Axial spondyloarthritis is a chronic inflammatory rheumatic disease that primarily affects the sacroiliac joints and the spine. Over years of inflammation, the vertebrae can develop syndesmophytes, erosions and eventual fusion, changes that are quantified radiographically with the modified Stoke Ankylosing Spondylitis Spinal Score, or mSASSS. Because structural damage is largely irreversible, predicting which patients will progress rapidly is a central clinical goal. The research team, led by investigators at Shengjing Hospital of China Medical University and the General Hospital of Northern Theater Command in Shenyang, set out to test whether artificial intelligence could extract prognostic information from routine radiographs that human readers and conventional clinical measures might miss.

The retrospective multicenter study enrolled 277 patients with axial spondyloarthritis from two hospitals: Shengjing Hospital of China Medical University and the General Hospital of Northern Theater Command of the Chinese People’s Liberation Army. For each patient, the team collected clinical variables and spinal radiographs, then constructed separate prediction pipelines for the cervical and lumbar segments of the spine. The target variable was the change in mSASSS over follow-up, meaning the model was trained to predict a continuous degree of progression rather than a simple yes-or-no outcome.

The clinical side of the model began with a two-stage feature selection process. Univariate regression analysis first identified candidate variables associated with progression, yielding 14 significant features in the cervical dataset and 18 in the lumbar dataset. These were then compressed using least absolute shrinkage and selection operator regression, a statistical technique that shrinks the coefficients of less informative variables to zero and thereby retains only the most predictive predictors. LASSO reduced the cervical set to 9 features and the lumbar set to 10, a pruning step designed to guard against overfitting in a dataset of modest size.

On the imaging side, the researchers trained three different deep learning architectures on the spinal radiographs and compared their validation performance. DenseNet121 emerged as the strongest network, achieving a mean squared error of 0.694 for the cervical spine and 0.663 for the lumbar spine, with mean absolute percentage errors of 0.779 and 0.742 respectively. DenseNet121 is a convolutional neural network architecture in which each layer receives feature maps from all preceding layers, a dense connectivity pattern that encourages feature reuse and stabilizes training on limited data. The selected network was used to convert each patient’s radiographs into a compact numerical fingerprint, which the authors call the DL2D signature, essentially a high-dimensional representation of the image content that encodes subtle patterns of structural change.

The DL2D signature was then merged with the selected clinical features to train two classical machine learning classifiers: a support vector machine and a random forest. The combined, multimodal models consistently outperformed models built on either data source alone, confirming that image-derived information and clinical variables carry complementary prognostic signals. The random forest-based combined model showed the strongest generalization, with validation and test mean squared errors of 0.539 and 0.405 for the cervical spine and 0.411 and 0.293 for the lumbar spine. Its mean absolute percentage errors were 0.540 and 0.390 for the cervical region and 0.438 and 0.390 for the lumbar region, and it also achieved higher coefficients of determination, stronger Pearson correlations and lower bias than the competing models.

Model performance was evaluated with a battery of complementary metrics. Mean squared error and mean absolute percentage error quantify how far predictions deviate from observed mSASSS changes, while the coefficient of determination, or R-squared, measures the proportion of variance in progression explained by the model. Pearson correlation assesses linear agreement between predicted and actual values, and bias captures systematic over- or under-estimation. Reporting multiple metrics matters in regression tasks because a model can excel on one measure while failing on another; the consistent advantage of the random forest ensemble across all five metrics strengthens the case that its performance is genuine rather than an artifact of metric selection.

Subgroup analyses by follow-up interval added an important nuance. Predictions were more accurate for patients followed for one to four years than for those followed four years or longer. This pattern is biologically plausible: over short horizons, current inflammatory activity and existing structural damage are strong determinants of near-term change, whereas over longer horizons, treatment changes, lifestyle factors and stochastic disease flares accumulate and erode predictive power. The authors conclude that the model is best suited to predicting short-term, one-to-four-year structural progression, which is precisely the window in which clinicians most often adjust therapy.

The practical implication is a workflow that fits naturally into rheumatology practice. Patients already undergo spinal radiographs at diagnosis and during follow-up, and the clinical variables used by the model, including measures such as C-reactive protein and erythrocyte sedimentation rate, are routinely collected. A trained version of this model could, in principle, process a standard X-ray alongside a handful of lab values and flag patients at high risk of rapid structural worsening, allowing clinicians to intensify monitoring or escalate treatment, for example with biologic agents targeting tumor necrosis factor or interleukin-17, before irreversible damage accumulates. The authors suggest the approach may facilitate risk stratification and timely intervention in axial spondyloarthritis.

The study also carries the usual caveats of retrospective machine learning research. With 277 patients from two centers in a single country, external validation in larger and more diverse cohorts will be needed before clinical deployment, and the model’s accuracy degrades for long-horizon predictions. The work was supported by the Basic Scientific Research Project of the Department of Education of Liaoning Province. Nevertheless, the results add to a growing body of evidence that deep learning features extracted from routine medical images, when fused with structured clinical data, can outperform either modality alone in prognostic modeling. For a disease in which every point on the mSASSS scale represents permanent structural change, a tool that reliably forecasts the next few years of spinal deterioration could meaningfully shift the window for intervention from after damage appears to before it takes hold.

Subject of Research: Deep learning-based prediction of structural spinal damage progression in axial spondyloarthritis

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

Article References: Integrating deep learning signature from 2D spinal radiographs with clinical features: a machine learning model to predict structural spinal damage progression in axial spondyloarthritis. (n.d.). https://doi.org/10.1186/s12880-026-02914-y

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02914-y

Keywords: axial spondyloarthritis, deep learning, machine learning, spinal radiographs, mSASSS, disease progression, prediction model, random forest, support vector machine, DenseNet121, risk stratification, radiology