The Prediction of Preterm Birth Using Time-Series Technology-Based Machine Learning: Retrospective Cohort Study

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Global preterm birth rates are rising, necessitating effective screening.
• Cervical-length ultrasonography is effective but costly for universal screening.

Purpose of the Study:

• To develop an improved preterm birth prediction model using time-series analysis of obstetric data.
• To leverage continuous electronic medical record (EMR) data for enhanced screening.

Main Methods:

• Utilized long short-term memory (LSTM) networks on EMR data from 5187 pregnant women.
• Analyzed over 25,000 obstetric records from early pregnancy to 28 weeks.
• Assessed model performance using Area Under the Curve (AUC), accuracy, sensitivity, and specificity.

Main Results:

• The time-series LSTM model outperformed traditional cross-sectional methods in prediction.
• Achieved an accuracy of 0.739, sensitivity of 0.407, specificity of 0.982, and AUC of 0.651.
• Identified blood pressure, glucose, lipids, and uric acid as significant predictors of preterm birth.

Conclusions:

• Time-series modeling offers advantages for preterm birth prediction.
• Findings can inform guidelines for preterm birth prevention and treatment.
• The model aids clinicians in making informed decisions during obstetric care.