0447: Foot-pressure distribution method for detection of foot joint arthritis in patients with rheumatic diseases

Background/Purpose: Rheumatoid arthritis (RA) and Seronegative Spondyloarthritis (SSA), the most common types of inflammatory arthritis, have significant ankle and foot joint involvement. However, these joints are underrepresented in routine disease activity tools. Conventionally, these joints are assessed either clinically or radiologically (radiographs and MRI). However, their utility in outpatient clinics is limited due to subjectivity, time constraints, and 2-dimensional limitations of radiographs. Foot pressure distribution dynamics is a novel approach to assessing foot and ankle joint-related arthritis, using a wearable insole with sensors to collect high-frequency foot pressure data during walking. This method identifies joints that are less weight-bearing due to synovitis, reflecting improper joint functionality. The aim of this study was to assess the differential foot pressure distribution using wearable devices in patients with foot arthritis and to develop a machine learning model for diagnostic use.

Methods: Adult patients with RA (ACR/EULAR 2010 criteria) or SSA (ASAS criteria) with foot involvement were enrolled. Inclusion criteria were disease duration ≥ 3 years and unilateral/bilateral ankle/foot pain/swelling. Only those patients who could walk for 5 minutes on a flat surface wearing the foot device were included. Patients with foot ulcers, corns, callosities, overriding/deformities of toes, and a history of ankle joint surgery were excluded. The foot device was constructed with 10 Force Sensing Resistors (FSRs). The data was separately collected for walking towards the corner (walk 1) and away (walk 2), each for 1 minute. A total of 20 subjects were recruited for the arthritis class, and 20 subjects as healthy controls. The number of samples for each subject, for each walk, was truncated at 3600 samples to maintain uniformity. Walk 1 and walk 2 of each subject were combined to create a 4-dimensional tensor containing subjects, feet, sensors, and recorded samples. Tensor decomposition was used for dimensionality reduction, thereby reducing the 4-dimensional tensor to a 2-dimensional tensor. Support Vector Machine (SVM) was used for classification. The performance of the model was analyzed using 10-fold cross-validation (CV) and leave-one-out (LOO) CV.

Results: The method successfully classified arthritis patients from normal individuals, achieving the best results when the decomposed tensor had three features after dimensionality reduction. Although SVM’s performance was mediocre with 75% accuracy with an 80-20 train-test split, tuning the hyperparameters inflated the results to 100%. The model also showcased its efficiency by achieving 97% accuracy in 10-fold CV and 100% in LOO CV. These results underscore the method's effectiveness in distinguishing arthritis patients from healthy individuals.

Conclusion: The foot pressure distribution method using a wearable device can detect foot arthritis with moderate to high accuracy. This may be useful in devising machine learning models to diagnose foot arthritis correctly and in preparing footwear for these patients according to the altered pressure distribution.