This randomized crossover study has demonstrated that a CL system with a pRBA achieves normoglycemia more than 95% of the time during the nocturnal period without increasing the risk of hypoglycemia. Postprandial results, although not inferior to those obtained with a standard premeal bolus, did not avoid an excessive glycemic excursion, at least in some patients. The percentage of time spent in target range is higher in our study compared with other CL studies. In fact, a very recently published multicenter outpatient study in close supervision (hotel, guest house, hybrid hospital-hotel) reported 72% of the overnight time to be within 3.9–10
Previous in-clinic CL trials using controllers that imitate the reasoning of diabetes caregivers15,25
have shown an increase in overnight glucose stability. However, CL comparison was done with CGM and not with venous glucose values.
In our study, reduction of hypoglycemia, measured as time spent with a BG level <3.9
mmol/L and as reduction in the number of episodes under the same threshold, is significant and better than that observed in other single-hormone CL studies.8,9,11
Because the nocturnal period corresponds roughly with one-third of the total lifetime, near-normalization of glucose values during the night without increasing hypoglycemia risk would have a worthwhile impact on the diabetes burden.
Although our postprandial results were not exceptionally good, the use of the algorithm allows for a more predictable glucose excursion with a maximum around 90
min after meal intake. Knowing this could be useful for future control strategies.
The strategy used by our controller is different from the methods used to build an artificial pancreas most widely reported in the literature (model predictive control and proportional integral derivative). However, solutions based on rule-based approaches have also yielded improvement in glycemic control,15
as confirmed by our experience with the pRBA controller. Despite the increasing number of clinical studies to test alternative control algorithms, it is unclear whether any one system has more advantages than another.
Personalization in line with each patient's characteristics is an essential element to arrive at a successful algorithm. Model-based algorithms appear to be the most widely accepted strategy, as they consider physiologic knowledge and personal conditions by ad hoc adjustment of several parameters to each patient. Even so, setting a specific patient's model is quite a complex process. The pRBA algorithm is also adjusted to each patient by using the previously prescribed CSII therapy and personalized parameters, complemented by a generic predictive neural network trained with a population of T1DM patients, which makes generalization to a wider set of patients easier.18
It requires no previous training with each patient's specific features. The idea of predicting either hypoglycemia or hyperglycemia episodes as a safety alternative has been proposed previously.26
Our approach is not only to take advantage of glucose prediction to estimate the future rate of change in glucose levels, which was used by the pRBA as a safety constraint, but also to determine the accepted range in glucose stability.
As mentioned earlier, an unexpected degree of glucose stability was observed under CL conditions. We hypothesized that our algorithm proposing a microbolus every 5
min, mimicking the physiological pulsatile insulin secretion,27,28
could contribute to insulin action optimization and favor stability. This pulsatile administration is different than the strategy used in other CL studies,7–9,13,15,29
in which a change in the insulin delivery rate is made so insulin is distributed throughout the actuation period.
Our study has some limitations. The sample size is small, although similar in number to other previously published CL studies. However, because the entire study was carried out at the same center, the risk of heterogeneity in procedures that can occur in multicenter trials is reduced. Communication between devices was done manually and could therefore be a source of errors, but stringent measures were taken to minimize the risk of error, and no omissions or discrepancies were observed between insulin doses proposed by the algorithm and those really administered. Two CGM devices were used: one as a reference for the experiment, whereas the second was kept as backup to be used in case of failure of the reference sensor. An outpatient study should address the management of the automatic detection of sensor failure. Future outpatient studies would require CGM calibration equal to that achieved by YSI to guarantee accuracy. The period between the breakfast and the end of the trial was too short, and probably a longer period of around 5–6
h would have clarified whether the pRBA had any beneficial impact on the postprandial glycemic control.
In conclusion, this study shows that the pRBA is a new, more physiological, and highly precise controller that achieves a significant increase in overnight normoglycemia and glucose stability in patients with a previously acceptable metabolic control.