Background Approximately 2.4 million patients in Japan would benefit from treatment for thyroid disease, including Graves’ disease and Hashimoto’s disease. However, only 450,000 of them are receiving treatment, and many patients with thyroid dysfunction remain largely overlooked. In this retrospective study, we aimed to develop and conduct preliminary testing on a machine learning method for screening patients with hyperthyroidism and hypothyroidism who would benefit from prompt medical treatment. Methods We collected electronic medical records and medical checkup data from four hospitals in Japan. We applied four machine learning algorithms to construct classification models to distinguish patients with hyperthyroidism and hypothyroidism from control subjects using routine laboratory tests. Performance evaluation metrics such as sensitivity, specificity, and the area under receiver operating characteristic (AUROC) were obtained. Techniques such as feature importance were further applied to understand the contribution of each feature to the machine learning output. Results The results of cross-validation and external evaluation indicated that we achieved high classification accuracies (AUROC = 93.8% for hyperthyroidism model and AUROC = 90.9% for hypothyroidism model). Serum creatinine (S-Cr), mean corpuscular volume (MCV), and total cholesterol were the three features that were most strongly correlated with the hyperthyroidism model, and S-Cr, lactic acid dehydrogenase (LDH), and total cholesterol were correlated with the hypothyroidism model. Conclusions We demonstrated the potential of machine learning approaches for diagnosing the presence of thyroid dysfunction from routine laboratory tests. Further validation, including prospective clinical studies, is necessary prior to application of our method in the clinic.

Approximately 2.4 million patients need treatment for thyroid disease, including Graves disease and Hashimoto's disease, in Japan. However, only 450,000 of them are receiving treatment, and many patients with thyroid dysfunction remain largely overlooked. In this retrospective study, we aimed to screen patients with hyperthyroidism and hypothyroidism who would greatly benefit from prompt medical treatment, and examined routine laboratory finding data and machine learning algorithms to investigate whether such accurate and robust screening is possible to prevent overlooking and misdiagnosing thyroid dysfunction. We succeeded in developing a machine learning method to construct the classification model for detecting hyperthyroidism and hypothyroidism in patients using 11 routine laboratory tests. We collected electronic health record and medical checkup data from four hospitals in Japan. As a result of cross-validation and external evaluation, we achieved a high classification accuracy for the hyperthyroidism and hypothyroidism models.