Integrating deep learning signature from 2D spinal radiographs with clinical features: a machine learning model to predict structural spinal damage progression in axial spondyloarthritis

Abstract Background To develop a machine learning model that integrates a two-dimensional deep learning (DL2D) signature derived from spinal radiographs with clinical features to predict the progression of structural spinal damage in axial spondyloarthritis (axSpA). Methods This retrospective study enrolled 277 axSpA patients from two centers and collected clinical and spinal radiograph data. Clinical features were first screened using univariate regression followed by least absolute shrinkage and selection operator (LASSO) regression. Deep learning models were built on spinal radiographs using three architectures; the optimal model was selected to extract the DL2D signature, which was then integrated with clinical features to train support vector machine (SVM) and random forest (RF) models for predicting cervical and lumbar Modified Stoke Ankylosing Spondylitis Spinal Score (mSASSS) progression. Model performance was assessed using mean squared error (MSE), mean absolute percentage error (MAPE), bias, coefficient of determination (R²), and Pearson correlation to assess linear agreement. Subgroup analyses were conducted by follow‑up intervals. Results Univariate analysis yielded 14 and 18 significant clinical features in the cervical and lumbar datasets, reduced to 9 and 10 by LASSO regression, respectively. DenseNet121 was chosen for DL2D signature extraction due to its superior validation performance (MSE: 0.694 for cervical, 0.663 for lumbar; MAPE: 0.779 and 0.742). The combined model integrating the DL2D signature with clinical features consistently outperformed single‑modal models. The RF‑based combined model showed strong generalization, with validation/test MSEs of 0.539/0.405 (cervical) and 0.411/0.293 (lumbar), MAPEs of 0.540/0.390 and 0.438/0.390, as well as higher R², stronger Pearson correlations, and lower bias than the other models. Subgroup analysis confirmed better prediction for 1–4 years than for ≥ 4 years follow‑ups. Conclusions The random forest model integrating the DL2D signature with clinical features accurately predicts short-term (1–4 years) structural spinal damage progression in axSpA, which may facilitate risk stratification and timely intervention.