The needle study: Machine learning as a new method for case-finding in celiac disease.

OBJECTIVES

Despite a well-defined diagnostic work-up, uncertainties persist regarding celiac disease (CeD) detection strategies in the general population. Machine learning (ML) algorithms offer promise in aiding medical decision-making on clinical data. This study aimed to utilize ML prediction models to identify uncommon features indicating the necessity for CeD screening in children.

METHODS

The discovery cohort comprised children with CeD exhibiting nonspecific clinical features, drawn from a single referral center, along with gender- and age-matched controls. Data collected included demographic details, symptoms, laboratory tests, and family history of CeD or other autoimmune diseases, excluding anti-transglutaminase immunoglobulin A (IgA) values. Various supervised ML models, utilizing input features to label CeD presence, were applied, with 73 features considered initially.

RESULTS

We collected data from 325 patients with CeD and 490 controls. A comprehensive evaluation using 10-fold cross-validation showed that a Ridge Classifier model achieved the best overall performance, with a mean area under the receiver operating characteristic curve of 0.763 (±0.070), F1-score of 0.662 (±0.067), sensitivity of 0.652 (±0.110), and specificity of 0.689 (±0.081). The least absolute shrinkage and selection operator (LASSO) model identified a stable set of 40 predictive features, with noteworthy features including muscle pain, gastroesophageal reflux disease-like symptoms, fatigue, and specific family histories of autoimmune disease.

CONCLUSIONS

ML models, particularly the LASSO model, exhibited accuracy in identifying CeD in children with nonspecific clinical features. The identification of 40 uncommon features predictive of CeD diagnosis could support a more effective case-finding strategy, potentially enhancing CeD detection.