Using Machine Learning to Predict Post-operative Complications of Total Ankle Arthroplasty

Research Type: Level 3 - Retrospective cohort study, Case-control study, Meta-analysis of Level 3 studies Introduction/Purpose: The risk profiles associated with complications following total ankle arthroplasty (TAA) are not fully understood. Previous efforts to assess complication risk have identified statistically significant risk factors, but small effect sizes limit these models’ clinical utility. Given its ability to process extensive and complex data, machine learning may be a clinically relevant predictive tool. We sought to evaluate the accuracy and effectiveness of four different models for predicting short term complications, extended length of stay, and mechanical failures. Methods: The National Readmissions Database (NRD) was queried for adult patients (≥ 18) who underwent TAA from 2015-2020. Primary outcomes were complications within 180 days, extended length of stay (LOS), and mechanical failure of hardware. For each outcome, four models were created (weighted logistic regression (LR), random forest classifier (RF), gradient boosting classifier (GBC), and an artificial neural network (ANN)) using Python v3.9. Model performance was assessed using accuracy and the area under the ROC curve (AUC). AUC was categorized into poor (AUC <.70), acceptable (0.70< AUC < 0.80), excellent (0.80< AUC < 0.90), and outstanding (0.90< AUC < 1.0), regarding the predictive capability of each model. Results: A total of 8,362 patients underwent TAA from 2015-2020. For predicting short term complications, random forest classification was marginally superior (RF- AUC: 0.58), though all models offered poor predictive capability. For predicting extended length of stay, weighted logistic regression proved most effective (LR-AUC: 0.66). All four models were relatively ineffective at predicting mechanical complications (AUC: 0.51-0.52). Conclusion: While the models created in this study offer relatively poor predictive capability, machine learning has the potential to accurately predict rare outcomes with heterogenous data; however, increasingly complex data requires much larger datasets. As national databases continue to expand, machine learning will become more accurate. In the meantime, simpler modeling techniques provide more accurate and interpretable predictions.