This study evaluated computer-assisted glucose control in routine practice over a 3-year period at three ICUs with different patient populations. With a reasonable glucose target level, we found that glucose control was adequate and safe, with low rates of hypoglycemia. Importantly, the protocol was feasible with six measurements per patient per day, and compliance to the protocol was high.
Safety is a major concern in contemporary health care. While the efficacy of glucose control is debated, implementation should be particularly safe [8
]. Reducing the possibility of human cognitive failure is a generic way to improve safety in health care [13
]. Prior to our study it was shown that standardization by paper protocol improves glucose control, and that computer-assisted protocols are feasible and lead to better compliance than protocols on paper [14
]. Recommendations of pump rates and measurement intervals and the notification of expiration of the interval, combined with the possibility of overruling, provide a flexible standardization of glucose control in the ICU.
Comparison with protocols from the literature is difficult because of wide variations in target glucose levels, patient populations, and metrics of control reported [9
]. We performed a literature search to find all computerized glucose protocols designed for ICU patients and tested in at least 20 patients. We summarize these findings in Table . Similar to what is reported for paper protocols, reported target ranges, patient populations, and measurement means vary greatly. However, we can conclude that compared with the studies published thus far, the present study is one of the largest to date and compares favorably in terms of occurrence of hypoglycemia and achievement of the target range. Interestingly, only one other published computer protocol [16
] aiming at achieving control with a low measurement frequency. This is remarkable, as glucose control with a high measurement frequency is associated with significant costs in terms of nursing effort and supplies [17
]. The glucose variability, which has recently been proposed as an important determinant of control [12
], was 1.22 mmol/l in our large cohort. This is lower than the figures published previously, which were 1.7 ± 1.3 in survivors and 2.3 ± 1.6 in nonsurvivors [12
Comparison with other computer-assisted glucose control protocols and intervention groups from clinical trials (BG, blood glucose)
The protocol GRIP used is published as open source, and the program can therefore be used freely by any hospital able to implement the bridge between GRIP and the hospital information system. We recognize that few hospitals have such resources, and we therefore provided a more detailed description of the algorithm in an online supplement, including typical recommendations for two example patients. Computer-based protocols published to date broadly fall into three categories: flow-chart based, empirical formula based, and model based. Most paper protocols that are directly computerized fall into the first category. GRIP uses a formula empirically derived from many protocols and expert knowledge. Model-based approaches model glucose dynamics in different compartments and try to predict future glucose levels and insulin need from that. There have been no direct comparisons of different computer algorithms to date, and more research is needed to identify the best approach.
Our study has a number of strengths. We studied a large number of patients treated in routine practice. This provides a less biased view on true performance and rate of hypoglycemia than studies of glucose protocols in a research setting. The low number of measurements required by GRIP makes the protocol less time consuming than protocols with hourly measurements. The low average number of measurements GRIP requires is mainly caused by the flexibility of the interval algorithm. GRIP performs frequent measurements when needed, but this is counterbalanced by identification of stable periods during a patient's stay (e.g., with steady glucose levels, full caloric intake and low insulin dose) during which much longer intervals can safely be recommended. Another strength is our analysis of actual compliance to the protocol. Prior publications of protocols rarely mention compliance metrics. Good performance attributed to a protocol that is violated often by nurses (by for instance measuring glucose earlier than recommended) can lead to disappointing results when strictly followed at another unit. In our analysis we find no structural violation of the protocol in terms of early and late measurements, and only late measurements were associated with more glucose values that were out of range.
Our study also has a number of limitations. Our choice of target glucose level (6.5 mmol/l) may be regarded as either low or high, as the optimal target glucose level is currently an open question [18
]. True adherence to normoglycemia would require a lower target range than we used. The number of measurements asked for by GRIP depends on the expected risk of hypoglycemia, which becomes higher when aiming for a lower target. We therefore expect that a lower target level would increase GRIP's recommended measurement frequency without a significant increase in hypoglycemic episodes, but we have not yet tested this. Clinical decision support tools may be less likely to be successful in a setting in which they were not developed [21
]. As the ICUs in our center function independently, and the development process has been limited solely to the surgical ICU, we expect that limitations of generalizability would have been uncovered in the implementation process, or would have been shown by inferior performance on the two other ICUs.
A computer program also brings benefits that we did not study, such as straightforward auditing of performance and easy incorporation of future changes in practice. For instance, when newly published evidence would favor a different target range, a quick adoption of the new preferred practice is possible without going through a learning period for nurses and physicians. In conclusion, our study shows that a computer-assisted protocol can safely and efficiently guide glucose control at intensive care units in routine practice.