First-trimester nuclear magnetic resonance-based metabolomic profiling increases the prediction of gestational diabetes mellitus

Background: Current strategies for predicting gestational diabetes mellitus demonstrate suboptimal performance.

Objective: To investigate whether nuclear magnetic resonance-based metabolomic profiling of maternal blood can be used for first-trimester prediction of gestational diabetes mellitus.

Study design: This was a prospective study of 20,000 women attending routine pregnancy care visits at 11 to 13 weeks' gestation. Metabolic profiles were assessed using a high-throughput nuclear magnetic resonance metabolomics platform. To inform translational applications, we focused on a panel of 34 clinically validated biomarkers for detailed analysis and risk modeling. All biomarkers were used to generate a multivariable logistic regression model to predict gestational diabetes mellitus. Data were split using a random seed into a 70% training set and a 30% validation set. Performance of the multivariable models was measured by receiver operating characteristic curve analysis and detection rates at fixed 10% and 20% false positive rates. Calibration for the combined risk model for all gestational diabetes mellitus was assessed visually through a figure showing the observed incidence against the predicted risk for gestational diabetes mellitus. A sensitivity analysis was conducted excluding the 64 women in our cohort who were diagnosed with gestational diabetes mellitus before 20 weeks' gestation.

Results: The concentrations of several metabolomic biomarkers, including cholesterol, triglycerides, fatty acids, and amino acids, differed between women who developed gestational diabetes mellitus and those who did not. Addition of biomarker profile improved the prediction of gestational diabetes mellitus provided by maternal demographic characteristics and elements of medical history alone (before addition: area under the receiver operating characteristic curve, 0.790; detection rate, 50% [95% confidence interval, 44.3%-55.7%] at 10% false positive rate; and detection rate, 63% [95% confidence interval, 57.4%-68.3%] at 20% false positive rate; after addition: 0.840; 56% [50.3%-61.6%]; and 73% [67.7%-77.8%]; respectively). The performance of combined testing was better for gestational diabetes mellitus treated by insulin (area under the receiver operating characteristic curve, 0.905; detection rate, 76% [95% confidence interval, 67.5%-83.2%] at 10% false positive rate; and detection rate, 85% [95% confidence interval, 77.4%-90.9%] at 20% false positive rate) than gestational diabetes mellitus treated by diet alone (area under the receiver operating characteristic curve, 0.762; detection rate, 47% [95% confidence interval, 37.7%-56.5%] at 10% false positive rate; and detection rate, 64% [95% confidence interval, 54.5%-72.7%] at 20% false positive rate). The calibration plot showed good agreement between the observed incidence of gestational diabetes mellitus and the incidence predicted by the combined risk model. In the sensitivity analysis excluding the women diagnosed with gestational diabetes mellitus before 20 weeks' gestation, there was a negligible difference in the area under the receiver operating characteristic curve compared with the results from the entire cohort combined.

Conclusion: Addition of nuclear magnetic resonance-based metabolomic profiling to risk factors can provide first-trimester prediction of gestational diabetes mellitus.