Controlled In Vitro Characterization of the Dynamic Response of Continuous Glucose Monitoring Systems: Adaptation of a Programmable Flow Platform and Decomposition of Dynamic Error

BackgroundContinuous glucose monitoring (CGM) systems are used not only for retrospective assessment of the glycemic profile but also for real-time decision-making, including automated insulin delivery. Accordingly, CGM performance characterization must capture not only the agreement of individual paired values but also the systems ability to reproduce the direction, rate, amplitude, and shape of glucose concentration change. Summary metrics, most notably MARD, cannot establish whether an observed deviation reflects an error in the formation of the test profile itself, a constant sensor offset, amplitude compression, a change in response rate, temporal misalignment, or hysteresis. ObjectiveTo adapt a programmable flow-based in vitro platform for the separate assessment of the experimentally delivered glucose profile and the dynamic response of CGM systems, and to propose a set of metrics that decomposes dynamic error into its components. MethodsGLU profiles were generated by programmable mixing of solutions at a constant total flow rate of 2 mL/min. Actual GLU concentration was independently measured with a SUPER GL2 glucose analyzer. Four static levels, three repeats of a 5.5[->]12.0[->]5.5 mmol/L profile, three repeats of a 6.0[->]3.0[->]6.0 mmol/L hypoglycemic profile, three 5.0[->]15.0[->]5.0 mmol/L profiles at different rates, one complex 4[->]18[->]3[->]12[->]5.5 mmol/L profile, and two proof-of-concept sensor experiments at 100- and 200-min transitions were investigated. Dynamic response was characterized by bias, MAE, RMSE, MARD, amplitude transfer coefficient K_A, rate transfer coefficients K_up and K_down, normalized shape RMSE, residual shift, and hysteresis loop area. ResultsAt the static levels, measured GLU exceeded the programmed value by 0.234-0.780 mmol/L. In the repeated 5.5[->]12.0[->]5.5 profiles, the ratio of actual to programmed rate was 0.978-1.083 on the rising phase and 0.987-1.157 on the falling phase, while the amplitude transfer coefficient was 0.967-1.066. In the hypoglycemic profile, minimum GLU was 2.55- 2.96 mmol/L, and time below 3.0 mmol/L was 15.2-72.6 min. The measured rates of 0.0519, 0.1045, and 0.2027 mmol/L/min preserved the intended ratio of approximately 1:2:4. In the complex profile, the programmed plateau of 18 mmol/L was not reached: mean measured GLU was 16.20 mmol/L. For CGM-A, K_A was 0.682 and 0.650, and K_up/K_down were 0.666/0.730 and 0.634/0.626; the corresponding values for CGM-B were 1.228 and 1.128, and 1.564/1.328 and 1.276/1.145. Hysteresis loop area differed 5- to 10-fold between the two sensor responses, exceeding an order of magnitude at the 100-min transition. ConclusionThe programmed concentration should be treated as a control setpoint, rather than as a reference measurement. The "programmed trajectory -- measured glucose -- CGM output" cascade first allows quantitative assessment of the agreement between the programmed and actually realized profile and only then separate characterization of sensor response. Decomposition of dynamic error into amplitude, rate, shape, and hysteresis components reveals differences that a single MARD value or correlation coefficient cannot capture.