Development and preliminary validation of a machine learning system for thyroid dysfunction diagnosis based on...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Endocrinology

Background:

• Millions in Japan require thyroid disease treatment, yet many remain undiagnosed or untreated.
• Thyroid dysfunction, including Graves' and Hashimoto's diseases, affects a significant population.
• A gap exists in identifying and treating patients with thyroid dysfunction.

Purpose of the Study:

• To develop and test a machine learning (ML) method for screening hyperthyroidism and hypothyroidism.
• To identify patients who would benefit from prompt medical treatment for thyroid disease.
• To leverage routine laboratory tests for efficient thyroid dysfunction screening.

Main Methods:

• Retrospective analysis of electronic medical records and checkup data from four Japanese hospitals.
• Application of four ML algorithms to build classification models for hyperthyroidism and hypothyroidism.
• Evaluation of model performance using metrics like sensitivity, specificity, and AUROC, with feature importance analysis.

Main Results:

• High classification accuracies achieved: 93.8% AUROC for hyperthyroidism and 90.9% for hypothyroidism.
• Key features for hyperthyroidism model: serum creatinine (S-Cr), mean corpuscular volume (MCV), and total cholesterol.
• Key features for hypothyroidism model: S-Cr, lactic acid dehydrogenase (LDH), and total cholesterol.

Conclusions:

• Machine learning shows significant potential for diagnosing thyroid dysfunction from routine laboratory tests.
• The developed ML models demonstrate high accuracy in identifying hyperthyroidism and hypothyroidism.
• Further prospective clinical validation is required before widespread clinical application.