A data-driven health index for neonatal morbidities.

Area of Science:

• Neonatal Medicine
• Artificial Intelligence
• Public Health

Background:

• Prematurity is a leading cause of neonatal mortality and long-term impairment.
• Current definitions of prematurity rely on gestational age or birthweight, which do not fully capture neonatal morbidity.
• A more comprehensive approach is needed to identify and manage risks associated with premature birth.

Purpose of the Study:

• To develop and validate a multi-task deep neural network model for predicting twelve neonatal morbidities.
• To establish a data-driven approach for defining prematurity based on predicted neonatal morbidities.
• To improve the assessment of prematurity and inform clinical decision-making.

Main Methods:

• Utilized a large dataset of 11,594,786 livebirths in California (1991-2012).
• Integrated maternal demographics, medical history, obstetrical complications, and prenatal fetal findings.
• Developed a multi-task deep neural network to simultaneously predict multiple neonatal morbidities.

Main Results:

• The proposed deep neural network model demonstrated superior performance compared to traditional gestational age and birthweight-based models.
• The model achieved a higher area under the precision-recall curve (0.326) than gestational age (0.229) and small for gestational age (0.156) models.
• Successfully predicted twelve distinct neonatal morbidities.

Conclusions:

• Machine learning techniques hold significant potential for predicting multiple prematurity phenotypes.
• This data-driven approach can enhance clinical decisions for preventing, diagnosing, and treating neonatal morbidities.
• A new, morbidity-focused definition of prematurity can be established using advanced predictive modeling.