An Explainable Hybrid Deep Learning Framework for Knee Deterioration Level Prediction and Surgery Recommendation

Osteoarthritis is a degenerative joint disease which primarily affects the knee and has a major effect on quality of life and mobility. The Kellgren-Lawrence (KL) grading system used in conventional diagnosis depends on the subjective and human error-prone manual interpretation of X-rays. To overcome this limitation, this study offers a multimodal, explainable framework that combines radiographic imaging and clinical information to automatically assess knee osteoarthritis using the Kellgren–Lawrence scale and suggest surgery when necessary. The clinical segment achieves 93.3% accuracy using a stacked ensemble of XGBoost, LightGBM, and CatBoost trained on functional, symptomatic, and demographic data such as age, BMI, KOOS pain score, edema, and crepitus. The imaging segment achieves 80% accuracy by fine-tuning EfficientNetV2L with a Transformer head on 5,000  knee X-rays. Decision-level weighted fusion of both  produces a hybrid system with 94% accuracy on a test set of 900 samples. Explainability is provided through SHAP for clinical feature  and LIME for image , allowing clinicians to examine both numerical and visual explanations for each prediction. The suggested framework gives a dependable, understandable, and clinically significant approach for OA severity classification and surgical recommendation, for real-time hospital implementation, longitudinal disease monitoring, multi-center validation, and integration of biomechanical parameters.