Machine learning to classify the focus score and Sjögren's disease using digitalised salivary gland biopsies: a retrospective cohort study.

Background: The classification of Sjögren's disease partly relies on focus score grading from a minor salivary gland biopsy. Expert regrading of the focus score leads to disease reclassification in half of cases. This study aimed to leverage machine learning to automatically classify the focus score and Sjögren's disease to identify new histological disease subtypes based on minor salivary gland biopsy.

Methods: This retrospective cohort study included minor salivary gland biopsy scanned haematoxylin and eosin slides from six expert centres (three centres in the UK and one each in Greece, Portugal, and France) of the European H2020 NECESSITY consortium. Participants with sicca but without Sjögren's disease and patients with Sjögren's disease and a focus score of either at least 1 or less than 1 where included. All patients with Sjögren's disease fulfilled the American College of Rheumatology-European League Against Rheumatism 2016 criteria. A deep learning model was trained on slides from five centres and validated on slides from the sixth centre. The primary outcome was the area under the receiver operator curve (AUROC) to classify the focus score and Sjögren's disease. Shapley values, an explainable machine learning technology, were computed to identify histological patterns driving the model's classification. People with lived experience of Sjögren's disease were involved in the decision to fund this research and in the dissemination of the findings.

Findings: The study was conducted between Oct 13, 2021, and Sept 5, 2024, and included 545 participants with a mean age of 54·2 (SD 13·5); 490 (90%) were female and 55 (10%) were male. After external validation, the model had an AUROC of 0·88 (95% CI 0·82-0·94) for the focus score classification task and an AUROC of 0·89 (0·82-0·94) for Sjögren's disease classification. The performance of Sjögren's disease classification for patients who were negative for anti-Sjögren's syndrome-related antigen A was 0·92 (0·87-1·00). Of histological patterns identified by the model, a new pattern of CD8⁺ T cells around acinar epithelial cells was associated with Sjögren's disease diagnosis.

Interpretation: This study showed that deep learning can reliably classify the focus score and Sjögren's disease using minor salivary gland biopsy exclusively. The study identified that CD8⁺ T-cell infiltration in acini was associated with Sjögren's disease. Further studies are needed to validate the models.

Funding: Société Française de Rhumatologie, European Alliance of Associations for Rheumatology.