Machine learning in knee arthroplasty: specific data are key—a systematic review

Artificial intelligence (AI) has already led to tremendous advancements in the aviation and automobile industries and is now rapidly gaining importance in healthcare [1, 3, 18]. Machine learning (ML) represents a distinct application of AI which emerged from studies on pattern recognition and learning theory. ML describes algorithms for automatic and incremental function optimization which can be used to make predictions by detecting non-linear relationships in large data sets [1]. In a nutshell, ML algorithms are able to recognize correlations in data sets which subsequently permits predictions based on these “learned” patterns. The aspect of “learning” is achieved by automated weighting of distinct parameters in the mathematical model [1, 19]. Ever more data are being compiled digitally in healthcare while processing capacities are increasing rapidly so that various ML algorithms have already been tested in ophthalmology, dermatology, radiology, and cardiology [3, 5, 6, 18]. Especially in medical image analysis, ML algorithms were developed and validated which are capable to outperform medical specialists [17]. In orthopaedics, the number of published studies on machine learning has increased tenfold since 2010 [1]. ML has the potential to become a powerful tool for patient-specific decision making regarding surgical indications and predicting outcomes in orthopaedics. Although the application of ML is in a preliminary phase in orthopaedics, several studies report well-performing ML models for gait classification, fracture detection, exoprosthesis control, osteoarthritis detection, spine pathology detection and bone age assessment [19]. Extensive data are necessary for well-performing ML models. In this context, knee arthroplasty is highly suitable for ML analysis as multiple sources are available: national and international registries including follow-up data, patient characteristics, and patient-reported outcome measures after knee surgery from the last 40 years. At the same time, recently introduced enabling technologies such as navigated and robotic surgery offer novel, patient-specific data sets, and a vast amount of imaging data are available in digital form. Furthermore, various other clinical data may be suitable for ML models. Recently, an ML approach using registry and healthcare data sets to provide an adjusted payment model was demonstrated [23]. Overall, however, literature on ML in knee arthroplasty is heterogeneous and scarce.

ML comprises different theoretical approaches and represents a tool that must first be adapted for specific tasks in a given field such as knee arthroplasty. However, no standard approach for implementing ML algorithms in orthopaedics has been described nor established yet so that the question arises to which extent knee arthroplasty currently benefits from ML and which approaches might be promising. We hypothesized that successful ML algorithms in knee arthroplasty depend on arthroplasty-specific data and their reasonable use. Therefore, the collaboration between a data scientist and an orthopaedic surgeon is of utmost importance. Hence, we aim (1) to investigate which predictions are already feasible using ML models in knee arthroplasty and (2) to describe prerequisites for the effective use of this novel approach. For this reason, we conducted a systematic review of ML algorithms for outcome prediction in knee arthroplasty.

A systematic review of the literature was performed to identify machine learning algorithms in knee arthroplasty on the basis of the PRISMA statement. Studies meeting the following criteria were included in this review:

In March 2021, a literature search through PubMed, Medline database and Cochrane Library was conducted. For the systematic search, the following search terms were used: “total knee arthroplasty” AND (artificial intelligence OR machine learning), “TKA” AND (artificial intelligence OR machine learning) from 1990 to 2021. The studies were screened and evaluated by an orthopaedic surgeon (I.L.) and a data scientist (F.H.) at our institution using the aforementioned selection criteria. The results were summarized and duplicates were discarded. The selection procedure is presented in Fig. 1. All articles were initially screened for relevance by title and abstract to assess the inclusion criteria. The two authors independently performed a careful reading of the studies and extracted the data. The following information was extracted from each article: author, year of publication, study design, follow-up, number of patients/cases, ML algorithm, metric, data screening, fine-tuning, mathematical and medical interpretation. For quality assessment, the Coleman Methodology criteria, which assess methodology using ten different aspects, were modified for the systematic review of machine learning algorithms in knee arthroplasty (Table 1). A score of 100 indicates that the study largely avoids chance, bias and other confounding factors. Hence, the modified Coleman Methodology score (mCMS) can be defined as excellent (85–100 points), good (70–84 points), fair (50–69 points), and poor (< 50 points). Both the orthopaedic surgeon and the data scientist scored the included studies giving a total mCMS between 0 and 100. In the event of divergent results, the data were double-checked and a consensus was reached. The level of evidence is level III.

The most important finding of the present review is that outcome prediction using ML models applying arthroplasty-specific data has already been successfully performed. Although ML applications for knee arthroplasty only gained particular popularity in the past years with all articles included in this review being from 2018 or more recent, several studies already demonstrated the high potential of ML. Exact study questions with well-described complications like renal failure or postoperative blood transfusions can already be particularly well-predicted [9, 14]. However, more complex and general predictions like revisions, which can have a variety of different causes, are more difficult to estimate [2]. Interestingly, registry data on knee arthroplasty were only scantly evaluated for ML analysis. In data science theory, the sheer quantity of patients and input parameters is of crucial importance. Registries include vast information with patient-related data and potentially specific patterns that are suitable for ML analyses. However, most studies (13/19) relied on local databases and only six studies applied registry data—including only two different medical registries. Multiple national registries on knee arthroplasty are being administered that contain relevant information from the last decades on this topic. Future studies for ML in knee arthroplasty should definitely investigate the amounts of data within knee arthroplasty registries.

A further finding of this review is that we could not demonstrate a correlation between the amount of input parameters or patients and the outcome metrics. From a data science perspective, the tabular form of clinical information is not well suited for ML analysis as it requires a tremendous amount of patients and various parameters to generate complex data sets. Six studies included more than 100,000 patients; however, ML is proficient to process substantially larger amounts of data. Image data contain a higher information density and is, therefore, more commonly applied to ML applications. However, the low data volume is only in part an explanation for poor outcomes as several studies yielded very accurate predictions, especially in confined, local data sets.

While the quantity of data remains critical to allow for repeating patterns, the data quality is likewise of high importance for ML algorithms: ML can only reveal patterns that can actually be derived from the data. While Karnuta et al. fed their algorithm with tabular data from 295,605 total knee arthroplasty and total hip arthroplasty patients, Rexwinkle et al. retrieved tissue samples from only 6 osteoarthritis patients in order to apply biomarker analysis as well as biomechanical testing [10, 24]. The latter study was based on much fewer cases, still the complexity of the study and specificity of the data allowed for ML application with promising results. We therefore assume that the inclusion of specific parameters is of paramount importance, especially if tabular data are applied for ML analysis.

Just as specific data are relevant, a multitude of irrelevant parameters may negatively impact the performance of ML models. Although an increased data width with the use of all accessible parameters may be beneficial in theory, such randomness may rather hinder the full potential of the ML algorithm within the limited scope of tabular data used so far for ML in knee arthroplasty. Especially in studies with a low proportion of subject-specific parameters, the input of parameters, which were chosen simply because they were available, may be overrated and confounding. Hence, the selection of data for ML applications is anything but trivial. In this review, only four studies included subject-specific parameters such as the administration of tranexamic acid for the prediction of postoperative blood transfusion after TKA. Therefore, further studies to investigate arthroplasty-specific parameters are encouraged.

Identifying and applying such specific parameters in knee arthroplasty is difficult, though. Currently, the parameter selection of ML approaches in orthopaedics is based on (1) known risk factors and clinical experience and (2)—to a significant part—on the accessibility of parameters in existing databases. Therefore, most studies apply administrative data including basic patient demographics. However, the data architectures of local healthcare management systems and registries were not constructed for ML evaluations. These data sources do not necessarily contain fully detailed and prospectively linked patient data, which, however, would be essential for predictions. In addition, most studies were retrospective so that the data architecture for the ML analysis had to be constructed post hoc. Hence, data screening and processing are inevitable if existing data sources should be utilized.

This retrospective approach requires the substantive discussion of the parameters to assure inclusion of relevant and exclusion misleading information, as well as the mere technical evaluation of the data with regard to incorrect inputs and missing data points. Further steps for enhancing the data set, such as data cleaning or completing a data set are recommended from a data science perspective. This data processing requires extensive human assessment and the coherent thinking of a medical professional and data scientist to identify relevant and applicable information. It has to be performed mostly by hand, which makes it time-consuming and futile. The work related to data screening and processing is not to be neglected: it dramatically slows down the potential progress that can be made using ML in knee arthroplasty. To avoid such an extensive data processing, we assume that prospective databases with an appropriate data architecture for ML analysis have to be established in the long term. Currently, we are unaware of a comprehensive and completed data integration of knee joint-specific data for AI applications. Besides the already discussed importance of the quality and amount of input data, the ML model itself has a significant impact on the outcome prediction. ML does not describe a single specific algorithm, but rather contains a variety of approaches which have to be modified to the addressed issue and data set. A wide range of algorithms was used in the studies discussed in this review. The most popular algorithm was an RF, which is closely related to decision trees and XGBoost. This may be attributed to the considerable easy implementation, application and stable results over different kinds of data sets. However, more sophisticated algorithms have been developed exploiting the potential of today’s data as well as hardware even more. With a considerable amount of data, even deep learning algorithms can be feasible. Ramkumar et al. demonstrated how to use an ANN for, e.g. length of stay prediction of 175,042 TKA cases using 15 preoperative variables [22]. To achieve the best results, the focus of most studies relied on testing various machine learning algorithms and picking the one with the highest metric value as final algorithm. However, an exclusive measurement of performance using only one metric can be misleading so that several metrics depending on the task at hand need to be examined. Metrics which weigh false-positive and false-negative predictions, such as sensitivity, specificity or Dice Score, are often more meaningful. Furthermore, prediction models based on ML algorithms are not explanatory models [7]. Therefore, the outcome metrics need to be elaborated to understand its use in the medical context. In this review, 16 of 19 studies did not extensively discuss the results in the mathematical or in the medical setting. For future investigations, we consider the cooperative analysis by a medical professional and a data scientist of ML models in knee arthroplasty as mandatory.

Only few studies managed to increase the performance of their ML models through regularization or random search of hyperparameters in order to obtain more meaningful results [2]. For deep learning algorithms, multiple techniques, i.e., regularization and data augmentation, loss weighting, hyperparameter search or data simulation, are commonly applied to boost performance. In machine learning, these techniques are used not as frequently while still being possible most of the time. A hyperparameter search for instance contains significant potential and is easily applicable with today’s hardware capacity. Another underrated example for performance increase are methods to tackle class imbalance such as loss weighting. While a hyperparameter search was applied once [2], none of the reviewed studies indicated the use of any kind of loss weighting. From data science perspective, numerous modifications of the ML models in knee arthroplasty have not been implemented and tested yet. Especially for limited data sets, fine-tuning of ML models is crucial to avoid overfitting. Figure 2 depicts an overview of the development of an ML model in knee joint surgery based on this discussion. This review has several limitations. All studies on ML in knee arthroplasty were not presumed to be included. Numerous studies from the field of medical informatics examined big data that included data sets related to knee arthroplasty. However, we evaluated specific studies which focused on outcome prediction in knee arthroplasty from the medical perspective. This may have disregarded other studies using promising ML approaches in orthopaedics with other purposes. Especially imaging data are suitable for ML models, but those are mostly utilized in ML models for diagnostic applications. We deliberately did not evaluate ML algorithms using imaging data for that reason.

In conclusion, the prediction of distinct outcomes with ML models applying specific data is already feasible in knee arthroplasty. However, the prediction of more complex and general outcomes is still inaccurate. Currently, the benefits for the clinical application may be small, but—from a data science perspective—the possibilities are not yet being fully utilized. To date, specific parameters of knee arthroplasty have not been adequately evaluated and large data volumes gathered in registries have not been fully considered nor analysed. The inclusion of specific input data as well as the collaborative development, modification and evaluation of ML models and its data by an orthopaedic surgeon and data scientist are essential prerequisites to fully utilize the capacity of ML in knee arthroplasty. Predictions suitable for clinical application must be built on solid data structures. We consider that a data architecture specifically for ML with prospective data is necessary to allow for more accurate and complex predictions. Future studies should, therefore, necessarily examine large-scale data with specific and longitudinally recorded parameters.