Related Experiment Video
Updated: Jan 15, 2026
Standardized Histomorphometric Evaluation of Osteoarthritis in a Surgical Mouse Model
Published on: May 6, 2020
Development and validation of a generalisable survival model to predict osteoarthritis progression
H H T Li1,2, L C Chan1, P K Chan3
1Department of Biomedical Engineering, The Hong Kong Polytechnic University, Hong Kong.
Objective:
To develop and validate a transfer-learning survival model to predict progression end-stage knee osteoarthritis (esKOA) or receive knee replacement (KR).
Method:
A DeepSurv model was trained on the Osteoarthritis Initiative (OAI) dataset with baseline clinical variables, including age, sex, BMI, comorbidities, smoking status, prior knee injury and surgery, pain medication use, use of walking aids, and activity level (9560 knees). A generalisable model was then developed by fine-tuning the OAI-derived model with data from the Multicenter Osteoarthritis Study (MOST) Centre 1 (3002 knees). This model was validated on an independent dataset from MOST Centre 2 (2972 knees). Model performance was evaluated using the concordance index from 1000 bootstrap resamples. SHapley Additive exPlanations (SHAP) were employed to assess changes in feature importance after fine-tuning.
Results:
The OAI-derived model performed well within OAI (C-index = 0.75) but fairly on MOST Centre 1 (C-index = 0.61, p < 0.0001), indicating domain shift or cross-cohort variation. Similarly, a model trained only on MOST Centre 1 data performed moderately within MOST (C-index = 0.63) but did not generalise to OAI (C-index = 0.60, p < 0.0001). After transfer learning, the generalised model maintained performance on OAI (C-index = 0.69) and improved on MOST Centre 1 (C-index = 0.64, p < 0.0001) and MOST Centre 2 (C-index = 0.67, p < 0.0001). SHAP analysis revealed that heart attack history, diabetes, smoking, and BMI became more influential predictors in the fine-tuned model.
Conclusion:
Transfer learning enabled the development of a generalised model for knee OA prognosis that performs consistently across cohorts. By adapting to population-specific risk patterns, this approach enhances model generalisability and reduces bias from ethnic or demographic overrepresentation in training datasets.
