A Stacked Long Short-Term Memory Approach for Predictive Blood Glucose Monitoring in Women with Gestational Diabetes Mellitus

Huiqi Y. Lu, Ping Lu, Jane E. Hirst, Lucy Mackillop, David A. CliftonSensors 2023 · OpenReview

TL;DR

This paper introduces a stacked Long Short-Term Memory (LSTM) model for predicting blood glucose (BG) levels in pregnant women with Gestational Diabetes Mellitus (GDM). By leveraging routine fingerstick BG data, the model achieves prediction accuracy comparable to more expensive continuous glucose monitoring (CGM) systems, paving the way for optimized monitoring schedules and improved patient care.

What Problem Does This Paper Address?

Gestational diabetes mellitus (GDM) is a common pregnancy complication characterized by high blood glucose levels. Effective management of GDM is crucial to prevent adverse outcomes for both mother and child. Current GDM management often relies on frequent, standardized blood glucose testing, which can be burdensome for patients and may not be optimal for individuals with stable glucose levels. While continuous glucose monitoring (CGM) offers more granular data, its high cost and limited evidence for GDM management make it inaccessible for many. This paper addresses the need for a cost-effective, accurate, and personalized approach to BG monitoring in GDM, aiming to optimize testing frequency and improve clinical decision-making.

Methods

The core of this study is the development and evaluation of a stacked Long Short-Term Memory (LSTM) recurrent neural network. LSTMs are well-suited for time-series data like blood glucose readings due to their ability to capture temporal dependencies and handle sequential patterns. The authors employed a three-layer stacked LSTM architecture to enhance the model's capacity to learn complex relationships and abstract representations from the data.

Key methodological aspects include:

  • Data Preprocessing: Rigorous cleaning of the raw BG data, including removal of invalid tags, out-of-range readings, and data outside a specific gestational age window. Missing data was imputed using a 5-day moving average.
  • Time-Series Windowing: The data was structured into observational windows (7 or 14 days of past BG readings) and prediction windows (predicting BG for the next 7 or 14 days).
  • Model Training: The stacked LSTM model was trained using batch learning with a batch size of 32 and optimized using the Adam optimizer. Mean Squared Error (MSE) was used as the loss function.
  • Evaluation Metrics: Performance was assessed using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), with models run five times to report mean and standard deviation.
  • Subgroup Analysis: The model's performance was evaluated separately for "Normal" and "Hyper" BG subgroups, defined by NICE guidelines.

Datasets

The study utilized a large dataset of 190,396 blood glucose readings collected from 1110 patients with GDM. This data was extracted from the GDm-Health platform, a smartphone-based, Bluetooth-controlled BG monitoring system used in UK hospitals. The data spans from August 2017 to May 2021 and was collected from a single site at the John Radcliffe Hospital, Oxford University Hospitals NHS Foundation Trust. No new benchmark datasets were introduced; instead, the study leverages existing real-world clinical data.

Key Insights and Ideas

  • Fingerstick Data is Sufficient: The study demonstrates that routine fingerstick BG data, when analyzed with advanced deep learning models like stacked LSTM, can achieve prediction accuracy comparable to CGM. This is a significant finding as it bypasses the need for expensive CGM devices.
  • Optimized Prediction Window: The research identifies a 7-day observational window to predict the next 14 days as the most effective timeframe for BG monitoring in GDM. This suggests a potential to reduce the frequency of blood tests for patients with stable glucose levels.
  • Pre-meal Predictions are More Accurate: The study consistently found that blood glucose levels before meals (BB, BL, BD) are more predictable than post-meal levels (AB, AL, AD). This is likely due to the more controlled nature of fasting states compared to the variability introduced by food intake.
  • Stacked LSTM's Power: The stacked LSTM architecture proved effective in capturing complex temporal patterns in BG data, outperforming simpler models and demonstrating robustness.

Key Results

  • The optimized 7-day observation to 14-day prediction window achieved the following RMSE values:
  • After-breakfast (AB): 0.958 ± 0.007 mmol/L
    • After-lunch (AL): 0.876 ± 0.003 mmol/L
    • After-dinner (AD): 0.898 ± 0.003 mmol/L
    • Before-breakfast (BB): 0.622 ± 0.003 mmol/L
    • Before-lunch (BL): 0.814 ± 0.009 mmol/L
    • Before-dinner (BD): 0.845 ± 0.005 mmol/L
  • The model's accuracy for fingerstick BG tests was found to be on par with benchmark 1-hour prediction models using CGM data.
  • Predictions for the Normal BG subgroup were more accurate (lower RMSE and MAE) than for the Hyper subgroup.
  • The study suggests that the optimized BG schedule based on the 7-day prediction of the next 14 days could potentially reduce the overall number of blood tests for stable patients.

Limitations and Future Work

  • Single-Site Data: The predictive models were developed using data from a single clinical site. This raises questions about generalizability to diverse patient populations and healthcare settings.
  • Fingerstick Discomfort: The authors acknowledge that fingerstick testing can be uncomfortable and inconvenient, especially for pregnant women.
  • Limited Patient Health Data: The model did not incorporate static patient health data (e.g., age, BMI) due to data missingness. While previous studies suggested limited impact on short-term predictions, this could be explored further.
  • Medication Data: The current model does not include time-series medication information, which could be crucial for personalized precision medication management.
  • Future Work: The authors plan to:
  • Develop models incorporating time-series medication data.
  • Explore different optimization methods and network architectures.
  • Expand the study by using federated learning or transfer learning on data from multiple hospitals.
  • Investigate how clinicians would integrate this algorithm into their decision-making and how to effectively communicate predictions to patients.

Summary updated

Abstract

Gestational diabetes mellitus (GDM) is a subtype of diabetes that develops during pregnancy. Managing blood glucose (BG) within the healthy physiological range can reduce clinical complications for women with gestational diabetes. The objectives of this study are to (1) develop benchmark glucose prediction models with long short-term memory (LSTM) recurrent neural network models using time-series data collected from the GDm-Health platform, (2) compare the prediction accuracy with published results, and (3) suggest an optimized clinical review schedule with the potential to reduce the overall number of blood tests for mothers with stable and within-range glucose measurements. A total of 190,396 BG readings from 1110 patients were used for model development, validation and testing under three different prediction schemes: 7 days of BG readings to predict the next 7 or 14 days and 14 days to predict 14 days. Our results show that the optimized BG schedule based on a 7-day observational window to predict the BG of the next 14 days achieved the accuracies of the root mean square error (RMSE) = 0.958 ± 0.007, 0.876 ± 0.003, 0.898 ± 0.003, 0.622 ± 0.003, 0.814 ± 0.009 and 0.845 ± 0.005 for the after-breakfast, after-lunch, after-dinner, before-breakfast, before-lunch and before-dinner predictions, respectively. This is the first machine learning study that suggested an optimized blood glucose monitoring frequency, which is 7 days to monitor the next 14 days based on the accuracy of blood glucose prediction. Moreover, the accuracy of our proposed model based on the fingerstick blood glucose test is on par with the prediction accuracies compared with the benchmark performance of one-hour prediction models using continuous glucose monitoring (CGM) readings. In conclusion, the stacked LSTM model is a promising approach for capturing the patterns in time-series data, resulting in accurate predictions of BG levels. Using a deep learning model with routine fingerstick glucose collection is a promising, predictable and low-cost solution for BG monitoring for women with gestational diabetes.
BibTeX
@inproceedings{lu2023stacked,
  author = {Huiqi Y. Lu and Ping Lu and Jane E. Hirst and Lucy Mackillop and David A. Clifton},
  title = {{A Stacked Long Short-Term Memory Approach for Predictive Blood Glucose Monitoring in Women with Gestational Diabetes Mellitus}},
  booktitle = {Sensors},
  year = {2023},
  url = {https://openreview.net/forum?id=y3jzDyphbU},
}
Figure