Ultrasound-based multimodal machine learning nomogram for predicting deep vein thrombosis after total knee arthroplasty

This study aimed to develop a multimodal, machine learning-based predictive model for individualized risk assessment of postoperative deep vein thrombosis (DVT) in patients undergoing total knee arthroplasty (TKA), integrating clinical data, handcrafted radiomic features, and deep learning-derived imaging biomarkers. We retrospectively analyzed a multicenter cohort of 1874 TKA patients for model training and internal validation, and an external cohort of 741 patients for independent validation. Doppler ultrasound images were preprocessed and analyzed to extract handcrafted radiomic features and deep learning-derived image biomarkers. Handcrafted radiomic features (m = 215) were extracted per IBSI standards, while deep features (m = 128) were derived via a convolutional autoencoder. To ensure consistency across centers, we applied statistical harmonization techniques to reduce scanner-related variability. Feature selection was performed using regularization methods, and five machine learning classifiers were trained and compared. The fused feature model using XGBoost achieved superior performance (external AUC = 93.8 %, accuracy = 92.6 %, sensitivity = 92.1 %), significantly outperforming clinical-only and unimodal models. Radiomic and deep features alone also demonstrated strong predictive value (AUCs >89 %). Decision curve analysis confirmed higher net clinical benefit across clinically relevant thresholds. The derived nomogram enabled individualized bedside DVT risk stratification with high predictive fidelity. The proposed multimodal model provides an accurate and clinically interpretable tool for individualized DVT risk prediction after TKA. By integrating imaging and clinical data into a machine learning nomogram, this approach enables more precise prophylaxis and postoperative management.