A learning system for error detection in subcutaneous continuous glucose measurement using Support Vector Machines

Current continuous glucose monitors have limited accuracy mainly in low level glucose measurements, being a sharply bounding factor in the clinical use. The ability to detect incorrect measurements from the information supplied by the monitor itself, would thus be of utmost importance. In this work, the detection of therapeutically wrong measurements of Minimed CGMS is addressed by means of Support Vector Machines (SVM). In a clinical study patients were monitored using the CGMS and during the stay at the hospital blood samples were also taken. After synchronization, a set of 2281 paired samples was obtained. Making use of the monitor's electrical signal and glucose estimation, the error detection is accomplished systematically through the study of classification problems using Error Grid Analysis for establishing accurate measurements versus benign errors and therapeutically relevant errors. Gaussian SVM classifiers were designed optimizing the σ-value iteratively. Validation was performed using 10×10 cross-validation together with permutation technique. An overall good performance is obtained in spite of the somewhat low sensitivity.