Algorithm predicts hyperglycemia in gestational diabetes

The researchers published the study covered in this summary on medRxiv.org as a preprint that has not yet been peer reviewed.

Key points to remember

  • Researchers used machine learning to develop an algorithm capable of identifying women previously diagnosed with Gestational Diabetes (GDM) who are at high risk of an imminent episode of acute hyperglycemia.

  • The algorithm refined by machine learning is based on the last 3 days of self-reported glucose readings, and it piggybacks on a system, GDm-Healthpreviously developed by some of the same researchers that gives women diagnosed with gestational diabetes a way to systematically record multiple daily self-measurements of blood glucose.

  • The goal of the new algorithm is to help clinicians identify women with DG who need urgent clinical review of their diabetes management and possible medication adjustment.

  • The algorithm is in the early stages of development and has several limitations that need to be resolved before it is ready for routine use in clinical practice.

why it matters

  • Appropriate assessment, management, and treatment of GDM can reduce maternal and fetal complications related to GDM.

  • This algorithm automates the process of identifying women who need urgent clinical examination and more care to manage their GDM, helping clinicians provide patient-centered care and make better use of limited resources at a time when the prevalence of GDM is increasing.

  • This algorithm could be a smart complement to the smartphone–GDm-Health based system to monitor women with DG or as a stand-alone system for any DG clinic if they have access to patients’ daily blood glucose data.

  • It is the first stratification system based on machine learning to quantify the risk of hyperglycemia in women with DG.

study design

  • The development and in-house validation of the predictive model used 272,712 blood glucose readings from 1148 women with DG who were seen at Oxford University Hospitals from April 2018 to May 2021. The development used data from 672 of these pregnancies, and the initial validation used data from 168 other pregnancies. . Further validation used data from 186 additional pregnancies managed at another hospital in England.

  • All participants used the GDm-Health smartphone application to record the blood glucose measurement and to communicate the information to the researchers. The goal for participants was to measure and record their blood sugar four to six times a day at least 3 times a week.

  • The researchers rated the regression models they developed based on their mean squared error (MSE), their R2 value (a measure of the goodness of fit of a model) and their mean absolute error (MAE).

Principle results

  • MSE,R2MAE and rank accuracy scores all suggest that the best-performing model was moderately accurate in predicting high blood glucose readings, but requires further refinement to be suitable for clinical practice.

  • Analyzes also showed that tree-based ensemble models significantly outperformed a linear model. This may be because tree models take into account the nonlinear effects of data inputs.

  • In general, the overall performance of the models did not change significantly by adding additional measures beyond blood glucose levels. This may be due to the size of the dataset: a larger number of inputs often requires a larger sample size for training a model.

Limits

  • The model was moderately accurate in predicting high blood glucose readings, but needs to be refined with a larger sample to improve accuracy and determine which confounders should be included to make it suitable for clinical practice.

  • Because each participant was responsible for collecting and recording her blood glucose data, the number of entries varied considerably, with some women entering more values ​​than others, which may have introduced bias.

  • The data collected did not specify when patients started or stopped taking GDM-related medications.

  • Many study participants had missing data. Deleting participants with missing values ​​left 70% of the internal validation cohort and 25% of the external validation cohort. Future work on the algorithm should use more data, and future models may consider adding additional metrics, such as body mass index.

  • The authors acknowledged that other measures not included in their analysis may perform better than a model based exclusively on blood glucose.

Disclosures

  • The study did not receive commercial funding.

  • One author reports personal fees from Oxford University Innovation, BioBeats, and Sensyne Health, outside of submitted work. Another author is a part-time employee of Sensyne Health.

This is a summary of pre-publication research study “Machine Learning–Based Risk Stratification for Gestational Diabetes Management” written by researchers primarily from the University of Oxford, England, on medRxiv and brought to you by Medscape. This study has not yet been peer reviewed. The full text of the study is available at medRxiv.org.

Follow Medscape on Facebook, Twitter, instagramand Youtube

Comments are closed.