Blood glucose level prediction model using non-clinical data for type 2 diabetic patients in Sri Lanka

Abstract

Diabetes is a chronic disease characterized by elevated blood glucose levels. Poorly managed diabetes can cause serious complications when blood sugar levels consis- tently exceed clinically acceptable ranges. Knowing the variations in blood glucose levels allows active participation in maintaining blood glucose levels of patients with diabetes within clinically acceptable limits. Effective management of blood glucose prevents serious complications. Early information on glucose changes helps diabetes patients manage their dietary plans, medication doses, and lifestyle. Frequent or reg- ular blood glucose monitoring can be challenging, as standard blood glucose meters require a small blood sample from a fingertip to measure glucose levels in patients with type 2 diabetes. These measurements are obtained after the patient’s blood glu- cose levels have already changed. Continuous Glucose Monitoring (CGM) devices offer an alternative for diabetic patients to monitor their blood glucose levels. These devices use a tiny sensor im- planted under the patient’s skin. Most existing techniques leverage recorded CGM measurements, insulin doses, and dietary intake of diabetic patients to predict blood glucose levels at 30 min, 60 min, and 75 min intervals in the future. Diabetic patients must be aware of their daily routines and meals to manage their blood glucose levels, which can be controlled using correct dosages. A personalized self-management tool, such as a mobile application, can help patients collect data accurately. CGM devices are usually used in type 1 diabetic patients integrated with insulin pumps. There is an additionalriskifanimplantedCGMisusedinpatientswithtype2diabetes. Thisstudy introducesamachinelearningalgorithmthatusesinputdata,includingcurrentglucose levels and previously predicted, medication dosage, food consumption, and physical activity of type 2 diabetic patients for short-term prediction of blood glucose level. A long-short-term memory model (LSTM) was developed to forecast blood glucose lev- els in type 2 diabetic patients at 30 min and 60 min intervals. Two methods were used tocollectbloodglucosedatafromdiabeticpatientsformodeldevelopment: continuous datafrom10subjectsover7daysandrandomdatafrom100subjectsoverafewhours. Thepredictionmodeldemonstratedanimprovedaccuracyforthepredictionof30min and 60 min. The root mean square error of this model at 30 min and 60 min was 18.8 mg/dl and 20.3 mg/dl, respectively, with accuracies of 81.49% and 79.86%. Further- more, the prediction data in the LSTM model fell within the clinically acceptable "A" zoneoftheClarkErrorGrid. Mostpreviousmodelspredictedbloodglucosevaluesfor type 1 diabetic patients using various physiological parameters. This study predicted blood glucose levels in type 2 diabetic patients using a minimal training dataset. The accuracyofthemodeltopredictbloodglucoselevelsintype2diabeticpatientscanbe improvedbytrainingitwithalargerdatasetfromamoreextensivepatientpopulation.