A multicentre validation study of a smartphone application to screen hand arthritis

Abstract

Background: Arthritis is a common condition, and the prompt and accurate assessment of hand arthritis in primary care is an area of unmet clinical need. We have previously developed and tested a screening tool combining machine-learning algorithms, to help primary care physicians assess patients presenting with arthritis affecting the hands. The aim of this study was to assess the validity of the screening tool among a number of different Rheumatologists.

Methods: Two hundred and forty-eight consecutive new patients presenting to 7 private Rheumatology practices across Australia were enrolled. Using a smartphone application, each patient had photographs taken of their hands, completed a brief 9-part questionnaire, and had a single examination result (wrist irritability) recorded. The Rheumatologist diagnosis was entered following a 45-minute consultation. Multiple machine learning models were applied to both the photographic and survey/examination results, to generate a screening outcome for the primary diagnoses of osteoarthritis, rheumatoid and psoriatic arthritis.

Results: The combined algorithms in the application performed well in identifying and discriminating between different forms of hand arthritis. The algorithms were able to predict rheumatoid arthritis with accuracy, precision, recall and specificity of 85.1, 80.0, 88.1 and 82.7% respectively. The corresponding results for psoriatic arthritis were 95.2, 76.9, 90.9 and 95.8%, and for osteoarthritis were 77.4, 78.3, 80.6 and 73.7%. The results were maintained when each contributor was excluded from the analysis. The median time to capture all data across the group was 2 minutes and 59 seconds.

Conclusions: This multicentre study confirms the results of the pilot study, and indicates that the performance of the screening tool is maintained across a group of different Rheumatologists. The smartphone application can provide a screening result from a combination of machine-learning algorithms applied to hand images and patient symptom responses. This could be used to assist primary care physicians in the assessment of patients presenting with hand arthritis, and has the potential to improve the clinical assessment and management of such patients.

Keywords: Artificial intelligence; Diagnosis; Early arthritis; Machine learning; Screening; Telemedicine.

Fig. 1

Example survey feature extraction using one-hot encoding of categorical features and scaling of numerical features. Positive survey responses are shown in bold

Fig. 2

Machine learning model training pipeline

Fig. 3

Online model prediction endpoint