Identifying the Clinical Signature of ACA-Positive Sjögren’s Syndrome Using Machine Learning

23 Apr 20264 min read0 viewsJournal Feed

GIST

ObjectiveWe sought to leverage machine learning algorithms to identify the complex clinical and serological signature of anti-centromere antibody (ACA) positivity in Sjögren’s syndrome (SS) patients.MethodsThis multicenter study analyzed clinical data from a cohort of 616 patients diagnosed SS, comprising 81 ACA-positive and 535 ACA-negative cases. To ensure robust model development, we randomly partitioned the dataset into training and validation subsets in a 7:3 ratio.

We implemented and compared six machine learning models after identifying optimal predictors using the LASSO regression. We mainly evaluate the performance of the model through the AUC and a series of comprehensive indicators.

To ensure clinical interpretability, we also employed the SHAP analysis method to quantify the influence of each feature on the model’s outcome.ResultsAmong the evaluated models, GBDT exhibited superior predictive efficacy. The model achieved an AUC value of 0.812 in the training set and maintained a robust AUC of 0.811 (95% CI: 0.699–0.906) in the validation cohort.

At the same time, the model has the highest sensitivity (0.750 in the validation test).

Clinical Editorial

Study Scope and Design

Model performance was primarily judged by AUC and additional metrics, with interpretability aided by SHAP analysis.

The investigation used machine learning to delineate the clinical and serological signature of anti-centromere antibody (ACA) positivity within Sjögren’s syndrome (SS).
This multicenter cohort included 616 SS patients, of whom 81 were ACA-positive and 535 ACA-negative.
Data were randomly split into training and validation sets in a 7:3 ratio to support model development and assessment.
Six machine learning models were compared after selecting predictors with LASSO regression.

Analytical Approach and Key Methods

The study employed gradient boosted decision trees (GBDT) as the principal model based on comparative performance.
Predictive performance was quantified with an AUC of 0.812 in the training set and 0.811 (95% CI: 0.699–0.906) in the validation cohort, indicating consistent discrimination.
Sensitivity in the validation set was reported as 0.750, representing the model’s ability to identify ACA-positive cases.
LASSO regression guided the selection of features prior to model fitting, and SHAP values were used to interpret feature importance and influence.

Principal Predictors and Interactions

SHAP analysis identified major predictors: serological markers (anti-SSA/Ro52, anti-SSA/Ro60, anti-AMA-M2, anti-SSB, and IgM), age, and Raynaud’s phenomenon (RP).
SHAP interaction analysis disclosed non-linear interactions, notably that the impact of RP on the ACA-positive prediction increases with advancing age, and that the predictive contribution of anti-AMA-M2 grows when IgM levels are lower.

Context and Implications

The ML framework demonstrated capacity to structure and quantify complex clinical-serological relationships distinctive to the SS-ACA+ subgroup.
The approach underscores non-linear patterns in clinical data and provides a quantitative basis for exploring ACA-associated profiles in future prospective work.

Limitations and Unreported Details

Specifics on model hyperparameters, exact feature counts, cohort demographics beyond age and RP, and external validation beyond the presented training/validation split are not detailed in the provided content.