Identifying Significant Structural Factors associated with Knee Pain Severity in Patients with Osteoarthritis using Hybrid Bio-BERT Bi-LSTM CNN Model

Abstract

Knee osteoarthritis (OA) is a progressive, long-term joint disorder marked by pain, stiffness, and impaired mobility, significantly diminishing patients’ quality of life. Accurately predicting the severity of knee pain is challenging due to the complex, multifactorial aetiology of OA and the heterogeneity in both structural changes and clinical symptoms among patients. This study introduces a hybrid deep learning model which integrates Bidirectional Encoder Representations from Transformers for Biomedical Text Mining (Bio-BERT), Bidirectional Long Short-Term Memory (Bi-LSTM), and Convolutional Neural Network (CNN) models to predict knee OA pain severity by combining clinical text and Magnetic Resonance Imaging (MRI) datasets. Using a multimodal Kaggle datasets which contains 163,064 clinical records and 2000 MRI scans, the model achieved a maximum test accuracy of 98.11%, with strong precision, recall, and F1-scores, particularly for severe pain cases. These results demonstrate its effectiveness in synthesizing textual and visual features, supporting potential clinical applications in diagnosis, early intervention, and treatment planning. Future work will focus on improving interpretability, validating external generalizability, enabling real-time Electronic Health Record (HER) integration, and ensuring ethical AI practices, including patient privacy and transparency.