Given the ever-increasing prevalence of diabetes around the globe, offering patients a way to easily and effectively monitor blood glucose may hold the key to managing disease-related morbidity. Pilot research by a North Dakota research team may help in the struggle: their personalized prediction model uses smartphone-collected patient data and aggregate population data to create personalized blood glucose predictions for individuals, with promising results.
As reported in a recent article in ICT Express, patients with diabetes often find it impractical to measure their blood glucose as many as four times each day, given the demands of daily routines. As a result, significant research has been conducted into automatic blood-glucose prediction models, using either population-based prediction or patient-based prediction. As the authors noted, however, both models have suffered “from issues of low accuracy and/or sparse data.”
To help address the problems associated with existing approaches, the researchers proposed a synthesized model that combines both population- and patient-based analysis and prediction. The system uses a smartphone to collect daily activity patterns, pool patients’ historical information, and analyze these metrics with in conjunction with historical population data.
With respect to personalized data, the smartphones collected regular blood glucose measurements as well as information on daily events that impact blood glucose levels, including insulin, meals, exercise, and sleep. Some of these data are collected automatically using the ambient-sensing features of a smartphone, while other data (such as insulin dose) need to be entered manually.
The model comprises a three-stage evolution model to make personalized blood glucose predictions. A time-series prediction model based on patient data forecasts the time series of an individual patient’s blood glucose measurements. A pooled-panel-data regression model makes predictions based on an individual patient’s historical data. Finally, a pre-clustered personalized regression model addresses the issue of heterogeneity that may occur in the pooled-panel-data regression model.
To help evaluate the performance of the proposed model, the investigators conducted a series of experiments using the diabetes data set from the University of California at Irvine’s Machine Learning Repository, which comprises 70 sets of data recorded on diabetes patients. The analysis used mean absolute error, root-mean-square error , and coefficient of determination to evaluate prediction performance.
The prediction models were all evaluated by extensive simulation experiments. As the authors noted, “[t]he experimental results demonstrate that the proposed model improves the prediction accuracy and remedies the data sparsity problem of the existing models.”
Equally important in a world of convenience is that it does so via an interface that most users will find easy and comfortable to use, with a variety of self-management functions. An Overview option gives a summarized view of patients’ historical health data. Several other tabs -- Glucose, Medications, Activity, Diet, and Blood Pressure – offer friendly interfaces where users can input or automatically collect their health-related data. A Warnings component uses the blood glucose prediction model and automatically sends warning messages to users based on their predicted blood glucose level.