Enhancing TSH-based congenital hypothyroidism screening using machine learning and resampling algorithms.

Area of Science:

• Medical screening
• Machine learning applications
• Neonatal health

Background:

• Congenital hypothyroidism (CH) is a leading cause of preventable intellectual disability.
• Current newborn screening for CH relies on thyroid-stimulating hormone (TSH), facing challenges with low positive predictive value (PPV).
• Previous machine learning attempts for CH screening were hindered by data imbalance and limited predictive features.

Purpose of the Study:

• To conduct a comprehensive evaluation of machine learning algorithms for congenital hypothyroidism screening.
• To address the limitations of current TSH-based screening methods, specifically low PPV.
• To develop a more accurate and efficient screening model for CH.

Main Methods:

• Analysis of data from 616,910 infants screened between 2019 and 2024.
• Training and evaluation of 576 distinct machine learning models using 12 classification and 12 resampling algorithms.
• Optimization for sensitivity and PPV using stratified 5-fold cross-validation and assessment of model explainability via SHAP values.

Main Results:

• A RUSBoost classifier with Gaussian Noise resampling achieved 100% sensitivity and 16.8% PPV.
• This represents a 60% improvement in PPV compared to current screening approaches.
• TSH remained the primary predictor, but the model incorporated additional features to enhance performance.

Conclusions:

• Machine learning models demonstrated no missed cases of CH and significantly improved screening performance.
• These algorithms offer a promising alternative to refine TSH-based CH screening.
• The findings suggest potential for reducing false positives, stress, and costs in global newborn screening programs.