使用例行血液檢查進行甲狀腺功能異常篩檢的機器學習方法

Background:
Thyroid dysfunction, including hypothyroidism and hyperthyroidism, is a
prevalent endocrine disorder that often goes undiagnosed due to non-specific
symptoms and the absence of routine screening. Although thyroid-specific
hormone testing is the diagnostic standard, emerging evidence suggests that
routine blood test results may contain early biochemical signatures of thyroid
imbalance, creating an opportunity for opportunistic screening using routinely
collected laboratory data.
Objective:
This study aimed to develop and validate machine learning models capable
of screening for thyroid dysfunction, excluding thyroid-specific hormones such
as TSH, FT3, and FT4, using only routine blood test data in a real-world, multihospital clinical network setting.
Methods:
We conducted a retrospective observational study using anonymized
patient data from the Taipei Medical University Clinical Research Database
(TMUCRD), which includes records from Taipei Medical University Hospital,
Wan Fang Hospital, and Shuang Ho Hospital. The source database spanned a 13-
year period from January 2008 to December 2020. The final cohort included
11,887 patients, comprising 803 cases and 11,084 controls with cases classified
as hypothyroidism (65.3%) or hyperthyroidism (34.7%). Models were trained
using routine laboratory features and derived ratios within a sex- and age-matched
case–control framework to mitigate class imbalance. To preserve clinical realism,
missing laboratory measurements were not imputed, reflecting the fact that test
availability depends on physician ordering decisions. Taking advantage of recent
implementations of native missing-value handling in Random Forest models, this
algorithm was included alongside XGBoost and LightGBM.
Model performance was assessed using 5-fold stratified cross-validation during
training and evaluated on a held-out test set using the area under the receiver
operating characteristic curve (AUC), sensitivity, and specificity.
xiii
Results:
In the combined thyroid dysfunction cohort, the Random Forest model
achieved the highest test performance, with an AUROC of 0.755, sensitivity of
0.820, and specificity of 0.575 at a screening-oriented operating point. For
hypothyroidism, Random Forest also demonstrated the best performance, with a
test AUROC of 0.765 and high sensitivity (0.857), supporting its suitability for
screening applications. For hyperthyroidism, XGBoost achieved the highest
discrimination, with a test AUROC of 0.749, sensitivity of 0.554, and specificity
of 0.750. Across cohorts, feature importance analyses consistently highlighted
renal, hematological, metabolic, and immune-related markers, including eGFR,
creatinine, liver enzymes, lipid profiles, and white blood cell differentials. The
inclusion of comorbidity indicators for diabetes mellitus, chronic kidney disease,
and heart failure did not materially improve model performance, indicating that
routinely available blood test parameters alone provide sufficient signal for
screening-level identification of thyroid dysfunction.