Author's response to reviews Title: Design and implementation of GRIP: a computerized glucose control system at a surgical ICU Authors: Mathijs Vogelzang (m.vogelzang@thorax.umcg.nl) Felix Zijlstra (f.zijlstra@thorax.umcg.nl) Maarten WN Nijsten (m.w.n.nijsten@chir.umcg.nl) Version: 2 Date: 3 November 2005 see over Author's response to reviews:
To The Editor of BMC Medical Informatics and Decision Making Dear editor, We hereby resubmit our manuscript entitled Design and implementation of GRIP: a computerized glucose control system at a surgical ICU for possible publication as a research article in BMC Medical Informatics and Decision Making. We appreciate the opportunity to revise our manuscript, and want to thank the reviewers for giving their suggestions to improve our paper. We have addressed the comments of the reviewers point-by-point below. We would like to emphasize that the primary focus of this article is the description and initial results of the GRIP computer system, suitable for a general public of both doctors dealing with glucose control, as well as medical informaticists interested in actual implementation of clinical decision support systems and their results. We therefore have not deleted figure 1, which was judged self-evident by one of the reviewers and was requested to be left out. We leave the removal of this figure at your discretion, as we think it will make the paper easier to understand for medical informaticists who are not familiar with the daily practice of glucose control. Another issue raised by a reviewer was the cost calculation in our introduction. We have explained below why we think our numbers are justified, but we also want to leave the inclusion of this section at your discretion As this paper deals with the description and initial results of GRIP, it can be seen as the equivalent of a safety and feasibility study. The focus is less on the clinical aspects and clinical results of our population. We think the results are very interesting and need further exploration, but this would make the paper overly lengthy and will be more appropriate to readers of a more clinically oriented journal. Therefore we have included the data asked by the reviewers, but kept the discussion concise. This work has not been submitted for publication nor has it been published in whole or in part elsewhere. All authors have read and approved the manuscript, and have agreed to its submission. Yours sincerely, Mathijs Vogelzang Felix Zijlstra Maarten Nijsten Please address correspondence to: M. Vogelzang Department of cardiology, Thoraxcenter P.O. Box 30 001 9700 RB Groningen the Netherlands Telephone number +31 50 361 61 61 Fax number +31 50 361 43 91 Email m.vogelzang@thorax.umcg.nl
Referee 1 (J Hans H DeVries): General The authors are to be commended for their efforts and their intention to make their computer software available to everyone. Major Compulsory Revisions How exactly the Leuven nurses do their thing indeed remains an unsolved mystery. A transparent computerized algorithm seems to be a major step forwards. The paper would benefit from a comparison to the Leuven data with regard to mean glucose at admission, mean glycaemia achieved, hypoglycaemia rate, etc. In our discussion we added a comparison of our numbers on glucose control with the numbers given by van den Berghe et al. (our reference #5). Decreases in enteral or intravenous glucose administration as filled out by the user automatically lead to a proportional reduction of the dose recommendation. How large would such a reduction typically be? We expanded upon the mentioned sentence by describing an example. How often was pump rate kept at the maximum level of 10U/hr, while glucose remained above target? In the results, we now state the number of times this happened. Please explain in more detail how the recommendation for the time of the next glucose measurement is derived from the mentioned risk factors for hypoglycemia. We have defined more precisely how the factors influence the risk for hypoglycemia. Background, final paragraph, please leave out only We modified the manuscript accordingly. Methods, design rationale, 3rd sentence: please clarify that earlier research has shown the importance of these factors, not the current research We replaced showed by has previously shown. Fig 1 is self evident and can be left out. We agree that figure 1 is a self-evident figure and that it may be superfluous to clinicians familiar with the subject. However, we considered that a substantial part of readers of BMC Medical Informatics and Decision Making may be less familiar with the clinical aspects of the subject, and figure 1 describes the process we are trying to optimize. We therefore prefer not to remove the figure, and leave this decision to the editor. Recommendation generation algorithm, 2nd par, 1st sentence: please leave out last part. We modified the manuscript accordingly. Sentence Second, the algorithm only uses data from the most recent four hours. Consequently, it takes at most 4 hours to completely adapt to a sudden change in condition, whatever that may be. Please imagine 4 hours to be replaced by 4 minutes and see that this sentence makes no sense at all. Please remove. We have rephrased this paragraph to make our choice of the four hour interval more clear. For 66 patients (61 %) glucose values were in the target range more than 75 % of time. The mean sampling frequency was less than 6 times a day in 76 patients (70 %). This kind of sentence is hard to grasp. Please give data for the whole group of 109 patients, not for an arbitrary 66 or 76 of them. These are not data on arbitrary numbers of patients, but the numbers actually define how many patients
satisfied the criteria we mention, which is equal to saying An adverse event occurred in 23 patients (70%). We have rephrased these sentences to Glucose levels met the target range for more than three-quarters of the time in 66 patients (61 %). The glucose sampling frequency was less than 6 times a day in 76 patients (70 %). Table 2: please provide number of diabetic patients, preferably type 1 and 2 We included the number of diabetic patients in the baseline table. Table 3 heading: Table 3 - Effect on hyperglycemia. Please clarify. We replaced the heading by Glycemic control. Conclusion: please add severe before hypoglycemia We modified the manuscript accordingly. Minor Essential Revisions P3, Par 2 please reference the last sentence. P3, Par 3 efficacy rather than effectiveness, tight rather than adequate The text of Fig 4 should be in English We modified the manuscript accordingly. P3, Par 3 penultimate sentence questionable. We replaced may by might. P10 For, rather than To the nurse and A number has, not have We appreciate the suggestion of the reviewer to improve the quality of the written english in our manuscript. However, we have not made these changes as we think the correct preposition in this case is to, and following rules for subject-verb agreement a number of studies should be followed by a plural verb. Discretionary Revisions (which the author can choose to ignore) P3, Par 3 Consider leaving this paragraph out, as it doesn t add much We think this paragraph defines the relevance of tight glucose control in a broader perspective than only intensive care units, and therefore chose to keep this paragraph in the manuscript.
Referee 2 (A Michael M Albisser): General The presentation from Vogelzang and colleagues offers evidence that glycemic control in ICU patients can be simplified. Stress hyperglycemia is common in this environment and the more usual paper protocols can be replaced with a computer assist which is readily integrated into the workflow of the unit. The automation component makes recommendations of an hourly flow rate that should be set to either meet or maintain a glycemic target. A recommendation of the time interval to the next measurement is also given. Clinical results are very acceptable from a population of ICU patients over 4 months. The key findings of this work are that facilitated glycemic control without hypoglycemia can be realized in a busy ICU with just 4-6 glucose measurements a day. Staff acceptance is high. Major Compulsory Revisions none Minor Essential Revisions none Discretionary Revisions (which the author can choose to ignore) The authors should indicate the proportion of patients who were diabetic as this sub population may be different in their abilities to counter- regulate against hypoglycemia than non-diabetics. We included data on the number of diabetics in table 2.
Referee 3 (Geoff G Chase): General This is currently a very important topic in critical care medicine and endocrinology. Perspective needs to be mentioned from the outset that there is great pressure on hospitals to set into motion ICU protocols for insulin drips which are facile for nursing, prevent hypoglyce mia and achieve target glycemic status as quickly as possible with the least amount of sampling. More complex protocols implemented via computer programs or paper-based protocols developed via computer simulation, are required. Thus, sound theory and mathematical modeling, with empiric verification is needed. This protocol addresses the above concerns. The reviewer considers the article and its contents to be very important. However, there is, as detailed below, a critical lack of clinical evaluation of the results. Particularly, in the area of putting context around what was achieved that would allow others to use their results to move the field forward. The engineering aspects are very solid. Thus, I have the following general comments, which are broken out into the segments below for specific address by the authors: Major Compulsory Revisions The results are very good in terms of 79% in the target 4-7.5 range. However, performance in this environment is directly proportional (or close) to the variability and insulin resistance of the patient cohort. These factors are generally proportional to measures of critical illness such as APACHE II Score or TISS score. Thus, we see that van den Berghe et als study had median APACHE of 9 and average below 6.1, while Krinsley (see refs below) achieved 7.7 with an average APACHE II of 16.9. Similar results can be seen in the cohort studied in Hann et al (also below) with average APACHE II of 22. So: 1. what is the patient cohort in age and APACHE II or similar score? Even an estimate will do. More specifically, the more critically ill the patient the more variable their glucose dynamics (Hann et al) and the more frequently one has to measure to achieve similar levels of control. Hence, a patient cohort description is required. We have added APACHE II scores, reasons of admission, and history of diabetes of our population to table 2. We comment on these values in comparison with van den Berghe's and Krinsley's work in the discussion. 2. Krinsley's works are a major part of the data in this field and should definitely be noted. I think they would add strength to the authors results along with comment 1 above in terms of relating their performance directly to patient cohort and the work of others across that range. It would also help them delineate, in this reviewers's opinion at least, the future directions to take. Level of critical illness and resulting stress-induced insulin resistance and hyperglycaemia are noted above. A follow-on effect is that the more insulin resistant a patient becomes the less able it becomes (all else equal) to reduce glucose levels with insulin alone. Hence, overall performance can be at higher levels of critical illness, directly related to the saturation of insulin effect as noted in several references. This effect is taken into account in the paper but only obliquely and could be stressed as it is an important feature of the protocol. First, we have included Krinsley's work in the introduction and comment on it in the discussion. Second, we have mentioned the saturation effect when we describe our choice to only give continuous infusions, and when describing the maximum insulin infusion amount. 3. the authors limit insulin rate to 10U/hour which is only a bit above typical saturation levels of 5-8U/hour (Prigeon et al, among others). This might be noted and expanded upon by the authors. They should definitely include the work of Prigeon et al in the references where this is discussed at the top of page 6 with added reference to the work of Chase et al (their ref #22) which also discusses this issue. I.e. it is more than just a safety issue. We appreciate this very interesting comment of the reviewer. We have added references to both mentioned papers in the discussion. We think that insulin saturation is a very interesting phenomenon, but a further discussion of this topic would divert too much from the main focus of this paper, which is to describe the GRIP computer system. We realize that we have taken a very pragmatic approach in designing our algorithm. In the sense of the classical computer expert system, we have converted our own expertise in performing glucose
control into the algorithm used by GRIP, and have made a number of arbitrary choices in the process. Not only the maximum recommended pump rate, but also the coefficients of 0.25, 0.2, 0.3 in our pump rate formula have been arbitrary, though well-considered decisions, for which evidence in the literature may or may not exist. We have added a section discussing this to the discussion. 4. Top of page 3. The authors note and use their reference #7 to say that bedside testing costs $2-$7 per test. This figure seems extraordinarily high. The work in, for example, Ref #22 uses standard glucose test strips which cost about $0.30 per test. Hence, the savings reported for cutting 6-7 test per day would be $2/bed/day or about $8-9k per year, which still seems high. The authors must expand on this and note that simpler, much cheaper measurements could be used. For example, an embedded CGMS sensor from Minimed Medtronic would cost, in bulk, no more than $75NZ = $45US per 2-3 day lifetime. This too is much cheaper than what is reported. This should be addressed. We think a peer-reviewed analysis of costs involved in point-of-care testing has more validity than adhoc calculations and we have to disagree with the reviewer at this point. We can try to explain the difference in numbers: the difference most likely arises from the fact that reviewer only takes consumable costs into account, whereas Howanitz. et al. (our ref #7) calculate all costs associated with a point-of-care blood glucose test, including time taken by the nurse, quality control, etc. Our own experience is that taking >12 measurements a day results in a noticeable pressure on an ICU nurse, which, all other things being equal, will require a higher nurse-to-patient ratio than with a protocol taking 5 measurements a day, and Howanitz et al. indeed find that the costs of labor are much higher than that of consumables. When applying the CGMS sensor figures to our data: We had 957 patient-days in our 4-month period. Extending this to a year, we get ~3000 patientdays, requiring 1000-1500 sensors, costing $ 45,000 $ 70,000, again not including labor/training, etc. We therefore think our numbers are justified, but when this is not acceptable, we suggest to leave this section out, and leave this decision to the editor. 5. An issue which might also be addressed: As insulin does not suppress endogenous glucose production in cytokine-mediated stress, and saturation of insulin effect occurs at higher concentrations (Prigeon et al), such critically ill patients would be significantly harder to manage than less acute patients with insulin alone. What are the average levels of glucose/nutrition administered to patients and how is it typically administered? These issues also have implications regarding glucotoxicity and may also contribute to hyperglycaemia. The issue deserves mention or discussion at the minimum. We added a paragraph on nutrition policy to our methods section. REFERENCES: ------------ [1] J. S. Krinsley, "Decreased mortality of critically ill patients with the use of an intensive glycemic management protocol," Crit Care Med, vol. 31, pp. A19, 2003. [2] J. S. Krinsley, "Effect of an intensive glucose management protocol on the mortality of critically ill adult patients," Mayo Clin Proc, vol. 79, pp. 992-1000, 2004. [3] C. E. Hann, J. G. Chase, J. Lin, T. Lotz, C. V. Doran, and G. M. Shaw, "Integral-based parameter identification for long-term dynamic verification of a glucose-insulin system model," Comput Methods Programs Biomed, vol. 77, pp. 259-70, 2005. [4] R. L. Prigeon, M. E. Roder, D. Porte, Jr., and S. E. Kahn, "The effect of insulin dose on the measurement of insulin sensitivity by the minimal model technique. Evidence for saturable insulin transport in humans," J Clin Invest, vol. 97, pp. 501-507, 1996. We have added all references except #1 to our paper. (Ref #1 is not a peer reviewed article, but an abstract). Minor Essential Revisions None, the paper is well written and accurately produced. Discretionary Revisions (which the author can choose to ignore) With regard to the questions about measurement frequency and critical illness level. A measurement density versus average and/or peak glucose level obtained plot over the patients in the cohort would be extremely convincing and also illuminate the clinical results better. This is a very interesting comment. We produced this plot, with on the X-axis the Hyperglycemic
Index (the time-averaged glucose above 6 mmol/l (hgi) and on the Y-axis the mean number of measurements per day (glucaantperdag)): In our opinion, this figure does not add in a meaningful way to the paper, as our main focus is to describe the workings and feasibility of the GRIP computer system. We therefore propose to leave this figure out of the paper.