Summary

Development and validation of a clinical point-of-care tool (DK3) to predict and monitor the risk of ACL reconstruction revision using machine learning analysis of the Danish Knee Ligament Reconstruction Registry.

Abstract

Background

Previous machine learning (ML) analysis of national ligament registries found moderate accuracy for predicting ACL reconstruction revision risk. We examined whether an enhanced ML-Cox regression approach can improve the prediction accuracy for ACL reconstruction revision using data from the Danish Knee Ligament Reconstruction Registry (DKRR). External validation is an integral step to evaluate model performance on a seperate cohort of patients to those used to develop the algorithm.

Methods

Data was extracted from the DKRR on all patients who underwent primary ACL reconstruction between 2005 and 2023. An enhanced ML approach using Cox regression with a least absolute shrinkage and selection operator (LASSO) penalised approach and stable iterative variable selection (SIVS) was applied using a multi-stage analysis. The most significant demographic, clinical and PROM data from this analysis of the DKRR were selected for the final Cox regression model. Data was randomly split in a 2:1 ratio into separate training and test cohorts to develop and internally validate regression models, respectively. External validation using other national ligament registries was performed to evaluate model performance and demonstrate the clinical utility of the DK3.

Results

The best performing Cox regression model for predicting ACL reconstruction revision risk incorporated age (at time of primary ACL reconstruction), Pain P1 and QoL Q2 and Q3, from 12-month Knee Osteoarthritis and Outcome Score (KOOS) data. This model demonstrated good prediction accuracy 1-, 2- and 5-years beyond 12-month follow-up assessment (C-index=0.73 ± 0.03, 0.74 ± 0.02, 0.74 ± 0.02 respectively).

The DK3 – Risk Monitoring Tool was developed for predicting and monitoring risk of ACL reconstruction revision at a patient-specific level. Using the DK3 to predict the risk of a 25 year-old patient with KOOS values of 2 for P1, Q2 & Q3 gives a predicted risk of ACL reconstruction revision of 2.8% (1 year), 4.8% (2 years) and 8.7% (5 years) after 12-month follow-up assessment. Using the risk modification option to adjust the KOOS values to 0 (i.e. normal) reduces the predicted risk to 0.8% (1 year), 1.5% (2 years) and 2.7% (5 years).

Conclusion

An enhanced ML-Cox regression using patient age and 3 KOOS items obtained 12-months post-surgery provided good prediction accuracy of ACL reconstruction revision risk at 1, 2 and 5 years relative to 12-month follow-up assessment. The current modelling demonstrates greater prediction accuracy, requiring fewer input variables, compared to previous ML studies incorporating pre-operative data. External validation using other national ligament registries ensures that the modelling performance is robust and supports the utility of this clinical point-of-care tool. The DK3 - Risk Monitoring Tool can be used to predict and monitor patient-specific ACL reconstruction revision risk.