This work contributes two elements toward the quest for closed-loop control of T1DM. First, the regulatory approval of the clinical trials was based entirely on in silico
experiments performed in a computer-simulation environment.23
Second, the control algorithm was tested at three centers in three different countries, which added external validity to the data. Fully automated CGM data transfer is another characteristic distinguishing this study from other reports.12
Key features of this MPC algorithm are therefore automated SC glucose monitoring, SC insulin delivery, and personalization for each study participant using routinely available characteristics.
An important feature of this study is its repeated- measures design: each participant was tested twice under identical conditions in a tightly controlled hospital setting. The only variable that differed between admissions 1 and 2 was, therefore, the control strategy: patient-directed open-loop control at admission 1 and algorithm-suggested closed-loop control at admission 2. This permitted an objective assessment of the performance of MPC. The major advantage of closed-loop control was the nearly five-fold reduction in the number of nocturnal hypoglycemic episodes, plus a greater percentage of time that blood glucose spent within the narrow target range of 3.9–7.8 mmol/liter overnight. On the other hand, the performance of the MPC algorithm before and after breakfast was generally inferior to the open-loop control.
A weakness of the study was that the order of open-loop versus closed-loop conditions was not randomized. Typically, such a randomization is required in order to avoid “learning” effects. In this study, we need to differentiate the effect of “algorithm learning,” i.e., the control algorithm “learning” about the subject and his/her meal profile from open-loop data, and the effect of “human learning,” which could potentially contaminate the results. The idea of algorithm learning is that, over time, a profile of a person’s characteristics and daily regiment can be estimated and then used for control. In order to test algorithm learning, in this first study we supplied the algorithm with information about meals from the open-loop trial. In order to do so, we departed from the gold-standard randomized-order trials and had open loop always first. However, this was a pilot study testing new technology and not necessarily aiming for perfection in study design. Our subsequent studies (now ongoing) employ randomized order. Human (patient, personnel) learning was possible in this trial, but during closed loop, the patient and the attending personnel had little influence on insulin dosing, which was entirely done by the closed-loop control algorithm. Thus the influence of the order of open- and closed-loop control experiments should be minimal.
In terms of technology advancement, two comments are important. First, the control algorithm used insulin boluses administered every 15 min instead of the continuous basal rate. This was done because, in preparation for this study, we found that, due to SC insulin transport and “smoothing” of the insulin boluses during their transit from SC space to the circulation, 15 min boluses result in blood insulin concentrations that are indistinguishable from those generated by continuous insulin administration.24
Other practical advantages of such a discrete insulin delivery include more precise insulin dosing and optimization of pump battery life. Second, although the overall CGM performance was satisfactory, the CGM devices suffered from transient loss of sensitivity, particularly overnight. Although the exact definition of such events is not possible, on approximately 15 occasions across all subjects, the CGM readings experienced rapid drops, which did not correspond to reference blood glucose changes. For example, in
, three such drops were observed—two during admission 1 and one during admission 2. Such events may have been caused by increased pressure on the sensor during sleep that recovered after the patient repositioned.