Currently submitted to: JMIR Medical Informatics
Date Submitted: Oct 23, 2025
(closed for review but you can still tweet)
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Screening Thyroid Dysfunction Using Machine Learning with Routine Blood Tests: A Retrospective study
ABSTRACT
Background:
Thyroid dysfunction is a prevalent endocrine disorder that often remains underdiagnosed due to non-specific symptoms and the absence of routine thyroid testing in standard health checkups. Although thyroid hormone assays are the diagnostic standard, routine blood tests may already contain early biochemical signals associated with thyroid imbalance. Leveraging these widely available markers with machine learning can enable earlier, low-cost detection.
Objective:
This study aims to develop and validate machine learning models to screen thyroid dysfunction using only routine blood test results, without including thyroid-specific hormones such as TSH or FT4. The study also investigates key laboratory predictors that may serve as early indicators of thyroid disorders.
Methods:
This retrospective study uses de-identified data from the Taipei Medical University Clinical Research Database (TMUCRD). Patients are categorized as hypothyroid, hyperthyroid, or euthyroid based on diagnostic codes, prescriptions, and thyroid hormone results. Thirty-nine routine laboratory features, including liver, kidney, lipid, hematologic, and metabolic markers that are used as predictors. Machine learning algorithms (logistic regression, random forest, and XGBoost) are trained on stratified datasets. Model performance is assessed using AUROC, precision, recall, specificity, and F1-score. Ethical approval was obtained from the Taipei Medical University Joint Institutional Review Board (TMU-JIRB).
Results:
Model development and validation are ongoing. Preliminary analyses using pilot data indicate that routine blood test features can distinguish thyroid dysfunction with high accuracy (AUROC range 0.85–0.90). Important predictors include hemoglobin, creatinine, and lipid markers. Final model validation and subgroup performance (by sex and age) will be presented in the completed study.
Conclusions:
This protocol outlines a machine learning–based framework to identify thyroid dysfunction using only routine laboratory data. Early results suggest that non-thyroid-specific biomarkers may provide reliable signals for preliminary screening, enabling broader and cost-effective approaches to thyroid dysfunction detection in primary care settings.
Citation
