Good afternoon, everyone. My name is Thomas. I'm a third-year medical student at Stanford. Topic today is which osteochondritis-sussekin subjects will heal with non-operative management and application of machine learning methods to the ROC cohort. Before I get started, I just want to thank ASSM, all my co-authors and mentors, and the ROC group, most importantly, for helping me get up here. So in terms of our disclosures, they're up to date on the ASSM website. This is just an overview of ROC. It's over 26 institutions, nationally and internationally, that study OCD of the knee. Just some background for our talk. The definition for OCD is thrown up on the slides. But the motivation for this study is that there are really limited evidence-based guidelines to predict which OCD lesions will heal with non-operative treatment. The most cited paper is by Wahl. They looked at 47 lesions, and they assessed progressive lesion reossification with serial radiographs every six weeks, up to six months. We wanted to try to replicate this but change the outcome, so we designed and trained a supervised algorithm to determine whether a patient with OCD of the knee would heal with non-operative treatment at the time of the baseline encounter. Now, to do this, we first identified patients in the ROC cohort that underwent non-operative treatment, and we applied some inclusion-exclusion criteria. We chose variables that were useful in predicting non-operative healing, and then we trained a set of, like a suite of classification algorithms and saw which ones best predicted the lesions healing with non-operative treatment. So specifically, we used something called the Baroud algorithm, which is a random forest-based algorithm for variable selection, and then validated that with univariate analysis. The predictors that we included in our final models were normalized lesion width and length, as well as lesion location on coronal and sagittal MRI. By contrast, the variables in the Wahl et al. model were age, pain versus mechanical symptoms, and also lesion size. And the study outcome in this study was successor failure of non-operative management, as assessed by full return to sport, as well as complete radiographic healing. So these are just some of our results in terms of our cohort characteristics. You can see their mean age, sex, lesion presence in coronal and sagittal planes, as well as lesion size. But most importantly, we studied 64 subjects, 40 in the failure group and 24 in the success group. We first trained logistic regressions, and as you can see, we applied the Wahl et al. model to our study data as an external validation step. And what you can find is that, this is kind of the inferential statistics portion that was discussed earlier, none of the variables actually used in the Wahl et al. model were independent predictors, independent significant predictors of our study outcome, which validated our decision to use other variables. However, it should be noted that their normalized lesion width variable did approach significance. By contrast in our study, and we can show the actual nitty-gritty results of it later, but normalized lesion length and width do matter. Bigger lesions are less likely to heal, and smaller lesions are more likely to heal. And additionally, where the lesion is with respect to the weight-bearing zones in the knee also matters. Less weight-bearing is good, more weight-bearing is bad. This is a receiver operating characteristic curve, which shows sensitivity on the y-axis and false positive rate on the x-axis across a variety of potential classification thresholds. If a curve is up and to the left, it's better, it has better sensitivity and specificity across multiple thresholds. And if it's towards this gray diagonal line, it represents a model that is basically guessing. So the Wahl et al. model had an AUC of 0.6 when applied to our data for a raw accuracy of 58%. This means that 58% of the time it could classify correctly whether a patient would heal with non-operative management or not. So it's close to the gray line, close to guessing. And we trained, the logistic regression that we trained was a bit better, had an AUC rock of 0.86 and accuracy of about 80%, so better. And then we made some, across the variety of models that we trained, a support vector machine with a radial kernel actually had the best AUC rock of 0.89, and a linear kernel had a raw accuracy of 84%, so 84% of the time you can predict which patient is going to heal with non-operative management. So in conclusion, machine learning models can predict which OCD lesions will heal with non-operative management. We identified some new predictors inferentially of non-operative treatment success such as lesion location in non-weight-bearing zones. And the models developed in this report represent a clinically meaningful improvement on prior efforts reported in the literature for this outcome. For example, our classification accuracy was 83.9% as compared to 58.2% achieved by various previous efforts. Our next steps will be to incorporate more data about patients with lesions at the trochlea and lateral femoral continental, as well as externally validating the model that we developed in this study and adding more data. There's more data to come from the rock cohort. And also to apply the same modeling techniques to OCD of the elbow with the rocket group led by Carl Nissen and Don Bey. I just want to thank the whole rock and rocket team for everything and helping me get up here. All right. Thank you.

osteochondritis dissecans

non-operative management

machine learning algorithm

ROC cohort

lesion healing