---
title: "Hybrid Mamba-Transformer Model Integrating Clinical and Genetic Features for GDM Prediction"
id: "plos-one-15-a-hybrid-mamba-transformer-architecture-fusing-clinical-and-genetic-features"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-15-a-hybrid-mamba-transformer-architecture-fusing-clinical-and-genetic-features"
content_type: "clinical_feed_article"
specialty: "Endocrinology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0355729"
published_at: "2026-08-12T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Hybrid Mamba-Transformer Model Integrating Clinical and Genetic Features for GDM Prediction
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-15-a-hybrid-mamba-transformer-architecture-fusing-clinical-and-genetic-features
- **Specialty:** [Endocrinology](https://medichelpline.com/clinical-feed/endocrinology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0355729)
- **Published At:** 2026-08-12T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- Gestational diabetes mellitus (GDM) is a pregnancy-related glucose tolerance disorder that elevates maternal and perinatal risk and increases later type 2 diabetes risk; early prediction can guide targeted interventions. - The study proposes a hybrid **Mamba-Transformer** architecture that fuses routine clinical and multi-gene features to improve early GDM risk prediction. - A correlation-driven weighted fusion method is introduced to combine clinical and genetic data, aiming to highlight interactive relationships between genetic susceptibility and clinical factors. - A sliding-window strategy reconstructs preprocessed samples into sequence-like, context-aware grouped instances to capture population-level heterogeneity and individual risk. - The architecture uses a modular encoder–decoder design: the **Mamba** module extracts efficient long-range dependencies and feature-level patterns, while the **Transformer** module performs deep semantic modeling and sequence abstraction. - The model was trained and evaluated on the public competition dataset (DMRPD) from the Alibaba Cloud Tianchi platform. - Reported performance on the test set: **AUC 0.825**, sensitivity 0.827 and specificity 0.729 at the optimal threshold (0.526), indicating competitive discriminative ability versus representative models. - Authors note the method may support earlier, more targeted screening strategies but emphasize the need for further validation on larger, independent cohorts to confirm generalizability and clinical applicability. - Contributions highlighted: (1) correlation-driven fusion of genetic and clinical features, (2) hybrid Mamba-Transformer architecture for heterogeneous data, and (3) a flexible, scalable model design adaptable to different scenarios.
## Clinical Analysis & Structured Key Points
A hybrid mamba-transformer architecture fusing clinical and genetic features for gestational diabetes mellitus prediction | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Gestational diabetes mellitus (GDM) is a common disorder of glucose metabolism during pregnancy. Early GDM prediction is crucial for reducing adverse maternal and neonatal outcomes. This paper proposes a hybrid Mamba-Transformer architecture that aggregates clinical and genetic features for GDM prediction. First, a correlation-driven weighted fusion method for clinical and genetic features is introduced. The integrated representation not only enhances feature representation but also highlights the interactive relationship between genetic susceptibility and clinical factors. Second, a sliding window approach is applied to reconstruct the sample sequences from the preprocessed data, generating augmented instances as model input. This transforms isolated individual features into context-aware group features, enabling the effective capture of both population-level heterogeneity and individual risk. Finally, the hybrid Mamba-Transformer architecture is constructed and trained on the publicly available competition dataset (DMRPD) from the Alibaba Cloud Tianchi platform. The model employs a modular and extensible encoder-decoder structure, where the Mamba module serves as an efficient feature extractor for dependencies, while the Transformer module performs deep semantic modeling and sequence abstraction. Experimental results indicate that the proposed method achieves competitive performance compared with other representative models. Specifically, the model attained an AUC of 0.825 on the test set, with sensitivity and specificity at the optimal threshold (0.526) of 0.827 and 0.729, respectively, suggesting reliable discriminative performance on the test set. These findings suggest that the proposed method may provide a useful approach for early GDM risk prediction and could potentially support more targeted screening strategies. However, further validation using larger and independent cohorts is required to confirm its generalizability and clinical applicability. Citation: Huang J, Shi W, Lin L, Wei Z, Li Q (2026) A hybrid mamba-transformer architecture fusing clinical and genetic features for gestational diabetes mellitus prediction. PLoS One 21(8): e0355729. https://doi.org/10.1371/journal.pone.0355729 Editor: Sefki Kolozali, University of Essex Faculty of Science and Engineering, UNITED KINGDOM OF GREAT BRITAIN AND NORTHERN IRELAND Received: August 30, 2025; Accepted: July 24, 2026; Published: August 12, 2026 Copyright: © 2026 Huang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: All relevant data are within the manuscript and its Supporting Information files. Funding: This work was supported by the National Natural Science Foundation of China (No.62476005) and the Open Fund of Fujian Provincial Key Laboratory of Medical Big Data Engineering (No. KLKF202302). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. 1 Introduction Gestational diabetes mellitus (GDM) is a glucose metabolism disorder of varying degrees that occurs during pregnancy. It is a glucose tolerance disorder that first appears or is diagnosed during pregnancy. The main feature of GDM is that pregnant women have glucose metabolism disorders during pregnancy, but their blood sugar levels have not yet reached the diagnostic criteria for overt diabetes [ 1 , 2 ]. GDM not only affects the health of pregnant women but may also lead to a series of complications, such as fetal malformation, macrosomia, and premature birth, increasing perinatal risks for mothers and infants. In addition, the probability of GDM patients developing type 2 diabetes mellitus (T2DM) after delivery is significantly increased [ 3 – 5 ]. Therefore, early GDM prediction has significant clinical value. It helps to develop personalized intervention measures to reduce the occurrence of pregnancy complications. In the GDM prediction, many studies at home and abroad have focused on the relationship between specific indices and GDM. For example, some studies have shown that factors such as the atherogenic index of plasma, gastrointestinal microbiome, and obesity are associated with GDM [ 6 , 7 ]. The epidemiological study of GDM mainly relies on epidemiological models and experimental data. It evaluates the GDM risk by analyzing factors such as population characteristics, geographical differences, and environmental pollution [ 8 , 9 ]. However, these methods require more prospective studies and high-quality in vivo and in vitro experiments to investigate their specific effects and mechanisms, and it is challenging to consider complex genetic interactions. In recent years, with the advancement of genomics and bioinformatics technologies, researchers have begun to focus on the role of genetic factors in the occurrence of GDM. Gene polymorphism studies have shown that certain specific gene mutations or variations are closely related to the occurrence of GDM. For example, recent studies have found that rs62069863 in the TRPV3 gene, rs2232016 in the PRMT6 gene, and rs10460009 in the LPIN2 gene are new genetic susceptibility loci for GDM in the Chinese Han population [ 10 ]. However, the impact of a single gene is limited. How to effectively integrate routine physical examination information and multi-gene information for intelligent GDM prediction is still a hot topic in current research. GDM can cause adverse consequences for mothers and their newborns. Pregnant women in some low and middle-income regions or countries often cannot receive early clinical intervention due to limited medical resources. Machine learning methods have good application prospects in the diagnosis of GDM. Algorithms such as SVM, Random Forest, AdaBoost, Decision Tree, XGBoost, and GBDT have been used to build GDM prediction models [ 11 , 12 ]. These methods can automatically learn complex patterns in data, but they still face problems such as high-dimensional feature selection and low generalization ability. With the development of deep learning, researchers have begun to use deep neural networks for GDM prediction. For example, patients at high risk of GDM are identified based on recurrent neural networks and Bayesian optimization [ 13 ]. It provides auxiliary decision support for clinicians and reduces unnecessary oral glucose tolerance tests. In addition, the Transformer architecture has also been applied in the fields of bioinformatics and medical prediction due to its outstanding performance in natural language processing tasks [ 14 , 15 ]. As a new sequence modeling method, Mamba has attracted widespread attention from researchers in the field of bioinformatics due to its efficient long sequence processing capabilities and computational efficiency [ 16 ]. The Mamba model can effectively model long-range dependencies. It has advantages over the traditional Transformer model in terms of computational complexity and training efficiency. Although these methods have made some research progress in their focused areas, there are still some limitations. Traditional statistical methods and machine learning methods have limited predictive ability when dealing with high-dimensional, non-linear, and complex clinical and genetic interactions. Deep learning is often considered a black box model. Its feature selection is opaque, or the algorithm lacks clinical logic, making it difficult for clinicians to understand its prediction basis. This has affected its trust and adoption in clinical practice. The traditional Transformer structure has a high computational complexity when processing long sequence data. Its attention mechanism may not be able to effectively capture long-distance dependencies, resulting in information loss or increased noise. Therefore, constructing efficient predictive models capable of capturing the intricate interactions between clinical and genetic features holds significant theoretical importance and practical value. In response to the above challenges, this paper proposes a hybrid Mamba-Transformer architecture fusing clinical and genetic features for GDM prediction. The proposed architecture is designed to capture the complex interactions between heterogeneous features while improving the modeling of long-range dependencies. By integrating genetic and clinical features, this method aims to capture complex feature interactions and improve GDM risk prediction. The effectiveness of this method is evaluated through experiments on a real-world GDM dataset. The main innovations and contributions of this study are summarized as follows: (1) We introduce a correlation-driven weighted fusion method for genetic and clinical features. This method enhances feature representation and emphasizes the interactive relationship between genetic susceptibility and clinical factors. In addition, it overcomes the limitation of existing GDM prediction models that often neglect genetic contributions. (2) We construct a hybrid Mamba-Transformer architecture for GDM prediction. By integrating the complementary strengths of Mamba and Transformer, the proposed framework enhances the modeling of complex dependencies and improves the representation of heterogeneous clinical and genetic information. (3) In architectural design, we develop a flexible and scalable model structure that can adapt to different application scenarios and varying task complexities, facilitating transferability and applicability across diverse prediction scenarios. The rest of this paper is organized as follows: Section 2 provides a brief review of related works on GDM prediction. Section 3 presents the overall architecture and detailed design of the proposed method. Section 4 describes the experimental procedures and results analysis. Finally, a discussion and conclusion are made in Sections 5 and 6. 2 Related works In recent years, various approaches have been proposed to analyze and predict the risk of GDM, including statistical and epidemiological analyses, machine learning techniques, and deep learning models. In this section, we review recent advances related to GDM risk analysis and prediction. 2.1 Statistical and epidemiological approaches Statistical and epidemiological approaches have been widely used to identify GDM-related risk factors, investigate associations among clinical and biological variables, and develop risk prediction models based on population-level medical data. These approaches have provided valuable insights into the associations between clinical characteristics, biological markers, environmental exposures, and GDM risk. For example, Man et al. [ 17 ] developed a diabetes risk prediction model using Cox proportional hazards regression based on baseline clinical variables from women with prediabetes and prior GDM. In addition to clinical indicators, observational studies have explored potential biological mechanisms underlying GDM. Hong et al. [ 18 ] investigated plasma amino acid profiles, while Faurø et al. [ 19 ] examined lipid-related biomarkers associated with GDM risk. Recent epidemiological studies have further expanded the understanding of GDM by incorporating environmental and lifestyle-related factors. He et al. [ 20 ] analyzed the association between residential green exposure and glycemic levels, while Mao et al. [ 21 ] and Li et al. [ 22 ] investigated the effects of chemical exposure and air pollution on GDM incidence. In predictive modeling, Manna et al. [ 23 ] integrated clinical biomarkers with metabolomics information to construct a multivariate prediction model, while Lyu et al. [ 24 ] demonstrated that simplified models based on routinely available clinical features could achieve practical risk prediction. Furthermore, psychosocial factors have also been considered in GDM assessment, as demonstrated by Kumar et al. [ 25 ]. Although statistical and epidemiological approaches remain valuable for identifying risk factors, their ability to capture complex nonlinear interactions among heterogeneous variables is limited. 2.2 Enhanced machine learning approaches for risk prediction To overcome the limitations of conventional statistical models, enhanced machine learning approaches have increasingly been applied to GDM and diabetes-related prediction tasks. These approaches incorporate ensemble strategies, feature engineering, optimization techniques, and interpretability methods, enabling them to model nonlinear relationships and complex feature interactions. Among these methods, ensemble learning algorithms, particularly gradient boosting models, have shown promising predictive performance. Kumar et al. [ 26 ] combined CatBoost with Shapley feature attribution to develop an interpretable GDM prediction model, while Belsti et al. [ 27 ] demonstrated the effectiveness of CatBoost compared with traditional statistical methods using routine prenatal data. Beyond model selection, recent studies have highlighted the importance of feature engineering, optimization strategies, and data quality enhancement. Olisah et al. [ 28 ] explored feature selection and missing value processing strategies, while Karuppasamy et al. [ 29 ] proposed a hybrid optimization-based stacking framework. Zhao et al. [ 30 ] and Wang et al. [ 31 ] further investigated ensemble learning and feature selection techniques to improve prediction performance. In addition, Xu et al. [ 32 ] focused on improving data reliability through label noise filtering, and Ramsingh et al. [ 33 ] explored a scalable Hadoop-based framework for analyzing large-scale social media data related to diabetes. Despite these advances, machine learning models generally depend on manually designed features and may have limited capability in automatically learning high-level representations from heterogeneous clinical and genetic information. 2.3 Deep learning approaches in medical prediction Deep learning approaches have recently attracted increasing attention in medical prediction because of their ability to automatically extract complex representations from high-dimensional data. Zheng et al. [ 34 ] evaluated multiple deep learning architectures for GDM prediction and reported the effectiveness of attention-based networks. Sumathi et al. [ 35 ] developed a deep-stacked autoencoder framework for automated GDM diagnosis. Other studies have focused on improving robustness and interpretability. Shaheen et al. [ 36 ] addressed class imbalance through sampling strategies, while Dhal et al. [ 37 ] optimized deep neural networks using hybrid optimization techniques. Singh et al. [ 38 ] and Cheng et al. [ 39 ] further incorporated explainability and semantic enhancement strategies to improve model transparency and representation capability. Recently, Transformer-based architectures have emerged as powerful approaches for medical prediction due to their ability to capture long-range dependencies through self-attention mechanisms. Oulhadj et al. [ 40 ] applied a vision transformer-based framework for diabetic retinopathy prediction and achieved competitive performance across multiple datasets. However, the quadratic computational complexity of self-attention limits the scalability of Transformers for long sequences or large-scale applications. To address this limitation, Mamba, a selective state space model, has recently been introduced as an efficient alternative for sequence modeling. Nataliani et al. [ 41 ] incorporated Mamba into a mobile U-Net framework for diabetic foot ulcer segmentation, demonstrating its potential to achieve a balance between computational efficiency and predictive performance. Overall, existing studies have significantly advanced GDM prediction through statistical analysis, machine learning, and deep learning techniques. However, several challenges remain. Statistical approaches are limited in modeling nonlinear feature interactions, while conventional machine learning methods often rely on handcrafted feature engineering. Although deep learning methods provide stronger representation capabilities, existing studies have rarely investigated the effective integration of heterogeneous clinical and genetic information for GDM prediction. Moreover, the potential of combining Mamba with Transformers remains underexplored in this field. Therefore, this study proposes a hybrid Mamba-Transformer framework that integrates clinical indicators and genetic features for GDM risk prediction, aiming to exploit complementary information from heterogeneous data sources and improve predictive performance. 3 Methods This section details the proposed methods of GDM prediction, which will be elaborated from two aspects: feature fusion and sample reconstruction, and the construction of the MT-GDM model. 3.1 Feature fusion and sample reconstruction The data source used in this study is the Diabetes Mellitus Risk Prediction Dataset (DMRPD), a publicly available dataset released through the Alibaba Cloud Tianchi Precision Medicine Competition [ 42 ]. Clinical features, such as age, body mass index (BMI), blood pressure, blood glucose levels, and lipid indicators, provide important information regarding an individual’s physiological status and metabolic health. In contrast, genetic features, including gene variants and susceptibility loci associated with GDM, reflect inherited risk factors that may influence insulin secretion, insulin resistance, and glucose metabolism. Since clinical and genetic information provide complementary perspectives, integrating these heterogeneous data sources enables a more comprehensive characterization of GDM risk factors. The overall workflow of data preprocessing, feature fusion, and sample reconstruction is illustrated in Fig 1 . Download: PNG larger image TIFF original image Fig 1. The process of data preprocessing, feature fusion, and sample reconstruction. https://doi.org/10.1371/journal.pone.0355729.g001 First, data cleaning was performed on the raw dataset, including outlier processing, missing value imputation, data type conversion, redundant data removal, and normalization, thereby ensuring data consistency and integrity. After data cleaning, the dataset was divided into training, validation, and test subsets at a ratio of 6:2:2. To avoid information leakage, all subsequent feature selection and parameter determination procedures were performed exclusively on the training set. Let the cleaned feature matrix and corresponding labels be represented by Eq 1 and Eq 2, respectively, where N denotes the number of samples and d denotes the number of features. (1) (2) Feature selection algorithms were subsequently applied to the training set to evaluate the correlation between individual clinical and genetic features and GDM status. Based on the obtained relevance scores, the most informative features were retained for subsequent fusion. To further enhance feature representation and capture the interaction between genetic susceptibility and clinical factors, a correlation-driven weighted fusion method was adopted. Specifically, the top-k selected clinical and genetic features were integrated to construct a fused feature termed CSF. The construction of CSF is defined in Eq 3, where cf i and sf i denote the selected clinical and genetic features, respectively. The Scaling function normalizes clinical features into the range [0,1], while
## Related Clinical Research

- [Pharma Friday: FDA OKs Rasonque for Metastatic Pancreatic Cancer and Mounjaro CV Approval — Sept 4](https://medichelpline.com/clinical-feed/endocrine-news-0-pharma-friday-september-4-2026.md)
- [Semaglutide Associated With Lower Bone Fracture Risk in Type 2 Diabetes](https://medichelpline.com/clinical-feed/endocrine-news-0-video-semaglutide-linked-to-lower-bone-fracture-risk.md)
- [Left ventricular hypertrophy and brain atrophy in type 2 diabetes: findings from the D2 cohort](https://medichelpline.com/clinical-feed/medrxiv-0-left-ventricular-hypertrophy-brain-atrophy-and-cognitive-decline-in-type-2.md)
- [New-onset type 2 diabetes mellitus and risk of obesity-related cancers: UK Biobank matched cohort](https://medichelpline.com/clinical-feed/medrxiv-3-new-onset-type-2-diabetes-mellitus-and-obesity-related-cancer-risk-a-matched.md)
- [GLP-1 receptor agonists and PMOS: evidence, mechanisms, and who may benefit](https://medichelpline.com/clinical-feed/medical-news-today-0-can-glp-1s-really-help-treat-pmos-experts-answer-3-key-questions.md)

## Navigation
- [← Back to Endocrinology Feed](https://medichelpline.com/clinical-feed/endocrinology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.