Artificial intelligence detects Sjögren disease in salivary gland biopsies

medwireNews: An artificial intelligence (AI) model can reliably screen minor salivary gland biopsies for signs of Sjögren disease based on histologic patterns, the PATHSAI study concludes.

A diagnosis of Sjögren disease typically requires the presence of anti-Sjögren syndrome-related antigen A (anti-SSA) or a result of at least 1 point on the focus score for lymphatic invasion of a minor salivary gland, defined as at least one lymphocytic aggregate of 50 cells or more per 4 mm² of gland surface.

However, accurate assessment of the focus score requires high expertise not found in all centers, and there is variability among pathologists, write Samuel Bitoun (Université Paris-Saclay, France) and colleagues in The Lancet Rheumatology.

They note that “expert regrading of the focus score leads to disease reclassification in half of cases.”

The team comments that the “artificial intelligence tool trained in this study robustly performs the unmet clinical need of focus score grading in patients with Sjögren’s disease, especially in patients who are negative for anti-SSA autoantibodies,” they say.

The study included digitized stained microscope slides of minor salivary gland biopsies from 545 patients treated at six centers in the European H2020 NECESSITY consortium. The patients were a mean of 54.2 years old and 90% were women.

The patients were classed into three groups: those with Sjögren disease and a focus score of at least 1 point (n=243); those with Sjögren disease and a focus score of less than 1 point (n=113); and participants with other causes of sicca (n=189). Among these groups, 64%, 100%, and 0% were anti-SSA positive.

AI accurate for focus score and Sjögren disease diagnosis

The researchers trained a deep learning model called cluster-constrained attention multiple instance learning (CLAM) on slides from 80% of the patients from five of the centers and validated the model on slides from the remaining 20% of patients in the sixth center. To check the usefulness of the model, they calculated the area under the receiver operating characteristic curve (AUROC).

The model was generally accurate at detecting biopsies with a focus score of at least 1 point, with an AUROC of 94%. The AUROC remained high at 88% when validated externally in the final center.

The model also had a specificity of 89% and 82% in the internal and external validations, respectively, and a sensitivity of 87% and 74%.

The internal and external AUROCs for the 100 anti-SSA-positive patients were 92% and 81%, respectively, and 92% and 91% for the 87 who were negative.

When training the model to predict Sjögren’s diagnosis, the investigators found AUROCs of 84% and 89% in internal and external validation, respectively. The AUROC was higher at 92% in both internal and external validation among anti-SSA-negative patients.

The authors note that their machine learning model needs to be verified in a trial before it could be approved for the assistance of pathologists in classifying Sjögren disease.

They add: “High accuracy in focus score grading will be crucial for new drugs for this disease. Increasing the pathologist’s confidence in the classification with heat maps showing the patterns used for the focus score or Sjögren’s disease tasks is crucial.”

Understanding the AI process

To find out how the algorithm detected Sjögren disease, the researchers analyzed the most important tiles from slide images and analyzed them using an explainable machine learning method called Shapley values, where positive and negative numbers indicated positive and negative links to Sjögren disease diagnosis, respectively.

When screened by an expert pathologist, six histologic patterns with positive Shapley values included dense lymphocytic aggregates of B cells surrounded by T cells, and lymphocytic aggregates with blood vessels and fibrosis.

A novel pattern of CD3+, CD8+, and CD103+ lymphocytes surrounding intact acinar epithelial cells was also identified.

“The hypothesis is that these cytotoxic CD8+ T cells might lead to the direct destruction of acinar cells, causing diminished salivary production,” the authors say, adding that the model could “unveil patterns relevant to the pathophysiology of Sjögren’s disease.”

A further three patterns with negative Shapley values included normal seromucous glands and fibrotic tissue with little inflammation.

The authors acknowledge limitations to the study including the focus on European centers and the lack of immunohistochemical slides preventing use of germinal centers as a diagnostic feature.

In a commentary on the article, Tamandeep Bharaj and Kathrine Skarstein, from the University of Bergen in Norway, write that the study “paves the way for a new method for minor salivary gland biopsy evaluation in Sjögren’s disease in which the deep learning model assists the pathologist in their diagnostic assessments.”

They conclude that “by sophisticatedly implementing the use of AI into Sjögren’s disease, this work marks a pivotal advancement, redefining the future landscape of diagnostics and health care in Sjögren’s disease.”

medwireNews is an independent medical news service provided by Springer Healthcare Ltd. © 2025 Springer Healthcare Ltd, part of Springer Nature

Lancet Rheumatol 2025; doi:10.1016/S2665-9913(25)00181-X
Lancet Rheumatol 2025; doi:10.1016/S2665-9913(25)00258-9