Machine Learning-Based Risk Stratification for Gestational Diabetes Management

Jenny Yang, David A. Clifton, Jane E. Hirst, Foteini K. Kavvoura, George Farah, Lucy Mackillop, Huiqi Y. LuSensors 2022 · OpenReview

TL;DR

This paper presents a machine learning model that uses daily blood glucose readings and electronic health records to predict a pregnant individual's risk of high blood sugar levels (hyperglycemia) in the coming days. This system aims to help clinicians prioritize patients for review, enabling more personalized and efficient management of Gestational Diabetes Mellitus (GDM).

What Problem Does This Paper Address?

Gestational Diabetes Mellitus (GDM) affects about 1 in 6 pregnant individuals worldwide and can lead to serious maternal and fetal complications, including preeclampsia and neonatal hypoglycemia. Early diagnosis and management are crucial, but GDM is often diagnosed late in pregnancy, leaving a narrow window for intervention. Current monitoring involves frequent self-testing and manual review of blood glucose logs by clinicians, which can be time-consuming and may not effectively identify individuals most at risk of dangerous glucose spikes. This paper addresses the need for a more efficient and proactive approach to GDM management by stratifying patients based on their predicted risk of hyperglycemia.

Methods

The researchers developed a machine learning model using regression techniques to predict the proportion of "high-reading" blood glucose measurements (above NICE guidelines) a patient is likely to have in the three days following a given observation period. They explored linear regression (MLR), Random Forest, and XGBoost models. The core idea is to use a three-day window of historical data (blood glucose readings and EHR data) to predict the risk for the subsequent three days. Feature engineering included creating "gradient" features (rate of blood glucose change) and "high-reading" features (percentage of high readings). XGBoost, a gradient boosting algorithm, emerged as the top performer due to its ability to handle complex relationships and its state-of-the-art performance in tabular data challenges.

Datasets

The study utilized two datasets:

  • Oxford University Hospitals NHS Foundation Trust (OUH) cohort: This was the primary dataset for training and internal validation, comprising 1148 pregnancies and over 272,000 blood glucose readings.
  • Royal Berkshire Hospital NHS Foundation Trust (RBH) cohort: This dataset, with 709 pregnancies, was used for external validation to assess the model's generalizability.

Key Insights and Ideas

  • Proactive Risk Stratification: The core innovation is shifting from reactive monitoring (reviewing past logs) to proactive prediction of future high-risk periods, allowing for timely interventions.
  • XGBoost's Superiority: Tree-based ensemble models, particularly XGBoost, significantly outperformed linear regression, highlighting their effectiveness in capturing the complex, non-linear relationships in GDM data.
  • Importance of Blood Glucose Readings: Daily blood glucose measurements, especially "pre-breakfast" readings, were found to be the most influential features for predicting hyperglycemia risk.
  • Generalizability: The model demonstrated similar performance on an independent external dataset, suggesting its potential for broad application across different clinical settings.

Key Results

  • Internal Validation (OUH): The XGBoost model achieved the best performance with a Mean Squared Error (MSE) of 0.021, an R-squared (R2) of 0.482, and a Mean Absolute Error (MAE) of 0.112.
  • External Validation (RBH): The model maintained strong performance, with the XGBoost model trained on "Tags" and "High-Readings" achieving an MSE of 0.020, R2 of 0.519, and MAE of 0.108.
  • Rank Accuracy: The models showed good accuracy in correctly triaging patients into risk categories (lower, middle, upper bounds), with the best models achieving accuracies around 60-65% for the upper bound.
  • Feature Importance: Blood glucose readings (Tags) and the "High-Readings" feature were consistently the most important predictors. EHR data like maternal age and gestational day showed less impact in this specific model configuration.

Limitations and Future Work

  • Dataset Size and Imbalance: The authors acknowledge that the dataset size, while substantial, could be larger. Data imbalance (more instances of lower risk) might also affect model performance. Future work could explore data imputation methods and weighted regression.
  • Medication Data Granularity: The precise timing of medication initiation or cessation was not clear, which could impact the model's ability to differentiate between patient groups. More detailed medication data could improve future models.
  • Clinical Practice Integration: While the model shows promise, further studies are needed to translate its risk scores into practical clinical decision-making tools and to assess its impact on patient outcomes and clinician workflows.
  • Additional Features: Exploring other known diabetes risk factors like BMI (which had missing data in this study) and other EHR data could potentially enhance model performance.

Summary updated

Abstract

Gestational diabetes mellitus (GDM) is often diagnosed during the last trimester of pregnancy, leaving only a short timeframe for intervention. However, appropriate assessment, management, and treatment have been shown to reduce the complications of GDM. This study introduces a machine learning-based stratification system for identifying patients at risk of exhibiting high blood glucose levels, based on daily blood glucose measurements and electronic health record (EHR) data from GDM patients. We internally trained and validated our model on a cohort of 1148 pregnancies at Oxford University Hospitals NHS Foundation Trust (OUH), and performed external validation on 709 patients from Royal Berkshire Hospital NHS Foundation Trust (RBH). We trained linear and non-linear tree-based regression models to predict the proportion of high-readings (readings above the UK’s National Institute for Health and Care Excellence [NICE] guideline) a patient may exhibit in upcoming days, and found that XGBoost achieved the highest performance during internal validation (0.021 [CI 0.019–0.023], 0.482 [0.442–0.516], and 0.112 [0.109–0.116], for MSE, R2, MAE, respectively). The model also performed similarly during external validation, suggesting that our method is generalizable across different cohorts of GDM patients.
BibTeX
@inproceedings{yang2022machine,
  author = {Jenny Yang and David A. Clifton and Jane E. Hirst and Foteini K. Kavvoura and George Farah and Lucy Mackillop and Huiqi Y. Lu},
  title = {{Machine Learning-Based Risk Stratification for Gestational Diabetes Management}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=2G34dJ5XwV},
}
Figure