JMIR Publications

ABSTRACT

Background:

Diabetes in pregnancy carries adverse outcomes extending to the later health of mothers and offspring. Although artificial intelligence (AI) has advanced medical care, its application to conditions requiring continuous management remains fragmented. Integrated care is essential from early pregnancy through the postpartum period, yet most AI studies focus on isolated tasks. No previous review has systematically categorized these AI applications by clinical purposes, methodologies, and points of care.

Objective:

We aimed to map AI research across clinical function and action stages to identify evidence gaps and inform strategies for integrated care in diabetes in pregnancy.

Methods:

Adhering to the PRISMA-ScR guidelines and JBI methodology, we searched MEDLINE, Embase, PubMed, and the Cochrane Library, supplemented by backward and forward citation searches. Eligible studies were peer-reviewed primary articles published in English between January 2023 and July 2026 that evaluated AI, including large language models (LLMs) and agentic AI, in women with or at risk of diabetes in pregnancy or in their offspring. Studies were categorized by clinical function (prediction, diagnosis, and management) and mapped across the pregnancy timeline into eight clinical action stages, from prevention of gestational diabetes mellitus (GDM) to long-term risk prediction. Evidence levels were graded according to the extent of clinical validation, from internal or controlled evaluation (L1) and external validation (L2) to clinical application of a fixed model (L3) and integration into routine care (L4). Two reviewers independently conducted full-text eligibility assessment, data charting, and classification, with discrepancies resolved by consensus.

Results:

From 4,849 records (1,192 from databases and 3,657 from citation searching), 128 studies were included. Prediction was the most frequent clinical function (77.3%), followed by diagnosis (13.3%) and management (9.4%). Across the eight clinical action stages, risk stratification and screening for GDM was the most frequent (57.8%), whereas no study addressed prevention of GDM or postpartum care as its primary stage. All studies remained at L1 or L2 (77.3% and 22.7%, respectively), none reached L3 or L4, and all management studies remained at L1. LLMs were evaluated in 5.5% of studies, and no study evaluated agentic AI. Although 7.8% of studies reported outputs for more than one stage, none evaluated whether the output of one stage informed the clinical action at the next.

Conclusions:

AI research in diabetes in pregnancy remains fragmented, prioritizing isolated clinical actions, chiefly risk prediction and diagnosis, while rarely integrating them into continuous care. This review identifies significant clinical gaps in early prevention, postpartum care, and advanced management evaluation for GDM, as well as AI applications for pre-existing diabetes. To achieve true continuity of care, future AI initiatives should prioritize integrating successive clinical actions across the perinatal continuum and validating their clinical implementation within real-world workflows.