68/100

Moderate evidence

Cite
More
EmailLinkedInX

Thyroid cancer risk model beat ultrasound scoring system in a 561-patient study

Study title: Efficacy of machine learning model and EU-TIRADS-PL reporting system in predicting thyroid malignancy risk: a comparison

Evidence type
Controlled or validated comparison
Sample
n = 561
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Efficacy of machine learning model and EU-TIRADS-PL reporting system in predicting thyroid malignancy risk: a comparison

A prospective study of 561 patients who had thyroidectomy built a machine learning model to estimate thyroid malignancy risk from ultrasound, fine-needle biopsy and laboratory data. The final logistic regression model reached a test-set ROC AUC of 74% and correlated better with malignancy than the EU-TIRADS-PL reporting system.

Read the study abstract

Thyroid nodules are common and mostly benign, but 5-15% are malignant. Increasing detection rates may lead to overdiagnosis and unnecessary fine-needle aspiration biopsies (FNAB). We aimed to develop a machine learning (ML) model for thyroid malignancy risk assessment using ultrasound, FNAB, laboratory data and to compare its performance with the EU-TIRADS-PL reporting system. This prospective study included 561 patients who underwent thyroidectomy based on preoperative ultrasound and FNAB results, including 261…

Efficacy of machine learning model and EU-TIRADS-PL reporting system in predicting thyroid malignancy risk: a comparison
"Doppler Ultrasound and Vascular Sonography Examination in Navi Mumbai, Maharashtra, India" by Goleisureintl, CC BY 4.0.

The gain

In a prospective study of 561 thyroidectomy patients, a logistic regression model reached a test-set ROC AUC of 74% and correlated better with thyroid malignancy than the EU-TIRADS-PL ultrasound reporting system.

In brief

Thyroid nodules are common and mostly benign, but 5-15% are malignant, and rising detection may lead to overdiagnosis and unnecessary biopsies. A tool that sorts risk more accurately could help limit unneeded procedures, although the study did not measure that outcome.

The model flagged several predictors: microcalcifications, younger age, male sex and a larger anteroposterior dimension. A larger longitudinal dimension was linked to lower risk. A separate anteroposterior-to-longitudinal index reached an AUC of 68%.

Main points

  1. The study enrolled 561 patients who underwent thyroidectomy after preoperative ultrasound and fine-needle biopsy.
  2. Of the nodules, 261 (46.5%) were malignant on histopathology.
  3. Data were split 3:1 into training and test sets, and several ML algorithms were compared.
  4. Logistic regression was chosen as the final model for its performance and interpretability.
  5. The model's test-set ROC AUC was 74%.
  6. Microcalcifications (p=0.002), younger age (p<0.001), male sex (p=0.001) and larger anteroposterior dimension (p=0.026) predicted malignancy.
  7. A larger longitudinal dimension was associated with lower malignancy risk (p<0.001).
  8. An anteroposterior-to-longitudinal dimension index had an ROC AUC of 68% (95% CI 62.0-71.0).

The full story

Researchers recruited 561 patients who underwent thyroidectomy based on preoperative ultrasound and fine-needle aspiration biopsy results. Malignancy was verified by histopathology and was present in 261 nodules (46.5%).

The team evaluated multiple machine learning algorithms using clinical, sonographic, cytological and biochemical variables, with data divided into training and test sets at a 3:1 ratio. Logistic regression was selected because it performed well and could be interpreted.

On the test set the model reached a ROC AUC of 74% and tracked malignancy better than EU-TIRADS-PL. The authors conclude it may improve risk assessment compared with systems based only on sonographic features.

Sources

1 on file
  1. Endocrine Practicedoi:10.1016/j.eprac.2026.10.016
    Peer-reviewed2026-11-13

Talk it over

Three questions to start on, two more if you have time
The model reached a test-set ROC AUC of 74% and tracked malignancy better than EU-TIRADS-PL. How would you weigh that gain against the researchers' own description of 74% as modest?

Start with a question

Thyroid cancer risk model beat ultrasound scoring system in a 561-patient studyhttps://truvace.com/item/efficacy-of-machine-learning-model-and-eu-tirads-pl-reporting-system-in-predicti

  1. The model reached a test-set ROC AUC of 74% and tracked malignancy better than EU-TIRADS-PL. How would you weigh that gain against the researchers' own description of 74% as modest?

    Answer in the thread
  2. Only patients who went on to thyroidectomy were included, and 261 of the 561 nodules (46.5%) were malignant. How might that group differ from people with thyroid nodules who never had surgery?

    Answer in the thread
  3. The researchers chose logistic regression for its performance and because it could be interpreted. Who should decide whether a clinic uses a model like this, and what would they need to see first?

    Answer in the thread
If you have more time
  • The study did not measure whether the model would cut unneeded biopsies. What kind of study could test that, and what result would settle it?

    Answer in the thread
  • The model flagged microcalcifications, younger age, male sex and a larger anteroposterior dimension. How would you check whether a clinic's own patients show the same pattern?

    Answer in the thread

What readers say