From clusters to clinical rules: unsupervised machine learning identifies four newborn hearing phenotypes with...

Area of Science:

• Neonatal Health
• Machine Learning in Pediatrics
• Hearing Screening Outcomes

Background:

• Hearing impairment affects 1-3 per 1000 newborns, with higher prevalence in high-risk infants.
• Early identification of risk factors is crucial for timely intervention and improved outcomes.
• Traditional risk stratification methods may not capture complex synergistic effects.

Purpose of the Study:

• To apply unsupervised machine learning to identify distinct risk patterns associated with hearing screening failure in high-risk newborns.
• To translate identified phenotypes into clinically applicable rules for risk stratification.

Main Methods:

• Retrospective cohort study of 447 high-risk newborns.
• Partition around medoids clustering using Gower distance on demographic, perinatal, and diagnostic data.
• Classification and Regression Trees (CART) decision tree and co-occurrence network analyses for rule generation and synergy exploration.

Main Results:

• Four distinct neonatal phenotypes were identified, with significant variations in hearing screening failure rates (20.5% to 48.2%).
• A decision tree generated simple bedside rules, predicting high failure risk (e.g., males < 1.795 kg with 83.3% failure).
• Co-occurrence network analysis highlighted a strong association between respiratory disorders, infection, and preterm/low birth weight, while jaundice remained isolated.

Conclusions:

• Unsupervised clustering effectively identified unique neonatal phenotypes and risk co-occurrence patterns in a high-risk cohort.
• Complementary analyses provided clinically actionable rules and revealed synergistic risk factors missed by traditional logistic regression.
• Pattern-based risk stratification using machine learning is a valuable addition to variable-centered approaches for neonatal care.