Original article
The control of blood glucose in the critical diabetic patient: a neuro-fuzzy method

https://doi.org/10.1016/S1056-8727(00)00137-9Get rights and content

Abstract

Conventional algorithms for regulating insulin infusion rates in those critical diabetic patients submitted to parenteral glucose and insulin infusions do not allow to approach near normal blood glucose (BG) levels since traditional control systems are not fully effective in complex nonlinear systems as BG control is. Thus, we applied fuzzy logic principles and neural network techniques to modify intravenous insulin administration rates during glucose infusion. Forty critically ill, fasted diabetic subjects submitted to glucose and potassium infusion entered the study. They were randomly assigned to two treatment regimes: in group A, insulin infusion rates were adjusted, every 4 h at any step between −1.5 and +1.5 U/h, according to a neuro-fuzzy nomogram; in control group B, insulin infusion rates were modified according to a conventional algorithm. In group A, BG was lowered below 10 mmol/l faster than in group B (8.2±0.7 vs. 13±1.8 h, P<.02). Mean BG was 7.8±0.2 in group A and 10.6±0.3 mmol/l in group B (P<.00001). BG values below 4.4 mmol/l were: A=5.8% and B=10.2%. BG values lower than 2.5 mmol/l had never been observed. In conclusion, the neuro-fuzzy control system is effective in improving the BG control in critical diabetic patients without increasing either the number of BG determinations or the risk of hypoglycemia.

Introduction

When a critical patient cannot meet his energy requirements, the parenteral administration of nutrients is needed. When the same patient is also diabetic, a good glycemic control is mandatory to avoid energy waste and optimize insulin treatment. The perioperative period is paradigmatic of such clinical situations and several algorithms to adjust insulin infusion rates have been proposed Alberti & Thomas, 1979, Hirsch et al., 1991, Pezzarossa et al., 1988, Watts et al., 1987. These algorithms operate through a feedback regulation Albisser, 1982, Westenskow, 1997 that follows the logic of an open-loop control system and adjust insulin administration rates according to a time schedule of blood glucose (BG) measurements. Since these conventional algorithms do not allow to reach and maintain near normal BG levels without increasing the frequency of BG measurements or the risk of hyper- or hypoglycemic events, target BG values are higher than desirable (between 8 and 11 mmol/l). Although, during the perioperative period, a near optimal BG levels have been advised, there is no clear-cut evidence to substantiate this approach in cases where the patient is allowed to be fed soon after surgery (Shade, 1988). Obviously, when the patient must be submitted to continuous assisted enteral or parenteral nutrition, BG should be maintained at near normal values to allow an efficient utilization of administered nutrients. To improve glycemic control, without increasing the number of BG determinations or the risk of hyper- or hypoglycemic events, we considered the last two BG determinations instead of the last one.

In this way, taking into account the degree and the direction of the glycemic variation, we could more accurately adjust the rate of insulin infusion. The many variables that characterize BG control in a critical diabetic patient make it very difficult to obtain a satisfying glycemic control. Since a mathematical model to manage such a complex system is not available, we appealed to fuzzy logic principles. Fuzzy logic, already used in the medical field Kosko, 1992, Kosko, 1993, allows to capture valuable information that could be lost if considered as distinctly categorized. A neural network (a software that simulates human neural processing capabilities) has been used to apply fuzzy logic theoretical principles and to extrapolate (based on previous knowledge) rules to be used to elaborate a nomogram for insulin infusion rates (neuro-fuzzy control system).

Section snippets

Fuzzy logic modelling

Glucose control can be conceived in terms of engineering control theory. Blood glucose level, the variable to be controlled, is the output of the controlled system, while the insulin infusion rate is the input to be adjusted to reach the desired BG value. In a conventional control system, observed and desired outputs are compared. The input is modified to reach the desired output according to a mathematical model that defines the relationships between the input and the output of the system.

Blood glucose

Mean starting BG levels were similar in the two groups (A: 15.1±0.8 vs. B: 15.8±0.8 mmol/l; P: ns) Fig. 2, Fig. 3. Subsequently, mean BG levels resulted lower when insulin infusion rates were adjusted with the neuro-fuzzy nomogram either (1) during the overall observation time (7.6.1±0.2 vs. 10.1±0.2 mmol/l (P<.00001); or (2) at the single time points (4 h: 11.1±0.6 vs. 14.8±0.8 mmol/l, P<.002; 8 h: 8.4±0.5 vs. 12.4±0.7 mmol/l, P<.001, 12 h: 7.0±0.5 vs. 10.1±0.6 mmol/l, P<.002; 16 h: 6.8±0.4

Sensitivity and effectiveness

The association between changes in BG values (Gt(n+1)Gt(n)) and the consequent variations in insulin infusion rates (It(n+1)It(n)) gives an evaluation of the sensitivity of control systems and results stronger in group A (r2=.60, P<.00001) than in group B (r2=.11, P<.00001). The association between the variations in insulin infusion rates (It(n+1)It(n)) and the consequent variations of BG values (Gt(n+2)Gt(n+1)) gives an estimate of the effectiveness of the control system and results

Discussion

When a critical diabetic patient is not permitted to eat and glucose and/or other nutrients are intravenously given, intravenous infusion of insulin is the best way to administer insulin Dunnet et al., 1988, Eldridge & Sear, 1996. A diabetic patient submitted to intravenous insulin infusion needs a continuous BG monitoring to control BG levels. Thus, in the past years, some algorithms that work with relatively few BG determinations were implemented to obtain an acceptable compromise between the

Acknowledgements

This research has been supported by 60% funds of MURST. The authors are indebted to Mr. Scott Hartman for his skillful revision of the manuscript.

References (28)

  • M.R Guerriere et al.

    Neural networks: what are they?

    Annals of Internal Medicine

    (1991)
  • I.B Hirsch et al.

    Perioperative management of surgical patients with diabetes mellitus

    Anesthesiology

    (1991)
  • I.B Hirsh et al.

    Diabetes management in special situations

    Endocrinology and Metabolism Clinics of North America

    (1997)
  • S.J Jacober et al.

    An update on perioperative management of diabetes

    Archives of Internal Medicine

    (1999)
  • Cited by (57)

    • Artificial intelligence applications in pharmaceutical technology

      2023, Computer-Aided Applications in Pharmaceutical Technology: Delivery Systems, Dosage Forms, and Pharmaceutical Unit Operations
    • Nephropathy forecasting in diabetic patients using a GA-based type-2 fuzzy regression model

      2017, Biocybernetics and Biomedical Engineering
      Citation Excerpt :

      In other words, cases within the neighborhood of the borderline indicate a vague status regarding the disease [3]. For these reasons, fuzzy models have been used in medical and especially diabetes research [4–10]. Some researchers have focused on diabetic complications with fuzzy models.

    • Tissue profiling by nanogold-mediated mass spectrometry and artificial neural networks in the mouse model of human primary hyperoxaluria 1

      2014, Journal of Applied Biomedicine
      Citation Excerpt :

      Generally, a structure of the ANN consists of an input layer, one or more hidden layers and output layer and its architecture reflects complexity of the analyzed system. Application of ANNs ranges from chemical kinetics (Amato et al., 2012) or biomolecular engineering and structural computing (Armano et al., 2005; Fawcett et al., 2013) to medical diagnosis e.g. in diabetes, cardiovascular diseases, epilepsy or cancer (Arnold, 1996; Dazzi et al., 2001; Yan et al., 2008; Elveren and Yumuşak, 2011; Chan et al., 2011; Barwad et al., 2012; Bradley, 2012; Spelt et al., 2012). Application of ANNs in medical diagnosis was recently reviewed by Amato et al. (2013).

    • Artificial neural networks in medical diagnosis

      2013, Journal of Applied Biomedicine
    • Integration of type-2 fuzzy clustering and wavelet transform in a neural network based ECG classifier

      2011, Expert Systems with Applications
      Citation Excerpt :

      In our previous works, we had proved that neural network and fuzzy system with feature extraction methods had better performances than the traditional clustering and statistical methods (Ceylan & Özbay, 2007; Özbay et al., 2006). Type-1 fuzzy c-means clustering in Jang, Sun, and Mizutani (1997), Liao, Celmins, and Hammell (2003), Fan, Zhen, and Xie (2003), Li, Mukaidono, and Turksen (2002), Castellano and Fanelli (2000a, 2000b), Dazzi et al. (2001), De, Basak, and Pal (2002), Meesad and Yen (2000) and Zarandi, Türkşen, and Kasbi (2006) was used to choose the best patterns belong to same class in implemented systems in our previous works. But, there are (at least) four sources of uncertainties in type-1 fuzzy logic systems (FLSs): (1) The meanings of the words that are used in the antecedents and consequents of rules can be uncertain (words mean different things to different people); (2) Consequents may have a histogram of values associated with them, especially when knowledge is extracted from a group of experts who do not all agree; (3) Measurements that activate a type-1 FLS may be noisy and therefore uncertain; (4) The data that are used to tune the parameters of a type-1 FLS may also be noisy (Karnik & Mendel, 2001; Mendel, 2000; Mendel & John, 2002; Rhee, 2007; Zarandi et al., 2006).

    View all citing articles on Scopus
    View full text