false
Catalog
AOSSM 2023 Annual Meeting Recordings no CME
Defining and Predicting the “Optimal Observed Outc ...
Defining and Predicting the “Optimal Observed Outcome” Following Surgical Treatment of Anterior Shoulder Instability: A Machine Learning Clustering Analysis of 200 patients with 11-year mean follow up.
Back to course
[Please upgrade your browser to play this video content]
Video Transcription
In the beginning, Lou, one of our residents, is gonna give the next talk, but I do owe him a huge debt of gratitude on this study as he was the real brain power behind the AI portion of this. So we'll talk a little bit about using machine learning and cluster analysis in shoulder instability. Disclosures are available on the website, none are really related to this talk. So here's sort of a quick overview. We'll talk about our motivation for doing the study, quick review of what cluster analysis is, give our aims, method, results, and conclusion. So when we talk about shoulder instability, I think we all know what we want for our patients after shoulder instability surgery. We all aim for them to have full range of motion, full strength, no pain, no apprehension or recurring instability, a full return to sports at the same level or maybe even a higher level, and normal shoulder biomechanics. However, some of these goals may be mutually exclusive of one another. For instance, if a patient has full range of motion, that probably increases their risk for having some apprehension. If they have a full return to sport, that may increase their risk for recurrent instability. So in some of these goals that we have, we may be trying to have our cake and eat it too, as many of these are probably mutually exclusive. So the question that we had with this is, is it appropriate for us to expect all of these perfect outcomes for all of our patients, or do these outcomes sort of compete with one another? It was really difficult for us to sort of sort through that and think about this on our own, so this is why we turned, I think this is another potential novel application for machine learning and artificial intelligence to sort of tell us what outcomes we actually should be looking for. So to do this, we utilized cluster analysis. So cluster analysis is an unsupervised machine learning tool that can identify hidden patterns in data in a less biased fashion than I would have on my own. It can then cluster the data into different groups based on these patterns that it finds. And then we can look at these groups to determine which of these are optimal or suboptimal based on what we want for our patients. So that's what we did in this study. Our specific aims were to determine what are the actual observed outcomes after anterior shoulder instability surgery. So not me saying, this is what I want to happen for my patient, this is specifically what I want to look for. No, just having the machine do the analysis and say, this is what you got. And then to sort of define what is the optimal observed outcome out of all those different clusters, see if we could pick out which one of those is the highest level of function or achieves the highest sort of optimal outcome. And determine what percentage of patients achieve that. And while we were at it, we were also curious to see is there a certain percentage of patients that can actually achieve a perfect outcome. So our methods, we used a retrospective cohort study out of our Rochester epidemiology project, included patients that had one or more anterior instability events, were less than 40 years old, treated surgically, had a minimum of two years of follow-up, excluded MDI and posterior instability patients. Our primary outcomes were to identify subgroups that had composite achievement of each of the following outcomes. So full motion within five degrees of normal of the contralateral side, no recurrent instability, no revision surgery, no pain, full return to sports, no OA, and no complications. And if they were able to achieve all of those, they were deemed to have a perfect outcome. So we identified a total of 228 patients that had all of those metrics available. The median follow-up was 11 years, so we had relatively long-term follow-up. No differences in gender and sports participation across our different groups. After putting the data through the unsupervised machine learning model, it was able to separate into multiple clusters. There are several ways you could look at this, but ultimately they really seem to distinguish themselves into two separate and distinct clusters. We looked at those and looked at the outcomes of each of those, and we then subsequently named one, the optimal outcome, and one, the suboptimal outcome, and we found that 64% of patients were in that optimal outcome cluster. 36 were in the suboptimal outcome. And interestingly, 41% of patients actually were able to achieve a perfect outcome on all of those outcome measures I just looked at. So here's a breakdown of our optimal outcome group and then the suboptimal group. So you see that first column, these are the patients that clustered together that had better performance than others in the suboptimal group. But I'll point out to you that if you'll just scan down that column, you'll see those are not zeros in that. So there's still 22% of patients that had some recurrent pain. 12% had recurrent instability. So this is not a perfect outcome, but these were the high achievers that patients tended to cluster around. So you can see that the optimal outcome group did outperform the suboptimal outcome group in all of those areas, but they weren't quite perfect. Patients who achieved the optimal observed outcome tended to be younger at the time of initial injury, presented sooner after initial instability. They went on to surgery sooner, and they had fewer instability events prior to surgery. So this kind of helps it pass as the common sense test. And this is backed up by all of our previous done research on this topic. So last slide here, conclusions. 64% of patients achieved the optimal observed outcome, which was defined as minimal post-operative pain, lower rates of recurrent instability or arthritis, and lower vision surgery rates and increased motion. Only 41% achieved a perfect outcome. I think this can be immensely helpful when we're trying to counsel our patients and set appropriate expectations for our patients, our physical therapists, and ourselves. And then the positive predictors for achieving that optimal observed outcome were shorter time of presentation, earlier surgery, and younger age. Thank you very much. Thank you.
Video Summary
In this video, the speaker discusses the use of machine learning and cluster analysis in shoulder instability. They raise the question of whether it is realistic to expect all patients to achieve perfect outcomes, as some goals may conflict with each other. To address this, they utilize cluster analysis, an unsupervised machine learning tool, to identify patterns in data and determine optimal and suboptimal outcome clusters. The study involves 228 patients who underwent anterior shoulder instability surgery, with a median follow-up of 11 years. The results show that 64% of patients were in the optimal outcome cluster, while 36% were in the suboptimal outcome cluster. Additionally, 41% of patients achieved a perfect outcome. Factors such as younger age, shorter time of presentation, and earlier surgery were associated with higher chances of achieving optimal outcomes. The findings can help guide patient counseling and expectations. No credits were mentioned.
Asset Caption
Christopher Camp, MD
Keywords
machine learning
cluster analysis
shoulder instability
optimal outcome
suboptimal outcome
×
Please select your language
1
English