Deep Learning Predicts Gestational Diabetes From Routine Clinical Data

Key Takeaways

• Among early-pregnancy records from a Santiago, Chile, hospital, an optimized gestational diabetes model showed a higher-specificity operating profile.
• At matched sensitivity, the optimized model had significantly higher specificity than conventional machine-learning models.

Cubillos, Monckeberg, Illanes, and Perez analyzed retrospective records from Hospital Parroquial de San Bernardo in Santiago, Chile, collected during 2019–2022 in a gestational diabetes prediction study published in Frontiers in Endocrinology. The cohort included pregnancies with and without subsequent GDM; preexisting diabetes was excluded. Visits occurred early in pregnancy, before the second-trimester diagnostic test. The reference diagnosis used fasting and later glucose measurements from a 75-g OGTT, without a one-hour reading. Standardized continuous features became images with column positions assigned by a genetic algorithm (GA). A pretrained ResNet-50 convolutional neural network (CNN) processed the images, and a multilayer perceptron produced the final prediction. Patient-level stratified partitions separated training, validation for configuration and threshold selection, and held-out testing. Conventional machine learning (ML) and deep learning (DL) without GA were same-dataset comparators. Investigators assessed bootstrap uncertainty, calibration, decision curves, and attention with gradient-weighted class activation mapping (Grad-CAM).

The higher-specificity configuration detected later GDM with sensitivity of 0.82 and specificity of 0.87. Its accuracy and discrimination were also reported, but these results do not establish a preferred clinical threshold. The alternative configuration increased sensitivity to 0.92 while specificity fell to 0.77. Greater detection therefore came with more false-positive classifications in this internal evaluation.

Validation-selected configurations and thresholds were applied unchanged to the held-out test set, which included pregnancies with and without GDM. Across matched observed sensitivity operating points, the GA models had a mean absolute specificity advantage of 13.93 percentage points over conventional ML. Comparisons with DL without GA also favored the optimized approach at most operating points, with paired bootstrap analyses supporting higher specificity at 7 of 9 and higher positive predictive value at 6 of 9. Brier scores were broadly similar, although the investigators’ recommended configuration departed from ideal calibration. Decision curves estimated net benefit rather than measured patient outcomes. Grad-CAM attended more to first-trimester fasting glycemia and parity for positive predictions and maternal weight for negative predictions. Models retrained after excluding pregnancies with high early fasting glucose retained discrimination.

Development and held-out testing took place at one Chilean hospital, with no independent external-population validation reported. Without a one-hour OGTT reading, the reference diagnosis only partly implemented International Association of Diabetes and Pregnancy Study Groups (IADPSG) criteria; early fasting glucose also overlapped a diagnostic threshold. The limited number of GDM cases in testing and imperfect calibration constrain interpretation. Grad-CAM attention neither identifies causal risk factors nor validates individual explanations, and decision-curve estimates do not demonstrate improved maternal or neonatal outcomes. The investigators reported early predictive performance and a matched-sensitivity specificity advantage in their internal comparisons.

Clinician Questions

Which glucose measurements defined gestational diabetes in the Santiago early-prediction cohort?

At Hospital Parroquial de San Bernardo in Santiago, Chile, gestational diabetes was defined by a second-trimester 75-g OGTT with fasting glucose of at least 92 mg/dL and/or two-hour glucose of at least 153 mg/dL. No one-hour measurement was obtained, so this only partly applied IADPSG criteria.

How did investigators assess whether early fasting glucose alone drove the GDM model’s predictions?

They excluded pregnancies with early fasting glucose at or above 92 mg/dL from the training, validation, and test partitions, then retrained the image model. They selected new configurations and thresholds on the filtered validation set and evaluated discrimination on the filtered held-out test set. The retrained models retained discrimination; this does not establish that fasting glycemia played no role.