Predicting anterior cruciate ligament: Reconstruction outcome machine learning analysis of National Knee Ligament Registries

Abstract

Introduction: Anterior cruciate ligament (ACL) injuries are common, and surgery is often performed to improve function. Many factors have been identified that may influence the risk of a poor outcome following ACL reconstruction (ACLR). However, putting those risk factors into context and applying them to an individual patient to accurately estimate their specific risk of a poor outcome is challenging. The ability to accurately quantify risk at a patient-specific level is desirable as it can lead to more informed discussions and surgical decision-making, and can guide efforts at decreasing risk.

Machine learning is a branch of artificial intelligence that enables the development of algorithms capable of predicting clinical outcomes based on analysis of large databases. These novel techniques can tease out relationships between variables that may be more complex than can be realized through traditional statistical analyses. The purpose of this thesis was to apply machine learning analysis to the Norwegian Knee Ligament Register (NKLR) and Danish Knee Reconstruction Registry (DKRR) to develop easy-to-use models capable of predicting postoperative outcomes (revision surgery and inferior patient reported outcome) for patients undergoing ACLR and identify the factors that are most important for making the outcome predictions. The hypothesis was that this analysis would lead to the development of accurate and externally valid clinical prediction tools that clinicians could use to predict the risk of revision surgery or inferior patient reported outcome for their patients undergoing ACLR.

[...]

Conclusion: The most significant findings from these studies are: 1) machine learning analysis of the NKLR and DKRR enabled the development and validation of prediction models that demonstrated moderate accuracy for predicting revision surgery and inferior outcome following ACLR and identified the most important factors used to predict these outcomes, 2) a rigorous approach to clinical prediction modeling has been described, laying the foundation for future innovation, 3) more work is needed to evaluate the performance of the prediction models on patients from outside Scandinavia and to determine the threshold for clinical relevance regarding ACLR outcome prediction, 4) the development and validation of clinical prediction tools may be limited by both the quality and quantity of the available data, and 5) the data collected by the registries should be expanded to include more variables that have been associated with outcome. Although these studies enabled the development of several risk estimation tools for patients undergoing ACLR, the performance of these models was limited by the data contained within the registries. More specifically, they were limited by the lack of some important relevant variables associated with outcome such as pre-operative knee laxity, posterior tibial slope, and rehabilitation factors. The choice of outcomes (revision surgery and low KOOS scores) may have also limited the model performance. In addition, external validation outside of Scandinavia was limited by poor data quantity in the STABILITY I cohort. Evolution of the national knee ligament registries to capture more variables is required to improve the ability to predict outcome using these databases. Overall, the processes outlined in these studies can serve as a guide for the pursuit of clinical prediction models in the future; however, the current clinical utility of the ACLR prediction models remains unknown. Prior to widespread adoption and implementation of these prediction algorithms, their performance relative to predictions made by surgeons must be ascertained. This represents an important next step because until it is known how well surgeons can predict outcome, it will never be known if prediction tools driven by artificial intelligence confer an advantage and, therefore, are clinically relevant.