Early Diagnosis of Knee Osteoarthritis With a Natural Language Processing-Driven Approach Based on Clinician Notes: Development and Validation Study.

Background: Knee osteoarthritis (OA) is a common form of knee arthritis that can cause significant disability and affect a patient's quality of life. Although this disease is chronic and irreversible, the patient's condition can be improved and the progression of the disease can be prevented if the disease is diagnosed early and the patient receives appropriate treatment immediately. Therefore, the prediction of knee OA is considered one of the essential steps to effectively diagnose and prevent further severe OA conditions. Knee OA is commonly diagnosed by medical experts or physicians, and the diagnosis of OA is mainly based on patients' laboratory results and medical images, including x-ray and magnetic resonance images. However, diagnosis through such data is often time-consuming. Moreover, the diagnosis results can vary among physicians depending on their expertise. Previous studies mostly focused on using approaches, such as those involving artificial intelligence, to automatically detect knee OA through such data. However, these studies did not incorporate clinicians' or doctors' notes (text data) into the analysis, although these data involving reported symptoms and behaviors are already available and easier to collect and access than laboratory data and image data.

Objective: We propose a novel natural language processing-driven approach based on clinicians' or doctors' notes of patient-reported symptoms (text data only) for diagnosing knee OA.

Methods: The textual information from clinicians' or doctors' notes was first preprocessed using text analysis algorithms with respect to natural language processing. We then incorporated deep learning models, including convolutional neural networks, bidirectional long short-term memory (BiLSTM), and gated recurrent units. Lastly, a disease-specific standard questionnaire called WOMAC (Western Ontario and McMaster Universities Arthritis Index) was taken into account to improve the overall performance of the models.

Results: Our experiment included 5849 records (OA: 3455; non-OA: 2394). Before applying our WOMAC-based processing approach, the best-performing model was BiLSTM (area under the curve, 0.85; accuracy, 0.87; precision, 0.85; sensitivity, 0.95; specificity, 0.76; F-score, 0.90), and there was an improvement in the results with BiLSTM after applying our approach (area under the curve, 0.91; accuracy, 0.91; precision, 0.91; sensitivity, 0.94; specificity, 0.87; F-score, 0.93).

Conclusions: Our proposed method for predicting the occurrence of knee OA showed better performance than other conventional methods that use image data and statistical laboratory data. The findings indicate the feasibility of using text data (symptom descriptions reported by patients and recorded by doctors) to predict knee OA. Medical notes of symptom reports can be considered a valuable data source for predicting whether a particular knee is likely to experience OA progression.