Machine Learning Approach with Harmonized Multinational Datasets for Enhanced Prediction of Hypothyroidism in...

Area of Science:

• Endocrinology
• Computational Biology
• Public Health

Background:

• Type 2 diabetes (T2D) is a growing global health issue.
• Comorbid hypothyroidism (HT) worsens T2D complications, but screening is infrequent, delaying diagnosis and treatment.
• Early HT detection in T2D patients can mitigate severe health risks.

Purpose of the Study:

• To develop a machine learning algorithm (MLA) for classifying T2D patients into low or high risk for developing comorbid hypothyroidism (HT).
• To improve the efficiency and timeliness of HT screening in T2D patients.
• To enhance clinical management and patient outcomes through early risk stratification.

Main Methods:

• Development of a machine learning algorithm (MLA) using multinational patient data from NIH All of Us (AoU) and UK Biobank (UKBB).
• Training and validation of the MLA on a combined dataset to assess its predictive performance.
• Evaluation of the MLA's Negative Predictive Value (NPV) and Area Under the Receiver Operating Characteristic Curve (AUROC).

Main Results:

• The MLA achieved a high Negative Predictive Value (NPV) of 0.989 and an AUROC of 0.762 on the combined dataset.
• MLA performance on the combined dataset surpassed models trained on individual datasets (AoU: 0.666, UKBB: 0.622).
• Increased dataset diversity during MLA training significantly improved its predictive accuracy.

Conclusions:

• The developed MLA is a high-NPV clinical decision tool for identifying T2D patients at low risk for HT, enabling focused lab testing.
• This automated tool can supplement standard care screening, allowing clinicians to prioritize high-risk individuals for timely intervention.
• Early identification of HT risk in T2D patients facilitates tailored management, ultimately improving patient health outcomes.