An Interpretable Experimental Data Augmentation Method to Improve Knee Health Classification Using Joint Acoustic Emissions

Abstract

The characteristics of joint acoustic emissions (JAEs) measured from the knee have been shown to contain information regarding underlying joint health. Researchers have developed methods to process JAE measurements and combined them with machine learning algorithms for knee injury diagnosis. While these methods are based on JAEs measured in controlled settings, we anticipate that JAE measurements could enable accessible and affordable diagnosis of acute knee injuries also in field-deployable settings. However, in such settings, the noise and interference would be greater than in sterile, laboratory environments, which could decrease the performance of existing knee health classification methods using JAEs. To address the need for an objective noise and interference detection method for JAE measurements as a step towards field-deployable settings, we propose a novel experimental data augmentation method to locate and then, remove the corrupted parts of JAEs measured in clinical settings. In the clinic, we recruited 30 participants, and collected data from both knees, totaling 60 knees (36 healthy and 24 injured knees) to be used subsequently for knee health classification. We also recruited 10 healthy participants to collect artifact and joint sounds (JS) click templates, which are audible, short duration and high amplitude JAEs from the knee. Spectral and temporal features were extracted, and clinical data was augmented in five-dimensional subspace by fusing the existing clinical dataset into experimentally collected templates. Then knee scores were calculated by training and testing a linear soft classifier utilizing leave-one-subject-out cross-validation (LOSO-CV). The area under the curve (AUC) was 0.76 for baseline performance without any window removal with a logistic regression classifier (sensitivity = 0.75, specificity = 0.78). We obtained an AUC of 0.86 with the proposed algorithm (sensitivity = 0.80, specificity = 0.89), and on average, 95% of all clinical data was used to achieve this performance. The proposed algorithm improved knee health classification performance by the added information through identification and collection of common artifact sources in JAE measurements. This method when combined with wearable systems could provide clinically relevant supplementary information for both underserved populations and individuals requiring point-of-injury diagnosis in field-deployable settings.

Keywords: Data augmentation; Knee health; Knee sounds; Machine learning.