Inferring Optimised Pregnancy Conception Dates from Multiple Gestational Age Estimates: A Hybrid Anomaly Detection and Linear Mixed Model Approach

Technical Summary: Inferring Optimised Pregnancy Conception Dates from Multiple Gestational Age Estimates

Problem Statement
Accurate estimation of gestational age (GA) is critical for perinatal research, pharmacovigilance, and clinical decision-making, particularly for classifying outcomes like preterm birth and small-for-gestational-age (SGA) infants. However, in low-resource settings such as Kenya, routine pregnancy data is often characterized by inconsistent records, limited access to early ultrasound, and conflicting estimates derived from multiple methods (e.g., Last Menstrual Period [LMP], fundal height, later ultrasounds, and postnatal assessments). Traditional deterministic rules or manual expert review used to resolve these conflicts do not scale to large observational registries and fail to adequately address systematic biases and data quality variations inherent in real-world datasets. The core challenge is how to synthesize multiple, often contradictory, and uncertain GA measurements into a single, reliable conception date for individual pregnancies without a universal "gold standard" for every case.

Methodology
The authors propose a two-stage hybrid approach combining machine learning anomaly detection with statistical modelling, implemented on data from the MiMBa Pregnancy Exposure Registry across 42 healthcare facilities in Homa Bay County, Kenya.

1. Stage 1: Anomaly Detection (Isolation Forest)

   • Objective: To identify and exclude implausible GA estimates before modelling.
   • Algorithm: The study employs the Isolation Forest (IF), an unsupervised anomaly detection algorithm. It operates on the principle that anomalies are "few and different," requiring fewer random splits to isolate in feature space compared to normal observations.
   • Implementation: Separate IF models were trained for distinct pregnancy outcome groups (Term Live Birth, Preterm Live Birth, Stillbirth, Miscarriage) to account for varying GA distributions.
   • Feature Engineering: Inputs included raw GA estimates (LMP, Ultrasound, fundal height, Ballard score, foot length) and derived features such as GA at expected delivery date, GA at actual delivery, z-scores relative to population distributions, and deviations from expected biological ranges.
   • Optimization: Hyperparameters (contamination rate, number of estimators, max samples) were tuned via grid search to maximize the silhouette score, ensuring optimal separation between normal and anomalous clusters.
   • Validation: A 5-fold cross-validation scheme was used, and flagged anomalies underwent manual review by clinicians to verify data entry errors.

2. Stage 2: Linear Mixed-Effects Modelling

   • Objective: To synthesize the remaining plausible measurements into a single, calibrated "best" GA estimate for each pregnancy.
   • Model Structure: A linear mixed-effects model was fitted with the observed GA at the outcome date as the dependent variable.
     • Fixed Effects: Included the GA estimation method (to capture systematic bias relative to first-trimester ultrasound) and pregnancy outcome category.
     • Random Effects: Included a pregnancy-specific random intercept to account for within-pregnancy correlation and individual variation.
   • Calibration: The model was anchored to first-trimester ultrasound (US1) as the reference standard.
   • Output: The model generates an Empirical Best Linear Unbiased Prediction (BLUP) for each pregnancy. This combines the fixed-effect bias correction with the random intercept, effectively "borrowing strength" from the entire dataset to refine estimates for individuals with sparse data. The final conception date is derived by subtracting this calibrated GA from the observed delivery date.

Key Results

• Data Quality & Anomalies: Anomaly rates were unevenly distributed, significantly higher in adverse outcomes (5.1% in miscarriages) compared to term live births (0.5%). The primary drivers of anomalies were data entry errors in antenatal clinic records (46.8%) and LMP recall errors.
• Method Precision and Bias:
  • Ultrasound: First-trimester ultrasound served as the gold standard. Later ultrasounds showed progressive systematic bias (increasing from 3.15 days in the second trimester to 6.41 days in the third) but retained high individual-level precision (Standard Deviation: 7.5–8.9 days).
  • Clinical/Postnatal Methods: Methods such as LMP, fundal height, Ballard score, and foot length exhibited minimal mean bias (< 2 days) at the population level but suffered from substantial individual-level uncertainty, with variability spreads exceeding 35 days.
• Model Performance: The hybrid approach successfully generated a unified, de-biased conception date for each pregnancy. By pooling information across the dataset, the model provided more reliable individual estimates than any single method could offer, even for women with limited or no early ultrasound data.

Significance and Claims
The paper claims that routine care datasets, often dismissed due to data quality issues, can yield reliable GA estimates when subjected to rigorous data cleaning and statistical modelling. The primary contribution is a scalable, data-driven framework that:

1. Resolves Conflicts: It synthesizes multiple conflicting dating methods into a single, calibrated metric without requiring a gold standard for every individual case.
2. Enables Pharmacovigilance: By providing accurate individual-level conception dates, the method supports the precise temporal alignment of drug exposures with fetal developmental windows, which is essential for pharmacovigilance registries and large-scale observational studies where clinical trial-level data controls are absent.
3. Bridges Resource Gaps: It offers a solution for low-resource settings where early ultrasound is scarce, allowing researchers to utilize existing digital health records for high-quality perinatal research.

The authors modestly note that while the methodology is robust for retrospective research, its translation into real-time clinical practice requires embedding these algorithms into interoperable Electronic Medical Records (EMRs) to provide point-of-care decision support. They also acknowledge limitations regarding the scarcity of early ultrasound for calibration in rare outcome groups and the need for external validation in diverse settings.