A radiomics-based machine learning model trained on 866 parotid gland ultrasound images from 202 participants distinguished Sjögren's disease (SjD) from healthy controls with AUC 0.99, 0.94 accuracy, 0.86 sensitivity, and 0.96 specificity. In this study, the model outperformed conventional radiologist visual assessment (accuracy 0.62 and 0.72), suggesting a more objective ultrasound adjunct for SjD evaluation.

Why It Matters To Your Practice

• Salivary gland ultrasonography is attractive in SjD because it is non-invasive, but interpretation can vary by operator and reader experience.

• An AI-radiomics approach may reduce subjectivity by quantifying intensity, texture, and micro-texture patterns that are difficult to assess consistently by eye.

• For clinicians using ultrasound in dry eye/dry mouth workups, this points to a potential decision-support tool rather than a replacement for established classification criteria.

Clinical Implications

• The model was trained on confirmed SjD and healthy controls, so its strongest performance applies to that binary distinction.

• Specificity of 0.96 suggests possible value when you want to strengthen confidence that abnormal parotid ultrasound patterns are truly consistent with Sjogren Syndrome.

• Sensitivity of 0.86 means false negatives remain possible; clinicians should not use this as a stand-alone rule-out test.

• Because non-Sjögren sicca and incomplete SjD showed substantial feature overlap with confirmed SjD, performance in real-world diagnostically ambiguous patients may be lower than the headline numbers imply.

Insights

• The dataset included 123 patients meeting 2016 ACR/EULAR criteria for SjD, 33 healthy controls, 24 non-Sjögren sicca patients, and 22 incomplete SjD cases.

• Investigators extracted 104 radiomic features and used a 5-fold soft-voting SVM ensemble.

• SHAP interpretability analysis highlighted intensity dispersion metrics, GLCM texture features, and LBP micro-texture patterns as the strongest contributors.

• The held-out non-Sjögren sicca and incomplete SjD cases are clinically important because they better reflect the tougher edge cases seen in practice.

The Bottom Line

• AI applied to parotid ultrasound can classify confirmed SjD versus healthy controls extremely well in this dataset.

• The likely near-term role is as an objective adjunct to clinician assessment, not an autonomous diagnostic test.

• Before changing practice, look for larger multicenter validation and evidence of performance in mixed sicca populations, not just clean case-control comparisons.