Category: Computers
Poster Session IV
Current early screening gestational diabetes mellitus (GDM) recommendations have yielded conflicting findings, suggesting that the current prediction parameters are insufficient. Machine learning is an analytic approach capable of processing complex interactions between variables that could yield improved predictive capabilities and enhance clinical decision making. We sought to fit a prediction model for GDM that could be implemented at the first prenatal visit.
Study Design:
Model development utilized prenatal data from the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b) cohort. Our primary outcome was a study criteria diagnosis of GDM, with pregestational diabetes and missing outcome data as exclusion criteria. Random forest (RF) models were developed using data commonly available during the first prenatal visit, including body mass index (BMI), laboratory data, and family history of diabetes, and evaluated using the area under the receiver operator curve (AUROC). The sensitivity and specificity of the RF model was then compared to current early screening parameters available in nuMoM2b (BMI over 30 kg/m2 or a first degree relative with diabetes).
Results: 8,242 nulliparous patients were included, with 359 (4.4%) developing GDM. RF modeling produced an AUROC of 0.71 [95% CI: 0.65, 0.77] on 26 features (Figure), with BMI at first prenatal visit, pre-pregnancy weight, age, and complete blood count results being the most predictive. The current early GDM screening criteria yielded a sensitivity and specificity of 0.63 and 0.61, respectively. In comparison, our model yielded a specificity of 0.69 (at a sensitivity of 0.63) and sensitivity of 0.70 (at a specificity of 0.61), thereby correctly identifying 11% more GDM patients than the current screening criteria (Table).
Conclusion: GDM can be predicted with moderate discrimination using machine learning models trained on nulliparous patient available at the first prenatal visit. Our model outperforms the current early GDM screening parameters and can ultimately assist screening to predict a diagnosis for earlier management.
Adesh Kadambi, BS
Delfina Care Inc.
San Francisco, California, United States
Isabel Fulcher, PhD
Vice President of Data Science
Delfina Care Inc.
San Francisco, California, United States
Jonathan Schor, PhD
Data Scientist
Delfina Care Inc.
San Francisco, California, United States
Ali Ebrahim, PhD
Delfina Care Inc.
San Francisco, California, United States
Senan Ebrahim, MD, PhD
Delfina Care Inc.
San Francisco, California, United States
Timothy Wen, MD,MPH (he/him/his)
Clinical Fellow
University of California, San Francisco
San Francisco, California, United States