Exploring Novel Data-Driven Clustering Methods for Uncovering Patterns in Longitudinal Neonatal Postoperative...

Area of Science:

• Pediatric Cardiology
• Neonatal Surgery
• Machine Learning in Medicine

Background:

• Postoperative temperature management is critical for neonates with congenital heart defects (CHDs) following cardiopulmonary bypass (CPB).
• Understanding distinct temperature trajectories can inform risk stratification and improve patient outcomes.

Purpose of the Study:

• To identify unique postoperative temperature patterns in neonates with CHDs after CPB using unsupervised machine learning.
• To compare the performance of different clustering methods (GBTM, SOM, k-means) in identifying these temperature trajectories.
• To evaluate the prognostic value of identified temperature clusters on postoperative complications.

Main Methods:

• A secondary analysis of prospective data from 450 neonates who underwent CPB was conducted.
• Group-based trajectory modeling (GBTM), self-organizing maps (SOM), and k-means clustering were used to identify 3 postoperative temperature clusters (persistent hypothermia, resolving hypothermia, normothermia).
• The association between temperature clusters and a composite outcome of postoperative complications was assessed using multivariable logistic regression.

Main Results:

• All three clustering methods identified distinct temperature trajectories: persistent hypothermia, resolving hypothermia, and normothermia.
• Strong agreement was observed between GBTM and SOM (κ=0.92), while agreement between GBTM and k-means was weaker (κ=0.41).
• Neonates in the persistent hypothermia group had significantly higher odds of postoperative complications compared to normothermic neonates in GBTM (OR 2.8) and SOM (OR 2.3) models.

Conclusions:

• Machine learning techniques effectively delineate distinct postoperative temperature trajectories in neonates after CPB.
• Persistent hypothermia post-CPB is a significant predictor of adverse outcomes in this vulnerable population.
• GBTM and SOM demonstrate strong concordance and prognostic value for identifying high-risk neonates based on temperature patterns.