Good evening, everyone. My name is Kendra Vasavada. I am a resident surgeon in general surgery at Yale New Haven Hospital, and a clinical researcher with NYU Langone Division of Sports Medicine. I'm here today to talk about a novel machine learning supermodel to predict revision ACLR reconstruction failure in the MARS cohort. Oh, there we go. We have a lot of authors through the MARS group. I want to thank all of them, and these are our conflicts of interest. So the MARS group and several others have implicated several factors in revision ACLR outcomes, ranging from graft type and age to preoperative knee hyperextension and posterior tubular slope. Although these findings have guided clinical decision-making, more accurate prediction of graft failure at a patient-specific level may improve preoperative counseling, operative management decisions, and cost of care. So machine learning is increasingly being utilized in the healthcare setting to harness insights from clinical datasets or big data to provide patient-specific predictions of an outcome of interest. A number of machine learning studies have built risk calculators with the potential for use in a variety of settings, including preop counseling to risk-based reimbursement. Yet there is a lack of standardization in machine learning methodology, limited awareness of ML techniques in the orthopedic community, and sometimes issues with data quality that limit the clinical utility of these results. So to that end, autoprognosis is a novel, validated ensemble algorithm that combines the strongest features of popular algorithms like Random Forest and XGBoost to create a single well-calibrated predictive supermodel. So AP simplifies ML methodology by eliminating the need for model selection and hyperparameter searches, reducing potential errors and variability. So the purpose of this study was to apply novel machine learning methodology to the MARS cohort data to determine an optimal predictive model of revision ACLR graph failure and the features that most strongly predict within the context of the model. We hypothesized that autoprognosis would have the most robust predictive and discriminative ability, and that important features identified would span radiographic, surgical, clinical, and patient-reported outcome variables. So as you all know, the MARS study cohort prospectively recruited between 2006 and 2011 and 1,233 revision ACLRs. Of those patients, we chose the patients who have completed their six-year postoperative follow-up and excluded those who were missing outcome variable data for primary outcome, which was graph failure after revision ACLR at the six-year post-op mark. So we used standard practices to encode variables by their data type and excluded categories or features that didn't meet quality thresholds. We handled missing values using imputation to finally have a dataset of pre-processed features that we then used to build five models, including the autoprognosis model. And we used standard performance metrics to evaluate the performance of these models, including the two here that are most important are area under the ROC curve and area under the precision recall curve. That metric is particularly good for imbalanced datasets like the MARS dataset, where, you know, I'll allude to this later, but we have a smaller number of patients who had the primary outcome of interest. So our cohort included 960 patients who completed a six-year follow-up with 5.7% or 55 experiencing graph failure. 55.2% were male, median age of 26, BMI of 25, 53.4 had non-traumatic mechanisms of injury with gradual onset, and about 88% of patients were undergoing their first revision surgery, a median of 3.7 years from their previous ACLR. AutoGraft was used in 49.4% of patients, followed by Allograft in 47.5, and combination graft at 3% for current revision surgery. So while all models had moderate to good discriminative power, autoprognosis demonstrated the highest discriminative power with well-calibrated scores. The most important features for a model predictive and discriminative ability across the various ranking methods and performance metrics that we used were the following. Prior ACLR femoral tunnel placement to position and size, prior ACLR tibial tunnel placement and size, and graft type in the current revision ACLR. So in summary, autoprognosis best predicts the revision ACLR graft failure with moderate discriminative power, and these were the top three features identified within the context of the model to have important contributions to predictive ability. This is the first study to do a machine learning analysis of the MARS cohort in order to create a patient-specific risk prediction model for revision ACLR outcomes. This work presents an opportunity to transform revision ACLR risk stratification from a subjective consideration based on surgeon knowledge and experience to an evidence-based quantitative measure on top of surgeon knowledge and experience. This work has the potential to guide preoperative counseling, shared decision-making, intentional risk mitigation, and fair risk-adjusted reimbursement. But we have a few more steps before we get there. The first of those is to study other outcomes that determine a successful ACL reconstruction or revision surgery. The next is validation studies with data sets outside of the MARS database, whether that's collaborating with a patient or whether that's collaborating with our partners in Norway, with the French folks, and with other data sets to validate these studies. And the last step is to build a bedside risk calculator that clinicians can then use that build on this work. So with that, I would like to thank the audience for their attention and I also would like to thank the sponsors who've made this possible. Dr. Wright, Laura, Amanda, Dr. Gisraoui, thank you for letting me come into your office with this study idea on a Post-it note. Dr. Haim, Dr. Devana, thank you for pushing me to build computational partners and have those folks on board. And Dr. Lee and Brenda, who are our computational minds behind this work and without them this project wouldn't have been possible. Thank you all. Thank you.

machine learning

predictive model

ACLR reconstruction

predictive ability

risk calculator