Gestational diabetes mellitus (GDM) is a disorder of glucose metabolism that arises during pregnancy and increases maternal and neonatal complications as well as future type 2 diabetes risk. This study presents a hybrid Mamba-Transformer architecture that fuses clinical and genetic features for early GDM prediction. Key components include a correlation-driven weighted fusion method for integrating clinical and genetic data, a sliding-window sample reconstruction to produce context-aware grouped instances, and a modular encoder–decoder architecture where a Mamba module extracts dependencies and a Transformer module performs deep semantic modeling. The model was trained on the public DMRPD dataset from the Alibaba Cloud Tianchi platform. On the test set the method achieved an AUC 0.825, with sensitivity 0.827 and specificity 0.729 at the reported optimal threshold (0.526). The authors conclude the approach shows competitive predictive performance but requires further validation on larger independent cohorts to assess generalizability and clinical applicability.
Gestational diabetes mellitus (GDM) is defined as a glucose tolerance abnormality first identified during pregnancy. It can lead to complications such as fetal malformation, macrosomia, and preterm birth, and is associated with an increased risk of later type 2 diabetes. Early identification of women at risk for GDM is clinically valuable for enabling personalized interventions that may reduce adverse outcomes.
Prior research has linked multiple clinical indices and environmental factors to GDM risk, including obesity, gastrointestinal microbiome changes, and various biochemical markers. At the same time, genomic studies have identified specific polymorphisms associated with GDM susceptibility in defined populations. Examples cited include loci such as rs62069863 (TRPV3), rs2232016 (PRMT6), and rs10460009 (LPIN2) reported in Chinese Han cohorts. However, a single gene often has limited predictive power; integrating routine clinical variables with multi-gene information remains an active research challenge.
Classic machine learning approaches (SVM, Random Forest, AdaBoost, XGBoost, GBDT, etc.) and emerging deep learning methods have been applied to GDM prediction. These approaches can learn complex patterns but face issues such as high-dimensional feature selection, limited generalizability, and opaque feature selection in deep models. Transformer architectures have been applied in bioinformatics for sequence modeling, while the Mamba model has been proposed to handle long sequences with improved computational efficiency. The authors position their hybrid approach as a way to leverage complementary strengths of Mamba and Transformer while addressing heterogeneous clinical and genetic interactions.
The manuscript summarizes prior approaches in three domains: epidemiological and statistical models that examine population and environmental risk factors; machine learning classifiers (SVM, Random Forest, etc.) applied to GDM risk stratification; and deep learning sequence models including Recurrent Neural Networks, Transformer variants, and the Mamba model for long-range dependency modeling. The authors note limitations across these areas: traditional methods struggle with high-dimensional, non-linear interactions; deep learning models may lack interpretability and clinical logic; and standard Transformers can be computationally expensive on long sequences, potentially missing long-distance dependencies.
Overall approach
The proposed pipeline combines feature fusion, data augmentation via sequence reconstruction, and a hybrid deep architecture. The method integrates clinical and genetic inputs into a joint representation, augments training instances using a sliding-window reconstruction, and feeds these into a modular encoder–decoder network composed of Mamba and Transformer modules.
Correlation-driven weighted fusion
A correlation-driven weighted fusion method was developed to merge clinical and genetic features. This fusion aims to enhance representation by weighting contributions according to their inter-feature correlations, thereby emphasizing interactions between genetic susceptibility and clinical factors. The paper frames this as a solution to models that often omit genetic contributions.
Sliding-window sample reconstruction
To convert isolated features into context-aware inputs, a sliding-window technique reconstructs sample sequences from preprocessed data. This produces augmented instances that reflect grouped, population-level heterogeneity as well as individualized risk context, enabling sequence-based modeling.
Hybrid Mamba-Transformer architecture
The architecture uses a modular encoder–decoder design. The Mamba module functions as an efficient feature extractor optimized for long-range dependency capture and computational efficiency. The Transformer module performs deeper semantic modeling and sequence abstraction. The design is described as flexible and scalable, intended to adapt to varied application scenarios.
Dataset and implementation
Model development and evaluation were performed using the publicly available DMRPD competition dataset hosted on the Alibaba Cloud Tianchi platform. The manuscript reports training and testing on that dataset; further dataset-specific preprocessing and hyperparameter details are provided in the full article and supporting information.
Experimental evaluation compared the proposed hybrid model with representative baseline approaches. Reported test-set performance for the proposed method included an AUC of 0.825. At the reported optimal decision threshold (0.526), sensitivity was 0.827 and specificity was 0.729. These metrics are presented as evidence of competitive discriminative performance compared with other methods evaluated in the study.
The authors highlight that the hybrid design and fusion strategy contributed to improved representation of heterogeneous clinical and genetic information and to more effective modeling of complex dependencies.
The study argues that fusing clinical and genetic features with a correlation-driven weighted approach and applying sequence reconstruction helps capture interactions relevant to GDM risk. The hybrid architecture is positioned as a balance between efficient long-range dependency extraction (Mamba) and deep sequence abstraction (Transformer). Advantages noted include improved feature representation and the potential to support more targeted screening strategies.
Limitations and future work
The authors acknowledge limitations: the analysis was conducted on a single public dataset (DMRPD) and the model requires further validation on larger, independent cohorts to confirm generalizability and clinical applicability. The study also recognizes ongoing challenges related to interpretability of deep models in clinical settings.
This work presents a hybrid Mamba-Transformer framework that fuses clinical and genetic information via correlation-driven weighted fusion and uses sliding-window reconstruction to create sequence-like inputs. Evaluated on the DMRPD dataset, the method achieved an AUC 0.825 with competitive sensitivity and specificity at the reported threshold. The authors suggest the approach may inform earlier, more targeted GDM screening strategies but emphasize the need for broader validation before clinical deployment.