New interpretation of GNRB® knee arthrometer results for ACL injury diagnosis support using machine learning

Abstract

Purpose:GNRB is an arthrometer and alternative diagnostic method less expensive than MRI and more accurate than KT-1000 in Anterior Cruciate Ligament (ACL) tears detection. Dynamic knee laxity tests are more complex to analyze and will require a new solution of universal interpretation. The hypothesis is that using a solution based on Artificial Intelligence (AI) will allow us to obtain a more accurate and robust non-invasive diagnostic method than the current solution with three laxity thresholds. Method:AI can enhance the reliability of this analysis by utilizing advanced algorithms and incorporating a wide range of additional parameters, leading to more precise diagnostics. The existing process solely rely on laxity differences obtained from the device, overlooking influential factors like clamping force. By considering a broader set of parameters and employing well-calibrated models a comparative study was performed between different Machine Learning (ML) models and Ensemble Learning to get the best compromise. The correction process will leverage statistical analysis of the current solutions. Results:Association of Voting, Stacking and threshold laxity methods results report a 6% increase in accuracy and approximately 13% improvement in tear detection compared to the current solution with 1384 GNRB® measurements. Predicted diagnoses are also more prone to new data from patients unknown to the model and confirmed using a validation database. Conclusion:A first ML model was introduced in ACL tears detection using GNRB device. GNRB coupled with ML was encouraging with better results than the current static diagnostic method. It could be integrated and recommended as a complementary solution to MRI.