Background

Juvenile idiopathic arthritis (JIA), a heterogenous group of childhood arthritides, is the most common childhood rheumatologic condition [1]. JIA causes significant morbidity across the globe and affects many joints and extra-articular organs [2].

While diagnosing JIA early in its disease process is crucial for disease modification and treatment [3,4,5], multiple diagnostic challenges limit ideal management, including significant shortages of pediatric rheumatologists in the US [6]. Magnetic resonance imaging (MRI), while more sensitive and specific than the clinical exam, is costly and time consuming, limiting its application [7]. Furthermore, no easy-to-use and inexpensive objective measurement exists for monitoring joints affected by JIA in the long term, apart from symptom-based questionnaires that can be affected by subjectivity [8].

A potential step in ameliorating these challenges is the development of new diagnostic modalities to define, diagnose, and longitudinally monitor articular involvement in JIA. Here, joint acoustic emissions (JAEs) have the potential to change this status quo. JAEs are the sounds that originate from the movement of the articulating surfaces of the joints, and therefore, vary based on the articulatory properties of the joint. JAEs are affected by joint diseases such as rheumatoid arthritis [9,10,11] or osteoarthritis [12], which change the tribological properties of the joint [13]. When novel machine learning algorithms are leveraged to analyze these sounds, JAEs can be used as digital acoustic biomarkers that are easy to obtain using non-invasive, inexpensive, and compact equipment applied to the surface of the skin. This information can be exceptionally valuable to inform pediatric rheumatologists toward efficient personalized treatment. Our group previously demonstrated that JAEs have merit in assessing knee joint health in adult populations [9, 14, 15]. We have also demonstrated the feasibility of JAEs to discriminate JIA from healthy controls [10, 11]. Although these proof-of-concept studies demonstrated the potential of using JAEs as a tool to identify and monitor JIA, they were conducted in a relatively small study population (n < 45).

The knees are the most involved joint in JIA [16,17,18]; in a study of 95 newly diagnosed children with JIA, knee involvement was seen in 74% at onset and 93% after 5 years [18]. The present study builds on our previous work on the use of JAEs as a biomarker to identify knee involvement in JIA. To validate our previous results, we recorded knee JAEs from a larger cohort compared to our previous studies. We also investigated the use of new machine learning classifiers. Specifically, we used JAEs to assess and differentiate knee health in children with active versus inactive JIA involvement. Here, we report on a more generalizable model for knee health assessment in children with JIA that offers clinical perspective compared to our previous work.

Findings

Methods

Subjects and data collection

JAEs from knees were recorded in Children’s Healthcare of Atlanta’s Center for Advanced Pediatrics. Children diagnosed with JIA were included in the study, while those with prior musculoskeletal injury were excluded. Each knee was analyzed separately from its pair. Knees were considered “active” if the knee had swelling alone, and/or tenderness and limited range of motion, at the time of recording on physical exam by an experienced attending rheumatologist. “Inactive” knees did not meet the active involvement criteria at the time of the recording.

All subjects performed ten seated, unloaded flexion-extension cycles, one cycle every four seconds guided by an instructional cartoon as previously described (10, 11). JAEs were recorded from both knees using two miniature uniaxial accelerometers (3225F7, Dytran Instruments Inc., CA, USA) attached 2 cm medial and lateral to the distal patellar tendon using double-sided tape (Rycote Microphone Windshields Ltd, Stroud, Gloucestershire, UK). These accelerometers have a broad bandwidth (2 Hz – 10 kHz), high sensitivity (100 mV/g), and low noise floor (0.0007 grms). A data acquisition system (USB-4432, National Instruments, TX, USA) sampled the joint sounds at 100 kHz. Figure 1 shows an overview of the described experimental setup to record JAEs.

Fig. 1
figure 1

Overview of the experimental setup that was used to record JAEs. Two uniaxial accelerometers were used to record the joint sounds from the knee. A data acquisition system tethered to a laptop (not shown in figure) was used to sample and save the analog acoustic emission signals

Full size image

Signal processing & machine learning model

The recorded JAE signals were divided into flexion-extension cycles. Wavelet denoising and bandpass filtering (100 Hz − 12 kHz) were used to remove noise from the signals. After denoising, we divided the JAE cycles into segments of 200 ms with 50% overlap. 273 time-frequency audio features were extracted from each segment and the average and standard deviation of each feature were calculated from all segments of the 10 flexion-extension cycles. A principal component analysis was performed on these 273 features to extract the 60 top principal components that described 95% of the variance of the data. These 60 principal components were used as input features for an Extreme Gradient Boosting (XGBoost) machine learning classifier. Knees that were labeled as active JIA were considered to have a stronger label than inactive JIA knees and were therefore used to train the classifier. This set of active JIA data was supplemented with 80% of the healthy controls to be our training data set. We tested our trained classifier on the remaining data; all inactive knees and the remaining 20% of the healthy controls. The classifier predicted a joint score between 0 and 1, which can be considered as the probability of having JIA. Leave-one-leg-out cross-validation (LOLO-CV) was used to assess the performance of the classifier on the training set. Joint scores, overall accuracy, confusion matrices, sensitivity / specificity, receiver operating characteristic (ROC) curve and corresponding area under the curve (AUC) are reported. All signal preprocessing was performed using Matlab (MathWorks, MA, USA) and the machine learning classifier was developed in Python using the scikit-learn toolbox. A two-sample Kolmogorov-Smirnov test was used to compare the distributions of the joint scores of the active and inactive JIA knees.

Results

A total of 116 children participated in this study, (86 with JIA diagnosed by a pediatric rheumatologist and 30 age and gender matched controls, Fig. 2and Table 1). Of the 172 knees of patients with JIA, 43 knees were considered active, and 129 knees were inactive recordings. The 30 healthy controls were examined by a pediatric rheumatologist and had no evidence of arthritis.

Fig. 2
figure 2

Schematic overview of the study population of the presented work

Full size image
Table 1 Overview of the subjects that were recruited for the presented study
Full size table

The LOLO-CV on the training set of the classifier resulted in an 81.1% accuracy, while the classifier resulted in an 87.7% accuracy for the testing data set (inactive JIA + healthy controls). The LOLO-CV and testing confusion matrices are shown in Fig. 3. The sensitivity and specificity of the LOLO-CV on the training data set were 88.6% and 72.3%, respectively. The sensitivity and specificity of the testing data set were 88.1% and 83.3%, respectively. Figure 4 shows the distribution of the joint scores of the correctly labeled healthy controls, active and inactive JIA knees separately. The distributions of the joint scores of the active and inactive knees were significantly different (p < 0.05). The ROC curve and corresponding AUC for the cross-validation is shown in Fig. 5.

Fig. 3
figure 3

(A) The confusion matrix of the cross-validation on the training data set and (B) the confusion matrix corresponding to the test data set

Full size image
Fig. 4
figure 4

Density distributions of the joint scores from healthy, inactive, and active JIA knees. From left to right for each group: median and 1st / 3rd quartiles, scatter plot of all joint scores, and corresponding joint score distributions

Full size image
Fig. 5
figure 5

Receiver Operating Characteristic (ROC) curve and corresponding area under the curve (AUC) of the classifier obtained for the cross-validation

Full size image

Discussion & conclusions

We present a study on a larger cohort of 116 children building further on our previous work on knee JAEs as a non-invasive and convenient tool to detect JIA [10, 11]. We introduced the use of a new machine learning classifier, Extreme Gradient Boosting (XGBoost), which generalized well on an extensive test data set. Moreover, this new classifier led to a better separation between children with JIA and healthy controls compared to the classifier (logistic regression) we used in our previous work. This is reflected by the joint scores of the healthy control knees which were consistently close to 0 (Fig. 4), representing the high sensitivity of the recorded JAEs.

The differences in joint acoustic emissions (JAEs) between inactive and healthy knees can be attributed to chronic inflammation and scarring in inactive knees, resulting in thickened synovium and altered JAEs. This is not observed in healthy knees, especially in children, who do not have these features associated with articular inflammation in JIA.

The presented technology utilizes JAEs as a digital biomarker for assessing joint health, which clinicians can easily implement it in outpatient settings for screening and monitoring diseases. This tool can aid general pediatricians in making referrals to rheumatologists, prompt early treatment [3,4,5], and facilitate decisions on diagnostic tests and treatment adjustments. Additionally, wearable devices incorporating this digital biomarker may serve as future tools for at-home disease monitoring.

Limitations of this study include the inherent subjectivity of active and inactive labeling by the physical exam of providers, potentially decreasing the accuracy of validation of the machine learning classifier. No significant difference was found between the median joint scores of active and inactive knees, even though their distributions were different. The subjects tested in our study also varied in their courses of treatment, potentially masking the accuracy of the classification labels made by patients and their doctors. More objective labeling by a validated gold standard method (e.g., radiologic diagnostics, including MRI, ultrasound, or inflammatory markers in the laboratory), and longitudinal monitoring could improve the classification algorithm and reveal significant differences between active and inactive joint scores. Future directions of JAE studies should also include disease and joint score correlation in other joints affected by JIA.

The presented work demonstrates the feasibility to use JAEs as a non-invasive digital biomarker for articular health assessment in JIA. The use of JAEs by clinicians in the outpatient setting, therefore, represents a potentially inexpensive and easy-to-use screening or disease monitoring tool to help decrease and quantify disease morbidity caused by JIA.