Enhancing Maternal Health Surveillance in the United States Through Natural Language Processing

Comprehensive Summary

This retrospective, multicenter study investigated whether natural language processing (NLP) applied to unstructured electronic health record (EHR) text can enhance the accuracy, timeliness, and scope of U.S. maternal and neonatal health surveillance compared to traditional administrative data sources. Rather than focusing on prediction alone, the authors propose transformer-based NLP and related machine-learning language models to perform automated clinical concept extraction and surveillance from free-text EHR documentation. The article utilizes evidence from prior studies analyzing tens of thousands to over a million clinical notes, radiology reports, and EHR records drawn from hospital systems, national databases, and federal health systems across varying time frames. Preprocessing approaches described across cited work include clinical note annotation, rule-based and probabilistic text mining, and fine-tuning transformer models to recognize diagnoses, severity, temporality, and social determinants of health. Models discussed include rule-based NLP systems, random forest classifiers with NLP features, BERT-based transformers, and generative LLM pipelines, typically benchmarked against ICD-10 coding, administrative datasets, or manual chart review as the reference standard. Across exemplars, best-performing NLP models achieved strong discrimination, with reported AUROCs ranging from 0.76–0.99, sensitivities often exceeding 90%, and substantial gains over ICD-based surveillance, particularly for nuanced or underreported conditions. The analysis showed that traditional surveillance systems systematically don’t fully capture obstetric complications, with studies demonstrating that only a minority of true events are identified through routine reporting. NLP-enhanced approaches consistently identified a larger proportion of clinically meaningful events, including severe disease subtypes and contextual factors absent from structured fields. Secondary analyses across cited studies included external validation, subgroup testing, precision–recall evaluation, and real-time deployment, with several systems maintaining high sensitivity when transferred to new institutions without retraining. Additional results highlighted improved sensitivity, PPV, and F1 scores when NLP-derived features were combined with structured EHR data, and superior performance in capturing disease severity, laterality, and social context. Limitations include reliance on retrospective data, variable annotation quality, potential selection bias, uneven demographic representation, and the fact that many models were validated outside obstetrics. External validation was performed in several referenced studies, though fairness and equity analyses were inconsistently reported.

Outcomes and Implications

This work suggests that transformer-based NLP and related language models can substantially enhance maternal health surveillance by converting unstructured clinical narratives into high-fidelity, real-time population health signals. Clinically, such systems could enable earlier detection of emerging trends in preeclampsia, hemorrhage, preterm birth, and stillbirth, while incorporating social determinants of health that drive inequities yet remain invisible in administrative data. However, translation to bedside and policy impact remains indirect and dependent on validation, privacy safeguards, workflow integration, and bias mitigation. With careful implementation, NLP-enhanced surveillance has the potential to shift obstetric monitoring from delayed, low-resolution reporting toward proactive, data-rich systems that better inform clinical care and public health decision-making.