In this report we present GRIP, a computer decision support system for glucose control by intensive insulin therapy. This system was successfully implemented at a surgical ICU, and was found to provide safe and efficient glycemic control. User acceptance, an important hurdle for successful implementation of a clinical decision support system, was excellent with minimal training. Nurses clearly rated GRIP as an improvement over the conventional sliding scale protocol. GRIP will be released under an open source license and will thus be free to use and improve by anyone in the future. We think especially GRIP's independence of a patient data management system (PDMS) is a strong point that makes it widely usable. At ICUs where a PDMS is already operational, GRIP might introduce double data entry. However, this can be easily resolved by either letting GRIP query the PDMS for information or by converting GRIP's advice module into a plug-in of the PDMS.
The efficiency and safety of control achieved by GRIP's current relatively simple algorithm was satisfactory. In the Leuven study, a mean glucose level of 5.7 mmol/L was achieved in a cohort of patients with a median APACHE II score of 9 [5
]. Krinsley achieved a mean of 7.3 mmol/L in a sicker cohort of patients (median APACHE of 15) [6
]. Our group had APACHE scores similar to Krinsley's, and a mean glucose level that was 0.4 mmol/L lower. With regard to hypoglycemia, in the Leuven study 39 out of 765 patients (5.1 %) had one or more values lower than 2.2 mmol/L, compared to 1 out of 179 (0.6 %) in our group. Krinsley describes the number of hypoglycemic episodes as the proportion of total number of glucose measurements, instead of as the proportion of patients at risk. In Krinsley's study, 0.34 % of measurements were lower than 2.2 mmol/L, compared to 0.02 % in our group.
During the design of the current recommendation algorithm a number of arbitrary decisions were taken. Based on experience, we chose to include glucose value and glucose difference as the two most important determinants of advised pump rate, and we chose to limit pump rate to 10 units per hour. Although evidence for the rationality of the latter decision exists [16
], most decisions were taken purely based on prior experience in controlling glucose levels. Our results show that this algorithm gives satisfactory results, but we are aware that GRIP has the potential for substantial improvements in the future, as our choices are unlikely to be the most optimal. For example, the current algorithm is very cautious and only takes small steps when increasing the insulin pump rate, sacrificing rapid control in favor of safety. Analysis of the data that have been collected thus far may reveal situations in which the algorithm can safely take larger steps to achieve more timely glucose control. It will be interesting to analyze differences between groups of patients with respect to factors such as reason of admission, presence of diabetes, and concomitant drug use. This will allow glucose control to transition from the current general "one size fits all" approach, which uses the same advice for every patient, to a more tailor-made approach, which makes use of as much information as possible to make its advice fit an individual patient's needs. As the advice generation module can be changed without affecting the user interface, algorithm improvements can be implemented without additional training of the nurses.
Prior studies have evaluated different types of computer systems for glucose control in critically ill patients. A number of studies have evaluated feasibility of continuous glucose sensors, which in theory provide ultimate efficiency and safety of control. However, apart from the obvious extra costs, continuous sensors have been found to lack reliability, requiring frequent replacements of the sensor, and report inaccurate data in the hypoglycemic range [22
]. We believe that for the foreseeable future, the vast majority of glucose control schemes will still employ sequential discontinuous measurements. Chase et al. have worked on developing models for accurate insulin dose prediction in critically ill patients, but with 2 measurements per hour, measurement frequency was much higher than GRIP's [19
]. Rood and colleagues have previously investigated computerizing guidelines for glucose control and found that glucose measurements were taken at the prescribed time more often, and glucose control improved in comparison with the preceding paper protocol [25
]. Unfortunately, a PDMS plug-in with an undisclosed 4-page flowchart was used, limiting widespread usability, and in our opinion a formula-based approach as taken by GRIP will be easier to optimize and customize to different patient groups than a flowchart-based set of rules. Naturally, in case the latter approach may prove to yield better control, GRIP's advice generation module can be easily altered to follow that approach.
We are currently in the process of starting implementation of GRIP at other ICUs. Furthermore, we are implementing GRIP in the coronary care unit. We consider a personalized algorithm to be the key to more efficient and safe glucose control in acute coronary care.