Pilot study of a machine-learning tool to assist in the diagnosis of hand arthritis

Background: Arthritis is a common condition, which frequently involves the hands. Patients with inflammatory arthritis have been shown to experience significant delays in diagnosis.

Aim: To develop and test a screening tool combining an image of a patient's hands, a short series of questions and a single examination technique to determine the most likely diagnosis in a patient presenting with hand arthritis. Machine learning techniques were used to develop separate algorithms for each component, which were combined to produce a diagnosis.

Methods: A total of 280 consecutive new patients presenting to a rheumatology practice with hand arthritis were enrolled. Each patient completed a nine-part questionnaire, had photographs taken of each hand and had a single examination result recorded. The rheumatologist diagnosis was recorded following a 45-min consultation. The photograph algorithm was developed from 1000 previous hand images and machine learning techniques were applied to the questionnaire results, training several models against the diagnosis from the rheumatologist.

Results: The combined algorithms in the present study were able to predict inflammatory arthritis with an accuracy, precision, recall and specificity of 96.8%, 97.2%, 98.6% and 90.5% respectively. Similar results were found when inflammatory arthritis was subclassified into rheumatoid arthritis and psoriatic arthritis. The corresponding figures for osteoarthritis were 79.6%, 85.9%, 61.9% and 92.6%.

Conclusion: The present study demonstrates a novel application combining image processing and a patient questionnaire with applied machine-learning methods to facilitate the diagnosis of patients presenting with hand arthritis. Preliminary results are encouraging for the application of such techniques in clinical practice.