Predictive performance of artificial intelligence algorithms for gestational diabetes mellitus in pregnant women: a protocol for systematic review and meta-analysis

Abstract

Background

Gestational diabetes mellitus (GDM) is a prevalent pregnancy complication that can pose numerous adverse health effects on both mothers and newborns. Accurate prediction of the risk of GDM serves as a valuable supplement to prenatal education and clinical decision-making. Compared with traditional prediction models, artificial intelligence (AI) algorithms have demonstrated higher predictive accuracy and stronger individualization capabilities. However, the application of AI models in GDM prediction is still in a developmental stage, and their performance and clinical utility have not been thoroughly evaluated. Therefore, this study aims to systematically review and critically appraise the published predictive performance of AI models for GDM prediction and to offer insights for future research and practical application.

Methods

A systematic literature search will be performed across six databases (PubMed, Web of Science, Cochrane Library, Scopus, EMBASE, and OVID). Screening of titles and abstracts, full-text review, and data extraction will be independently completed by two authors. Qualitative data on the characteristics of the included studies, methodological quality, and the applicability of models will be summarized through narrative descriptions and tabulated formats. For models with predictive performance data from multiple studies, a random-effects meta-analysis or meta-regression will be employed to synthesize the findings, considering potential heterogeneity.

Ethics and dissemination

Ethical approval is deemed not applicable for this systematic review and meta-analysis. The findings will be based on published literature, disseminated through publication in a peer-reviewed journal, and presented at major conferences focused on clinical healthcare.

Systematic review registration

PROSPERO registration number CRD42025645913

Supplementary Information

The online version contains supplementary material available at 10.1186/s13643-026-03167-0.

Introduction

Gestational diabetes mellitus (GDM) is defined as glucose metabolism abnormalities of varying severity that first emerge during pregnancy, representing a prevalent metabolic condition associated with pregnancy [1]. According to the 10th edition of the “Diabetes Atlas” published by the International Diabetes Federation, 16.7% of women of reproductive age experience elevated blood glucose levels during pregnancy, with approximately 80.3% of these cases attributed to GDM [2]. GDM significantly increases the risk of adverse pregnancy and delivery outcomes, including macrosomia, preterm birth, preeclampsia, and the future development of type 2 diabetes [3, 4]. Additionally, offspring of women with GDM are more susceptible to neonatal hypoglycemia, respiratory distress syndrome, and long-term metabolic disorders [5–7]. Therefore, early identification and management of GDM are critical for improving maternal and neonatal health outcomes.

Accurate prediction of GDM risk is crucial for timely intervention and effective management. Traditional risk assessment methods, which are usually based on clinical risk factors such as age, body mass index, and family history of diabetes, have limitations in predictive accuracy and accounting for individual variability [8–11]. However, the advent of artificial intelligence (AI) algorithms has markedly enhanced the precision of GDM prediction by autonomously detecting patterns within vast datasets and integrating multiple data sources—spanning clinical, biochemical, and lifestyle factors—to provide a holistic assessment of GDM risk [12–14].

Currently, the application of AI algorithms in GDM risk prediction extends across two major areas: machine learning and deep learning [15, 16]. Machine learning algorithms, including Gaussian Naive Bayes, decision trees, random forest (RF), gradient boosting machines (GBM), and support vector machines (SVM), have proven effective in predicting GDM during early pregnancy, achieving remarkable predictive performance [12, 17–19]. However, the predictive performance of these algorithms varies significantly across different studies. For example, a machine learning model developed by Gallardo et al. based on routine early-pregnancy examination data demonstrated high predictive accuracy in a specific population but may perform poorly in other GDM populations due to differences in data characteristics [20]. Meanwhile, deep learning algorithms, like artificial neural networks, also exhibit noticeable fluctuations in performance across different populations [21–23]. This discrepancy not only reflects the intrinsic constraints of the algorithms but also highlights the profound impact of dataset characteristics and population diversity on model performance [24, 25]. This underscores the need for further validation of the applicability and generalizability of AI algorithms across diverse populations, despite their promising potential in GDM prediction [14].

Despite the growing interest in AI-driven predictive models for GDM, the field remains in its developmental stage [18, 26, 27]. The performance and clinical applicability of these models vary widely across studies, and there is a need for a systematic evaluation of predictive accuracy, generalizability, and potential impact on clinical practice [28, 29]. To address this gap, we propose a systematic review and meta-analysis of published studies on AI algorithms for GDM prediction. This protocol outlines the methodology for identifying, appraising, and synthesizing evidence on AI prediction models, aiming to provide a comprehensive overview of their status and potential for future application in a clinical setting.

Research aims

This study aims to systematically evaluate available evidence on AI algorithms for predicting GDM. It will identify existing diagnostic prediction models and establish the most effective ones to inform clinical decision-making. The specific objectives of this systematic review and meta-analysis are:

1. To identify and catalogue existing AI-based diagnostic prediction models for GDM and qualitatively describe their features in the included studies.

2. To summarize and compare the predictive performance of current AI-based diagnostic prediction models for GDM.

3. To critically appraise the methodological quality and reporting standards of included studies.

4. To identify the most effective and highest-performing diagnostic prediction model for GDM and provide evidence for clinical decision-making.

Methods and design

Protocol registration and reporting

The present study protocol has been registered within PROSPERO under registration number (No. CRD42025645913) and is being reported in accordance with the reporting guidance provided in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Protocols (PRISMA-P) guidelines (see checklist in Additional file 1). The methodology for data extraction and management will be guided by the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHRAMS) checklist and the recommendations reported by Debray et al. [30, 31].

Eligibility criteria

The study selection will be based on predefined eligibility criteria established using the PICOTS system [31], as detailed in Table 1. PICOTS is a modification of the established PICO system, tailored to meet the specific requirements of systematic reviews of prediction models, and additionally considers time (prediction period and timing for model use) and clinical setting [31].

Table 1.

Eligibility criteria for the systematic review framed using the PICOTS system

	Inclusion criteria	Exclusion criteria
Population	Pregnant women from around the globe who are in the first or second trimester and have not been diagnosed with GDM	Women with pre-gestational diabetes (type 1 and type 2 diabetes)
		Women with diagnosis of GDM
Index	Development or external validation of a diagnostic prediction model for GDM (e.g. prediction model for the diagnosis of GDM)	Prognostic prediction model for women (e.g., prediction model for women with GDM to predict pregnancy complications)
Comparator	Not applicable
Outcome	GDM (any diagnostic criteria)
Timing	The period of gestation before being diagnosed with GDM	The period of gestation following a confirmed diagnosis of GDM
Setting	Diagnostic prediction models intended for use by healthcare professionals in prenatal clinical settings prior to the diagnosis of GDM during pregnancy, aimed at informing clinical decision-making for physicians	Diagnostic prediction models, which are intended to be used after the diagnosis of GDM

GDM gestational diabetes mellitus

Population

Studies proposing diagnostic prediction models for pregnant women from around the globe who are in the first or second trimester and have not been diagnosed with GDM will be considered for inclusion. Studies proposing models for women with pre-gestational diabetes (type 1 and type 2 diabetes) will be excluded. Additionally, studies proposing models for women who have been diagnosed with GDM will also be excluded.

Intervention

Studies involving the development of prediction models (with or without external validation) and external model validation studies (with or without model updating) that aim to clinical decision-making regarding the diagnosis of GDM or prediction of risk will be considered for inclusion.

Outcomes

The primary outcome to be predicted in the included studies will be the diagnosis of GDM (based on any diagnostic criteria). Secondary outcomes and effect measures will be defined by the study authors. Predictions for all these outcomes will be conducted before the diagnosis of GDM.

Timing

Included studies need to report on prediction models for timing before pregnancy or during the period before being diagnosed with GDM. Prediction models for the period of gestation following a confirmed diagnosis of GDM will be excluded.

Setting

Diagnostic prediction models intended for use by healthcare professionals in prenatal clinical settings prior to the diagnosis of GDM during pregnancy, aimed at informing clinical decisions for physicians will be considered for inclusion. Models intended to be used after the diagnosis of GDM will be excluded.

Types of studies

All quantitative studies (including cohort studies, case-control studies, and clinical trials) that investigate the application of AI algorithms to predict the occurrence of GDM in pregnant women and involve the development and/or validation of at least one prediction model will be eligible for this study, without any restrictions on publication date. Provided that the full text of the literature is accessible, a comprehensive review and analysis of eligible studies from any region of the world will be conducted, with pregnant women as the subjects. The initial literature search will cover the period from the inception of the database to 1 June 2025. The literature search process will remain ongoing until the study is completed, with supplementary searches conducted periodically to incorporate the latest developments.

Condition/domain being studied

This review will focus on GDM as the condition of interest. Given the variations in GDM diagnostic criteria across different regions and studies, we will standardize all GDM diagnostic data from all included studies to align with the International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria [32]. According to the IADPSG criteria, GDM is diagnosed using the 75-g oral glucose tolerance test (OGTT) conducted during 24–28 weeks of gestation, with the diagnostic criteria being: fasting blood level ≥ 5.1 mmol/L, a 1-h glucose level ≥ 10.0 mmol/L, or a 2-h glucose level ≥ 8.5 mmol/L [32]. For original studies using diagnostic thresholds that differ from the IADPSG criteria, conversions will be made by leveraging the correlation between blood glucose levels and the risk of GDM. Additionally, for studies that cannot be standardized to the IADPSG criteria, a sensitivity analysis will be performed to evaluate the impact of the diverse diagnostic criteria on the pooled effect size.

Search strategy

A comprehensive search will be conducted across six databases, including PubMed, Web of Science, Cochrane Library, Scopus, EMBASE, and OVID. To enhance the accuracy of the search results and minimize the possibility of missing relevant studies, the author team has devised a rigorous search strategy by combining Medical Subject Headings and keywords. Additionally, the bibliometric networks have been visualized using the VOSviewer software tool for analysis (see Fig. 1). The tailored search strategies for each database are outlined in Table 2, summarizing the key search terms related to the target population and AI algorithms. Detailed descriptions of the search strategies for all databases are described in Additional file 2.

Fig. 1.

Cluster analysis of keywords from databases

Table 2.

Key terms for developing search strategy

Population (P)	Intervention (I)
‘Pregnancy induced diabetes’ ‘Diabetes in pregnancy’ ‘Gestational diabetes mellitus’ ‘Maternal diabetes’ ‘Pregnancy diabetes mellitus’ ‘GDM’	‘Artificial Intelligence’ ‘Deep Learning’ ‘Machine Learning’ ‘Knowledge Acquisition (Computer)’ ‘Hierarchical Learning’ ‘Ensemble Learning’ ‘Transfer Learning’ ‘Knowledge Representation (Computer)’ ‘Diagnostic prediction model’ ‘Risk prediction’ ‘Risk assessment’

To evaluate the effectiveness and accuracy of the keywords, search strategies, and search scope in identifying relevant literature, the preliminary literature search will be conducted in the PubMed database. Subsequently, the search strategies will be adjusted accordingly based on the results obtained, and comprehensive searches will be carried out in the remaining electronic databases. In addition, we will also review the reference lists of relevant literature, particularly systematic reviews related to the topic of this study, and conduct additional searches in the electronic databases to minimized the omission of the key literature as much as possible. All searches will be conducted under the supervision of an academic librarian.

Screening and selection procedure

Following the completion of the initial search based on the search strategies, all retrieved records from the databases will be exported into the Endnote 21 reference management software, and then the “Find Duplicates” function of the software will be used to remove duplicates. Each literature will undergo a dual screening by two authors based on the title and abstract to initially identify studies that are relevant. Following this initial screening, the same process will be applied to the full-text screening, with a rigorous evaluation conducted according to the exclusion. The rationale for excluding each literature will be recorded. Any disagreements between the two authors will initially be resolved through discussion. If consensus cannot be achieved, a third author will be consulted. The PRISMA flow diagram (see Fig. 2) will be used to depict the literature screening process.

Fig. 2.

Preferred reporting items for systematic reviews and meta-analysis flow diagram of the identification, screening, and eligibility of included articles

Data extraction and management

Data extraction from selected studies will be guided by CHARMS checklist [30]. Two authors will independently perform data extraction using a pre-formatted extraction template in Microsoft Excel (Microsoft, WA, USA). The following data details will be extracted: (1) title of the study; (2) author of the study; (3) time of publication; (4) country of the study; (5) type of study; (6) dataset size (including sample size and any information on missing data); (7) source of participants; (8) predicted outcomes; (9) diagnostic criteria for GDM; (10) potential predictors; (11) incidence of GDM; (12) odds ratio or risk ratio for predictor; (13) type of AI model (and its algorithms); (14) model performance (properties of discrimination with confidence intervals, calibration, classification, and overall performance); (15) model validation method. If multiple models are used in a study, we will extract performance data from all models to enable comparative analysis. In cases where the information presented in the literature is ambiguous, the researchers will proactively contact the corresponding author to acquire the relevant information. If the required data cannot be obtained from the author, the relevant literature will be excluded from further consideration. After the data extraction is completed, the two authors will cross-check each other’s extracted information. Any discrepancies will be resolved by a third-party arbitrator.

In addition, the following data related to AI models will be extracted:

• Type of AI model. The AI models are classified into machine learning and deep learning. Machine learning models encompass algorithms, including SVM, decision trees, RF, k-nearest neighbors, GBM, and other relevant algorithms. Deep learning models mainly consist of algorithms based on neural networks.

• Model validation method. The approaches for validating the predictive performance of AI models mainly consist of cross-validation, K-fold cross-validation, external validation, and additional specific validation methods.

• Sample size and missing data. The sample size of each study and any reported information on missing data will be recorded, including the extent of missingness and how the authors addressed these missing data.

Critical appraisal

The prediction model risk of bias assessment tool (PROBAST) will be employed to evaluate the methodological quality (risk of bias) and relevance to the review question (applicability) of the included studies [33]. PROBAST is composed of four key domains: participants, predictors, outcomes, and analysis. The risk of bias for each domain is categorized as “high”, “low”, or “unclear”. The evaluation of applicability is employed to determine whether the model development/validation studies align with our systematic review question regarding the target population, predictors, or outcomes of interest. The above process will be conducted independently by two authors. Any discrepancies will be resolved through discussion and consultation with a third author.

Data synthesis and analysis

Data synthesis

After the completion of data extraction, a narrative synthesis method will be utilized to systematically elucidate the quantitative data derived from the included studies, focusing on the characteristics and features of the data. This will cover key aspects such as predictive factors, performance metrics, classification indicators, and descriptive analyses of critical items. To facilitate comparison, the findings of each included study will ultimately be presented in tabular form. For studies reporting multiple AI models, all models will be included in the narrative synthesis to provide comparative insights.

Meta-analysis and investigation of heterogeneity

If the AI models identified in the included studies demonstrate sufficient homogeneity, data synthesis will be conducted through meta-analysis stratified by the type of AI modelling study. The following conditions will be deemed to satisfy homogeneity:

• AI model development studies where the target population, predicted outcomes, and anticipated time frames for model application are comparable or

• Several validation studies targeting the identical AI model.

During the meta-analysis, a random-effects model will be applied to pool and analyze the performance metrics of AI models (such as discrimination and calibration), thereby estimating the overall average performance of the models included in the studies. The performance of the diagnostic prediction model will be based on the following metrics [34–36], detailed in Table 3. To further conduct a comprehensive assessment of the models’ overall performance, multivariate meta-analysis will be employed for the joint analysis of discriminative ability and calibration, while considering their correlation. Besides, the restricted maximum likelihood estimation and the Hartung-Knapp-Sidik-Jonkman methods will be applied to assess between-study heterogeneity and the 95% confidence intervals for average model performance.

Table 3.

Summary of measuring performance of diagnostic prediction models

Items	Explanation	Performance measures/statistics	Values for better performance	Visualization
Discrimination	To measure the models’ ability to distinguish between cases and non-cases	AUROC	Higher	ROC curve
		AUPRC	Higher	PRC curve
		ACC	Higher
		BER	Lower
		D statistics	Lower
		MCC	Higher
		F1 score	Higher
		Log-rank	Lower, p > 0.05
Calibration	To evaluate consistency between predicted probabilities of the model and the actual observed results	Calibration curve, slope, intercept	Slope closer to 1 and intercept closer to 0	Calibration plot
		Reliability-deviation	Lower
		Reliability-within-bin variation	Lower
		Reliability-within-bin covariance	Higher
		Resolution	Higher
		Predictive range	Higher
		Hosmer–Lemeshow test	Higher
		Total O:E ratios	Higher
Classification	To measure the models’ ability to correctly classify individuals as cases or non-cases	Sensitivity	Higher	Reclassification scatter plot
		Specificity	Higher
		PPV	Higher
		NPV	Higher
		RMSE via neighborhood estimate	Lower
Overall performance	To comprehensively evaluate the overall performance of the models, including its accuracy and practicability	Brier score	Lower
		Brier skill score	Higher
		Prediction squared error	Lower
		Decision curve analysis	/	DCA plot

AUROC area under receiver operating characteristic curve, AUPRC area under precision-recall curve, ACC accuracy, BER balanced error rate, MCC Matthews correlation coefficient, PPV positive predictive value, NPV negative predictive value, RMSE root-mean-squared error

Heterogeneity in performance metrics is anticipated due to differences in study design and population. The potential range of model performance across different populations can be estimated by calculating an approximate 95% prediction interval. This will help assess the generalizability of AI models across diverse settings, addressing concerns related to population diversity and language bias raised in the limitations. Variations in case-mix within each study will be quantified by estimating the standard deviation of the linear predictor [31]. When performance metrics or uncertainty measures are unreported, we will estimate these values based on sample size, event rate, or confusion matrices using normal approximation [31]. If estimation is not feasible, the missing data will be recorded and included only in the narrative synthesis. Statistical heterogeneity will be assessed using the I² test, with I² values above 50% indicating moderate to high heterogeneity. Potential sources of heterogeneity will be explored through meta-regression analysis (P < 0.05).

Sensitivity analysis

To assess the robustness of our primary findings, we will conduct a series of sensitivity analyses. First, to address potential unit-of-analysis issues arising from studies that report multiple AI models, our primary meta-analysis will include only the primary model (as specified by study authors) or the model with the highest accuracy. We will then perform a sensitivity analysis by including all models in a separate meta-analysis, where appropriate, to evaluate the impact of our model selection strategy on the pooled estimates.

If enough studies are included, subgroup analyses will be performed based on

• Region—categorized according to the Organization for Economic Co-operation and Development classification into low/middle-income countries and high-income countries.

• Population characteristics—including ethnicity, medical history, and pregnancy trimester—to assess differences in model performance across various subgroups.

• Type of AI modelling study—development or validation.

• Type of AI models—categorized into machine learning (e.g., RF, SVM, gradient boosting machine) and deep learning (e.g., artificial neural networks), with further comparisons across specific algorithm families if data permit.

• Predictors—type or number of predictors to assess their impact on model performance.

• Study quality—risk of bias.

This overall process will be conducted using Stata 17.0 software, in accordance with the Meta-analysis of Observational Studies in Epidemiology (MOOSE) guidelines [37].

Reporting and presentation of findings

The findings of this study will be presented in accordance with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement and the PRISMA guideline [38, 39]. The GRADE approach (grading of recommendations, assessment, development, and evaluation) will also be applied to assess the certainty of the evidence [40]. The certainty of the evidence will be categorized into “high”, “moderate”, “low”, or “very low”.

Discussion

This systematic review and meta-analysis aims to synthesize the current evidence on AI algorithms for predicting GDM and to explore the potential of AI-driven models in clinical practice. By systematically evaluating available studies, this review seeks to identify the most effective AI algorithms based on their predictive accuracy, which may ultimately facilitate the early detection of high-risk populations for GDM. The findings are expected to advance the application of AI algorithms in GDM prediction and contribute valuable insights to the broader field of medical AI.

Moreover, for these algorithms to achieve meaningful clinical impact, several critical factors beyond predictive accuracy should be considered. First, clinical integration requires that AI models be seamlessly embedded into antenatal care workflows, with clear guidance on when and how clinicians should act on predictions. Second, model interpretability is essential for clinician trust. Many high-performing algorithms, particularly deep learning models, operate as “black boxes,” obscuring the rationale behind risk assessments. Emerging explainable AI techniques, such as SHapley Additive exPlanations and Local Interpretable Model-agnostic Explanations, can help address this challenge by quantifying the contribution of individual risk factors to each prediction, thereby enabling clinicians to understand why a specific patient is classified as high-risk.

However, this protocol also has several limitations. First, by restricting our search to studies published in English, we may introduce language bias, potentially excluding relevant studies published in other languages. Consequently, our findings may not fully capture the global distribution of AI-based GDM prediction research, particularly regarding research progress in non-English speaking regions. Second, as this review will primarily focus on model development and validation studies, evidence on clinical implementation, interpretability, and cost-effectiveness may remain limited. Third, included studies may rarely report on ethical considerations such as data privacy, fairness, or bias mitigation, which could limit our ability to assess their readiness for clinical deployment. Nevertheless, this review is expected to systematically synthesize the current state of research on AI algorithms in GDM prediction and provide a reference for future studies and clinical practice.

Abbreviations

GDM

Gestational diabetes mellitus

AI

Artificial intelligence

RF

Random Forest

GBM

Gradient boosting machines

SVM

Support vector machines

PRISMA-P guidelines

Preferred Reporting Items for Systematic Reviews and Meta-Analyses Protocols guidelines

CHRAMS checklist

Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies checklist

IADPSG criteria

International Association of Diabetes and Pregnancy Study Groups criteria

OGTT

Oral Glucose Tolerance Test

PROBAST

Prediction Model Risk of Bias Assessment Tool

MOOSE guidelines

Meta-analysis of Observational Studies in Epidemiology guidelines

TRIPOD statement

Transparent Reporting of A Multivariable Prediction Model for Individual Prognosis Or Diagnosis statement

Authors’ contributions

YN.L and MY.L designed and developed the research question. YN.L and JY.S developed the search strategy. YN.L registered the protocol and wrote the first draft of the manuscript. YH.S and ZY.L are the guarantors. YP.Y and JY.S developed the risk of bias assessment strategy. AR.D and ZL.Z revised the manuscript. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (No. 32070189), Hunan Province Health Commission Scientific Research Project (No. D202314038701), and Postgraduate Scientific Research Innovation Project of Hunan Province (No. CX20251466).

Data availability

The datasets during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

All authors declare that they have no competing interests.