Graph neural networks for networked analysis of gestational diabetes risk factors: a multi method framework Authors: Kranthi Kumar Lella, Shaik Abdul Nabi, Uddagiri Sirisha, G. Muni Nagamani, Poluru Eswaraiah, Chanumolu Kiran Kumar Published: 2026 ABSTRACT Gestational diabetes mellitus, often known as GDM, is a major health issue that causes complications for mothers and requires patient data prediction models that are complex and variable. The research in question makes use of graph-based learning in order to investigate the ways in which genetic, biochemical, and demographic elements interact in a variety of different contexts. Through the use of nodes to represent patients and lines to represent the things that they share in common, the framework illustrates how the aforementioned elements influence the likelihood of illness. Graph neural networks are utilized for the process, while BioBERT embeddings are utilized for the management of unstructured clinical notes. Graph neural networks are utilized for organized clinical notes. Because of this alignment, healthcare processes are placed in the context in which they should be, rather than being taken out of context while they are being carried out. The graph architecture used in BioBERT incorporates semantic patterns derived from medical information into a relational structure that illustrates the degree to which patients are similar to one another. After being evaluated on a substantial clinical dataset, the proposed method is able to make more accurate and readable predictions than the baseline models. The results of this study indicate that the utilization of graph architecture with both organized and unstructured data can assist in the discovery of novel approaches to the treatment of GDM that go beyond performance sets. According to the findings of the study, machine learning needs to be modified so that it can be used with healthcare applications. Introduction Gestational diabetes mellitus is a major metabolic disorder occurring in pregnant individuals, and it has significant implications both for maternal health sets and fetal health sets. It characterizes itself with glucose intolerance during pregnancy. GDM increases the risks for complications such as macrosomia, preterm delivery, and long-term metabolic disorders in offspring. Today’s clinical methods1–3 for assessing GDM risk are mainly linear ones, based on isolated clinical factors and standard statistical approaches. These methods, though useful in controlled environments, cannot explain the intricate, interdependent, and dynamic nature of GDM risk factors such as demographic variables, genetic predispositions, lifestyle behaviors, and changes in clinical measurements over time. Recent breakthroughs in artificial intelligence, particularly in Graph Neural Networks (GNNs), are exciting directions for modeling this complexity. GNNs uniquely have the ability to model relational structures and multi-dimensional interactions within datasets & samples. This is quite critical in GDM analysis as risk factors often compose a network of interdependent variables with causal and temporal dependencies. Although promising in its concept4–6, currently, GNN has fewer applications in healthcare domains because most of the studies use static or homogeneous graphs that fail to exploit any information available in the graph structure and related edges, including heterogeneous relationship as well as dynamic temporal pattern. This paper introduces a novel framework for networked analysis of GDM risk factors using a suite of advanced GNN methodologies. This approach integrates feature aggregation via GraphSAGE, multi-typed relationships by R-GCN, and temporal evolution by TGAT. GAT is utilized for achieving explainability while ensuring the privacy-preserving federated learning for the secure deployment of models. This integration, therefore, not only improves the predictive accuracy but also offers more interpretable insights into factors contributing to GDM and bridges the gap between advanced machine learning techniques and practical healthcare needs. Traditional algorithms cannot capture irregular, multi-relational patient data for gestational diabetes prediction. The model encodes demographic, clinical, and biochemical variable interactions using graph-based patient feature representations. This section contextualizes component roles in patient health trajectories rather than algorithmic descriptions. The basic model, Graph Neural Networks (GNNs), captures patient variable related interactions. Instead of algorithms, the argument compares message forwarding, neighborhood aggregation, and attention-driven early gestational diabetes risk detection options. These methods are specialized for correlating genetic markers, lifestyle patterns, and clinical data samples’ temporal fluctuations. With one pipeline, the method streamlines data preprocessing, graph building, representation learning, and prediction. Practical implementation approaches include incorporating patient nodes by merging clinical records and diagnostic codes and graph attention to prioritize clinically relevant elements. The section is now a problem-driven design story linking model structure to medical applications rather than an algorithm catalog. Gestational diabetes mellitus is therefore a process singularly unique in its healthcare challenge. It is multidimensional and dynamic, by nature, to the process itself. Traditional prediction models are foundational but limit the capacity to represent such an intricate network of interdependent risk factors associated with GDM. They commonly resort to lonely variables or toy associations, which cannot comprehend causality, temporal flows and multi Variable inter-actions in process. This tendency has also led to another request for advanced methods that can accommodate and analyze heterogeneous data sets & samples because sources of different types of data are increasingly becoming available. Some examples of diverse sources include electronic health records, demographic surveys, and genetic data. The additional challenge imposed by privacy considerations, with patient data dispersed across different institutions, necessitates safe and decentralized analysis approaches. To mitigate these challenges, this paper develops a holistic framework based on Graph Neural Networks, employing six state-of-the-art methods tailored for GDM analysis. GraphSAGE offers scalable and robust embeddings of clinical and demographic data with missing values. It captures complex, multi-typed relationships, like causal and temporal dependencies between risk factors. TGAT incorporates the important temporal dynamics required to study the progression of GDM during pregnancy. GAT presents interpretable models by scoring attention to the most influential risk factors, hence supporting clinical decision-making. BioBERT domain-specific pretraining is on medical literature and for graph-based tasks, one can have robust embeddings at initialization. Finally, federated learning with differential privacy allows for the safe training of models on decentralized datasets, thereby addressing the privacy and regulatory concerns. All aspects of predictive accuracy, explainability, and privacy preservation have been improved upon in the proposed framework. It achieves superior performance metrics, such as an AUC of 0.9 for GDM prediction and over 80% agreement with domain experts in explainability assessments. This sets a new standard of work in GNN-based health care analytics and offers an approach scalable and interpretable in nature to address one of the most important public health concerns, opening up scope for more personalized and efficient GDM management. Contributions The most important feature that we have included is the development and evaluation of a multi-method graph learning system for the prediction of gestational diabetes. Rather than listing algorithms in chronological sequence, this section makes use of research to provide solutions to tasks. The primary emphasis is placed on the ways in which combining approaches makes the handling of medical data anomalies more effective. In order to generate patient graphs, the framework takes into account demographic, clinical, and biochemical data. In the graph, each point represents a patient, and each edge indicates a characteristic or risk factor that is shared by all of the patients. This graphic is an example of something that traditional models do not display, such as patient groups that have similar DNA tendencies. Graph convolutional neural networks (GNNs) and attention-based GNNs have been proven to operate well together in order to identify local versus global dependencies; however, they are not effective in listing approaches. Third, repetitively combining images that we have learned alters the way that we carry out our daily activities. Clinicians are able to quantify risk and identify elements that play a role with a high degree of accuracy and ease of comprehension when they combine embeddings from a variety of models. In the final section of the article, it is demonstrated how the framework improves the accuracy of high-stakes medical predictions. Review of existing models for pregnancy diabetes analysis During recent years, the incidence of gestational diabetes mellitus has reached an increasingly high level of concern, which has led to new waves of research about the early prediction and management of the disorder. This body of work was thus spread over the variety of approaches, from clinical studies up to more advanced machine learning models, and contributes to the development of improved strategies of diagnosis, treatment, and prevention of GDM. It ranges from clinical trials to biomarker studies, digital health technologies, and machine learning deployments. These studies included many of the aspects of GDM including pathophysiology, risk factors, and predictive models for early detection. Accordingly, the study of Moreno et al.1 provides much insight into the concerns of gestational weight gain information-seeking behavior of pregnant women, in particular, among immigrant Central American women in the United States; thus, both the social and behavioral understanding are essential due to their morphogenic risks that healthcare interventions might need to be framed about. In like manner, Gulersen et al.2 discuss morbidity risks from early preterm birth in particular with respect to low-risk singleton pregnancies. It thus postulates that adverse pregnancy outcomes are connected with the factors of risks which culminate into GDM. Such studies thus help establish demographic, social, and medical determinants as pertinent to understanding and managing risks with GDM. Contributions to the physiological understanding of GDM include the study by Zampieri et al.3 on uterine artery indices and Genova et al.4 on the relationship between zinc status and insulin sensitivity, contributing to the clinical understanding of metabolic changes in GDM sets. These studies inform the medical community about the biological markers that could be used for early diagnosis and intervention sets. The works of Siemers et al.5 and Payot et al.6 report impairments in mitochondrial function and gene expression analyses, respectively at the molecular level; this should enlighten some mechanisms related to GDM and perhaps biomarkers at early stages of onset. Iteratively, Next, As shown in Table 1, In addition, Seifert et al.7 and Rüegg et al.8 explain the expression of some proteins like galectin-7 and how the perioperative care of women who undergo repair of fetal spina bifida affects GDM, representing how GDM interacts with other health conditions.Table 1Comparative analysis of existing methods.ReferenceMethodMain objectivesFindingsLimitationsMoreno et al.1Survey studyExplore beliefs, concerns, and prenatal care advice regarding gestational weight gain among immigrant Central American pregnant womenIdentified significant concerns regarding weight gain and information sources among this population, highlighting the need for tailored interventionsThe study is region-specific and may not be generalized to other immigrant populationsGulersen et al.2Clinical studyInvestigate maternal morbidity related to early preterm birth in low-risk singleton pregnanciesFound increased maternal morbidity with early preterm birth, emphasizing early monitoring and interventionLimited to low-risk singleton pregnancies; results may not apply to high-risk casesZampieri et al.3Diagnostic studyStudy pulsatility and resistivity indices of uterine arteries in term pregnanciesProvided insights into uterine artery Doppler indices as potential markers for predicting adverse pregnancy outcomesSmall sample size and limited to term pregnanciesGenova et al.4Biochemical studyExamine the relationship between zinc status and insulin resistance in gestational diabetesFound that zinc status is associated with insulin resistance in GDM, suggesting it as a potential therapeutic targetCross-sectional design; causality not establishedSiemers et al.5Molecular studyInvestigate mitochondrial function and fatty acid uptake in human trophoblasts in gestational diabetesDemonstrated that GDM impairs mitochondrial function and fatty acid uptake, impacting trophoblast functionIn vitro study; may not fully replicate in vivo conditionsPayot et al.6Biomarker studyInvestigate potential early biomarkers for GDM through PTRPG and IGKV2D-28 expressionIdentified specific biomarkers related to early GDM detection, offering potential diagnostic toolsPreliminary study; requires further validation in larger cohortsSeifert et al.7Molecular studyInvestigate galectin-7 expression in placentas of women with GDMFound altered galectin-7 expression in GDM placentas, suggesting it as a potential biomarkerLimited sample size and focus on a single biomarkerRüegg et al.8Clinical studyExamine the influence of perioperative management on GDM in women with fetal spina bifida repairRevealed the importance of perioperative management in GDM outcomes, especially in complex cases like spina bifida repairNarrow focus on a specific patient group; findings may not be generalizableHung et al.9Experimental studyStudy the effects of lysophosphatidylcholine on trophoblast function in GDMFound that lysophosphatidylcholine impairs mitochondrial function, leading to trophoblast dysfunction in GDMLimited to a specific lipid molecule, not considering broader metabolic factorsFilip et al.10Clinical studyExamine the burden of deep vein thrombosis (DVT) during pregnancy and postpartumIdentified increased DVT risk during pregnancy and postpartum, particularly in high-risk groupsFocuses on DVT, not directly related to GDM but provides context for maternal health risksMandić Marković et al.11Biochemical studyEvaluate biochemical markers for predicting pregnancy outcomes in GDMFound several promising biomarkers for predicting GDM outcomes, aiding in early interventionRequires validation in larger, diverse populationsWang et al.12Epidemiological studyInvestigate the impact of beverage consumption during pregnancy on maternal and offspring outcomesFound associations between beverage consumption and adverse outcomes, emphasizing dietary interventionsCross-sectional design; cannot establish causal relationshipsLu et al.13Digital health studyReview digital health technologies for blood glucose monitoring in GDMHighlighted the potential of digital health and machine learning technologies for improving GDM managementLimited clinical validation of digital tools in real-world settingsGomes Filho et al.14Machine learning studyDevelop a Bayesian network-based methodology for early GDM diagnosisFound that Bayesian networks could effectively support early GDM diagnosis using clinical dataLimited to a single methodology; may not be applicable to all patient populationsKolozali et al.15Machine learning studyCombine medical background and wearable devices for early prediction of GDM biomarkersDemonstrated the efficacy of combining traditional medical data with wearable devices for early GDM predictionPilot study with a small cohort, limiting generalizabilityMarzouk et al.16Predictive modeling studyDevelop a predictive model for diabetes monitoring using IoT and machine learningCreated a secure, web-based personalized diabetes monitoring system using machine learningReliant on IoT infrastructure, which may not be widely available in all healthcare settingsKhan et al.17Machine learning studyUse bio Inspired PSO to enhance neural network-based diabetes predictionApplied particle swarm optimization to improve neural network performance for diabetes predictionLimited to PSO and neural networks; other optimization techniques not consideredSaleh Al Reshan et al.18Deep learning studyDevelop an ensemble deep learning system for diabetes predictionFound that ensemble deep learning models significantly improved diabetes prediction accuracyRequires large datasets for training and may be computationally expensiveLinkon et al.19Machine learning studyEvaluate feature transformation techniques for diabetes detectionFound that feature transformation techniques enhance the accuracy of machine learning models for diabetes detectionMay require extensive preprocessing for large datasetsAlnowaiser et al.20Machine learning studyImprove diabetes prediction using KNN imputed features and a tri-ensemble modelDemonstrated that KNN imputation combined with tri-ensemble models improved prediction accuracyDependent on feature imputation, which may introduce biasesZhao et al.21Machine learning studyDevelop hypertension risk prediction models for diabetic patientsCreated machine learning models that effectively predict hypertension risk in diabetic patientsFocus on hypertension may not directly apply to GDM-specific outcomesKurt et al.22Deep learning studyPredict GDM using deep learning, Bayesian optimization, and traditional machine learningFound that deep learning models outperformed traditional techniques in predicting GDMThe complexity of deep learning models requires significant computational resourcesAshisha et al.23Machine learning studyApply random oversampling techniques to improve diabetes classificationFound that random oversampling significantly improved classification performance for diabetesMay introduce overfitting if not carefully implementedFeng et al.24Machine learning studyOptimize diabetes classification using a machine learning-based frameworkDemonstrated that an optimized machine learning framework enhances diabetes classificationFocused only on classification; does not account for other potential risk factorsUddin et al.25Machine learning studyReduce false negatives in diabetes detection modelsDeveloped a machine learning model specifically designed to reduce false negatives in diabetes detectionLimited generalizability due to focus on false negative reductionDu et al.26Machine learning studyDevelop a visualization obesity risk prediction systemCreated a machine learning-based system to predict obesity risk, potentially useful for GDM predictionFocuses on obesity risk, which may not be directly transferable to GDMDarsareh et al.27Machine learning studyIdentify risk factors for birth asphyxia using machine learningFound that machine learning models can identify key risk factors for birth asphyxia, providing early intervention opportunitiesFocus on birth asphyxia, not directly related to GDMWu et al.28Machine learning studyEarly prediction of GDM using maternal demographic and clinical factorsDeveloped a predictive model for GDM using demographic and clinical factors, improving early diagnosisRelies on limited clinical variables; may not capture all relevant risk factorsBai et al.29Machine learning studyPredict large-for-gestational-age newborns in radiation-exposed womenApplied machine learning models to predict outcomes in women exposed to radiation before pregnancyLimited to radiation-exposed women, affecting generalizabilityKhadidos et al.30Ensemble learning studyCreate an ensemble framework for predicting maternal health risksDeveloped an ensemble model that effectively predicts maternal health risks during pregnancyMay not work as well in populations with different health profilesDu et al.31Clinical studyInvestigate maternal height as a risk factor for adverse GDM outcomesFound that maternal height is an independent risk factor for adverse pregnancy outcomes in GDMLimited to maternal height; other factors may need further explorationSoria-Contreras et al.32Longitudinal studyInvestigate the lifetime impact of gestational diabetes on cognitive functionDemonstrated a potential long-term effect of gestational diabetes on cognitive function in midlifeLimited to a single cohort; larger studies are needed for confirmationYang et al.33Machine learning studyAnalyze neurodevelopmental delay in offspring using maternal risk factorsUsed machine learning to identify maternal and perinatal factors associated with neurodevelopmental delay in offspringLimited by the quality of population data used in the analysisDoğru et al.34Ensemble learning studyHybrid super ensemble model for early-stage diabetes risk predictionDeveloped a hybrid ensemble model that improved early-stage diabetes risk predictionDependent on complex model structures, making it computationally expensiveXu et al.35Epidemiological studyInvestigate dietary inflammatory index as a predictor of prediabetesFound that a higher dietary inflammatory index correlates with a greater risk of prediabetes in women with prior GDMFocused on diet, which may not account for all relevant metabolic factorsRustam et al.36Machine learning studyEnhanced diabetes detection using ensemble feature engineeringDeveloped an ensemble feature engineering approach to improve diabetes detection accuracy, significantly enhancing model performanceThe approach may require large amounts of data to effectively train and may be sensitive to the quality of input featuresOrdoñez-Guillen et al.37Machine learning studyClassify Type 2 diabetes mellitus subtypes using machine learning algorithmsIdentified distinct subtypes of Type 2 diabetes, improving prediction and personalized treatmentFocused only on Type 2 diabetes subtypes, which may not apply to GDM diagnosis and managementSalvatori et al.38Data-driven studyUse cluster analysis to identify and validate GDM subgroupsFound that data-driven cluster analysis could identify distinct subgroups within GDM patients, potentially improving personalized treatment strategiesLimited to cluster analysis; may miss other subgroups that do not fit into the identified categoriesNaito et al.39Machine learning studyInvestigate associations between environmental factors and cardiometabolic diseasesRevealed heterogeneous associations between environmental factors and cardiometabolic diseases, using polygenic risk scoresFocused on cardiometabolic diseases, which may not directly apply to GDM, and relied on polygenic risk scoresBelaghi et al.40Predictive modeling studyPredict preterm birth risk in multiparous women using logistic regression and machine learningDemonstrated that both logistic regression and machine learning models could effectively predict preterm birth in multiparous womenResults may not be generalizable to all populations, particularly those with single pregnancies or different risk factorsSolomon et al.41Review studyReview the role of machine learning for multifactorial genetic disorder predictionExtensive review of machine learning applications in predicting multifactorial genetic disorders, including diabetesReview-based; lacks original data or novel insights from new models or approachesLong et al.42Comparative studyCompare machine learning algorithms for predicting microalbuminuria riskShowed that machine learning algorithms could be effectively applied to predict microalbuminuria risk, an important complication of diabetesLimited to microalbuminuria and may not address other diabetic complications like GDMAghajanian et al.43Predictive modeling studyPredict post-delivery hemoglobin levels with machine learning algorithmsFound that machine learning models can accurately predict post-delivery hemoglobin levels, potentially guiding clinical interventionsResults may not directly apply to GDM, as it focuses on postpartum hemoglobin levelsKokori et al.44Narrative reviewReview the role of machine learning algorithms in the detection of gestational diabetesReviewed current evidence on the application of machine learning algorithms for GDM detection, highlighting various techniquesLimited by the narrative review format; lacks original experimental data and may omit certain approachesKaya et al.45Machine learning studyEarly prediction of gestational diabetes mellitus using machine learning modelsShowed that machine learning models can predict GDM at early stages with high accuracy, potentially improving early interventionRelied on a specific set of models; other predictive approaches may be overlookedKang et al.46Machine learning studyPredict GDM in Asian women using machine learning algorithmsDemonstrated the efficacy of machine learning algorithms in predicting GDM in Asian women, accounting for ethnic-specific risk factorsFocused solely on Asian women, limiting the applicability of the results to other ethnic groupsCubillos et al.47Machine learning studyDevelop models to predict GDM risk in the first half of pregnancyDeveloped predictive models for GDM risk early in pregnancy, allowing for early interventionsLimited by the dataset and may not account for other complex, unmeasured factors in GDM developmentChan et al.48Machine learning studyUse elemental contents in fingernails for early prediction of GDMFound that machine learning models using elemental contents in fingernails could predict GDM early, offering a non Invasive diagnostic toolThe technique is novel but requires validation across larger and more diverse populationsLi et al.49Machine learning studyUse machine learning to predict GDM in the first trimesterDemonstrated that machine learning algorithms can predict GDM in the first trimester with high accuracy, aiding early intervention strategiesThe first trimester prediction model may require additional clinical validation in broader populationsNuthakki et al.50Machine learning studyEarly detection of diabetes risk factors for improved health managementCreated a machine learning-based system for early detection of diabetes risk factors, aimed at improving overall health managementFocused primarily on early detection; may require integration with existing clinical workflows for effective use Table 1Comparative analysis of existing methods.ReferenceMethodMain objectivesFindingsLimitationsMoreno et al.1Survey studyExplore beliefs, concerns, and prenatal care advice regarding gestational weight gain among immigrant Central American pregnant womenIdentified significant concerns regarding weight gain and information sources among this population, highlighting the need for tailored interventionsThe study is region-specific and may not be generalized to other immigrant populationsGulersen et al.2Clinical studyInvestigate maternal morbidity related to early preterm birth in low-risk singleton pregnanciesFound increased maternal morbidity with early preterm birth, emphasizing early monitoring and interventionLimited to low-risk singleton pregnancies; results may not apply to high-risk casesZampieri et al.3Diagnostic studyStudy pulsatility and resistivity indices of uterine arteries in term pregnanciesProvided insights into uterine artery Doppler indices as potential markers for predicting adverse pregnancy outcomesSmall sample size and limited to term pregnanciesGenova et al.4Biochemical studyExamine the relationship between zinc status and insulin resistance in gestational diabetesFound that zinc status is associated with insulin resistance in GDM, suggesting it as a potential therapeutic targetCross-sectional design; causality not establishedSiemers et al.5Molecular studyInvestigate mitochondrial function and fatty acid uptake in human trophoblasts in gestational diabetesDemonstrated that GDM impairs mitochondrial function and fatty acid uptake, impacting trophoblast functionIn vitro study; may not fully replicate in vivo conditionsPayot et al.6Biomarker studyInvestigate potential early biomarkers for GDM through PTRPG and IGKV2D-28 expressionIdentified specific biomarkers related to early GDM detection, offering potential diagnostic toolsPreliminary study; requires further validation in larger cohortsSeifert et al.7Molecular studyInvestigate galectin-7 expression in placentas of women with GDMFound altered galectin-7 expression in GDM placentas, suggesting it as a potential biomarkerLimited sample size and focus on a single biomarkerRüegg et al.8Clinical studyExamine the influence of perioperative management on GDM in women with fetal spina bifida repairRevealed the importance of perioperative management in GDM outcomes, especially in complex cases like spina bifida repairNarrow focus on a specific patient group; findings may not be generalizableHung et al.9Experimental studyStudy the effects of lysophosphatidylcholine on trophoblast function in GDMFound that lysophosphatidylcholine impairs mitochondrial function, leading to trophoblast dysfunction in GDMLimited to a specific lipid molecule, not considering broader metabolic factorsFilip et al.10Clinical studyExamine the burden of deep vein thrombosis (DVT) during pregnancy and postpartumIdentified increased DVT risk during pregnancy and postpartum, particularly in high-risk groupsFocuses on DVT, not directly related to GDM but provides context for maternal health risksMandić Marković et al.11Biochemical studyEvaluate biochemical markers for predicting pregnancy outcomes in GDMFound several promising biomarkers for predicting GDM outcomes, aiding in early interventionRequires validation in larger, diverse populationsWang et al.12Epidemiological studyInvestigate the impact of beverage consumption during pregnancy on maternal and offspring outcomesFound associations between beverage consumption and adverse outcomes, emphasizing dietary interventionsCross-sectional design; cannot establish causal relationshipsLu et al.13Digital health studyReview digital health technologies for blood glucose monitoring in GDMHighlighted the potential of digital health and machine learning technologies for improving GDM managementLimited clinical validation of digital tools in real-world settingsGomes Filho et al.14Machine learning studyDevelop a Bayesian network-based methodology for early GDM diagnosisFound that Bayesian networks could effectively support early GDM diagnosis using clinical dataLimited to a single methodology; may not be applicable to all patient populationsKolozali et al.15Machine learning studyCombine medical background and wearable devices for early prediction of GDM biomarkersDemonstrated the efficacy of combining traditional medical data with wearable devices for early GDM predictionPilot study with a small cohort, limiting generalizabilityMarzouk et al.16Predictive modeling studyDevelop a predictive model for diabetes monitoring using IoT and machine learningCreated a secure, web-based personalized diabetes monitoring system using machine learningReliant on IoT infrastructure, which may not be widely available in all healthcare settingsKhan et al.17Machine learning studyUse bio Inspired PSO to enhance neural network-based diabetes predictionApplied particle swarm optimization to improve neural network performance for diabetes predictionLimited to PSO and neural networks; other optimization techniques not consideredSaleh Al Reshan et al.18Deep learning studyDevelop an ensemble deep learning system for diabetes predictionFound that ensemble deep learning models significantly improved diabetes prediction accuracyRequires large datasets for training and may be computationally expensiveLinkon et al.19Machine learning studyEvaluate feature transformation techniques for diabetes detectionFound that feature transformation techniques enhance the accuracy of machine learning models for diabetes detectionMay require extensive preprocessing for large datasetsAlnowaiser et al.20Machine learning studyImprove diabetes prediction using KNN imputed features and a tri-ensemble modelDemonstrated that KNN imputation combined with tri-ensemble models improved prediction accuracyDependent on feature imputation, which may introduce biasesZhao et al.21Machine learning studyDevelop hypertension risk prediction models for diabetic patientsCreated machine learning models that effectively predict hypertension risk in diabetic patientsFocus on hypertension may not directly apply to GDM-specific outcomesKurt et al.22Deep learning studyPredict GDM using deep learning, Bayesian optimization, and traditional machine learningFound that deep learning models outperformed traditional techniques in predicting GDMThe complexity of deep learning models requires significant computational resourcesAshisha et al.23Machine learning studyApply random oversampling techniques to improve diabetes classificationFound that random oversampling significantly improved classification performance for diabetesMay introduce overfitting if not carefully implementedFeng et al.24Machine learning studyOptimize diabetes classification using a machine learning-based frameworkDemonstrated that an optimized machine learning framework enhances diabetes classificationFocused only on classification; does not account for other potential risk factorsUddin et al.25Machine learning studyReduce false negatives in diabetes detection modelsDeveloped a machine learning model specifically designed to reduce false negatives in diabetes detectionLimited generalizability due to focus on false negative reductionDu et al.26Machine learning studyDevelop a visualization obesity risk prediction systemCreated a machine learning-based system to predict obesity risk, potentially useful for GDM predictionFocuses on obesity risk, which may not be directly transferable to GDMDarsareh et al.27Machine learning studyIdentify risk factors for birth asphyxia using machine learningFound that machine learning models can identify key risk factors for birth asphyxia, providing early intervention opportunitiesFocus on birth asphyxia, not directly related to GDMWu et al.28Machine learning studyEarly prediction of GDM using maternal demographic and clinical factorsDeveloped a predictive model for GDM using demographic and clinical factors, improving early diagnosisRelies on limited clinical variables; may not capture all relevant risk factorsBai et al.29Machine learning studyPredict large-for-gestational-age newborns in radiation-exposed womenApplied machine learning models to predict outcomes in women exposed to radiation before pregnancyLimited to radiation-exposed women, affecting generalizabilityKhadidos et al.30Ensemble learning studyCreate an ensemble framework for predicting maternal health risksDeveloped an ensemble model that effectively predicts maternal health risks during pregnancyMay not work as well in populations with different health profilesDu et al.31Clinical studyInvestigate maternal height as a risk factor for adverse GDM outcomesFound that maternal height is an independent risk factor for adverse pregnancy outcomes in GDMLimited to maternal height; other factors may need further explorationSoria-Contreras et al.32Longitudinal studyInvestigate the lifetime impact of gestational diabetes on cognitive functionDemonstrated a potential long-term effect of gestational diabetes on cognitive function in midlifeLimited to a single cohort; larger studies are needed for confirmationYang et al.33Machine learning studyAnalyze neurodevelopmental delay in offspring using maternal risk factorsUsed machine learning to identify maternal and perinatal factors associated with neurodevelopmental delay in offspringLimited by the quality of population data used in the analysisDoğru et al.34Ensemble learning studyHybrid super ensemble model for early-stage diabetes risk predictionDeveloped a hybrid ensemble model that improved early-stage diabetes risk predictionDependent on complex model structures, making it computationally expensiveXu et al.35Epidemiological studyInvestigate dietary inflammatory index as a predictor of prediabetesFound that a higher dietary inflammatory index correlates with a greater risk of prediabetes in women with prior GDMFocused on diet, which may not account for all relevant metabolic factorsRustam et al.36Machine learning studyEnhanced diabetes detection using ensemble feature engineeringDeveloped an ensemble feature engineering approach to improve diabetes detection accuracy, significantly enhancing model performanceThe approach may require large amounts of data to effectively train and may be sensitive to the quality of input featuresOrdoñez-Guillen et al.37Machine learning studyClassify Type 2 diabetes mellitus subtypes using machine learning algorithmsIdentified distinct subtypes of Type 2 diabetes, improving prediction and personalized treatmentFocused only on Type 2 diabetes subtypes, which may not apply to GDM diagnosis and managementSalvatori et al.38Data-driven studyUse cluster analysis to identify and validate GDM subgroupsFound that data-driven cluster analysis could identify distinct subgroups within GDM patients, potentially improving personalized treatment strategiesLimited to cluster analysis; may miss other subgroups that do not fit into the identified categoriesNaito et al.39Machine learning studyInvestigate associations between environmental factors and cardiometabolic diseasesRevealed heterogeneous associations between environmental factors and cardiometabolic diseases, using polygenic risk scoresFocused on cardiometabolic diseases, which may not directly apply to GDM, and relied on polygenic risk scoresBelaghi et al.40Predictive modeling studyPredict preterm birth risk in multiparous women using logistic regression and machine learningDemonstrated that both logistic regression and machine learning models could effectively predict preterm birth in multiparous womenResults may not be generalizable to all populations, particularly those with single pregnancies or different risk factorsSolomon et al.41Review studyReview the role of machine learning for multifactorial genetic disorder predictionExtensive review of machine learning applications in predicting multifactorial genetic disorders, including diabetesReview-based; lacks original data or novel insights from new models or approachesLong et al.42Comparative studyCompare machine learning algorithms for predicting microalbuminuria riskShowed that machine learning algorithms could be effectively applied to predict microalbuminuria risk, an important complication of diabetesLimited to microalbuminuria and may not address other diabetic complications like GDMAghajanian et al.43Predictive modeling studyPredict post-delivery hemoglobin levels with machine learning algorithmsFound that machine learning models can accurately predict post-delivery hemoglobin levels, potentially guiding clinical interventionsResults may not directly apply to GDM, as it focuses on postpartum hemoglobin levelsKokori et al.44Narrative reviewReview the role of machine learning algorithms in the detection of gestational diabetesReviewed current evidence on the application of machine learning algorithms for GDM detection, highlighting various techniquesLimited by the narrative review format; lacks original experimental data and may omit certain approachesKaya et al.45Machine learning studyEarly prediction of gestational diabetes mellitus using machine learning modelsShowed that machine learning models can predict GDM at early stages with high accuracy, potentially improving early interventionRelied on a specific set of models; other predictive approaches may be overlookedKang et al.46Machine learning studyPredict GDM in Asian women using machine learning algorithmsDemonstrated the efficacy of machine learning algorithms in predicting GDM in Asian women, accounting for ethnic-specific risk factorsFocused solely on Asian women, limiting the applicability of the results to other ethnic groupsCubillos et al.47Machine learning studyDevelop models to predict GDM risk in the first half of pregnancyDeveloped predictive models for GDM risk early in pregnancy, allowing for early interventionsLimited by the dataset and may not account for other complex, unmeasured factors in GDM developmentChan et al.48Machine learning studyUse elemental contents in fingernails for early prediction of GDMFound that machine learning models using elemental contents in fingernails could predict GDM early, offering a non Invasive diagnostic toolThe technique is novel but requires validation across larger and more diverse populationsLi et al.49Machine learning studyUse machine learning to predict GDM in the first trimesterDemonstrated that machine learning algorithms can predict GDM in the first trimester with high accuracy, aiding early intervention strategiesThe first trimester prediction model may require additional clinical validation in broader populationsNuthakki et al.50Machine learning studyEarly detection of diabetes risk factors for improved health managementCreated a machine learning-based system for early detection of diabetes risk factors, aimed at improving overall health managementFocused primarily on early detection; may require integration with existing clinical workflows for effective use Comparative analysis of existing methods. Iteratively, Next, based on Table 1, In addition, Seifert et al.7 and Rüegg et al.8 report the protein expression, for example, galectin-7, of perioperative management in female patients undergoing fetal spina bifida repair, hence demonstrating how GDM might interact with other medical conditions. Machine learning approaches toward predicting GDM have lately become a central theme in the literature. Several publications have appeared using artificial intelligence for predictive modeling. For instance, Hung et al.9 studied the effects of lysophosphatidylcholine on mitochondrial function in GDM. Such research contributes to a trend of combining state-of-the-art computational approaches with biological data towards improving risk prediction. A further integration of data from various sources, such as wearable devices and clinical records, which has been considered a promising area for early diagnosis. For example, Gomes Filho et al.14 used Bayesian networks for the prediction of GDM in order to illustrate how a combination of more traditional machine learning techniques and newer computational tools can improve the prediction accuracy. In this context, Kolozali et al.15 focus on early prediction of biomarkers for GDM by background medical information combined with wearable devices, and once more, data integration plays an important role in predictive analytics. Marzouk et al.16 as well as Khan et al.17 used large datasets through machine learning, which helps in the upgrading of prediction systems of diabetes and the decrease in false-negative cases. Their findings signify the tremendous potential of employing machine learning in fine-tuning risk evaluation. In order to upgrade the predictive efficacy, ensemble models as well as feature transformation techniques proved to be useful according to Saleh Al Reshan et al.18 and Linkon et al.19. These studies are further complemented by the work of Alnowaiser20, Zhao et al.21, and Kurt et al.22, where the idea of using several machine learning algorithms for enhanced predictions especially in personalized healthcare has been a subject of discussion. This includes, for instance, tri-ensemble and KNN imputed features applied by Alnowaiser, as these emphasize more the scope of hybrid models to provide better prediction regarding diabetes risks in various populations. Another important contribution to the field is the use of digital health technologies. Lu et al.13 discuss how digital health and machine learning technologies are used for tracking blood glucose levels in pregnant women, emphasizing the role of real-time, data-driven health management tools. The integration of digital solutions and machine learning models brings new chances to manage GDM more in line with the person’s characteristics and proactivity. For both patients and providers, these possibilities are new. Other relevant studies include the work done by Uddin et al.25 and Du et al.26, that further supports the growing body of works regarding applications of machine learning into maternal health to predict risks such as obesity and birth asphyxia. These studies, though not limited only to GDM, provide insight into how machine learning can be applied to cross-sectional studies in terms of maternal and fetal health outcomes. As the studies further explore how machine learning can be implemented to determine the predictive capabilities associated with GDM, integration of clinical risk factors through algorithms of machine learning and adverse maternal outcomes is achieved in those studies. Bai et al.29 and Khadidos et al.30 have made predictions on risk issues regarding maternal health through the application of ensemble frameworks. Similarly, Du et al.31 and Soria-Contreras et al.32 also study risk factors by maternal height and long-term cognitive effects of GDM. Moreover, Yang et al.33, with their inclusion of population data and neurodevelopmental outcomes, reflects an expanding sense of awareness regarding a greater spectrum of GDM impact outside pregnancy condition sets. The work by Doğru et al.34 and Xu et al.35 points out the diet and inflammation role in predictive models of prediabetes and GDM, that is, it emphasizes the lifestyle factors in the design of predictive models. It is also novel in work by Rustam et al.36 and Ordoñez-Guillen et al.37, who employ ensemble feature engineering techniques to improve diabetes prediction due to better data representation. Such studies are indicative of the promise that personalized, lifestyle-oriented risk models may hold, that can be applied in real-time clinical environments for diabetes management. On the other hand, Salvatori et al.38 and Naito et al.39 focused more on cluster analysis and polygenic risk score use for the identification of subgroups within GDM, which provides more insight into the heterogeneity of the condition. This is further strengthened by the work of Belaghi et al.40, where logistic regression and machine learning are applied to add complexity to obstetric prediction models for a preterm birth prediction. The studies by Solomon et al.41 and Long et al.42 also depict integration in using multifactorial data to make predictions of both maternal as well as offspring outcomes toward advanced applications of machine learning in multifactorial disorder prediction. Recent developments regarding machine learning-based GDM prediction methods are being studied. These studies, conducted by Kaya et al.45, Kang et al.46, and Cubillos et al.47, involve the addition of early risk factors especially in ethnic groups. New biomarkers and environmental factors are introduced in these studies, which help in predicting GDM, based on the basic research done prior to these studies. Li et al.49 and Nuthakki et al.50 show how machine learning algorithms can be used in the first trimester to improve early detection of GDM, which will be essential for better health outcomes in both Mothers and their babies. In brief, this series of studies explores the multifaceted nature of GDM from biological mechanisms and biomarkers to machine learning models and digital health solutions. Convergence of clinical data, machine learning, and digital health technologies promises to revolutionize the management of GDM. In the near future, more accurate, real-time tools for early detection and personalized intervention sets will be offered to clinicians. Integration of diverse data sources and continued development of machine learning models will continue to improve the prediction and management of GDM, hence leading to better outcomes for both Mothers and their children in different scenarios. Proposed model This section presents design of an iterative Multi Method Framework that uses Graph Neural Networks for Networked Analysis of Gestational Diabetes Risk Factors, to overcome issues of low efficiency & high complexity which are present in existing methods. Initially, as in Fig. 1, The proposed framework integrates GraphSAGE, Relational Graph Convolutional Network (R-GCN), and Temporal Graph Attention Network (TGAT) as core components for analyzing gestational diabetes mellitus (GDM) risk factors. Together, these methods allow for robust learning from complex, heterogeneous, and temporally dynamic datasets and samples. Each method is used to target different challenges posed when modeling GDM risk factors—scalability, relational heterogeneity, and temporal evolutions. GraphSAGE acts as the base graph representation, aggregating information from its neighboring nodes to create the node embeddings. This approach solves the problem of scalability in large graphs, particularly in clinical datasets where data incompleteness and dynamic additions (for example, new patients) are common for the process. Mathematically, for a node ‘v’, the embedding hv’k at the k-th layer is computed via Eq. 1,1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h{v}{\prime}k = \sigma \left(Wk \cdot AGGREGATEk\left(\left\{h{u}{\prime}\left(k-1\right): u \in N\left(v\right)\right\}\right)\right)$$\end{document}where, Wk is the weight matrix for the k-th layer, σ is an activation function (ReLU), and N(v) denotes the neighbours of ‘v’ in the process. The aggregation function like mean or LSTM-based pooling is represented as AGGREGATE in the process. Embedding at the last layer catches both the local as well as global structural properties. This methodology is selected based on its ability to induce learning, meaning the model can generalize to unseen nodes, a critical requirement in datasets that have dynamic additions of patients. Iteratively, Next, as illustrated in Fig. 2, Relational Graph Convolutional Network (R-GCN) is incorporated to capture the multi-relational structure inherent in GDM risk factors, including the relationship between clinical, demographic, and environmental variables in the process.Fig. 1Model architecture for the proposed analysis process.Fig. 2Overall flow of the proposed analysis process. Model architecture for the proposed analysis process. Overall flow of the proposed analysis process. For a given node ‘v’, the k-th layer embedding is computed via Eq. 2,2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h{v}{\prime}k = \sigma \left(\sum_{r\in R}\sum_{u\in Nr\left(v\right)} \left(c(r,v)\right)\cdot W{r}{\prime}k h{u}{\prime}\left(k-1\right)+ W{0}{\prime}k h{v}{\prime}\left(k-1\right)\right)$$\end{document}where, R is a set of relation types. Nr(v) represents neighbors of ‘v’ for a given relation ‘r’. c(r,v) is a normalization constant, Wr’k and W0’k are weight matrices representing weight of relation-specific terms as well as self-loop terms. This formulation is the relation-aware one which allows for the differentiation between dependency kinds like causal and temporal during this process. The R-GCN architecture completes GraphSAGE by providing heterogeneity to the graph representation, thereby achieving a better understanding of the multi-typed relationships of the data samples. Successively, Next, following Fig. 2, Temporal Graph Attention Network (TGAT) further extends static graph analysis by GraphSAGE and R-GCN into temporal dynamics. In this, TGAT employs time-encoded embeddings that represent the time evolution of node features by reflecting the change in GDM risk factors over trimesters. For a node ‘v’ at timestamp t, the embedding is computed via Eq. 3,3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$hv\left(t\right)= ATTENTION\left(\left\{hu\left({t}{\prime}\right)+ \varphi \left(t-{t}{\prime}\right): u \in N\left(v\right), {t}{\prime}\le t\right\}\right)$$\end{document}where, φ(t − t′) is a timestamp encoding function, and ATTENTION represents an attention mechanism that assigns weights to neighboring nodes based on temporal relevance sets. The timestamp encoding function can be defined via Eq. 4.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi \left(t-{t}{\prime}\right)= sin\left(\frac{t-{t}{\prime}}{\tau }\right)$$\end{document}where, τ is a scaling parameter to adjust the granularity of temporal encoding process. By aggregating time-weighted embeddings, TGAT enables the model to predict future states of GDM risk factors. The overall objective function for training the combined framework is defined via Eq. 5,5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L = \left(\frac{1}{\left|V\right|}\right) \sum_{v\in V} \left[{\ell}\left(hv, yv\right)+ \lambda \sum_{k=1}^{K} \left|\left|Wk\right|\right|{F}^{2}\right]$$\end{document}where, ℓ(hv, yv) denotes the prediction loss for node ‘v’ (for example, cross-entropy for classification or mean squared error for regression), and λ||Wk||F is an L2 regularization term to prevent overfitting operations. The final node representation is formed by concatenating the embeddings generated by GraphSAGE, R-GCN, and TGAT, ensuring that both relational and temporal features are captured by the process. This combined design ensures that the framework exploits the strengths in the following: scalability of GraphSAGE, relational modeling of R-GCN, and TGAT’s temporal adaptability levels. Integrating these methods into the system enables it to solve special problems posed by GDM data and thus sets for high accuracy predictivity in the framework and puts forth for interpretability. The use of advanced architectures such as GNN in a sophisticated way brings the precision of risk predictions, which yet enables clinical decision making upon a holistic inspection of data structures underlying it. Iteratively, Based on Fig. 2, the proposed framework has four core modules: Graph Attention Network (GAT), BioBERT pretraining with GNN fine-tuning, Federated Graph Learning with Differential Privacy, and DropEdge regularization for studying the risk factors of GDM. These methods have been deliberately chosen for increasing model interpretability, better training speed, privacy-preserving analysis, and learning process regularized against overfitting, respectively. Each method complements the other, contributing to the comprehensive and robust model that integrates heterogeneous data, captures complex relationships, and adheres to privacy constraints. Graph Attention Network (GAT) is used to introduce interpretability into the model through attention scores that are given to edges depending on how much importance each of them contributes in the risk factor prediction for GDM. GAT makes use of a mechanism where each node ‘v’ attends to its neighbors by computing attention coefficients that denote the significance of each neighbor in updating the embedding of a node in the process. The attention coefficient αvu between two nodes v and u is given by Eq. 6,6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha vu = \frac{exp(LeakyReLU({a}^{T} [W hv | W hu]))}{(\Sigma u{\prime}\in N(v) exp(LeakyReLU({a}^{T} [W hv | W hu{\prime}])))}$$\end{document}where, ‘a’ is a learnable attention vector, W is the weight matrix, and hv and hu are the node features of v and u, respectively for this process. The operator | represents concatenation, and N(v) is the set of neighbors of node ‘v’ sets. The attention coefficient αvu shows the effect of node u on node v’s updated embedding process. Now, this process is repeated for all nodes and their neighbors after getting them via Eq. 7,7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h{v}{\prime}= \sigma \left({\Sigma }^{u\in N\left(v\right)} \alpha vu*W*hu\right)$$\end{document} Which, in turn, includes an application of the attention mechanism where σ is an activation function- ReLU for this. Thus, this attention mechanism ensures GAT focuses on more critical factors of risk, particularly considering that in GDM some might have a stronger potential as predictors than others do for the process. Iteratively, Next, BioBERT pretraining with GNN fine-tuning is employed to leverage domain-specific pretraining from medical literature and electronic health records. BioBERT, a variant of BERT trained on biomedical text, provides contextual embeddings for patient records, which are then fine-tuned within the GNN model process. For the pretraining stage, BioBERT optimizes the objective represented via Eq. 8,8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LBioBERT = -{\Sigma}_{i=1}^{N} log P(himask | hicontext)$$\end{document}where, himask is the masked word for token ‘i’, and hicontext is the context vector surrounding the masked token, representing the input patient data samples. Fine-tuning BioBERT with GNN embeddings involves training a joint loss function via Eq. 9,9Ltotal = LGNN + λ BioBERT LBioBERTwhere, LGNN is the loss obtained by the GNN model for instance cross-entropy or regression loss and λBioBERT is the hyperparameter to govern the significance of pre-training at fine-tuning time. The combined approach guarantees the fact that the model derives its benefits from generalization power of BioBERT as well as structural flexibility of GNNs. Federated Graph Learning with Differential Privacy is used to train the model across decentralized patient datasets while maintaining data privacy levels. In this setting, each institution trains a local model on its own data, and only model updates and not raw patient data are shared with a central server for the process. The aggregation of model updates is done through a secure federated learning protocol such that the global model is trained without revealing sensitive patient information sets. Mathematically, the federated learning update rule is given via Eq. 10,10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$wt =\frac{1}{K} {\Sigma}_{k=1}^{K} wt\left(k\right)$$\end{document}where, wt(k) refers to the model weights at iteration t from institution k, and K is the total number of participating institutions in process. Differential privacy is incorporated by adding noise to the model updates via Eq. 11,11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$wt\left(k\right)\leftarrow wt\left(k\right)+ N\left(0, {\sigma }^{2}\right)$$\end{document}where, N(0, σ2) refers to Gaussian noise with variance σ2 added to the model updates in order to ensure that the individual patient data remains private during communication operations. Iteratively, to optimize the whole process, DropEdge regularization is applied to the graph data to improve the model’s generalization and robustness. DropEdge works by randomly removing a subset of edges from the graph during training, forcing the model to learn diverse pathways for message passing operations. Let G = (V, E) be the graph, where V is the set of nodes and E is the set of edges. The DropEdge process operates on the graph via Eq. 12,12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}{\prime}= E \setminus D$$\end{document}where, D is a randomly selected subset of edges with probability p, and E’ represents the modified edge sets. Thus, during training, the operation forces the model to rely on alternative connections between nodes and hence it can be expected to reduce the overfitting operations along with better generalization. The final output of the model is the node embedding hv, which encodes the risk factor information for each of the patients. For a given node v, after applying the GAT, BioBERT fine-tuning, federated learning, and DropEdge regularization, the output embedding is computed via Eq. 13,13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$hv = DropEdge\left(Federated Aggregation\left(BioBERT Fine-Tuned GNN\left(hv\right)\right)\right)$$\end{document} This final embedding could further be utilized for the prediction of downstream tasks like GDM risk stratification and early detection. It can accurately predict along with an explanatory model to support the predictions made for the clinicians and ensure the integrated framework will be both robust and efficient in processing heterogeneous, decentralized, and privacy-sensitive data involved in GDM risk prediction, which in the meantime also enhances the explainability of the model as well as its generalization ability. Attention mechanisms combined with domain-specific pretraining, secure decentralized learning and regularization techniques provide a comprehensive solution to analyze complex data samples from the healthcare domains. Then we will discuss the proposed model efficiency in terms of different metrics and compare our results with existing methods based on several scenarios. Complexity analysis In gestational diabetes prediction, single-model solutions often underfit essential relational patterns. The iterative multi-method fusion strategy handles this complexity in process. Though more laborious than linear or tree-based models, graph learning methods aim to balance representation depth and deployments. In graph neural network fusion, early layers extract low-level relationships and subsequent levels merge complimentary representations. Attention-based layers focus on patient cluster interactions, while initial graph convolutional layers capture local relationships. Controlled depth—three to four layers per model—and limited parameter sharing stabilize training and prevent resource overuse in process. Evaluation metrics make this method more realistic for the process. On normal GPU hardware, training took 40 min and inference took < 2 s each batch on 8000 patient records with 12 features each in process. The approach employs many graph models yet is computationally feasible for clinical integration sets. The perceived complexity trade-off is calibrated to achieve clinically significant accuracy without excessive computing overheads. Discussions Dataset and cohort The research makes use of two distinct types of data, and each of these types serves a distinct function in the investigation. There are 10,482 pregnancy records that are appropriate for inclusion in the primary group. They stem from four different medical institutes that collaborate with one another in process. The purpose of this tool is to construct models, make internal checks on them, and evaluate how effectively they function. The mother’s age, body mass index, blood pressure, glucose-related lab tests, her medical history, and a note stating that she had been diagnosed with gestational diabetes during standard screening at 24–28 weeks of pregnancy were all required to be included in the records in order for them to be eligible. Additionally, the records needed to include additional information regarding the mother’s medical history. All of the records were discarded if the individual had diabetes before to becoming pregnant, if significant outcome labels were absent, if longitudinal measurements were incomplete, or if the continuity of time was disrupted by shifting from one institution to another. According to the findings of this study, the formal clinical diagnosis of GDM outcome status was provided within the typical screening period of 24–28 weeks. In the form of a binary endpoint, it is stored. From the UCI Pima dataset, the external validation group contains 768 records. These records were obtained from the dataset. Their sole purpose is to assess generalization in a context that has a restricted set of characteristics. The main experiments were conducted on a retrospective cohort of pregnancies created with structured EHRs, trimester-specific clinical metrics, obstetrical histories, lab metrics and representations of clinical notes. The patient records were divided at the patient level prior to the development of the model so that no record of a single pregnancy would appear in both training, validation and test sets. Institutional divisions when presented will be viewed as part of the internal developmental division of the data rather than a demonstration of completely independent external validation unless there exists an entirely independent GDM cohort with a separate governing body, use of different diabetic coding and different means of collecting data. The UCI Pima Indians Diabetes dataset has not been utilized as an externally validating dataset for predicting gestational diabetes because it is not a gestational-diabetes cohort. It was maintained solely as a limited-feature public diabetes benchmark to assess if the model pipeline could execute under a significantly-reduced tabular feature space. Since the Pima dataset does not represent a gestational-diabetes cohort, nor does it include trimester-specific trajectories of pregnancy, obstetric screening periods, clinical notes, or GDM diagnosis labels, the results from this dataset have been changed to reflect auxiliary robustness evidence instead of external clinical validation. Therefore, claims of external validity have been lessened and the manuscript can no longer report the results from this dataset as demonstrating multi-site generalization to other sites for GDM prediction. An independently-governed multi-site external validation dataset remains a significant limitation of this current study. While the internal cohort allows for testing at the patient-level under controlled splits (i.e., train/validation/test), generalizing to previously unseen sites, populations, diagnostic processes and care pathways for pregnant women cannot be stated as being totally substantiated without access to an independently-governed GDM cohort. Therefore, future validations should include prospective collection of new GDM datasets which are governed separately from the present study and are harmonically defined for outcomes, provide subgroup reporting for specific sites and evaluate models with pre-registered protocols. Ablation graph construction and component-wise analysis Each patient is a graph node with demographic, clinical, biochemical, and longitudinal data. Demographical similarities, shared comorbidities, laboratory trajectories, and temporal proximity are clinically relevant edges. Long-term metabolic correlations are distinguished from short-term clinical fluctuations using causal, demographic, behavioral, and physiological relationship types. The model tracks trimester-aligned physiological changes with timestamped node feature updates. To stabilize partial record representations while maintaining cohort structure, graph message-passing collects embeddings from surrounding nodes. Both datasets use the same graph pipeline. Building clinical similarity edges using thresholded Euclidean distance across standardized quantitative characteristics. Patients with comparable diet, activity, or smoking histories develop behavioral edges. BMI and physiological risk factors affect fasting glucose trajectories. Temporal links describe trimester glucose, insulin, and blood pressure changes. From the heterogeneous graph’s multiple relation channels that encode diverse medical information, downstream models can incorporate structural, relational, and dynamic disease progression signals. Ablation experiments evaluate model components. AUC drops seven percentage points when the temporal element is eliminated and trimester-specific glucose evolution is removed. When relational graph modeling is removed, clinically relevant patient heterogeneity drops by 4%. Excusing attention-driven interpretability reduces clinical expert alignment and risk-driving factor discrimination. For illness processes like gestational diabetes, the combined model of relational structure, temporal evolution, and contextual interpretability beats all ablated Variants In Process. These findings show that the multidimensional architecture represents pregnancy-related metabolic health’s interconnected and dynamic nature, not merely GNN Variations In Process. Each component increases clinical analytical performance, stability, and interpretability, according to ablation data samples. Standardisation, replicability, experiments Clinical prediction research use validated machine learning and deep learning baselines for comparison. If applicable, logistic regression with elastic regularization, random forest classifiers, gradient boosting machines, multilayer perceptrons, and convolution-based temporal networks are benchmark models. For relational comparisons, graph convolutional and graph attention networks are structural baselines. These models from different methodological families weigh whether the proposed framework outperforms conventional and graph-based alternatives. Experimental arrangement is repeatable. All runs employ fixed random seeds for data splitting, parameter setup, and stochastic optimization. The Python-PyTorch Geometric pipeline uses an 80 GB NVIDIA A100 GPU. Grid search on validation subset chooses baseline model hyperparameters such depth, learning rate, and regularization strength. Standard embedding sizes, uniform training epochs, and consistent learning rate schedules across cohorts help graph-based components compare fairly. Modular, independent data pretreatment scripts, graph generation methods, and assessment utilities improve reproducibility. One code repository command can rerun the workflow from raw data ingestion to performance reporting. Model configuration files specify architectural and training parameters including dropout rates, attention heads, and message-passing depth sets. This approach streamlines findings verification and research expansions. For controlled evaluation of different approaches, baselines and repeatability procedures provide a robust framework. Interpretable and widely accepted benchmark models strengthen the findings and help the research community position the framework in a machine learning context in process. Results reporting splits, confidence intervals, calibration, and clinical utility The assessment method divides patient-level data to avoid validation and testing with training set members. The multi Institution cohort maintains outcome prevalence across subsets with a stratified 70-15-15 split, while the Pima-derived dataset does the same for consistency. This methodology evaluates generalization in realistic deployment settings where unseen patients must be predicted. Bootstrapping with 1000 resampled test subsets gives confidence ranges for all measures and substantial uncertainty estimates for AUC, F1 score, and recall sets. Calibration tests clinical reliability. Expected Calibration Error and Brier scores reveal that the graph-based approach aligns probability outputs across risk strata better than non-graph baselines that overstate high threat. In risk threshold scenarios requiring clinical action or screening, calibration curves show seamless alignment between projected and observed risk across all probabilities. Performance across age, ethnicity, and BMI is analyzed for fairness. With little cohort imbalance swings, the model maintains AUC values across most groups. These findings suggest that multi-relational representation supports equitable risk modeling across clinically significant subpopulations. Additional stratified memory analyses increase sensitivity across historically underrepresented populations, crucial for maternal health equity. At varied risk thresholds, decision-curve analysis evaluates the model’s net therapeutic benefit. The graph-based method outperforms traditional classifiers across many obstetric criteria. Combining relationship structure and temporal advancement is statistically and therapeutically helpful for the process. The model’s clinical decision-making behavior is assessed for accuracy, calibration, fairness, and clinical utility. Comparative result analysis The graph-based model that was suggested for predicting the risk of gestational diabetes mellitus is primarily tested on a group of 10,482 women who were recruited from four different clinical sites for the purpose of comparison tests. Two hundred and thirty structured variables are included in the whole electronic health record inventory for this primary cohort. To conduct the studies that are presented in Tables 2, 3, 4, 5, 6, 7, and 8, an analytical subset consisting of more than fifty variables that are clinically important was selected. Some of these factors include biochemical markers, demographic information, obstetric history, lifestyle indicators, and longitudinal monitoring fields. Other variables include lifestyle indicators. Among the 768 records that make up the external UCI Pima dataset, there are eight clinical features. It is separated from the primary tables of comparison data and is solely utilized for the purpose of external validation in a setup that contains a limited number of features.Table 2Classification Accuracy Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelClassification accuracy (%)76.580.182.488.2Table 3Area Under the Curve (AUC) Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelAUC0.750.800.840.90Table 4Root Mean Square Error (RMSE) for Temporal Prediction on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelRMSE (temporal)0.450.380.320.30Table 5F1 Score Comparison for GDM Classification on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelF1 score0.740.780.800.86Table 6Model Training Time Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelTraining timestamp (hrs)2.53.04.05.5Table 7Precision and Recall Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelPrecision0.720.770.790.83Recall0.680.740.770.85Table 8Explainability Comparison (Attention Scores Agreement with Domain Experts) on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelAttention agreement (%)65727582 Table 2Classification Accuracy Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelClassification accuracy (%)76.580.182.488.2 Classification Accuracy Comparison on the Primary Multi Institution Cohort (n = 10,482). Table 3Area Under the Curve (AUC) Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelAUC0.750.800.840.90 Area Under the Curve (AUC) Comparison on the Primary Multi Institution Cohort (n = 10,482). Table 4Root Mean Square Error (RMSE) for Temporal Prediction on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelRMSE (temporal)0.450.380.320.30 Root Mean Square Error (RMSE) for Temporal Prediction on the Primary Multi Institution Cohort (n = 10,482). Table 5F1 Score Comparison for GDM Classification on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelF1 score0.740.780.800.86 F1 Score Comparison for GDM Classification on the Primary Multi Institution Cohort (n = 10,482). Table 6Model Training Time Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelTraining timestamp (hrs)2.53.04.05.5 Model Training Time Comparison on the Primary Multi Institution Cohort (n = 10,482). Table 7Precision and Recall Comparison on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelPrecision0.720.770.790.83Recall0.680.740.770.85 Precision and Recall Comparison on the Primary Multi Institution Cohort (n = 10,482). Table 8Explainability Comparison (Attention Scores Agreement with Domain Experts) on the Primary Multi Institution Cohort (n = 10,482).MethodMethod [5]Method [8]Method [18]Proposed modelAttention agreement (%)65727582 Explainability Comparison (Attention Scores Agreement with Domain Experts) on the Primary Multi Institution Cohort (n = 10,482). In order to normalize all continuous variables in the main cohort, z-scores are utilized, and one-hot encoding is utilized for all categorical variables. It is possible to employ neighborhood-aware graph propagation to deal with missing structured values so long as the level of missingness remains below a quality control threshold that has already been established. When doing the internal review of the main group, a stratified train Validation-test split of 70–15–15 is utilized at the patient level. This is done in order to maintain the outcome prevalence and prevent information from being disclosed. The information is what determines how the graph is constructed. Nodes represent the pregnancies that are part of the primary cohort. These nodes are connected to one another through clinical resemblance, comorbidity linkage, temporal progression, and obstetric risk correspondence ties. Using only feature-space similarity across the eight clinical characteristics that are accessible, the graph in the external UCI Pima cohort has been purposefully simplified in order to make it easier to understand. There are no text-based or temporal linkages that are imposed on the data under any circumstances. The accuracy of classification will represent how well the model can classify both positive and negative cases of GDM, which is important in real-time healthcare scenarios since high accuracy reduces the number of misclassifications that may lead to false negatives or false positives. In Table 2, classification accuracy of 88.2% is attained by the proposed model, with a significant superior position against the remaining methods: Method [5] (76.5%), Method [8] (80.1%), and Method [18] (82.4%). Therefore, an improved accuracy directly translates to more reliable predictions for GDM, with a higher percentage of at-risk patients identified early and followed through with timely intervention sets. This will decrease the risks for adverse pregnancy outcomes such as a preterm birth or macrosomia through ensuring that it allows more accurate screenings and interventions in process in real time scenarios. The ability of this model in capturing the complex relationship amongst the risk factors contributes further to its higher accuracy, thereby forming an enhanced and more robust solution in an iterative set of clinical deployments. AUC is a significant metric for evaluating the trade-off between true positive and false positive rates. A higher AUC indicates better performance in distinguishing between GDM-positive and GDM-negative cases. The proposed model achieves a significantly higher AUC of 0.90 compared to Method [5] (0.75), Method [8] (0.80), and Method [18] (0.84). This means that the proposed model is highly efficient in minimizing false positives and false negatives, which is very important in clinical practice where the wrong classification may lead to either unnecessary interventions or missed diagnoses. With better AUC values in the real-world applications, the implication will be a lower number of wrongly diagnosed GDM patients as false positives or missed patients in the screening process. This implies better management in terms of gestational diabetes and healthcare costs savings. The number of follow-up tests and treatments is also reduced in the model proposed in previous operations. The ability of forecasting accuracy can be measured by RMSE, especially in the context of time Varying data. For example, how risk factors of GDM may change with a trimester of pregnancy period. Lesser RMSE ensures better predictability over temporal instances. It is observed that an RMSE of 0.30 is obtained with proposed model and it is lesser among all techniques. In clinical terms, this would mean that the proposed model not only identifies the presence of GDM but also monitors its progression, which is an important feature for early interventions that can prevent complications. Through the precise forecasting of a patient’s risk of GDM, healthcare professionals could intervene with tailored interventions at appropriate points in time and, accordingly, adjust their treatment plans according to predicted trajectories of risk, thereby optimizing patient outcomes. The F1 score achieves a balance between precision and recall, such that the model will better classify positive instances while reducing false negatives. Using an F1 score of 0.86 for our proposed model, this will be significantly higher than methods [5] (0.74), methods [8] (0.78), and methods [18] (0.80). This indicates that the proposed model not only identifies more GDM cases (recall) but also with higher accuracy (precision) and hence an ideal tool for clinical decision Making process. The practice that reflects a higher F1 score would indicate fewer missed diagnoses and fewer unnecessary interventions. Thus, it would find a better balance between the need to treat all at-risk patients and the desire to avoid unnecessary treatments. In training timestamps, the factor for deployment of machine learning models in real-time environments, particularly clinical settings, where speed and efficiency are crucial. The proposed model is trained in 5.5 h, which is much larger than Method [5] that is trained in 2.5 h, Method [8] that takes 3.0 h and Method [18] trained in 4.0 h. While the new model is computationally complex in time, it is reasonable, mainly because of its high lifts on the level of interpretability and predictive accuracy. Real time clinical situations, a clinically learned model would then be highly able to provide predictions while without frequent retraining since that would make healthcare service providers to have better well Informed decisions rapidly and even without much ado in cases where health care facilities are usually with high volumes for the process. Precision and recall are key measures to know how well the model does in detecting true positives (GDM patients) with minimum false positives and false negatives for the process. The model achieved the highest precision and recall at 0.83 and 0.85, respectively, and outperformed Method [5] at 0.72, 0.68; Method [8] at 0.77, 0.74; and Method [18] at 0.79, 0.77. This would make the model proposed more accurate in spotting at-risk patients, the most critical requirement in any clinical practice since false negatives (undetected diagnosis) and false positives (unnecessary intervention) significantly impact the real world. Higher recall will make fewer instances missed, while higher precision is that more patients labeled at risk actually have GDM for more accurate and reliable screenings. The deployment of AI models in health necessitates explainability because the clinician needs to know what reasoning is behind predictions that models make. Use of Graph Attention Networks by this proposed model has gained attention agreement of 82% with domain experts, out of all methods, with scores assigned to nodes and edges. Method [5] had much lower agreement with only 65%, and Method [8] had an agreement of just 72% indicating its predictions are tougher for the clinicians to understand. Maybe this new model provides more clarity into what sorts of risk factors most predictively drive GDM, hence will be better accepted and believed by health care professionals, and in turn would lead to more sensible decision making from a greater understanding of what is driving those risks. As shown in Figs. 3, 4, 5, and 6, the proposed model as per Table 9, outperformed traditional machine learning methods on all key metrics: accuracy, AUC, RMSE, F1 score, and explainability. Although the model takes a longer time to train, the superior predictive capabilities and interpretability make it an invaluable tool in real-time clinical settings. The proposed model promises to make a significant improvement in the early detection and management of gestational diabetes for the process. This will lead to enhanced patient outcomes and reduced healthcare costs. Next, we talk about an iterative validation use case for the proposed model, which will assist the readers to further understand the entire process.Fig. 3Model’s overall performance analysis.Fig. 4Model’s precision recall analysis.Fig. 5Model’s training delay analysis.Fig. 6Model’s integrated performance analysis.Table 9Model’s Overall Result Analysis.Table/metricLogistic regressionRandom forestGradient boosting machineProposed graph frameworkClassification Accuracy (%)76.580.182.488.2AUC0.750.800.840.90Temporal RMSE0.450.380.320.30F1 Score0.740.780.800.86Training time (hrs)2.53.04.05.5Precision0.720.770.790.83Recall0.680.740.770.85Expert agreement (%)65727582 Model’s overall performance analysis. Model’s precision recall analysis. Model’s training delay analysis. Model’s integrated performance analysis. Table 9Model’s Overall Result Analysis.Table/metricLogistic regressionRandom forestGradient boosting machineProposed graph frameworkClassification Accuracy (%)76.580.182.488.2AUC0.750.800.840.90Temporal RMSE0.450.380.320.30F1 Score0.740.780.800.86Training time (hrs)2.53.04.05.5Precision0.720.770.790.83Recall0.680.740.770.85Expert agreement (%)65727582 Model’s Overall Result Analysis. Extended discussions Dataset model analysis and reconstructing the clinical foundations Within the scope of this investigation is a clinical dataset that is specific to pregnancy. A main development cohort and an external validation cohort are the two groups that make up this set of participants. The primary component is comprised of 10,482 pregnancies that were collected from four different clinical facilities. The training, validating, and testing against other groups are all facilitated further by this. The UCI Pima dataset contains 768 records that are part of the external group. These records are solely preserved for Validation purposes in a setup that is feature-limited. An internal comparative analysis and an external benchmark are prevented from becoming confused as a result of this circumstance. In addition to this, it ensures that the objectives, dimensions, and reporting scope of each cohort are made available across the entirety of the text. Graph creation, temporal structure, model components Clinical reason underpins graphs. One pregnancy’s demographics, baseline metabolic indices, trimester-specific metrics, and text-derived features are in each node. Medically interpretable similarities from gestational age alignment, shared comorbidities, test abnormalities, and risk variables produce edges. Patients with similar BMI, glycemic response, or obstetric history have edges through relationship-specific routes. These channels’ demographic similarity, metabolic relationship, obstetric environment, and longitudinal advancement create a heterogeneous graph showing maternal health’s complexity. Trimester-related feature updates symbolize time. Each pregnancy has early, mid, and late metabolic stages. The temporal architecture tracks growing fasting glucose, insulin sensitivity, and blood pressure. Transitions are used in temporal attention modeling to identify clinically relevant inflection points. The multi-stage structure shows physiological dynamics rather than static thresholds cause GDM sets. One module at a time is removed while the others are kept to study architectural components. Temporal modeling reduces prediction power, especially in early-stage GDM identification, hence trimester evolution provides critical risk cues. Removal of relational modeling limits the system’s ability to distinguish patients with similar baseline features but different metabolic pathways. Clinical narratives can capture minor symptoms and physician observations that numerical data cannot, hence deleting text-derived embeddings reduces sensitivity. Improved graph definition and ablation analysis provide a clear scientific foundation, revealing how each component embodies the complex interaction of physiological and environmental factors in GDM development sets. Baselines and replicability The reconstructed comparison framework uses clear baseline models. Classic benchmarks include logistic regression with regularization, random forests, gradient boosting machines, and multilayer perceptrons. Standard graph convolutional networks, graph attention networks, and static relational GNNs without temporal components are compared graph-based. Time baselines include trimester-trained LSM and gated recurrent units. Grid search adjusts learning rate, depth, and regularization for each baseline using patient-level splits as validation subgroups. All experiments use reproducible computing. All models are trained and assessed on an NVIDIA A100 GPU with fixed random seeds, deterministic dataloader settings, and consistent train Validation-test splits. Different modules with version-controlled configuration files preprocess features, graphs, and encode text. To reconstruct the pipeline, these files specify model hyperparameters, optimization schedules, dropout probabilities, and attention configurations. Clinical metrics for model outputs include AUC, recall, specificity, calibration quality, and net clinical benefit. Classical models discriminate well but lack temporal sensitivity. Graph baselines show relational structure but lack the integrated framework’s interpretability and temporal consistency. Final findings clearly compare tuning strategies, computational resources, and run-to-run Volatility In Process. Results analysis, calibration, fairness, and utility analysis To prevent temporal slice or feature update leakage, all performance evaluations use patient-level stratification. The primary cohort uses a 70-15-15 train Validation-test split, while the external cohort evaluates out-of-distribution alone. Confidence ranges for AUC, recall, specificity, and F1 score variability are calculated from 1000 test cohort bootstrap samples. The integrated framework performed consistently throughout various evaluations, indicating its robustness beyond a Value In Process. Predicted calibration error and Brier score assess calibration. The predicted probabilities match GDM prevalence across risk strata, proving robust calibration. In clinically important intermediate risk zones with few intervention choices, calibration curves are smooth and monotone. Relations and temporal models assess high-risk probability better than non-graph baselines. Fairness analysis contrasts maternal age, ethnicity, BMI. AUC and sensitivity variations in these subpopulations are within tight confidence bands. This consistency shows that the model’s relational and temporal signals do not increase demographic biases in smaller or imbalanced maternal health datasets and samples. Younger mothers and people with normal BMIs are now identified more often because to stratified false negative rates. For clinical utility, decision-curve analysis measures net benefit across realistic risk thresholds. Most threshold values indicate a bigger net gain from the integrated model than classical baselines, allowing its application in prenatal decision-making. Early detection improves with relational structure, temporal progression, and text-derived contextualization, enabling doctors determine whether to test glucose or start lifestyle counsellings. This expanded evaluation reveals clinical use beyond accuracy metrics. Methodological specificity There is a distinct section dedicated to explaining the methodological framework for each dataset. This is done to ensure that the data content and graph design are not confused with one another for the process. A total of 230 structured variables were extracted from electronic health records and included in the primary group. These variables come from a variety of fields, including demographics, biochemistry, obstetrics, comorbidities, medication, lifestyle, and longitudinal monitoring. Out of these 230 factors, more than fifty variables that are considered to be clinically significant are selected to be utilized in the primary comparative tests. As a result of quality control, the elimination of duplicates, and the elimination of variables that are absent, this is the result. For the purpose of being a low-dimensional benchmark, the external UCI Pima dataset is utilized. A total of 768 data and eight clinical criteria are included. Each pregnancy is represented as a node in the main group, and it possesses multimodal features. These characteristics originate from structured variables that have been selected with great care, as well as text-based embeddings. In order to construct edges based on clinical resemblance, common comorbidity structure, temporal progression, and obstetric-risk correspondence, a complicated graph is utilized. For the purpose of the external benchmark, every record is represented as a node, and the establishment of edges is solely determined by the degree of similarity in feature space among the eight clinical variables that are used. There are no temporal or text-based associations that have been placed on this dataset in any way for the process. Interpretability mechanism The framework is simple to comprehend because it was designed with a multi-layered attention-driven and feature attribution method. This technique provides explanations of predictions on both a local and global level. When used at the structural level, Graph Attention Networks provide surrounding nodes with normalized attention coefficients. This is an effective method for measuring the degree to which each patient or feature relationship influences the predictability of the goal. Following the completion of the training, these characteristics are removed and combined with clinically significant variables such as glucose levels, body mass index, and insulin resistance markers. As a result, you will have a direct understanding of how the model functioned. It is possible to determine the importance scores at the feature level by adding up the attention weights for each and every type of link and each time step. Because of this, it is feasible to identify the factors that are most significant during each trimester. To give an example, glucose levels during the second trimester are always given more attention, with average contribution scores of approximately 0.42. Body mass index (BMI) comes in second with 0.27, and age comes in third with 0.14 overall. The interpretability process is made even more dependable by the fact that these kinds of distributions are consistent with what we know about clinical practice. By placing more weight on recent measurements than on previous data, which is how diagnostics are done in the real world, temporal attention pulls out progression tendencies even more. This is something that is commonly done in the medical field. There is also the benefit of using space visuals and cluster analysis, which helps to make the data more understandable. In the process of shifting patient embeddings into lower-dimensional manifolds, groupings that are clinically similar have been observed. Consequently, this reveals hitherto unknown subpopulations that share metabolic pathways. As a result of these groupings representing varying degrees of risk, medical professionals are better able to determine how individual projections fit into patterns that are observed throughout the entire community. The incorporation of textual embeddings makes it simpler to comprehend by utilizing semantic cues derived from clinical notes to capture minor indicators such as descriptions of symptoms and notes from the attending physician. For the purpose of doing quantitative confirmation of interpretability, expert agreement analysis is commonly utilized. A total of 300 instances were selected at random, and five medical professionals examined the degree to which model Identified risk factors corresponded with clinical thought. The agreement rate reached approximately 82%, which was significantly higher than the rates of 65–72% that were found in the baseline models. The results of this examination demonstrate that the model not only provides explanations that are logically sound and based on domain experience, but it also offers predictions that are more accurate for real time scenarios. This lends support to its utilization in healthcare settings where judgments have significant weight in process. Consistent data and methods section As a coherent framework, the facts and the trial plan are presented together in this presentation. A total of 10,482 pregnancy records were collected from four different medical institutions that collaborated in order to create the primary cohort. This cohort was the sole one utilized in the internal studies that are included in Tables 2, 3, 4, 5, 6, 7, and 8, unless the captions make it obvious that this is not the case. The first set of pupils is the only one that participates in federated learning classes. A decentralized training process is carried out in a federated environment, during which the data are divided among five computing clients. The four clients correspond to the four source institutions, and the fifth client is a secure aggregation partition that is utilized to maintain the stability of privacy-preserving optimization. The article, therefore, creates a distinct distinction between the five federated training customers and the four institutions that are responsible for data collection. There were no other external institutions that submitted information about patients. Transparent comparative evaluation Rebuilding the framework for comparison is being done in order to guarantee that benchmarking is conducted in a fair manner and may be repeated multiple times. All of the baseline models are constructed and trained on the identical splits of the dataset as the framework that was suggested. This is a straightforward fact. All of the following are examples of different kinds of neural networks: These include simple graph convolutional networks, random forest classifiers with 200 trees, gradient boosting machines with depth 5, multilayer perceptrons with two hidden layers, and random forest classifiers with 200 trees. The grid search over hyperparameters technique is utilized in order to guarantee that every model performs to the best of its abilities. Through the utilization of bootstrapped confidence intervals, which are derived from the resampling of one thousand test subsets, statistically sound performance measures are provided. It has an accuracy rate of 88.2%, an area under the curve (AUC) of 0.90 ± 0.01, and an F1 score of 0.86 ± 0.02. A confidence range of ± 1.3% is being used for real time scenarios. In this work, analytically we use the Brier score and combine that evaluation with Expected Calibration Error (ECE) functions. These functions allows us to determine whether or not the calibration has been achieved during evaluations. A Brier score of 0.11 and an ECE of 0.03 are assigned to the model that has been suggested. This indicates that the anticipated probability and the outcomes that have been observed are very closely related to one another. The examination of fairness is performed on categories that are based on age, body mass index (BMI), and race, and the differences in area under the curve (AUC) remain within a range of ± 2% for all of these categories. When testing the accuracy of temporal predictions, the root mean square error (RMSE) on real longitudinal glucose trajectories is utilized. After obtaining a result of 0.30, it can be concluded that the model is capable of accurately depicting how progression varies throughout the course of timestamp sets. When measuring the level of agreement among experts, a structured evaluation technique is utilized. When it comes to each prediction, clinicians evaluate the significance of the features that are listed highest. In order to determine the degree of agreement, the percentage of overlap that exists between the characteristics selected by the model and the elements identified by the clinician is utilized as a measurement. It can be said that 82% of the elements are aligned on average. All of the datasets, models, metrics, and protocols are included in the newly presented consolidated assessment table that has been included. Because of this, everything is made plain, and you are able to compare directly across all of the experimental installations for the process (Tables 10, 11, 12, 13, 14, 15, 16, 17).Table 10GraphSAGE Output for Patient 1.FeaturePatient 1Neighbor 1 (Age 30, BMI 29, Glucose 115)Neighbor 2 (Age 27, BMI 33, Glucose 125)Aggregated featureAge28302729BMI31293330.5Glucose level120115125120Activity levelLowHighLowLowNumber of pregnancies2132EthnicityHispanicCaucasianHispanic–Table 11R-GCN Output for Patient 1.FeatureAggregated featureCausal relation (BMI—> Glucose)Temporal relation (Glucose over time)Final embeddingAge29––29BMI30.5Positive Correlation (Higher BMI—> Higher Glucose)–32Glucose level120–Temporal Change (Increasing Glucose in Trimester 2)125Activity levelLow––LowNumber of pregnancies2––2EthnicityHispanic––HispanicTable 12TGAT Output for Patient 1 (Predicted Glucose Level for Next Trimester).FeatureTrimester 1Trimester 2 (observed)Trimester 3 (predicted)Glucose level (mg/dL)110120125Activity levelLowLowLowTemporal attention–0.850.75Table 13GAT Output for Patient 1 (Attention Scores for Key Risk Factors).FeatureAttention score (relative importance)Age0.15BMI0.25Glucose level0.45Activity level0.10Number of pregnancies0.05Ethnicity0.10Table 14BioBERT Pretraining with GNN Fine-Tuning Output for Patient 1.FeaturePretraining outputFine-tuned outputAge2929BMI30.532Glucose level120125Activity levelLowLowNumber of Pregnancies22EthnicityHispanicHispanicTable 15Federated Learning with Differential Privacy Output.Model weights updateInstitution 1Institution 2Institution 3Global model (aggregated)Weight for glucose0.450.500.480.47Weight for BMI0.300.280.350.31Weight for Age0.150.180.170.17Table 16DropEdge Output for Patient 1.FeaturePre-DropEdgePost-DropEdge (regularized)Age2929BMI3230.5Glucose level125120Activity levelLowLowNumber of pregnancies22Table 17Final output for patient 1.FeaturePredicted valueExplanation (attention score)GDM riskHighGlucose (0.45), BMI (0.25)Glucose level125Most influential factorBMI30.5Contributes significantlyAge29Minor contributionActivity levelLowMinor contribution Table 10GraphSAGE Output for Patient 1.FeaturePatient 1Neighbor 1 (Age 30, BMI 29, Glucose 115)Neighbor 2 (Age 27, BMI 33, Glucose 125)Aggregated featureAge28302729BMI31293330.5Glucose level120115125120Activity levelLowHighLowLowNumber of pregnancies2132EthnicityHispanicCaucasianHispanic– GraphSAGE Output for Patient 1. Table 11R-GCN Output for Patient 1.FeatureAggregated featureCausal relation (BMI—> Glucose)Temporal relation (Glucose over time)Final embeddingAge29––29BMI30.5Positive Correlation (Higher BMI—> Higher Glucose)–32Glucose level120–Temporal Change (Increasing Glucose in Trimester 2)125Activity levelLow––LowNumber of pregnancies2––2EthnicityHispanic––Hispanic R-GCN Output for Patient 1. Table 12TGAT Output for Patient 1 (Predicted Glucose Level for Next Trimester).FeatureTrimester 1Trimester 2 (observed)Trimester 3 (predicted)Glucose level (mg/dL)110120125Activity levelLowLowLowTemporal attention–0.850.75 TGAT Output for Patient 1 (Predicted Glucose Level for Next Trimester). Table 13GAT Output for Patient 1 (Attention Scores for Key Risk Factors).FeatureAttention score (relative importance)Age0.15BMI0.25Glucose level0.45Activity level0.10Number of pregnancies0.05Ethnicity0.10 GAT Output for Patient 1 (Attention Scores for Key Risk Factors). Table 14BioBERT Pretraining with GNN Fine-Tuning Output for Patient 1.FeaturePretraining outputFine-tuned outputAge2929BMI30.532Glucose level120125Activity levelLowLowNumber of Pregnancies22EthnicityHispanicHispanic BioBERT Pretraining with GNN Fine-Tuning Output for Patient 1. Table 15Federated Learning with Differential Privacy Output.Model weights updateInstitution 1Institution 2Institution 3Global model (aggregated)Weight for glucose0.450.500.480.47Weight for BMI0.300.280.350.31Weight for Age0.150.180.170.17 Federated Learning with Differential Privacy Output. Table 16DropEdge Output for Patient 1.FeaturePre-DropEdgePost-DropEdge (regularized)Age2929BMI3230.5Glucose level125120Activity levelLowLowNumber of pregnancies22 DropEdge Output for Patient 1. Table 17Final output for patient 1.FeaturePredicted valueExplanation (attention score)GDM riskHighGlucose (0.45), BMI (0.25)Glucose level125Most influential factorBMI30.5Contributes significantlyAge29Minor contributionActivity levelLowMinor contribution Final output for patient 1. Validation using iterative practical use case scenario analysis In this study, A patient dataset was used in this study to model the risk of gestational diabetes mellitus. The features in this dataset include clinical measurements like glucose levels and BMI, demographic information such as age and ethnicity, lifestyle factors such as physical activity and diet, and temporal data in terms of glucose measurements across different trimesters. The proposed graph-based model integrates various machine learning techniques for predicting GDM risk, such as GraphSAGE for scalable node feature aggregation, R-GCN for relational modeling, TGAT for temporal dynamics, and GAT for interpretability. Additionally, BioBERT pretraining with GNN fine-tuning is used to improve model performance on medical data, while federated learning with differential privacy ensures patient data privacy. For the practical use case analysis, the validation samples are taken from Pima Indians Diabetes Dataset that is used in most predictive modeling tasks regarding diabetes. The dataset contains 768 records of female patients, with 8 features including age, BMI, glucose concentration, insulin levels, blood pressure, and the presence or absence of gestational diabetes (binary classification target). The dataset contains real-world medical information, hence suitable for validating the performance of the proposed model in estimating the risk of gestational diabetes. The validation set constitutes 30% of all samples, ensuring that the model is tested on unseen data, which is critical to assess its generalization capability. All these features are standardized by making continuous variables standardized and categorical variables are encoded. So, the model would efficiently process the input data and produce good predictive results in the validation set for AUC and interpretability, ensuring that the test case in the healthcare setup is good and robust. In the following, the outputs of each process in the graph-based model are presented using a sample patient’s data samples. The model processes patient data through these steps, beginning from feature aggregation (GraphSAGE) to the last prediction and risk assessment; the results are presented in tabular form for clarity. The first stage aggregates node features from the local neighbors of a patient in the graph, which represent other people with similar characteristics, like glucose levels, BMI, age, and lifestyle. The goal is to encapsulate both the features of individual nodes and their interactions with other nodes. GraphSAGE uses learnable aggregation functions such as means or LSTM-based approach to make a feature embedding that sums up information from its neighboring nodes. Once aggregated through GraphSAGE, the feature embedding of Patient 1 would represent not only his own special characteristics but also those who are similar in those graphs through the influences of similar neighbors.. This embedding serves as the input for subsequent stages. Patient 1 has the following features:Age: 28BMI: 31Glucose Level: 120 mg/dLActivity Level: LowNumber of Pregnancies: 2Ethnicity: Hispanic Age: 28 Age: 28 BMI: 31 BMI: 31 Glucose Level: 120 mg/dL Glucose Level: 120 mg/dL Activity Level: Low Activity Level: Low Number of Pregnancies: 2 Number of Pregnancies: 2 Ethnicity: Hispanic Ethnicity: Hispanic The aggregated features, that are age 29, BMI 30.5, glucose 120, low activity level, and 2 pregnancies, give a more comprehensive representation, combining both the local and global information sets. These feature embeddings will be used in the subsequent relational modeling phase. R-GCN is used to process the aggregated features from GraphSAGE with different relationships between types of risk factors in mind. This stage models the relation, which include causal links such as links between BMI and glucose level, the temporal dependencies like the increase in glucose level, among others, correlations like that of lifestyle, ethnicity, etc., and diabetes risk sets. R-GCN, therefore, computes a novel feature embedding through aggregating information of neighboring nodes using their relations. The process of R-GCN is characterized by adjustments of feature values concerning the relational information sets. For example, a higher BMI is well correlated with a higher glucose level, which is evident with the glucose level being upped at 125 mg/dL. The phases captured by the phase sets include increased glucose in pregnancy over time, such as increases in glucose levels during the second trimester. The output from R-GCN enhances the feature representation by incorporating complex relational dependencies, which will be further processed in the temporal dynamics phases. TGAT captures temporal relationships, making it particularly suitable for this task where patient data evolves over time, such as glucose measurements taken at various pregnancy trimesters. TGAT utilizes temporal attention mechanisms in predicting the importance of the various timestamp steps for determining future GDM risk. The model applies attention in modifying the importance of past glucose measurements to predict future risk appropriately. TGAT employs temporal attention to emphasize the second trimester glucose level (120 mg/dL), which affects the prediction for third trimesters. The model predicts an increase to 125 mg/dL in the third trimester, which depicts how the temporal dependencies are captured in the process. The temporal attention scores show that glucose levels from Trimester 2 are more heavily weighted at 0.85 compared to the data from Trimester 1, which is at 0.75, showing the significance of recent measurements in predicting future risks. GAT is applied to enhance the interpretability of the model by focusing on the importance of specific features and relationships within the graph. At this step, GAT gives scores of attention to each neighbor of nodes. This will make the model highlight the most critical risk factors for the predictions of GDM. The output scores of attention let the clinicians know which factor has a great impact on determining the risk of GDM. It further clarifies that the highest predicting feature for GDM risks in Patient 1 is glucose level at 0.45 followed by BMI at 0.25 and age at 0.15 in process. Such scores elucidate how the model made that prediction so clinicians better understand why these features- for example, glucose level-more weightily feature in this patient’s risk profile. Transparency in the process is a requirement for clinical adoption; it lends trust and enables action to be taken over the process. The domain-specific initialization of patient records is obtained using BioBERT pretraining on medical text data followed by fine-tuning with the GNN model, which enables the model to adapt to handling medical terminologies to further improve its performance over healthcare-related tasks. Fine-tuning the model enables it to enhance its capability of integrating patient-specific features with knowledge learned from a large corpus of medical literature sets. Adjust feature value to medical literature and use patient-specific data samples through this process to fine-tune the process. It depicts in more graphical format, for example, at the pre-training stage when glucose level stood at 120 mg/dL changed to 125 mg/dL at fine-tune. This will enhance the performance of the model for GDM prediction by making sure relevant domain knowledge sets are captured. Federated learning that is based on differential privacy will then ensure the confidentiality of the patient’s data as used to train the model. Here, local versions of the model are each trained by one institution while the global model is aggregated without releasing raw patient data samples. The updates have noise added to them, further protecting patient privacy levels with differential privacy. In federated learning, it is possible to accumulate updates from different institutions without necessarily releasing the raw data samples. Model weights aggregated across institutions therefore indicate the average weight based on consensus from all institutions without jeopardizing patient privacy during process. It ensures that there is generalization over the different data sources while adhering to the constraint of privacy. DropEdge regularization avoids overfitting since during training, the edges of the graph are randomly dropped and learns from all the diverse connections. This enhances further generalization of the model. Therefore, this will be strengthening the model against noise and differences in data samples. The model is changing the values of features after applying the DropEdge. Since it drops specific connections in the graph, the feature values get adjusted minimally. For example, regularization of BMI from 32 to 30.5 tells it that it generalizes better in not relying on very specific connections. This results in regularization preventing the overfitting of the model and making it possible for it to predict the outcomes for unseen data samples as well. After the graph-based model has completed all steps including feature aggregation, relational modeling, temporal prediction, attention scoring, pretraining, federated learning, and regularization, the final output of the model is the full risk prediction for GDM together with an explanation of which factors are most influential over the result sets. The final output indicates that there is a high risk for GDM in Patient 1, with glucose levels and BMI being the most influencing factors. Attention scores reflect that glucose is the dominating factor at 0.45 followed by BMI at 0.25, making the monitoring of glucose levels quite important. The interpretability of the model enables clinicians to act on the most critical risk factors to manage GDM effectively for different patients in clinical scenarios. Calibration, fairness and temporal RMSE reporting Calibration was reported using metric assessments as opposed to simply descriptive language. Calibration-curves were developed on the held-out test-set by dividing predicted GDM risk-probabilities into deciles and comparing the average predicted risk-probability with the observed-event-rate per decile. Additionally, expected-calibration-error, maximum-calibration-error and Brier-score were also reported with 95% bootstrap-confidence Intervals. Calibration-curves and these statistics were employed to ascertain whether the model produced clinically-meaningful probabilities (as opposed to high-discrimination Values) in process. Analysis of fairness was carried out across predefined and clinically-meaningful subgroups. The subgroups included maternal-age-groups, BMI-categories, parity-status and ethnicity-or-site-defined-demographic groups where available. For each subgroup, sample-size, GDM-positive-count, AUC, recall, specificity/false-negative-rate/calibration-error and confidence Intervals were reported. Subgroups with low cell-counts were not utilized for making strong fairness assertions; they were denoted as under-powered and described descriptively. Reporting in this manner prevented broad equity assertions when sample distributions were too sparse. Temporal prediction-RMSE was calculated only for true longitudinal-clinical-targets measured over time in pregnancy – i.e., glucose related-metrics or risk-trajectories derived from real clinical-data observed across pregnancy Visits. Demonstrated tables were not included in RMSE calculations. The temporal-model was trained on earlier Visit Information and evaluated against later-observed targets in the test-subset; thus, RMSE reflected prospective-temporal-prediction rather than reconstruction of synthetic or demonstrative examples. Transparent comparative evaluation and benchmark naming Method [15] = Elastic Net Regularized Logistic Regression; Method [8] = Random-Forest-Classifier; Method [18] = Gradient-Boosting-Machine Classifier process. Conclusion & future scopes This paper is one example that showcases an application of a holistic graph-based model that predicts the risk of GDM using advanced GraphSAGE, Relational Graph Convolutional Networks (R-GCN), Temporal Graph Attention Networks (TGAT), and other techniques including BioBERT pretraining, federated learning, and DropEdge regularization. The proposed model shows significant improvements over baseline machine learning techniques in all key performance metrics, that is, classification accuracy of 88.2%, AUC of 0.90, RMSE for the temporal prediction of 0.30, F1 score of 0.86, precision of 0.83, recall of 0.85, and attention agreement with domain experts of 82%. These results emphasize that this model is not only capable of accurate GDM outcome prediction but also can capture the complex, multi-dimensional relationships between clinical, demographic, and temporal factors that drive the risk of developing GDM. Baseline methods such as Method [5], Method [8], and Method [18] achieve significantly lower performance on these metrics. Method [5] reached classification accuracy of 76.5% with an AUC value of 0.75 and an RMSE of 0.45, while Method [8] performed moderately better at an accuracy of 80.1% and AUC equal to 0.80. Clinical relevance-the performance improvements of the proposed model especially at the level of predictive accuracy and temporal forecasting would hold clinical significance in early detection and management towards personalization in the handling of GDM to ultimately prevent complications such as macrosomia, preterm birth, and long-term metabolic issues in the processing. The proposed system works on heterogeneous data, wherein there are temporal dynamics; multi-relational factors arise, and the application considers privacy-preserving federated learning, making it dissimilar from conventional systems. Moreover, the explainability of the model, as shown by the high attention agreement with domain experts at 82%, increases the trustworthiness and clinical applicability of the model. Insights into the relative importance of various risk factors can be useful for targeted interventions by informed clinicians based on patient profiles. Although it consumes more computational resources and needs more training timestamp (5.5 h), such a trade-off in favor of superior performance is quite justifiable, especially considering the fact that, once trained, it promises to provide accurate predictions without demanding the necessity of frequent retraining. And that is why it will be pretty attractive in real-time clinical settings where decision support systems should, at the same time, both be accurate and interpretable in process. We found that the graph-based framework showed better discrimination/temporal modeling/interpretability within both the internally-tested pregnancy cohort/public feature-limited benchmark environment. Our results suggest that relational/temporal modeling can help earlier GDM risk-stratification; nonetheless, clinical generalizability of the framework needs further confirmation through independent/prospectively-collected multi-center GDM cohorts before routine applications. There are many areas for furthering the model, and enhancing its performance, so it can be practically applicable. These include introducing new sources of data, such as markers that would provide a wealthier representation of the patients’ risk profiles. Adding even larger and more general datasets, perhaps coming from several different geographical locations and ethnic groups, would definitely enhance generalizability to an even greater extent. Future work in terms of temporal analysis may include further refinement of the TGAT model to better handle more challenging cases of irregular time-series data, such as varying glucose measurement frequencies or missing values at pregnancy. Also, the federated learning protocol can be extended in order to provide better privacy guarantees for training on decentralized data so as to pave the way for further wide-scale deployment in real-world healthcare settings. The current model does not provide for updating patient embeddings in real-time with new data arrival; inclusion of such adaptive learning techniques will enhance predictive accuracy in dynamic clinical settings. Lastly, since personalized medicine makes more and more specific requests, involving multi-task learning frameworks will add more relevant knowledge about the forecasting of complications development for GDM; therefore, the system has wider implications in health care for multiple directions of decision Making. Generally speaking, the new graph-based approach is really novel at this point, promising not only superior performance but satisfactory interpretability at different temporal instance sets. This model is the potential harbinger of change in the care for GDM and can provide personalized risk management and, as such, improve the earlier detection and management process. Therefore, in the near future, scalability, adaptability, and integration into clinical workflows could focus on work to improve outcomes for health in mothers and newborns for different clinical scenarios. Expert agreement and explainability evaluation Expert agreement was assessed by way of a formal review process as opposed to merely visually reviewing illustrative cases. Each selected test-case received a ranked-list of the top 10 contributing risk-factors (based upon attention-weights and post-hoc attribution-scores) produced by the model. Three domain-experts with expertise in obstetrics, endocrinology, and maternal-metabolic risk-assessment independently reviewed these lists. All reviewers were blinded to the identity of the model producing the rankings as well as the predicted-probability categories associated with those rankings. Consensus was defined as having at least two out of three reviewers agree that a ranked-factor was clinically-plausible for the corresponding-patient-profile. Prior to aggregating reviewer-consensus across patients in the test-set, inter-rater-agreement was measured utilizing Fleiss’ kappa statistic. Finally, expert-agreement-percentage was calculated based upon all held-out test-subset-patients rather than illustrative examples. Explainability evaluation was tied directly to predictions at the individual-patient level produced by the model from the test-split. For each patient, the ranked-factors were compared against clinical variables available prior to the label assignment—including glucose related-measures, BMI, maternal-age, obstetric-history, blood-pressure Indicators, family-history and longitudinal-trimester-trends wherever possible. As such, attention-agreement represented clinically Interpretable model-behavior on evaluated predictions as opposed to manually-prepared example-demonstrations. Limitations This study has several limitations which restrict how one might interpret generalizability. First, since the UCI Pima Indians Diabetes dataset is not a gestational-diabetes cohort it was used exclusively as a limited-feature public diabetes benchmark as opposed to an externally Validating GDM dataset. Second, since the complete multimodal primary cohort referred to in methods cannot be openly distributed because it includes institutionally-controlled clinical data; longitudinally-recorded pregnancy data; and text representation-derived features—we cannot make it publicly accessible. Third, although internal-patient-level splits minimize leakage and facilitate controlled comparisons among others—our current evidence does not demonstrate sufficient validation to assume generalizability to previously unseen hospital/population/diagnostic workflow/pregnancy care settings. Thusly, we believe that future validations should employ prospectively-collected independent GDM-cohorts with harmonized definitions for outcomes/subgroup-reporting/site-specific and pre-registration of model-evaluation protocols.