Automatic classification of glycaemia measurements to enhance data interpretation in an expert system for gestational diabetes
Introduction
As in other types of diabetes, the prevalence of Gestational diabetes mellitus (GDM) is increasing throughout the world (IDF, 2015). If the new International Association of Diabetes Study Group diagnosis criteria (IADPSG, 2010) –recently proven to be cost effective (Duran, Saenz, & Torrejon, 2014) – are adopted, the prevalence could be doubled. Several adverse outcomes are associated with hyperglycaemia in pregnancy, as foetal macrosomia, shoulder dystocia or caesarean section (Metzger, Lynn, & Lowe, 2008). Although most cases resolve with delivery, both mother and foetus are at a higher risk of developing type 2 diabetes in the future (Boney, 2005, Franks et al., 2006).
Maternal glycemic control reduces adverse GDM outcomes (Hartling, Dryden, & Guthrie, 2013) so patients are prescribed to self-monitor their blood glucose (BG) levels with a glucose meter around main meals. Although measurements are stored in the glucose meter memory file, patients usually note down their results in a paper logbook, structuring measurements in relationship to meals. They indicate the specific meal which the measurement is related to (breakfast, lunch or dinner) and whether it was made before (preprandial) or after the meal (postprandial). Clinicians evaluate the patients’ measurements each week or every other week to determine the appropriate treatment, which consists of nutritional prescription, physical activity and, if necessary, insulin therapy. It has been observed that patients commit errors when manually reporting their BG levels, being the mean values significantly higher than the logbooks’ ones (Given, O’ Kane, Bunting, & Coates, 2013). Although this could mask a bad glycemic control and make clinicians establish a wrong therapy, they still prefer to examine logbooks instead of meter memory files (Polonsky, Jelsovsky, & Panzera, 2009). The reason might be that logbooks are easier to be reviewed, as they provide structured information that glucose meter memory lacks, like associations of measurements to meals, which are essential to make therapy adjustments. Logbooks also provide additional information such as food intakes, insulin doses or exercise.
Telemedicine allows patients to send their BG data to the system to be remotely evaluated, which avoids unnecessary displacements (Carral et al., 2015, Pérez-Ferre et al., 2010) and improves access to specialized care in rural areas (Mohan & Pradeepa, 2014). Furthermore, by a more exhaustive and frequent evaluation of accurate data, telemedicine is capable of improving glycemic control (Wojcicki, Ladyzynski, & Krzymien, 2001) and reducing GDM adverse outcomes (Dalfra et al., 2009, Ferrara et al., 2012). Monitoring data in telemedicine systems should be presented to clinicians organized as they appear in paper logbooks to facilitate their interpretation.
The use of telemedicine could increase clinician's workload as it favors the generation of a greater amount of data to be evaluated by clinicians. Expert systems can solve the potential increment of clinicians’ workload (Klonoff & True, 2009) by automatically analysing patients’ monitoring data according to expert specifications (Hernando, Gómez, Corcoy, & del Pozo, 2000). The automatic analysis of monitoring data could optimize clinician's time by notifying which patients are evolving satisfactorily and which ones need a deeper examination. Expert systems, like clinicians, need to analyze glycaemia data in relation to meals to be able to determine patients’ condition and to generate specific recommendations about therapy adjustments.
Glycaemia data entry in expert systems can be performed by patients either manually or by uploading the data stored in their glucose meter (El-Gayar, Timsina, Nawar, & Eid, 2013a). The automation of data entry is preferred as it minimizes transcription errors (Given et al., 2013), results in more data captured, simplifies the date entry process and increases patients’ satisfaction (El-Gayar, Timsina, Nawar, & Eid, 2013b). One of the problems that expert systems in diabetes have to face is the management of incomplete glycaemia measurements. A measurement is considered incomplete if it lacks its association with a meal or with a moment of measurement. Newer glucose meters can include the functionality to allow registering these data manually, but even if they do, it is a time-consuming task and patients sometimes forget to introduce it. Without an accurate method to manage incomplete glycaemia data, expert systems cannot determine if patients’ metabolic condition is altered due to a specific meal that should be adjusted or due to an extended fasting period.
The majority of studies available in literature about expert systems do not explicit describe the method used to retrieve the associated meal and moment of measurement of glycaemia data or how they manage the lack of this information. Some expert systems allow patients to add a meal tag to glycaemia measurements after uploading data from the glucose meter (Cafazzo et al., 2012, Lim et al., 2011, Quinn et al., 2008), but it has been observed that sometimes patients forget to label some of the measurements (Mackillop et al., 2014) so they cannot be automatically analyzed. To solve this problem, the expert system can preselect a meal tag for each measurement downloaded and allow patients to modify it. Bromuri et al, pre-select glycaemia meal tags based on previous measurements, so, if the last measurement was taken before dinner, an after dinner period is preselected (Bromuri, Puricel, & Schumann, 2016). However, this method might present problems when dealing with repeated or missing measurements, for example if the patient forgets to measure her glycaemia after dinner, she measures it the following day before breakfast and she, by mistake, accepts the preselected meal tag. Some commercial applications automatically classify the glycaemia measurements downloaded by patients according to patient's predefined mealtimes (Sanofi Diabetes, 2015), but this method might present an elevated rate of errors as we will see in the following sections. We propose an innovative method for glycaemia meal tag preselection using machine learning techniques.
This paper presents the methodology to design an automatic classifier to associate the appropriate meal and moment of measurement to each glycaemia data downloaded from a glucose meter, its integration within the Sinedie expert system for GDM and the classification results obtained in a pilot study at Hospital de Sabadell with 47 patients for 8 months.
Section snippets
Material and methods
This section describes the Sinedie expert system and how the automatic classifier is integrated with the BG levels uploading procedure. We explain the two different classification strategies studied to design the classifier: a simple algorithm based on the patient's mealtime schedules, measurements’ time and BG level; and a more complex algorithm based on machine learning techniques. Finally, we describe the design of the clinical evaluation experiment.
Results
In this section we present the results obtained in the machine learning techniques evaluation, considering FS and learning algorithms tests, the classification module structure specification and its performance in the Sinedie expert system where it was integrated.
Discussion
Machine learning techniques have confirmed our initial hypothesis about the need of using more variables than the 3 initially proposed (mealtime schedules, BG level and measurement time) to obtain better accuracy results that optimize the automatic classification process. Increasing the number of input features improved classification accuracy in both learning algorithms tested, but it is by the application of feature selection methods how we achieved the highest performance. The common 6
Conclusions
Machine learning tools have proved to be effective to design an automatic classifier for glycaemia measurements, required to provide automatic recommendations based on an expert system for gestational diabetes. All methods tested obtained accuracy results above 80%, but the highest performance was achieved with the C4.5 decision tree learning algorithm with 7 inputs selected by a wrapper evaluator and the Genetic search algorithm. The selected classifier not only performed well with the
Acknowledgment
This work has been funded by the Spanish grant Sinedie (PII0/01125), co-funded by FEDER. We would like to thank the patients of Hospital de Sabadell for their collaboration and support in this research.
References (45)
- et al.
An intelligent mobile based decision support system for retinal disease diagnosis
Decision Support Systems
(2014) - et al.
A systematic review of IT for diabetes self-management: Are we there yet?
International Journal of Medical Informatics
(2013) - et al.
Referral to telephonic nurse management improves outcomes in women with gestational diabetes
American Journal of Obstetrics and Gynecology
(2012) - et al.
Evaluation of DIABNET, a decision support system for therapy planning in gestational diabetes
Computer Methods and Programs in Biomedicine
(2000) - et al.
Feature selection and classification model construction on type 2 diabetic patients’ data
Artificial Intelligence in Medicine
(2007) - et al.
A Telemedicine system based on Internet and short message service as a new approach in the follow-up of patients with gestational diabetes
Diabetes Research and Clinical Practice
(2010) - et al.
Comparison of NN and LR classifiers in the context of screening native American elders with diabetes
Expert Systems with Applications
(2013) - et al.
Metabolic syndrome in childhood: association with birth weight, maternal obesity, and gestational diabetes mellitus
Pediatrics
(2005) - et al.
An expert Personal Health System to monitor patients affected by Gestational Diabetes Mellitus: A feasibility study
Journal of Ambient Intelligence and Smart Environments
(2016) - et al.
Design of an mHealth app for the self-management of adolescent type 1 diabetes: A pilot study
Journal of Medical Internet Research
(2012)