Noise spectral analysis and error estimation of continuous glucose monitors under real-life conditions of diabetes patients

Abstract We present an analysis of the nature of noise and measurement errors for two popular continuous glucose monitors (Medtronic OneTouch UltraSmart and Abbott FreeStyle Libre) under daily, real-life conditions. We use three different techniques to extract the noise in the time series of the interstitial glucose values provided by these sensors: local regression, singular spectrum analysis, and wavelets. Likewise, we characterize the noise using five different methods: power spectral density, normality tests for noise values, the generalized Hurst exponent, Bayesian information criterion, and the fit of the noise to an autoregressive model. To analyze measurement errors, we use five different metrics: the instantaneous error, the minimum error in a 60-min timeframe, the lag of the minimum error in the 60-min timeframe, and both the average and the standard deviation of the absolute value of the instantaneous error. We perform the analysis on data provided by 20 patients totaling 198162 samples that account for more than 1732 days of data measured by patients during their daily lives. The results conclude that continuous glucose monitors have either pink or red noise; the fit of autoregressive models to the noise is lower than previously reported, and the elective model, if any, has a high order; the medians of the measurement errors have a strong presence of outliers with values above 80 mg/dl, lags higher than reported in hospital, controlled trials, and considerable intra- and inter-patient variability; and that glucose monitor accuracy is dependent upon the patients’ glycemic variability.