(188z) Sparse Kernel Filtering Algorithms for Online Glucose Prediction in T1D

Development of predictive models for nonlinear systems with time-varying parameters is challenging. It is difficult to present a personalized physical model based on first principles, as the parameters of such models are usually time-variant or cannot be estimated easily. The alternative is data-driven modeling methods to estimate parameters and capture the relationships between inputs and outputs. But one of the challenges to achieve a real-time prediction based on empirical models is the dynamic changes related by many physical or environmental elements over time, which necessitate online updating of model parameters. Kernel-based modeling approaches with sparsification criteria, Approximate Linear Dependency (ALD) Criterion and Surprise Criterion (SC) are proposed in this work to enhance the accuracy of data-driven models. The approach is illustrated for real-time prediction of glucose concentration. With information from continuous glucose monitoring (CGM) systems, accurate online prediction of glucose concentration is becoming feasible for type 1 diabetes mellitus (T1D). In artificial pancreas (AP) systems, online glucose prediction is a key issue for proactive blood glucose regulation and for producing early alarms for hypo- and hyperglycemia. To obtain reliable and accurate glucose prediction values, various approaches based on physical or empirical models have been proposed.

Kernel-based modeling with sparsification criteria, ALD and SC is investigated for online glucose prediction for patients with T1D. The model is trained by current and historical CGM data and the kernel filtering algorithms are proposed to characterize the glycemic variability and serve as the online learning machine for glucose prediction. By combining sparsification criteria based on information theory, the learning dictionary online could be updated recursively by checking if the new kernel function is appropriate to be add into the subset. For ALD criterion, the distance of the new coming data to the linear span of the present dictionary in the reproducing kernel Hilbert spaces (RKHS) is indicated and the thresholds are set to classify if the data are learnable, redundant or abnormal. For SC criterion, the uncertainty of a new input-output pattern relate to the current knowledge of the learning system is quantified. It uses an information theoretic method that captures the surprise of the new exemplar and allows us to add or discard it in the previous learning system. Therefore, the dictionary growth of the online filter could be effectively curbed and the overfitting is avoided.

In AP systems, the time and space complexity of the glucose prediction model could be reduced by setting an appropriate threshold to ignore the redundant data without harming the prediction performance. Meanwhile, the abnormal readings of the CGM sensor could also be removed by setting the overfitting threshold. This is critical to compensate for CGM measurement noise and CGM inaccuracy. Based on above, the proposed real-time update scheme permits the prediction models to properly address time-varying characteristic of the glucose dynamics.

The validation of online glucose prediction is investigated from both in silico data (UVA/Padova metabolic simulator) and clinical data. Model predictions are compared with actual CGM measurements. The sampling time of the sensor is 5 minutes. Prediction accuracy is assessed on the basis of the metrics: mean absolute relative deviation [MARD (%)], root mean-square error [RMSE (mg/dL)], prediction-Error Grid Analysis (PRED-EGA) and network size. The proposed sparse kernel-based algorithms show better accuracy, especially in cases where random sensor noise is added and CGM sensor calibration procedure is performed.