Development and Validation of Machine-Learning-Based Prediction Models for Thyroid Diseases During Pregnancy
Clinical Snapshot
PICO Framework
| P — Population | Pregnant women who gave birth at a single tertiary centre in Beijing, China (n = 5,461) |
| I — Intervention | Eight machine-learning prediction models (logistic regression, Bayesian, k-nearest neighbours, support vector machine, neural network, CART, XGBoost, random forest) trained on nine clinical feature variables selected via the Boruta algorithm |
| C — Comparator | No explicit external reference standard for thyroid disease diagnosis is described; models are compared against each other using internal train/test split performance metrics |
| O — Outcomes | Discriminative performance (AUROC, AUPRC), accuracy, Matthews correlation coefficient, log loss, Brier score, calibration, and decision curve analysis for prediction of thyroid disease during pregnancy |
Bottom Line
This retrospective single-centre study from Beijing develops and internally validates eight machine-learning models to predict thyroid disease during pregnancy, identifying a random forest model as the best performer. However, the study has critical methodological flaws that preclude clinical adoption. Most importantly, the reference standard for thyroid disease diagnosis is never defined — a fundamental requirement for any diagnostic accuracy study. The reported performance metrics (AUROC 0.9999 on the test set) are biologically implausible for a complex multifactorial condition and raise serious concerns about data leakage or overfitting, particularly given the absence of cross-validation and confidence intervals. The single-centre tertiary hospital design introduces substantial spectrum bias, and no external validation is performed. Sensitivity, specificity, and likelihood ratios at any clinical threshold are absent, making it impossible to assess real-world clinical utility. The nine predictor variables identified largely replicate existing guideline-based risk factors, raising questions about incremental value over current practice. Senior clinicians should not adopt this model in practice. Independent external validation in a representative prospective cohort with a clearly defined reference standard is required before this approach can be considered for clinical translation.
Key Findings
P Value: Not reported
Effect Size: Random forest AUROC: 0.9988 (training set), 0.9999 (test set); Accuracy: 0.9839 (training), 0.9960 (test); MCC: 0.9679 (training), 0.9778 (test)
Primary Outcome: Prediction of thyroid disease during pregnancy using a random forest model
Nnt Or Sensitivity: Sensitivity and specificity at any operating threshold not reported; AUROC values are near-perfect and clinically implausible without confidence intervals or external validation
Confidence Interval: Not reported for any metric
Clinical Application
The nine predictor variables (age, height, pre-pregnancy weight, gravidity, parity, primiparity/multiparity, hypertensive disorders of pregnancy, scarred uterus, autoimmune disease) are routinely collected in antenatal booking visits and are feasible to obtain. However, the model itself is not available as a validated clinical tool, and the near-perfect performance metrics cannot be accepted at face value without independent external validation. In Australia, the RANZCOG and RACGP do not recommend universal thyroid screening in pregnancy but support targeted screening based on risk factors — an approach consistent with the ATA 2017 guidelines. The nine predictor variables identified in this study broadly overlap with RACGP-recognised risk factors. However, this model cannot be recommended for Australian clinical practice without: (1) external validation in an Australian or comparable Western population; (2) TGA consideration if deployed as a software-based medical device (SaMD); (3) demonstration of clinical utility beyond existing guideline-based risk stratification. Thyroid function testing (TSH, FT4) is Medicare-rebatable (MBS Item 66719) when clinically indicated, and TPO antibody testing is available. The PBS lists levothyroxine for hypothyroidism management. This model does not replace biochemical testing but would theoretically identify who to test — a role currently served by clinical risk factor assessment. Pregnant women undergoing antenatal care, particularly those with identifiable risk factors for thyroid disease (advanced maternal age, autoimmune disease history, prior hypertensive disorders of pregnancy, multiparity, scarred uterus)
Abstract
BACKGROUND International guidelines recommend early screening based on targeted risk factors to identify thyroid disease during pregnancy. The complexity of these risk factors makes accurate prediction challenging. This study aimed to develop and compare multiple machine-learning-based predictive models for thyroid disease during pregnancy. MATERIAL AND METHODS This retrospective study analyzed the clinical characteristics of 5461 women who gave birth at a single center. The dataset was divided into training and test sets. In the training set, feature variables associated with thyroid disease during pregnancy were selected using the Boruta algorithm. Eight models were developed: logistic regression, Bayesian approach, k-nearest neighbors, support vector machine, neural network, classification and regression tree, extreme gradient boosting, and random forest (RF). Model performance was evaluated using the receiver operating characteristic (ROC) curve, precision-recall curve (PRC), calibration curve, and decision curve analysis. RESULTS Nine feature variables were identified: age, height, pre-pregnancy weight, gravidity, parity, primiparity or multiparity, hypertensive disorders of pregnancy, scarred uterus, and autoimmune disease. The RF model demonstrated the best performance, with accuracy of 0.98387819 and 0.99597990, Matthews correlation coefficient of 0.96794139 and 0.97781292, log loss of 0.12670703 and 0.09442025, Brier score of 0.02495798 and 0.01921069, area under the ROC curve of 0.99877170 and 0.99991140, and area under the PRC of 0.99864486 and 0.99922572 in the training and test sets, respectively. CONCLUSIONS The RF model demonstrates excellent discriminative performance, accuracy, consistency, and generalizability in predicting thyroid disease during pregnancy.
References
- 1.Yang, G., Gao, Y., Liu, P., Jiang, J., Qiao, H., & Sheng, W. (2026). Development and validation of machine-learning-based prediction models for thyroid diseases during pregnancy. Medical Science Monitor: International Medical Journal of Experimental and Clinical Research. https://doi.org/10.12659/MSM.953235
Related Research
Acta obstetricia et gynecologica Scandinavica
ChatGPT in urogynecology: Comparing large language model responses to human experts
3 Aug 2026
Ultrasound in medicine & biology
A Systematic Review: The Application of Attention Mechanisms in Medical Ultrasound Image Processing
2 Aug 2026
Preventing chronic disease
Reducing Rates of Cesarean Delivery in Rural US Communities: A Systematic Review of Interventions and Approaches to Care
31 July 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service