Reviewer 1 : Specific Review Responses

“Results” section: No significant difference in incidence of hypoglycemia during long-term trials was observed (0.25% vs 0.25%, p = 0.51). Percentage of blood glucose within the target range was increased by 83% compared to the retrospective cohort (54.1% vs 29.5%, p < 0.01). Results section, Paragraph 3: “...there was no significant difference in the low incidence of hypoglycemia between the long-term trials and retrospective control data at a rate of 0.25% of resampled measurements each (p = 0.51, Fisher’s Exact test).” Results section, Paragraph 4: “The percentage of BG within the target 4.0 – 7.0 mmol/L range was 83% higher for the entire long-term cohort compared to retrospective control (54.1% vs. 29.5%, p < 0.01, Chi-squared test), and BG within the wider 4.0 – 8.0 mmol/L band was 42% higher (68.7% vs. 48.5%, p < 0.01, Chi-squared test). Increased time within target ranges was consistent across patients, where the per-patient medians of BG within the 4.0-7.0 mmol/L and 4.0-8.0 mmol/L bands were consistently higher for model-based control (61.7% vs. 29.1% for 4.0 – 7.0 mmol/L band, p < 0.01, Mann-Whitney test and 80.4% vs. 47.7% for 4.0 – 8.0 mmol/L band, p <0.01, Mann-Whitney test).” 3. Care needs to be taken to provide complete descriptions of what was done. For example, the manuscript reports that 3168 hours of computer control and 1003 blood-glucose measurements were obtained using computer control but does not report how many hours of retrospective control was evaluated, or how many blood samples were used to assess hypoglycemic rates. These data are essential. The number of hours assessed for the retrospective group was presented in Table 2 as 3,571 hours and 1,091 blood glucose measurements. We have added the following statement to the Results section, Paragraph 2 to also include the number of retrospective hours available: “The control system was used for a total of 226 hours in short-term trials and 3,168 hours in long-term trials, and 3,571 hours of control were available for the retrospective cohort.” 4. Similar care needs to taken to define outcome measures and make statistically valid comparisons. Is the “incidence of hypoglycemia” defined “per day”, “per patient”, or “per blood sample” and is the number or frequency of blood samples the same in the “long-term computer control” and “slidingscale” groups. The manuscript reports that: All blood glucose results are resampled hourly to aid comparisons across studies that had different BG measurement frequencies. [Note, scientific results should be reported as past tense – the verb here should be “were”]. It is not clear which – if any – results were obtained hourly. More importantly, the authors should consider whether any bias was introduced in the comparison of hypoglycemia in the two groups by any difference that might have been present in the number or frequency of blood samples obtained. The reported blood glucose results in the original manuscript submission used linear interpolation between BG measurements. This linear interpolated line for each BG trace was then sampled hourly to provide a consistent time-basis for comparisons, and results report BG control on a per-hour-of-control basis. Thus, the percentage of BG measurements within a certain range will provide an estimate of the amount of time spent in that range. Hence, the incidence of hypoglycemia reported is an estimate of the total hours spent in a hypoglycemic range across the cohort. Such linear interpolation methods are implicit in other commonly used metrics used to assess glycemic control such as the Hyperglycemic Index (Vogelzang, Horst et al. 2004), but do have implications on the interpretation of results (Eslami, de Keizer et al. 2008). We have added the following statements to the Results section, Paragraph 2 to more fully explain the rationale behind linear interpolation and resampling: “Linear interpolation between BG measurements was used to provide hourly estimates of BG concentrations. Thus, percentages of measurements within reported ranges represent an estimate of the time spent within the specified range when comparing across datasets with different measurement frequencies.” The BG measurement frequencies between the three groups of the study were 2.0 hours/measurement for the short-term trials, 3.0 hours/measurement for the long-term trials and 3.2 hours/measurement for the retrospective data as reported in Table 2. Thus, as the retrospective group had the lowest BG measurement frequency, any individual clinical hypoglycemic measurement would represent a higher percentage of total BG measurements, and thus bias the results to over-reporting hypoglycemia in this group. The low rate of BG sampling typical in neonates suggests that many episodes of hypoglycemia go undetected (Harris, Battin et al. 2010). Additionally, the higher frequency of sampling in the computercontrolled groups (particularly the short-term control group) means it is more likely a hypoglycemic episode would have been detected clinically in these groups. Due to the differences in BG measurement rates between groups there will always be a bias whether results are reported on a per-patient, per-sample, per-day or per-hour basis. The linearly interpolated per-hour basis was chosen by the authors as a conservative choice for presenting data from the new computer-controlled method as the inherent biases are against the new method by both a) limiting over-reporting hypoglycemia in the less-measured retrospective group, and b) potentially over-reporting hypoglycemia in the more frequently measured computer controlled groups. 5. Once the actual incidence rate is established – which the reviewer expects to be very low if normalized to the number of blood samples – the authors should provide a statistical power analysis. Generally, per patient rates as low as 3-4% have been reported using other computer-based algorithms suggesting that the present study may be severely underpowered to detect changes in the overall rate. The incidence of hypoglycaemia can be expressed using several difference bases. Differences in the number of patients with a hypoglycemic measurement was not significant in long-term control vs. retrospective groups (5 patients vs. 2 patients, p > 0.40, Fisher’s Exact test). Hypoglycaemia expressed as a percentage of time was very similar at 0.25% of total control in both long-term control and retrospective groups. This study was not designed to detect any differences in hypoglycaemia rates between groups and thus cannot conclude whether or not the computer-model system affected the rate of hypoglycemia. The data presented in this study appear to indicate that time spent at hypoglycemic levels was similar between retrospective and computer-controlled group, so it is possible that the hypoglycaemic episodes in computer control episodes appeared to be brief. We have added the following data to Table 2 to provide more detail of the distribution of BG in the hypoglycaemia region: Short-term Long-term Retrospective %BG < 4.0 mmol/L [IQR] 0.0 [0.0 0.0] 4.8 [3.1 9.3] 2.8 [0.0 5.4] %BG < 3.0 mmol/L [IQR] 0.0 [0.0 0.0] 0.0 [0.0 2.0] 0.0 [0.0 0.2] %BG < 2.7 mmol/L [IQR] 0.0 [0.0 0.0] 0.0 [0.0 0.7] 0.0 [0.0 0.0] 6. The manuscript reports that Infants received most nutrition via parenteral solutions containing 1012.5% dextrose. This is rather vague. The data on which patients received nutrition via gastric tube (or regular feeding) versus IV should be available. Moreover, the algorithm is described as using nutrition data to determine insulin infusion rate. The reviewer’s concern is that many infants – particular infants that are sick – are given food which they do not manage to keep down (i.e. they throw-up). This creates problems whenever insulin is administered after the food is taken but before the carbihydartes are absorbed into the blood stream . The manuscript should make clear whether the present algorithm can be used in an infant who is receiving nutrition via the gastro-intestinal tract and if so whether insulin is delivered in advance of that nutrition being absorbed into the blood stream (in advance of the rise in glucose). If so, how does the algorithm deal with instances in which insulin is delivered and thefood purged. Foremost, the presented system is able to be used on infants that are receiving expressed breast milk (EBM) or milk-powder formula in addition to parenteral dextrose sources. The amount of milk given per day is relatively low with a per-patient median at approximately 16 mL/kg/day (as reported in Table 2 of the manuscript). The glycemic effect of the milk is assumed by the algorithm to accumulate over an hour when dosing insulin to assume a conservative glycemic rise (where some infants will not absorb the milk, and those that do absorb can do so at varying rates). The attending clinician can adjust the assumed milk intake if it is not being tolerated, and the system can recommend an adjustment to the insulin rates given this change in information. It should be noted that all insulin infusions in this study were continuous IV insulin infusions, and thus no extra insulin was given specifically for infants that received milk beyond accounting for the estimated extra glucose uptake. The kinetics of IV insulin are relatively swift compared to subcutaneous routes, and adjustments to decrease insulin usage if the infant did not appear to be tolerating feeding would expect to take effect quickly. We have added the following information to Table 2 to show the proportion of dextrose that was fed via the enteral route, and the number of patients that had enteral feedings: Short-term Long-term Retrospective Patients that received EBM 6 (75%) 19 (86%) 18 (86%) Proportion of dextrose via EBM (%) 1.2 [0.3 1.7] 11.0 [1.5 21.4] 5.5 [1.9 7.2] 7. A more complete description of the algorithm should be provided together with some discussion about the relative merits of combining the insulin infusion with glucose infusions to prevent or correct hypoglycemia. Recognizing that the manuscript is already above 3000 words this added material could be accompanied by a reduction in the length of the Discussion, which includes topics strictly related to the study’s objective or results (e.g., the discussion as to what the appropriate target is). We have added the following paragraph to the Methods section after Paragraph 5 to provide a brief overview of the control system. It should be noted that the control system is a modelbased and adaptive, using a data-driven stochastic model to forecast parameter values, and thus there does not exist any closed-from equation to represent the method of insulin rate selection: “The computer system uses entered information on insulin rates, dextrose input rates from IV and enteral sources and prior blood glucose concentrations to determine the insulin sensitivity level of the patient [30]. This insulin sensitivity parameter represents the level of glycemic response the infant has been exhibiting to exogenous insulin over the last few hours. Thus, lower observed insulin sensitivity would result in recommendations of relatively higher insulin rates and vice-versa. A stochastic model is used to determine the potential changes in sensitivity to insulin in the upcoming hours based on observed changes in sensitivity to insulin in retrospective data [24]. The current BG, rates of dextrose inputs, level of sensitivity to insulin and forecasted changes in sensitivity to insulin are used by the computer model to select an insulin rate that balances the goals of achieving BG within the target range whilst limiting the potential for hypoglycemia [25, 32]. Incorporating information about nutrition inputs allows insulin dosages to be scaled accordingly to provide control for infants receiving higher and lower amounts of calories.” It should be noted that this system does not specifically combine/mix insulin and glucose infusions together, as suggested by the reviewer. Each component is controlled separately by different pumps and/or feeding mechanisms. Mixing insulin and glucose infusions together would deliver a fixed ratio of insulin to administered dextrose. The work presented in the paper suggests that the sensitivity of infants to insulin varies greatly and sometimes quickly, as highlighted in Figure 2 and the insulin sensitivity metrics in Table 1, and that the ratio of insulin/glucose administration would need to be adjusted accordingly. 8. The authors should discuss the use of blood glucose values obtained from different sources (arterial line if present; capillary otherwise). Did the authors conduct preliminary tests to ensure there was no statistical difference in these measures or quantify the variance. Both arterial and capillary sampling was used during the trials. The breakdown for the sampling methods are: Arterial samples Capillary samples Retrospective 45% 55% Long-term 27% 73% Short-term 100% 0% It is generally accepted in this unit that arterial and capillary glucose levels are very similar (to within about ~0.1-0.2 mmol/L, (Liu, Moberg et al. 1992)). Virtually all trial patients were measured using the blood gas machine, together with the vast majority of retrospective samples, with use of other devices only occurring during rare technical issues with the blood-gas analyser. We have added the following line to Table 2 stating the measurement site breakdown: Short-term Long-term Retrospective BG sampling site: arterial/capillary 100%/0% 27%/73% 45%/55% We have also added the following statement to Results, Paragraph 2 to call out the sampling site data in Table 2: “A mixture of arterial and capillary BG samples were used in the long-term and retrospective data sets.” Quantifying the variance as suggested between arterial vs capillary samples in these patients would be difficult to interpret particularly as the use of an arterial line is not random (eg: baby with arterial line is often younger, and/or sicker) so it would not be possible to determine whether any differences in variance between the two sampling methods are due to the baby being in a different condition, or whether there was an actual difference between the two sampling methods. Also, with large sample numbers a small difference (eg: 0.1 mmol/L) would be statistically significant, but potentially not clinically relevant. Minor Essential Revisions 1. There are numerous typing and grammatical errors. The aforementioned use of present tense where past tense is called for (see item 4 above). We have modified the manuscript to present all results in past tense as requested by the reviewer. 2. The sentence This study presents the first trials of model... trials should read trial. We have modified the specified sentence in the Introduction to remove the plural. 3. “Moribund” and “not expected to survive” man the same thing. The sentence Infants who were moribund or not expected to survive were excluded is redundant. We have modified the specified sentence in the Methods to read: “Infants who were not expected to survive were excluded.” 4. The sentence including ...and BG within the wider 4.0 – 8.0 mmol/L band was 42% higher (p < Chisquared test) should report a p-value. We have now included a p-value for this result and the sentence now reads: “...and BG within the wider 4.0 – 8.0 mmol/L band was 42% higher (68.7% vs. 48.5%, p < 0.01, Chi-squared test).” 5. The sentence including ...with the 4.0-7.0 mmol/L and 4.0-8.0 mmol/L bands was consistently higher for model-based control (p < 0.01 for both banes, Mann-Whitney test), what are “banes”? We have corrected the typo and replaced “banes” with “bands”. 6. A careful proofreading is required. We have reviewed the tense and grammar aspects of the manuscript and made modifications as necessary throughout the document. Discretionary Revisions 1. The authors should consider reporting that a pilot study on 8 subjects was performed in advance of implemented the algorithm to assess safety of 24 hours but not report the actual results. These 8 subjects are not treated sufficiently like the subsequent 22 to combined with that group, and do not constitute a sufficiently large enough group to draw independent conclusions on. Excluding the data would allow more room to describe both the sliding scale algorithm and the computer algorithm, and explore the results obtained once the computer algorithm was actually implemented. The authors agree that the short-term study data represents a unique cohort and the blood sampling frequency was not typical of routine neonatal care. However, we also feel that such densely sampled data is rare in neonatal publications and represents an interesting insight into neonatal response to insulin, particularly in quantifying inter-patient variation in sensitivity to insulin. Thus, the authors also feel that the inclusion of the short-term study provides a key justification for using adaptive control we used in long-term study patients as it clearly demonstrated the wide variation in response to insulin observed during intensive monitoring. We have added the following statement to Discussion, Paragraph 2, to highlight the insulin sensitivity variation data that is a feature of the short-term trials: “Additionally, the inter-patient variation in sensitivity to insulin observed in the shortterm trials, presented in Figure 2, highlight this wide range of response between individuals.” 2. The author’s should also consider presenting the results in Figure 2 as a daily average with 95% CI. This could be obtained by first averaging al the blood values obtained for a given subject on a given day into one number , and then averaging across subjects (N=22 or 21 with N decreasing as time goes on). This would allow differences between protocols to be more easily identified, clarify how long each algorithm takes to achieve target etc. We have modified Figure 2 to present the time course of BG and insulin sensitivity as boxplots that show daily median, inter-quartile range and outliers. The authors feel that this modified presentation of the data makes trends much clearer for this long-term data. For the short-term patients we have chosen to keep the individual time-traces as the smaller data set of this study means the data is still readable in this form, and highlights the wide insulin sensitivity variation and relatively uniform BG response. 3. The present study does not provide any outcome data on the underlying causes of hyperglycemia, it relationship to co-morbidities, whether tight glycemic control should be done and if so what the target should be etc. Thus, space devoted in both the Introduction and Discussion to these items might better be devoted to introductory comments highlighting the underlying objective of the present study and a discussion of the results showed. To accommodate the additions to the manuscript as a result of the revision process, we have removed the first sentence of the final Discussion paragraph and the final sentence of the first paragraph of the Introduction provide more focus on the main aspects of this study.