CORR Insights®: Machine-learning Models Predict 30-day Mortality, Cardiovascular Complications, and Respiratory Complications After Aseptic Revision Total Joint Arthroplasty

Where Are We Now? Aseptic revision arthroplasties carry more postoperative morbidity and mortality than primary joint arthroplasties do, particularly in the short term, and recent studies have shown a high degree of dissatisfaction and functional limitations among patients undergoing revision arthroplasty [2, 8]. Risk stratification is important before revision arthroplasty, even more so than in primary joint arthroplasty. Machine-learning and artificial intelligence (AI) programs have emerged in the past decade; AI tools use mathematical predictive models that run substantial amounts of data through defined algorithms. When these tools are designed to modify predictions in light of processed data, what results is called machine learning. Proofs of concept have included a model that predicted the risk of and time to TKA [5] and models that determine the risk of 30-day complications and mortality after primary THA and TKA using preoperative clinical and biochemical parameters [3, 4, 7], among others. The current study in Clinical Orthopaedics and Related Research® by Abraham et al. [1] expands on this concept by predicting 30-day postoperative morbidity and mortality in patients undergoing aseptic revision THA and TKA. Previous studies used machine learning to assess factors that predict 30-day mortality and morbidity after primary THA and TKA based on medical comorbidities and laboratory parameters. The present study is one of the first I know of to use this model in revision joint arthroplasty. The open-source XGBoost tool used in the present study was temporally validated and will be helpful for surgeons to preoperatively plan and stratify risk for patients undergoing aseptic revision arthroplasty of the knee and hip. The present study demonstrates the utility of AI-integrated machine learning and sets a precedent for its use in joint arthroplasty specifically and in major surgery more broadly. And, most importantly, this study—which provides a freely available online risk calculator that allows users to input patient data and easily calculate the postoperative risk of 30-day mortality and cardiac and respiratory complications after aseptic revision TKA or THA (http://nb-group.org/rev2/)—will help surgeons educate patients about their specific risk of adverse outcomes and guide appropriate preoperative medical management. Where Do We Need To Go? The use of AI and machine learning in surgical risk stratification seems like the next logical step in applying this technology in the field of surgery. Surgical risk stratification depends on various demographic factors including age, BMI, pre-existing medical comorbidities, and biochemical markers of the patient’s physiologic state [6]. Hence, intrinsically, risk stratification is a function of multiple variables, some of which are dynamic, thus making this type of multivariable analysis especially suitable for computation using AI-integrated machine-learning programs. The use of arthroplasty registries has become common in orthopaedic surgery in many countries. These are rich repositories of patient demographic data. However, medical comorbidity quantifiers, the most widely used of which is the American Society of Anesthesiologists physical score, have not been routinely recorded in many of these major registries. Indeed, the registries have only started to include these data recently. The maintenance of a database that integrates variables that have implications on patient risk and outcomes will be vital going forward in terms of formulating accurate risk stratification and predictive models. In the present study [1], the XGBoost tool was used to create a scoring tool for 30-day adverse outcomes. This tool is freely available and very accessible. Computation of multiple variables and the ability to discriminate between patients with the outcome of interest and those without it is a marked strength of this tool. If the current pattern of technological improvements is any indication, tools similar to the XGBoost will only continue to get better through more iterations. This should allow clinicians and researchers alike access to predictive models of risk stratification that incorporate a greater number of variables into a more-nuanced analysis. In the present study [1], using the XGBoost tool, the training dataset used data from 2014 to 2018, and the validation dataset used data from 2019. A post hoc analysis showed that the use of 2020 data did not improve the calibration of the 2019 validation dataset. By design, machine-learning programs tend to become more predictive with a greater amount of data available for computation. The inclusion of data from more years should improve calibration and the predictive model overall. How Do We Get There? I suggest a two-pronged approach. First, the model in this study that used XGBoost, as well as other similar tools using AI and machine-learning algorithms, needs to be more widely applied to ensure they generalize well across diverse study populations. Applying these tools to multiple arthroplasty registries will corroborate the utility of these machine-learning algorithms across populations. In countries where arthroplasty registries are unavailable, the repeatability of the present study’s results should be externally validated using large-scale hospital-based studies, which can be done retrospectively using available data. Of course, these studies need to carefully consider confounding factors. I suggest regression modeling for this purpose because this will allow us to mitigate confounding and identify the most relevant variables for clinical prediction. However, because not all registries collect the same data, the availability of all data needed for such an analysis poses a potential problem. This brings me to the second prong of the approach: By conducting studies and identifying the variables that are most predictive in risk stratification, patient-related variables can be identified. This can then provide a platform on which useful recommendations can be made regarding patient variables that will be included in the joint registry databases as standard practice. By integrating relevant predictive information in the registries, risk stratification using machine-learning algorithms can be more universally and uniformly applied. I think a happy middle ground will be to suggest that all registries collect the same variables to ensure uniform reporting among studies.