Machine Learning Estimation of Gestational Age at Delivery Using Linked Mother-Infant Electronic Health Records Across Two Health Systems

This study demonstrates that supervised machine learning models trained on linked mother-infant electronic health records can accurately and generalizably estimate gestational age at delivery across different health systems, providing a robust framework to support large-scale maternal and neonatal health research.

The Gist

Imagine you are trying to guess exactly how long a pregnancy lasted just by looking at a patient's medical file. Sometimes, the file is missing the specific date the baby was due, or the notes are messy and hard to read. This is a big problem for researchers trying to study how drugs affect babies, because if they get the timing wrong, their whole study could be flawed.

This paper is about building a smart digital detective that can figure out the length of a pregnancy (gestational age) even when the official record is missing or unclear.

Here is how the researchers did it, explained simply:

1. The Training Ground: Two Big Libraries

The researchers didn't just look at one hospital's files; they used two massive libraries of medical records:

• Vanderbilt University Medical Center (VUMC) in Nashville.
• University of Michigan (UMich) in Ann Arbor.

Think of these as two different "training gyms." They took millions of mother-and-baby records from these gyms to teach their computer program how to spot patterns.

2. The Detective's Toolkit: What Clues Did It Use?

The computer program (a machine learning model) didn't just guess. It looked for specific clues in the files, kind of like a detective piecing together a puzzle. They tested three different "toolkits" to see which worked best:

• Toolkit A (Mom Only): Just looked at the mother's history (her age, race, past pregnancies).
• Toolkit B (Mom + Hospital Notes): Added general hospital data (like ICD codes, which are like shorthand labels for medical conditions).
• Toolkit C (The Full Package): Added the baby's data too! This included the baby's birth weight, their "Apgar score" (a quick health check right after birth), and the baby's own medical labels.

The Analogy: Imagine trying to guess how long a cake was baking.

• Toolkit A is like guessing based only on the baker's experience.
• Toolkit B is like looking at the recipe card.
• Toolkit C is looking at the baker, the recipe, and the finished cake's size and texture. Unsurprisingly, the Full Package (Toolkit C) was the most accurate.

3. The "Smart Guess" vs. The "Average Guess"

Before using their fancy AI, the researchers tried a simple method: just guessing the average pregnancy length for everyone.

• The Result: The simple average was often way off, like guessing every cake takes exactly 45 minutes regardless of size.
• The AI Result: The machine learning models were much sharper. They could predict the pregnancy length within one week of the true date about 85% to 93% of the time. Within two weeks, they were right 94% to 98% of the time.

4. The "Cross-City" Test

To make sure their detective wasn't just memorizing the Nashville library, they sent the same rules to the Michigan library.

• The Outcome: It worked just as well, and actually performed even better in Michigan. This proves the "detective" isn't just a local expert; it's a generalist that can work in different hospitals.

5. Where the Detective Stumbles

The paper is honest about where the system isn't perfect yet:

• Preterm Babies: The system is great at guessing the length of full-term pregnancies (babies born at the "right" time). However, it struggles a bit more with babies born very early (preterm). It's like the detective is good at solving standard cases but gets confused by rare, complex mysteries.
• Older Data: The system performed better on records from recent years. This might be because older records (from before 2015) used different coding systems or had less precise ultrasound technology, making the clues harder to read.

The Bottom Line

The paper concludes that we now have a reliable, portable "calculator" that can fill in the missing pregnancy dates in medical records. By using a mix of mother's history, hospital notes, and baby's details, this tool can help researchers study pregnancy safety with much more accuracy than before.

Important Note: The authors specifically state this is a tool for research to fix missing data in studies. They do not claim this tool should be used by doctors to make immediate clinical decisions for individual patients in a hospital setting right now. It is a way to clean up the data so scientists can learn more about maternal and baby health.

Technical Summary

Technical Summary: Machine Learning Estimation of Gestational Age at Delivery Using Linked Mother-Infant Electronic Health Records

Problem Statement
Accurate estimation of gestational age (GA) is critical for pregnancy and offspring health research, particularly in pharmacoepidemiologic studies where defining exposure windows and fetal susceptibility depends on precise pregnancy dating. However, electronic health record (EHR) and administrative claims databases frequently suffer from missing, incomplete, or erroneous GA data due to provider entry errors, lack of ultrasound documentation, or system infrastructure limitations. These inaccuracies can lead to misclassification of drug exposures, introducing bias and undermining study validity. While prior research has focused on estimating GA using administrative claims data, there is a need to leverage the richer, linked mother-infant data available in EHR systems to improve estimation accuracy and generalizability across diverse healthcare settings.

Methodology
This retrospective cohort study developed and validated supervised machine learning algorithms to estimate GA at delivery using linked mother-infant EHR data from two academic health systems: Vanderbilt University Medical Center (VUMC, 2005–2025) and the University of Michigan (UMich, 2012–2024).

• Data Sources and Cohort: The study utilized the Research Derivative at VUMC and the UMich EHR database. The initial cohorts included 91,471 (VUMC) and 55,733 (UMich) mother-infant dyads. After applying exclusion criteria (retaining one infant per mother, excluding non-live births, and requiring adequate prenatal care), the final analytic cohorts consisted of 54,344 dyads at VUMC and 34,345 at UMich.
• Reference Standard: Clinical Gestational Age (CGA) estimates recorded in maternal records at the time of delivery served as the ground truth. These textual patterns (e.g., "39 weeks 2/7") were parsed into numeric days.
• Predictor Variables: Features were derived from three sources:
  1. Maternal Records: Demographics (age, race/ethnicity), ICD codes for preterm/term/postterm birth, fetal growth restriction, and excessive fetal growth.
  2. EHR System Data: Standard encounter information and ICD code frequencies (encoded as distinct days or visits).
  3. Infant Records: Birthweight, 1-minute and 5-minute APGAR scores, and infant-specific ICD codes (preterm, SGA, LGA).
     Missing values for birthweight and APGAR scores were imputed using medians.
• Model Configurations: To evaluate performance across different data environments, three configurations were tested:
  • C1 (Maternal): Maternal predictors only.
  • C2 (Maternal + EHR): Maternal + standard EHR features.
  • C3 (Maternal + EHR + Pediatric): All available predictors, including infant records (representing linked data environments).
• Algorithms: Random Forest (RF), Gradient Boosting (GB), and an Ensemble model (averaging RF and GB predictions) were trained. Hyperparameters were optimized via 5-fold cross-validation. A quantile transformation was applied to address the left-skewed distribution of CGA.
• Evaluation: Performance was assessed on held-out test sets using the proportion of predictions within ±1 week and ±2 weeks of the reference standard, and Mean Absolute Error (MAE). Bootstrapping (1,000 resamples) was used to calculate 95% confidence intervals.

Key Results

• Overall Performance: The ensemble model achieved the highest accuracy. On the VUMC test set, it predicted GA within ±1 week for 85.2% of pregnancies and within ±2 weeks for 94.3%, with an MAE of 4.4 days. On the UMich test set, performance improved to 93.1% (±1 week) and 97.8% (±2 weeks), with an MAE of 2.8 days.
• Comparison to Baselines: Machine learning models substantially outperformed simple mean and median baselines. For instance, on the VUMC test set, the ML models improved MAE by 6.8 to 7.9 days compared to the mean baseline (12.3 days).
• Predictor Impact: Performance consistently improved as more data sources were added (C1 < C2 < C3). The inclusion of infant-specific data (birthweight, APGAR, pediatric ICD codes) provided the most significant boost in accuracy.
• Temporal and Stratified Analysis:
  • ICD Transition: Models performed better for pregnancies delivered on or after October 1, 2015 (ICD-10-CM era) compared to the ICD-9-CM era.
  • Delivery Status: The models performed best for full-term and late-term deliveries. Performance was notably lower for preterm deliveries.
  • Generalizability: The models trained on VUMC data and replicated on UMich data demonstrated strong portability, with the UMich replication showing even higher agreement with the reference standard.

Significance and Claims
The authors claim that this study presents a robust and generalizable machine learning framework for estimating gestational age at delivery using linked mother-infant EHRs. The primary significance lies in:

1. Leveraging Linked Data: Demonstrating that combining maternal and infant records (C3 configuration) significantly outperforms models relying solely on maternal or administrative data.
2. Portability and Reproducibility: Successfully replicating the modeling pipeline across two distinct health systems with different data histories and patient demographics, suggesting the learned associations are broadly generalizable rather than site-specific.
3. Utility for Research: Providing a reliable tool to impute GA in large-scale real-world clinical studies where GA data is missing or invalid, thereby reducing bias in pharmacoepidemiologic and perinatal research.

The paper acknowledges limitations, including suboptimal performance for preterm deliveries and the exclusion of non-live birth outcomes (stillbirth, termination) from the current model scope. The authors emphasize that while the framework is robust, future work is needed to address performance gaps in preterm estimation and to evaluate potential biases regarding health disparities.