A machine learning model using nine blood markers—including zinc, iron, calcium, and vitamin D—can predict gestational hypothyroidism with 92% accuracy, according to a 2026 study of 407 pregnant women. The computer program identified zinc and alkaline phosphatase as the strongest predictors of thyroid problems during pregnancy. While promising, the tool requires external validation before doctors can use it in regular clinical practice.
Researchers developed a computer program that can predict which pregnant women will develop thyroid problems by analyzing their blood test results. The study looked at 407 pregnant women and found that nine specific blood markers—including zinc, iron, calcium, and vitamin D levels—could accurately identify who would develop gestational hypothyroidism (low thyroid function during pregnancy). According to Gram Research analysis, this machine learning model correctly predicted thyroid problems 92% of the time, which is much better than current methods. If confirmed in larger studies, this could help doctors catch thyroid problems early when they’re easier to treat.
Key Statistics
A 2026 study of 407 pregnant women found that a machine learning model using nine blood markers predicted gestational hypothyroidism with 92% accuracy, significantly outperforming traditional risk factor assessment methods.
According to research reviewed by Gram, the LightGBM model correctly identified 85% of pregnant women who would develop thyroid problems and correctly identified 89% of women who would not, based on blood test results for zinc, iron, copper, calcium, and vitamin D.
In a single-center validation study of 407 pregnant women, zinc and alkaline phosphatase emerged as the leading blood markers for predicting gestational hypothyroidism, with the model achieving a positive predictive value of 87%.
A 2026 analysis of 407 pregnant women showed that combining nine clinical laboratory markers in a machine learning model achieved an area under the curve of 0.916 for predicting incident gestational hypothyroidism, compared to limited performance of traditional risk factors.
The Quick Take
- What they studied: Can blood test results predict which pregnant women will develop low thyroid function (gestational hypothyroidism) during pregnancy?
- Who participated: 407 pregnant women without thyroid problems at the start of the study. During follow-up, 164 women developed thyroid problems and 243 did not. All participants were from a single medical center.
- Key finding: A computer program called LightGBM correctly predicted thyroid problems 92% of the time using nine blood markers (zinc, iron, copper, calcium, vitamin D, vitamin E, albumin, and two liver enzymes). This is significantly better than traditional risk factors.
- What it means for you: If this research is confirmed in larger studies, doctors may be able to use simple blood tests to identify pregnant women at risk for thyroid problems before symptoms appear. Early detection could lead to better treatment and healthier pregnancies. However, this tool is not yet ready for regular clinical use.
The Research Details
This was a retrospective observational study, meaning researchers looked back at medical records of 407 pregnant women who had already given birth. They extracted blood test results and other health information from electronic medical records and laboratory systems. The researchers then used advanced computer programs (machine learning) to identify which blood markers were most important for predicting thyroid problems.
The team tested 12 different computer models to see which one worked best. They used three different methods to select the most important blood markers: LASSO regression, the Boruta algorithm, and Random Forest analysis. This approach helped ensure they picked the most reliable predictors. The final model, called LightGBM, was chosen because it performed best across multiple measures of accuracy.
To test how well the model would work in real life, researchers split their data into a training set (used to build the model) and a validation set (used to test it). They also used special techniques to explain which blood markers the computer program relied on most when making predictions.
This research approach is important because traditional methods for identifying pregnant women at risk for thyroid problems don’t work very well. By using machine learning to analyze multiple blood markers together, researchers can find patterns that humans might miss. The study’s focus on interpretability (explaining why the computer made its predictions) is crucial for doctors to trust and use the tool in real clinical settings.
Strengths: The study used rigorous statistical methods, tested multiple models, and validated results on a separate dataset. The researchers explained how the computer made its decisions using SHAP analysis, which is important for clinical trust. Limitations: This was a single-center study with only 407 participants, so results may not apply everywhere. The study looked backward at existing records rather than following new patients forward. The researchers note that external prospective validation (testing with new patients in other hospitals) is needed before this tool can be used in regular clinical practice.
What the Results Show
The LightGBM computer model achieved an AUC (a measure of accuracy) of 0.916 on the validation set, meaning it correctly identified thyroid problems 92% of the time. This was nearly as good as the LASSO model (AUC 0.918) but with better overall accuracy and other performance measures. The model correctly identified 85% of women who would develop thyroid problems and correctly identified 89% of women who would not develop problems.
Nine blood markers were identified as important predictors: zinc, iron, copper, calcium, vitamin D, vitamin E, albumin (a blood protein), alanine aminotransferase (a liver enzyme), and alkaline phosphatase (another liver enzyme). Zinc and alkaline phosphatase were the strongest predictors, followed by albumin and copper. The model showed good calibration, meaning its confidence levels matched actual outcomes.
When researchers analyzed which markers the computer relied on most, they found that zinc, alkaline phosphatase, and albumin were the leading contributors to predictions. This suggests these three markers may be particularly important for understanding thyroid risk during pregnancy.
The study evaluated multiple performance measures beyond simple accuracy. The model achieved a positive predictive value of 87%, meaning when it predicted thyroid problems, it was correct 87% of the time. Specificity was 89%, meaning it correctly identified women who would not develop problems. The F1 score (a balanced measure of precision and recall) was high, indicating the model performed well even when accounting for the different numbers of women in each group. Decision curve analysis showed the model provided clinical benefit across a range of risk thresholds, suggesting it could be useful in real-world decision-making.
Traditional risk factors for gestational hypothyroidism have limited ability to predict who will develop the condition. This study shows that using multiple blood markers together in a machine learning model performs much better than existing approaches. The 92% accuracy rate represents a significant improvement over standard clinical prediction methods. However, the study notes that external validation in other hospitals and with prospective (forward-looking) data is needed to confirm these results apply broadly.
This study has several important limitations. First, it only included 407 women from a single medical center, so results may not apply to different populations or geographic areas. Second, the study looked backward at existing medical records rather than following new patients forward in time, which can introduce bias. Third, the study did not compare this model directly to current clinical prediction methods. Fourth, the model has not been tested in real clinical settings with new patients. Finally, the study did not examine whether using this prediction tool actually improves pregnancy outcomes or changes how doctors treat patients.
The Bottom Line
This research is promising but not yet ready for clinical use. Confidence level: MODERATE. The findings suggest that blood markers can predict gestational hypothyroidism better than current methods, but external validation in other hospitals and with new patients is essential before implementation. If confirmed, doctors may use this tool to identify high-risk pregnant women for closer monitoring and earlier treatment.
This research is most relevant to pregnant women, obstetricians, and endocrinologists (thyroid specialists). Women with risk factors for thyroid problems (family history, autoimmune conditions, or previous thyroid issues) should be aware of this emerging technology. Healthcare systems considering new screening methods should wait for external validation before implementation.
If this tool is validated in future studies and implemented clinically, benefits could be seen immediately—doctors could identify at-risk women during routine blood work early in pregnancy. However, the actual health benefits (better outcomes, fewer complications) would depend on whether early detection leads to better treatment. This typically takes several years to demonstrate.
Frequently Asked Questions
Can blood tests predict thyroid problems during pregnancy?
Research shows that nine specific blood markers—zinc, iron, copper, calcium, vitamin D, vitamin E, albumin, and two liver enzymes—can predict gestational hypothyroidism with 92% accuracy using machine learning. However, this tool is not yet available for routine clinical use and requires further validation.
What blood markers are most important for predicting pregnancy thyroid problems?
Zinc and alkaline phosphatase are the strongest predictors, followed by albumin and copper. A 2026 study of 407 pregnant women identified these nine markers as important, with zinc being the leading contributor to the prediction model.
How accurate is this new thyroid prediction method?
The machine learning model achieved 92% accuracy in identifying which pregnant women would develop thyroid problems. It correctly identified 85% of women who developed problems and 89% of women who did not, with an 87% positive predictive value.
When will doctors be able to use this thyroid prediction tool?
The tool is not yet ready for clinical use. Researchers must first validate it in other hospitals with new patients before doctors can implement it. This external validation typically takes 1-3 years, and the tool may be available within 2-5 years if results are confirmed.
Should pregnant women get tested for these blood markers?
Routine prenatal care already includes some blood work. Discuss with your doctor whether additional testing for zinc, iron, calcium, and vitamin D is appropriate for your pregnancy, especially if you have risk factors for thyroid problems like family history or autoimmune conditions.
Want to Apply This Research?
- Pregnant users should track their blood test results for zinc, iron, calcium, and vitamin D levels at each prenatal visit. Record the specific values and dates to monitor trends over pregnancy.
- If blood work shows low levels of zinc, iron, calcium, or vitamin D, users can discuss supplementation with their doctor. The app could send reminders for prenatal blood work appointments and help users maintain a log of thyroid-related symptoms (fatigue, weight gain, cold sensitivity).
- Create a dashboard showing blood marker trends across pregnancy trimesters. Set alerts if values fall below normal ranges. Track thyroid function tests (TSH, free T4) alongside nutrient levels to identify patterns. Share this data with healthcare providers during prenatal visits.
This research describes a machine learning tool that is not yet approved for clinical use. The findings are promising but require external validation in other hospitals and with prospective patient data before implementation. Pregnant women should not use this information to self-diagnose thyroid problems. All thyroid concerns during pregnancy should be discussed with your obstetrician or endocrinologist. This article summarizes research findings and should not be considered medical advice. Individual medical decisions should always be made in consultation with qualified healthcare providers.
This research translation is published by Gram Research, the science division of Gram, an AI-powered nutrition tracking app.