Closing the Gap in High-Risk Pregnancy Care Using Machine Learning and Human-AI Collaboration

Hussein Mozannar, Yuria Utsumi, Irene Y. Chen, Stephanie S. Gervasi, Michele Ewing, Aaron Smith-McLallen, David A. Sontag

Introduction

High-risk pregnancies, characterized by complications such as pre-eclampsia or gestational diabetes, are significant contributors to maternal and infant morbidity. While clinical interventions can mitigate these risks, timely identification remains a critical bottleneck. Health insurers typically rely on standardized medical billing codes to identify pregnant members; however, these codes are often reactive, appearing only after a patient is well into their second trimester or already experiencing complications. This research presents a machine learning-based system designed to reduce the latency of pregnancy identification and provide interpretable risk stratification for nurse care managers.

Methods

The researchers developed a two-stage pipeline integrated into a human-centric interface to address the challenges of identification and interpretability.

HAPI: Hybrid Algorithm for Pregnancy Identification

The core of the identification system is HAPI, a hybrid model that combines high-certainty clinical "anchor" codes (e.g., positive pregnancy tests) with a machine learning classifier. When anchor codes are unavailable, the system utilizes a Lasso-regularized logistic regression model trained on over 62,000 features, including medical history and prescriptions. To ensure stable predictions, an Exponential Moving Average (EMA) is applied to smooth probability scores over time.

The HAPI model architecture and its ability to predict pregnancy earlier than standard clinical codes.

Complication Prediction and Human-AI Interface

Once a pregnancy is identified, a second classifier (Lasso) predicts the risk of Gestational Diabetes (GDB) and Gestational Hypertension (GHT). To bridge the gap between automated scores and clinical action, the team developed a Patient Dashboard. Using a GROUP-Lasso approach, the dashboard highlights specific ICD codes and historical data that contributed to a patient's risk score, providing nurses with actionable evidence rather than a "black box" number.

The Patient Dashboard designed to provide nurses with interpretable evidence for risk stratification.

Datasets

The study utilized longitudinal data from Independence Blue Cross (IBC):

  • Identification Dataset: 36,735 female patients (ages 18–48), including healthy pregnancies, those with complications, and a non-pregnant control group.
  • Complications Dataset: 12,243 pregnant patients, with a 26.4% complication rate (GDB or GHT).
  • Features: 62,734 features derived from claims, pharmacy, and lab data, aggregated into 5-day and 10-day temporal windows.

Key Results

  • Predictive Performance: The complication classifier achieved an AUC of 0.76 and an accuracy of 76.8%.
  • Earlier Detection: In cases where the ML model outperformed standard billing codes, it identified pregnancies an average of 54 to 66 days earlier.
  • Workflow Impact: User studies with nurses indicated that the AI-augmented dashboard improved the accuracy of identifying high-priority patients and significantly reduced the time required to review patient histories.
  • Fairness Audit: The model demonstrated a higher AUC for Black patients (0.787) compared to White patients (0.740), though raw accuracy varied due to differing base rates of complications.

Limitations and Future Work

A primary challenge is data sparsity; approximately 60% of complications are not flagged early enough because patients lack sufficient claims data at the start of pregnancy. Future research will focus on moving beyond telephonic care management toward integrated clinical interventions and ensuring model robustness as billing patterns and healthcare technologies evolve.

Summary updated

Abstract

High-risk pregnancy (HRP) is a pregnancy complicated by factors that can adversely affect outcomes of the mother or the infant. Health insurers use algorithms to identify members who would benefit from additional clinical support. We aimed to build machine learning algorithms to identify pregnant patients and triage them by risk of complication to assist care management. In this retrospective study, we trained a hybrid Lasso regularized classifier to predict whether a patient is currently pregnant using claims data from 36735 insured members of Independence Blue Cross (IBC), a health insurer in Philadelphia. We then train a linear classifier on a subset of 12,243 members to predict whether a patient will develop gestational diabetes or gestational hypertension. These algorithms were developed in cooperation with the care management team at IBC and integrated into the dashboard. In small user studies with the nurses, we evaluated the impact of integrating our algorithms into their workflow. We find that the proposed model predicts an earlier pregnancy start date for 3.54% (95% CI 3.05-4.00) for patients with complications compared to only using a set of pre-defined codes that indicate the start of pregnancy and never later at the expense of a 5.58% (95% CI 4.05-6.40) false positive rate. The classifier for predicting complications has an AUC of 0.754 (95% CI 0.764-0.788) using data up to the patient's first trimester. Nurses from the care management program expressed a preference for the proposed models over existing approaches. The proposed model outperformed commonly used claim codes for the identification of pregnant patients at the expense of a manageable false positive rate. Our risk complication classifier shows that we can accurately triage patients by risk of complication.
BibTeX
@misc{mozannar2023closing,
  author = {Hussein Mozannar and Yuria Utsumi and Irene Y. Chen and Stephanie S. Gervasi and Michele Ewing and Aaron Smith-McLallen and David A. Sontag},
  title = {{Closing the Gap in High-Risk Pregnancy Care Using Machine Learning and Human-AI Collaboration}},
  year = {2023},
  eprint = {2305.17261},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.17261},
}
Figure