn the video frames. And then we actually use the transformer model, which is a new kind of model that ChatGPT uses, and then use that model to stitch the frames together. And learn from each frame to frame all broken down and then learn on top of that all these separate asks and applications. None of this was done with ejection. So and the point Mike is really in real world settings, we don't have perfect views, perfect videos. And the ability to do this with deep learning is that we will adapt to those variations that exist. Well, that's why I pointed it out because there's a lot of companies out there using edge detection, but not really AI. This is true AI. Yes, yeah. And and again, it demonstrates its performance by being applied to, you know, lower quality acquisitions to acquisitions and public resources repositories and actually beating benchmarks whatever they are publicly reported. And what did you tell us what you found? So we found that the model is, you know, you know, with deep learning models, all we have are performance metrics. So we were able to use a full like, you know, all studies including passing long access, short access to study is taken and this whole. The user classify and we can for 21 estimation tasks where like LVOT diameter, EFF N diastolic diameter, we have mean absolute errors which are extremely small compared with the measurement errors that are present in these measures themselves. And we we have we do this for 21 tasks and where the median and mean absolute are normalized mean apparel only .134 tasks that are classification based where we actually are using information from these echocardiograms for detecting a positive or a negative signature. Our model had a performance a median ARC of .91 across a range of estimation and classification tasks. Amazing. So and then in terms of practicalities, does this help you save time and throughput? I think it it, it, it just automates the entire process is under under a couple of minutes. You have a full estimation of an echocardiographic, single echocardiographic study. Everything is a tabular data element. So it's not just a report that somebody has to manually parse through. You can actually you get a tabular output for every single element and somebody can manually cross check each update these numbers. From a workflow perspective, this is really well aligned with how we read echo studies. We see the images, we updated table element and that turns it into a report and that we're able to do fully. End to end. So from an efficiency perspective, it helps to triage which echoes would be over at first. It helps to really do the normal echoes really quickly and somebody could just read them later and get them through the system. And I think we have shown that even if you had like 5 views like you know, personal long and the ethical views, you are still able to get very good estimation suggesting without doppers able to translate it to handheld echocardiography where very few images are captured. So Justin, we can do a lot more. So it'd be great, you know, if a provisional report was put right into the chart after this street from the straight from the echo machine and then it gets an overrated, you know, that's kind of what we see more and more of. And do you know how this would compare if you compared it to or have you compared it to those algorithms that use edge detection, automated edge detection? So some of the models are in public domain and we have tried to compare them. We are models able to for those that. UEFA assessment using Azure action were able to beat the performance of those models on detecting those elements. Also for LB other LV diameter measurements, we are able to beat them successfully. Again really just showing that our models are are able to capture whatever these instruction models do and do it better. Right. Unbelievable. I knew this is where we would get I was always a little skeptical of all the people claiming they were using AI when they were really just using repurposed edge detection from the 90s. This is really an advance. Thanks for sharing with us today. I'd like to get back together after H maybe do a full segment where you can demonstrate this to us, you know, some videos and things. Would that be? Thanks for looking forward to, thanks for joining us, and thanks to all of you for joining us here live from H 2024.