engleski

The Use of Artifical Intelligence in Assessing Glucose Variability in Individuals with Diabetes Type 2 from Routine Primary Care Data

Background: The continuous glucose monitoring technique is recommended for follow-up of individuals with diabetes type 1. For those with diabetes type 2, glucose variability measures, either performed automatically or by visit-to-visit method, can be used to complement glycosylated haemoglobin (HbA1c) in predicting long-term outcomes. Methods: A-proof-of-concept study, conducted in primary care. A total of 63 variables were used from electronic health records to describe clinical characteristics of 110 individuals with diabetes type 2 of both gender, 40-86 years old (average 62.69), and on treatment with oral hypoglycaemic drugs. The artificial neural networks (ANN) of machine learning techniques was used to model inter-day glucose variability based on the estimation of variances (the square of the standard deviation) of sporadically recorded fasting and postprandial (2h after breakfast) glucose measurements as the outcome measures. Model of increased HbA1c (≥7%) was used as the benchmark. The number of variables for modelling was reduced by using the pre- processing method. Multiple linear regression (MLR) models were performed on the prepared subsets to compare to the predictive accuracy of ANN models. Results: A higher glucose variability, for both fasting and postprandial glucose variances, was associated with higher HbA1c values (Q1-Q4 differences, p = 0.002 and 0.006, respectively). The two top- ranked variables in ANN models of glucose variability were the same, indicating HbA1c and glomerular filtration rate, a measure of chronic renal impairment. MLR models of glucose variability did not give the significant predictors. Conclusions: For created models of glucose variability, to become practically useful, their outcome measures should be dichotomised and standardised according to the thresholds of HbA1c or some standardised measures of glucose variability, such as the coefficient of variation.

Diabetes mellitus type 2, Glucose variability, Primary care, Routine data, Machine learning, Pilot study

nije evidentirano