This Florida Orthopedic Society 2025 presentation focuses on a pilot study exploring predictive and causal models to analyze the lateral approach to total knee arthroplasty (TKA). This approach is noted for high patient satisfaction but is technically demanding, particularly regarding exposure of the posteromedial tibial corner. A notable, though uncommon (less than 2%), complication is an MCL (medial collateral ligament) tear, which can significantly impact surgical workflow, increase costs by necessitating a stabilized implant conversion, and affect surgeon adoption. The goal of the study is to use observational data to identify predictive and treatment signals that might reduce the risk of MCL injury and enhance procedural safety. The study collected data from 24 consecutive cases out of approximately 1200 lateral approach TKAs performed at a surgery center in Orlando Florida. The dataset was small and highly imbalanced, with the adverse outcome (MCL tear) occurring in about 2% of cases. Despite the small dataset, the objective was to build a hypothesis-generating framework rather than making definitive claims. To uncover predictive signals, an optimized Random Forest classifier was trained. While the overall performance was modest (an F1 score of just 0.3), the model was useful as a flagging tool for feature importance. To pinpoint high-risk patients, the study also used logistic odds ratio. This analysis identified female patients age >74 with valgus alignment and BMI above 35 as the highest-risk group. These findings, combined with SHAP (Shapley Additive Explanations) values which represent feature importance, were intended to give clinicians clear targets for risk mitigation. To facilitate the interpretation of the model, we applied SHAP for both feature importance and interactions. Furthermore, an interactive dashboard was built in Google Colab. This dashboard allows clinicians to move sliders representing important predictive features (age, sex, BMI, knee alignment, surgeon subgroups) to simulate patient profiles and observe real-time risk predictions. This interactive UI is more intuitive for orthopedic clinicians than standard SHAP summary plots or force plots, bridging the gap between clinicians and data science. It uses SHAP values to extract importance and simulate effects with sliders, allowing the recreation of 'high-risk' profiles and observing their risk behavior under different model assumptions, providing a reproducible interface for understanding patient vulnerability. Layered on this interactive UI is a causal framework (S-Learner). This framework aims to estimate how surgeon subgroup, treated as a 'treatment', would change the risk of MCL tear (the treatment effect) while adjusting for patient characteristics like age, gender, and pre-alignment. This allows exploration of "what if" scenarios (counterfactuals) and helps in benchmarking different surgeon subgroups. Initial analysis stratified MCL tear rates by surgeon, showing surgeon subgroups had tear rates between 1.4% and 3%. Specifically, for high-risk patients with valgus knee alignment, Surgeon 1 had the lowest tear rate. This simple stratification suggested that Surgeon 1 excels, even in tough cases. To investigate this further, a causal analysis was run, treating "surgeon" as the treatment intervention after creating a dynamic Directed Acyclic Graph. Key confounders were controlled for, as they affect both which surgeon a patient sees and their tear risk. The Treatment Effect was estimated by repeatedly resampling the full 1,200-patient cohort within each surgeon and outcome stratum, computing the tear-rate difference, and building a bias-corrected confidence interval. The result of this analysis indicated that Surgeon 1 lowered MCL-tear risk by about 2 percentage points versus Surgeon 0, with a confidence interval entirely below zero, suggesting the result is both statistically significant and clinically meaningful. The practical takeaway is that if the baseline MCL tear rate is around 2%, a treatment effect of -2 pp means you might expect Surgeon 1's patients to have essentially negligible risk on average. However, it's noted that these estimates depend on the sample of 24 events, and uncertainty remains even with bias-corrected bootstrap, so this is strong but not definitive evidence. The study concludes that Surgeon 1’s low tear rate, especially in high-risk cases, suggests their technique is a benchmark for safer surgery. The recommendation is to adopt the benchmarked surgeon's technique when presented with high-risk patients (females with severe valgus and high BMI) and train others on their methods to help scale the lateral approach safely. Anecdotally, a novel MCL retractor appears promising, but a dedicated study is needed to measure its real-world benefit. Future research should include collecting prospective data on retractor usage and outcomes and testing whether combining optimized technique with custom MCL retractors can drive tear rates further toward zero. In summary, the study utilized a Random Forest risk model with SHAP and logistic odds ratios and an interactive dashboard to pinpoint highest-risk patients. A stratified causal analysis then revealed that Surgeon 1’s technique can lower tear rates by roughly 2 percentage points, essentially moving risk towards zero, though this estimate comes with uncertainty due to the sample size. Although a 2 percent reduction appear small, this is a relative risk greater than 80% when the baseline rate of MCL tear is < 2%. By adopting and disseminating Surgeon 1’s best practices, using a well-defined protocol, and an MCL retractor, the aim is to develop policy and make the lateral-approach TKA safer and more widely achievable. The study found that the high-risk patient profile identified by the predictive model is the same group that showed a significant causal treatment effect. This profile should be flagged for extra vigilance, but even in the best hands, higher-risk patients remain vulnerable and share outcome risk.
