270164 Early Hypoglycemia Alarm Systems Based On Autoregressive and Recursive Partial Least Squares Models
Hypoglycemia, having unusually low glucose concentration in blood (usually defined as ≤ 60 mg/dL), caused by intensive insulin therapy is a major challenge of artificial pancreas systems. Patients with diabetes regulate their glucose concentrations by insulin administration. Since patients with Type 1 diabetes (T1D) do not have natural means to reduce plasma insulin levels and their glucagon response is often impaired, they are often unable to prevent low blood glucose levels. Hypoglycemia is prevented or treated by eating and/or reducing or stopping insulin infusion. Hence, early detection and prevention of hypoglycemia is essential for insulin pump users and necessary for the implementation of fully automated closed-loop control. The frequency of hypoglycemia among people with T1DM is variable, and may occur as often as 2-3 times within a day. Hypoglycemia may lead to unconsciousness and seizures if untreated. Availability of continuous glucose monitoring sensors (CGMS) has enabled collection of frequent glucose concentration data and monitoring, prediction and control blood glucose in T1D.
Early alarms to warn the potential of hypoglycemia are essential and should provide enough time to take the necessary action to avoid hypoglycemia. Many alarm systems proposed in the literature are based on interpretation of recent trends in glucose values. In the present study, subject-specific linear models are introduced as a better alternative to capture glucose variations and predict future blood glucose concentrations. These models are then used in early hypoglycemia alarm systems that notify patients to take required action to prevent hypoglycemia before it happens.
An autoregressive and recursive partial least squares (ARPLS) modeling algorithm is proposed to analyze the frequent data from CGMS and predict future glucose values. Partial Least Squares (PLS) is a multivariate regression method, especially convenient for large number of highly correlated data set. PLS summaries the original data matrix to extract the most predictive information and maximize the covariance between input variables and response variables. Current or future glucose concentration can be expressed as a function of previous glucose measurement with autoregressive models. Here in this model, previous glucose concentrations are used as input to model and future glucose concentrations, response variables, are predicted.
In order to adapt the changes in the system, it is necessary to avoid accumulation of old data as new data become available. Autoregressive PLS models developed are recursively updated at each sampling step with a moving window size of one day. Different prediction horizons are analyzed to keep the prediction error minimum while providing a long enough horizon for early detection. An early hypoglycemia alarm algorithm using the developed models is proposed and evaluated.
ARPLS models are tested retrospectively using T1D subject data collected in University of Illinois at Chicago. The algorithm is also tested in silicoprospectively using an FDA approved metabolic simulator. The data from 17 T1D subject is first filtered to avoid adverse effect of noise on prediction. Even though continuous glucose sensors have analog filters embedded, the output signal is still noisy and should be filtered in order to enhance their signal to noise ratio. A Savitzky-Golay filter is used to reduce noise in the CGMS data.
The alarm algorithm is developed based on the future glucose predictions from ARPLS algorithm. The algorithm first checks the current data and if the value (actual glucose reading) is under the hypoglycemic threshold, immediate hypoglycemia alarm is triggered. Then it checks further for future predictions to trigger early hypoglycemia alarms. When the n-step predicted value crosses the hypoglycemia threshold (70mg/dl), an early hypoglycemia alarm is issued. Only early hypoglycemia alarms are considered to evaluate our algorithm’s performance; alarms held during the event are not counted as true positive as the continuous sensors are already equipped with immediate alarms for the current data point. Sensitivity, false positive ratio and time to detection are reported to determine alarm performance.
ARPLS models are evaluated in terms of root mean squared error (RMSE) and sum of squares of glucose prediction error (SSGPE) and compared with previously proposed time series algorithms. For a CGM sampling rate of 5 minutes, 4.45 RMSE and 4.04% SSGPE are obtained for 6 step-ahead glucose predictions for recursive ARPLS. For nonrecursive ARPLS, RMSE is 6.17 and SSPGE is 4.55%. For nonrecursive and recursive time series models (ARMA (3,1)) SSPGE’s are reported as 10.32% and 5.56% respectively. Alarm systems developed are tested retrospectively both using the T1D patient data and in silico simulations. The performance of alarm systems based on autoregressive recursive PLS is evaluated. Out of 69 hypoglycemia events, 62 of them are detected with average detection time of 28.25 min. Sensitivity of 90% is reported for the early alarm based on 6 step-ahead predicted glucose values. False alarm rate is also reported as 0.36 false positive/day.
The recursive and autoregressive algorithm proposed enable the dynamic adaptation of the models to inter-/intra-subject variation and glycemic disturbances. The ARPLS models developed provide satisfactory glucose prediction with relatively less error than other proposed algorithms and are good candidates to detect and prevent hypoglycemia.