Early prediction of gestational diabetes mellitus using deep learning applied to routinely collected clinical data

Background Prediction of gestational diabetes mellitus (GDM) in early pregnancy enables timely, targeted interventions to reduce risks for both the mother and fetus. We aimed to improve early GDM prediction and clinical risk stratification using a novel deep learning (DL) approach that transforms tabular clinical data into images using a genetic algorithm. Methods A dataset of 1,611 pregnancies was used. The dataset contained routinely collected early-pregnancy data, including clinical measurements, blood tests, and medical history. A novel method to transform tabular data into images was applied to enable the use of deep learning models for GDM prediction. A genetic algorithm optimized the spatial arrangement of variables during image generation by determining their spatial positions. Interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM) to identify variables the model preferentially attended to for positive and negative predictions. Results Multiple model configurations were evaluated by varying the spatial arrangement of variables within the transformation. Prioritizing high sensitivity, one model achieved an accuracy of 0.86, sensitivity of 0.82, specificity of 0.87, and an AUC-ROC of 0.83. Another model achieved an accuracy of 0.79, sensitivity of 0.92, specificity of 0.77, and an AUC-ROC of 0.85. Compared with previously published machine learning models using the same dataset, our approach demonstrated improved performance. Our best models increased specificity by an average of 13.93 percentage points compared with traditional ML models at matched sensitivity levels, thereby reducing false-positive classifications that may lead to additional clinical evaluation. Grad-CAM analysis revealed that attention patterns of the model aligned with established clinical risk factors for GDM, with first trimester fasting glycemia and parity preferentially attended for positive predictions and maternal weight for negative predictions. Conclusions This study demonstrates that transforming tabular clinical data into image representations, optimized using a genetic algorithm, improves early GDM prediction. Representing tabular variables as images enables more effective application of DL models, which have demonstrated enhanced pattern recognition in image-based tasks compared with traditional machine learning methods. Performance improvements over state-of-the-art approaches were achieved. Accurate DL models can be developed using a limited number of input variables, supporting early identification of women at increased risk of GDM and facilitating timely clinical management.