- Email: [email protected]
Use of the DIAS model to predict unrecognised hypoglycaemia in patients with insulin-dependent diabetes
ELSEVIER
Computer MethodsandPrograms in Biomedicine 50(1996)241-246
Use of the DIAS model to predict unrecognised hypoglycaemia in patients with insulin-dependent diabetes D.A. Cavan”“, R. Hovorkab, O.KL. Hejlesenc, S. Andreassenc, P.H. Siinksen” “Department of Endocrinology, St Thomas’ Hospital, London, UK bDepartment of Systems Science, City University, London, UK ‘Department of Medical Informatics, Aalborg University, Aalborg, Denmark
Abstract The Diabetes Advisory System (DIAS) is a model of human glucose metabolismimplementedin a causal probabilistic network. It handlesdata on insulin dose,carbohydrate intake and blood glucoseconcentration to predict hourly blood glucoseconcentrationsand thus provide an indication of blood glucosevalues betweenhome blood tests. DIAS was used to predict blood glucoseprofiles in eight patients with well-controlled insulin-dependent diabetes, who are at increasedrisk of hypoglyca.emia(abnormally low blood glucoselevels). DIAS predicted nocturral hypoglycaemia in six patients and daytime hypoglycaemiain one patient. The occurrenceof nocturnal hypoglycaemiawas not recognisedby the patient or suspectedby their doctor but was subsequentlyconfirmed by blood testing in five patients. It is known that unrecognisednocturnal hypoglycaemiais common in patients with well-controlled diabetes.The ability of DIAS to identify such periods of hypoglycaemiawith reasonableaccuracy illustrateshow the advancedtechnology it employsmay provide reliable decisionsupport to clinicians. Keywords:
Diabetes;Insulin; Computer model
1. Introduction
Insulin-dependent diabetes (IDDM) is a disease in which autoimmune destruction of insulin-secreting cells in the pancreas leads to lifelong dependency on exogenous insulin for survival. The main burden of the disease on both the patient * Corresponding author. Medical Unit, St Thomas’ Hospital, London SE1 7EH, UK. Tel: 0171 9289292; fax: 0171 9288289; e-mail: [email protected]
and national resources is attributable to the complications of the disease. These include blindness, renal failure, amputations, heart disease and stroke. The Diabetes Control and Complications Trial (DCCT) has established a clear correlation between levels of blood glucose (as reflected in glycosylated haemoglobin measurements) and development of complications, such that any improvement in blood glucose concentration towards normality carries a reduced risk [l]. The advent of home blood glucose monitoring has greatly improved the ability of patients to achieve good diabetic control. The main draw-
0169-2607/96/$15.00 0 1996 Elsevier Science Ireland Ltd. All rights reserved PZZ 0169-2607(96)01753-l
242
D.A.
Cavan
et al. i Computer
Methods
and Progrnms
back of good diabetic control, clearly demonstrated in the DCCT trial, is an increased frequency of hypoglycaemia, when blood glucose concentration falls to abnormally low levels. In its mildest form, hypoglycaemia manifests as sweating but more profound hypoglycaemia may lead to coma and rarely death. For many patients it is the most feared aspect of their condition and this fear may conflict with attempts at achieving good diabetic control. This fear is particularly pronounced in those patients who have lost the ability to recognise impending hypoglycaemia. Such ‘hypoglycaemic unawareness’ is more common in long-standing diabetes and while its precise cause is unknown, recent research has shown that avoidance of hypoglycaemia may restore awareness of any future episode in some individuals [2]. Thus the challenge in managing patients with IDDM is to maintain near normal blood glucose concentrations while minimising the risk of hypog1ycaemi.a. Experience to date suggests that this is best achieved by intensive blood glucose monitosing, regular carbohydrate intake and adjustment of insulin dose to meet changing requirements. Despite this, the increased risk of hypoglycaemia in patients with well-controlled diabetes and the loss of warning symptoms associated with recurrent hypoglycaemia is potentially hazardous; reduced alertness caused by hypoglycaemia while driving, for example, may have tragic consequences. The Diabetes Advisory System (DIAS) is a model of human glucose metabolism implemented in a causal probabilistic network (CPN). This gives it the ability to handle uncertainty, for example, in blood glucose measurements or physiological variations in glucose metabolism. The model has elements describing the blood glucose concentration and the amount of carbohydrate in the gut as functions of hepatic glucose output, carbohydrate uptake from the gut, insulin-independent glucose uptake (e.g., central nervous system), insulin-dependent glucose uptake (e.g., muscles), loss of carbohydrate from renal excretion and carbohydrate intake. DIAS includes two parameters, insulin sensitivity and time-to-peak of NPH (long acting) insulin absorption, to represent individual characteristics
in Biomedicine
50 (1996)
241-246
‘of glucose metabolism and insulin absorption from a subcutaneous depot. Standard modelling Imethodology offered by the CPN approach is employed as implemented in the HUGIN package 1131.Glucose metabolism is represented using a discrete-space (finite number of states for each variable), discrete-time (finite numlber of time instances), stochastic model. The relationships between variables are specified using conditional probability tables. Given a set of measurements, e.g., blood glucose concentrations,, carbohydrate intake and insulin doses, a posteriori (marginal) probability distributions for the remaining model variables are calculated using the 13ayes theorem. Thus, unlike in the deterministic modelling approach, unknowns are estimated as probability distributions rather than as point-estimates. The precision of parameter estimates and subsequently precision of any predictions are inherently obtained. DIAS can be run in three modes, the learning mode, the prediction mode, and the advisory mlode. In the learning mode, relevant data (blood glucose measurements, food intake represented as grams of carbohydrate, and insulin doses) from a single day are entered (instantiated) into the model and the joint probability distribution for the two parameters is calculated. This procedure is repeated for up to four days giving four daily estimates of the joint distribution. Assuming a bivariate normal form of the joint distribution, the means and the standard deviations of the two parameters are estimated 141. This hierarchical modelling approach is similar to that adopted in ph.armacokinetics [5]. In the prediction mode, the probability distribution of parameters calculated in the learning mode is used to make predictions of blood glucose given the carbohydrate intake and insulin regimen of the subject. It is this mode which is the focus of the present paper as the ability of DIAS to predict nocturnal hypoglycaemia was examined. In the advisory mode, a utility measure is minimised to find therapy which gives the least overall (predicted) risk of too low or too high blood glucose concentration [6]. DKAS was not operated in this mode for the purposes of the present study.
D.A. &van et al. / Computer Methods and Pr#ograms in Biomedicitze 50 (1996) 241-246
243
Table 1 Patient characteristics NO.
2 3 4 5 6 7 8 Mean S.D.
Sex
F F M F M M M F
Age (years)
Duration of diabetes (yr)
Weight (1%)
Body mass index (wt/ht*)
HbA, (“/) (ref 5.4-7.6X)
Insulin dose Way)
42 42 24 33 26 36 67 30 37.5 13.6
21 36 15 24 2 13 3s 14 20 11.5
73.3 68.2 80.6 56.7 82.2 74.2 70.8 70 12 7.91
26.3 24.7 27.6 24.9 21.6 23.2 23.1 27.3 24.8 2.14
9.6 7.8 9.0 9.9 8.4 8.8 7.3 7.4 8.5 1.0
42 44 56 37 54 76 48 56 52 12
The aim of this study was to determine whether DIAS can identify periods of unrecognised hypoglycaernia in subjects with well-controlled dliabetes, and whether of any predictions hypogl,ycaemia could be verified by blood testing. We therefore asked eight patients with insulin-dependent diabetes to collect data for four days to determine whether the DIAS simulations of their glucose profiles suggested unrecognised hypoglycaemia.. Patients in whom hypoglycaemia was SUSpetted underwent further study during which they measured blood glucose concentrations at the times hypoglycaemia was predicted. This paper reports the predictions of hypoglycaemia by DIAS and their subsequent confirmation in these patients.
2. Methods Eight patients with insulin-dependent diabetes mellitus (IDDM) were recruited from the diabetic clinic at St Thomas’ Hospital, London. Patient characteristics are shown in Table 1. The inclusion criteria were HbA, < 10% (representing good glycaemic control, reference range 5.47.6%) insulin-dependent from the time of diagnosis, and ability and willingness to perform home blood glucose measurements using a meter. Patients with serious concomitant illness, anorexia or other eating disorder or psychiatric illness were excluded. On recruitment they were asked to col-
lect the following data for four consecutive weekdays: four blood glucose measurements per day (before meals and before bed) using BM strips and a Reflolux meter (Boehringer Mannheim, Mannheim, Germany) and full details of insulin injections (time, type and dose). They were asked to record any hypoglycaemic attacks. Patients were seen by a dietitian and asked to keep a (detailed diary of all food intake during this period. Patients were seen within a week following the data collection period. Their records were checked for completeness and the food diary reviewed by a dietitian who assessed the carbohydrate content of each meal. Recorded blood glucose values, insulin injections and carbohydrate intake with times were entered into DIAS and a simulation of the blood glucose profile over the 4-day study period was generated. Patients for whom the model predicted hypoglycaemia (blood glucose below 3 mmol/l) which was not documented by blood glucose measurernent were asked to record four daily blood tests for a further four day period and, in addition, to test their blood at the time hypoglycaemia was predicted. In order to avoid introducing expectation of hypoglycaemia, patients were not advised of the reason for the tests at these times, but only that they would assist in verifying the model. Clinic computer records of the patients were searched for evidence of previous problems with hypoglycaemia.
244
D.A.
Cauan
rt al. I: Computer
Methods
and Programs
ill Biomedicine
50 (1996)
241-246
B meal-data g 2501 200. 150. 100. 50. O0 ins-A-data 1 it-M&data il 3224. 16. 8O-
1
8
12
16
20
24
4
8
fl bg-data f bg-sim mmol/l 16. 12. 84OIUI IO 95
Tue Jut 11 95
Fig. I. Example of simulated blood glucose profile in a patient showing episodes of nocturnal hypoglycaemia. The simulation covers :wo days (July IO-1 1). Meals expressed as grams of carbohydrate (meal-data) are shown in the upper graph. The lower panel shows insulin doses (soluble insulin; ins-A-data) as open bars and isophane (NPH) insulin (ins-N-data) as black bars. Measured blood glucose values (bg-data) are shown as black squares. The sinulated blood glucose profile (bg-sim) suggests nocturnal hypoglycaemia at around 0200 h on the 11th. Note the simulation suggests further periods of hypoglycaemia before lunch on the 10th and at 1200 h and 1800 h on the I Ith, although measured blood glucose values range from 5 to 12 mmol/l. 3. Results
DIAS predicted unrecognised hypoglycaemia in seven of the eight patients. In one (subject 8), hypoglycaemia was predicted at around midday and documented by home blood tests on one occasion during the initial study period, although no symptoms were reported. In six patients, hypoglycaemia was predicted at night, at times ranging between 0200 h and 0500 h. None had been aware of nocturnal hypoglycaemia and only four had reported any hypoglycaemia during the study period. Seben patients, however, recorded blood glucose measurements below 3 mmol/l. In only one patient was no hypoglycaemia predicted by DIAS or recorded by the patient. Study of previous records showed that four patients in whom recurrent hypoglycaemia was predicted had reported loss of symptoms and one had reported nocturnal hypoglycaemia. Fig. 1 shqws a DIAS simulation showing nocturnal hypoglycaemia and all data regarding reported and unrecognised hypoglycaemia are summarised in Table 2.
The six patients in whom nocturnal hypoglycaemia was predicted by DIAS proceeded to further study. When asked to set an alarm clock and measure their blood glucose concentration at the exact time hypoglycaemia was predicted by DIAS, all did so on at least one night. In five of the six patients, these tests confirmed hypoglycaemia on at least one occasion. Table 3 shows the mean bedtime and fasting blood glucose values for these patients during the initial study and the time at which hypoglycaemia was predicted. The number of night-time tests performed is shown together with their values.
4. Discussion
This paper describes the first clinical use of advanced technology as a management tool in insulin-dependent diabetes, The DIAS model of human glucose metabolism, implemented in causal probabilistic networks, was used to identify
D.A.
Cavan
et al. / Computer
Methods
and Programs
in Biomedicine
50 (1996)
241-246
245
Table 2 Relationship between reported hypoglycaemia, unrecognised hypoglycaemia and history of hypoglycaemia unawareness or nocturnal hypogfycaemia in each subject Patient
No. of reported hypoglycaemic episodes over 4 days
Unrecognised hypoglycaemia predicted by DIAS
Previously reported hypoglycaemia unawareness (HU) or nocturnal hypoglycaemia (NH)
I
3 0 4 4 0 1 0 0
Yes Yes Yes Yes Yes Yes No Yes
HU NH Nil Nil Nil HU
2 3 4 5 6 7 8
periods of unrecognised hypoglycaemia in patients with well-controlled insulin-dependent diabetes. The patients studied were a small sample with good overall control and thus not representative of the clinic population as a whole. A high frequency of hypoglycaemia was not unexpected: the DCCT trial suggests that such patients are prone to hypoglycaemia [l] and three had experienced hypoglycaemia of some degree in the period leading up to the study. The significance of this study is the ability of DIAS to identify periods of hypoglycaemia, which were unrecognised by the patient, yet which were subsequently confirmed in all but one of the patients studied. While previous studies have demonstrated high frequencies of nocturnal hypoglycaemia in patients with insulin-dependent diabetes, hourly measurement of blood glucose concentration in a hospital setting has usually been employed in such studies [7,8]. A number of our patients had previously reported loss of awareness of hypoglycaemia. This may result from recurrent hypoglycaemia [2] and may further exacerbate the difficulties in detecting it in an outpatient setting. This is illustrated in the case of patient 5 who had previously been asked by a specialist nurse to test his blood glucose concentration at 0300 h to check for hypoglycaemia but none was detected. The ability of DIAS to suggest a time, specific for each patient, when hypoglycaemia is likely to occur, may increase the likelihood of its being detected.
It should be emphasised that DIAS provides blood glucose simulations which are applicable only for the days for which data have been collected and it should not be assumed that hypoglycaemia would always occur at the time predicted. Day to day changes in levels of physical activity and food intake will affect blood glucose concentrations and may explain the differences between patients in the frequency with ,which hypoglycaemia was confirmed, from one (patient 5) to Ihree (patient 3) out of four measurements being under 3 mmol/l. Such factors may also explain why the prediction of hypoglycaemia could not be confirmed in patient 2. In summary, this small study concurs with previous findings that hypoglycaemia; is common in patients with well-controlled insulin-dependent diabetes [7,8] and suggests that such hypoglycaemia may be difficult to identify clinically. These preliminary data suggest that DIAS may aid detection of hypoglycaemia. DIAS is a prototype which requires much further developmental work and clinical validation; confirmation of these findings in larger studies may, however, suggest a role for DIAS as a useful clinical tool for outpatients. Fifteen years ago, the management of patients with insulin-dependent diabetes was revolutionised by the introduction of home blood glucose monitoring, which for the first time gave an indication of blood glucose levels between clinic visits [9]. Now, the advanced technology incorporated in DIAS may provide an indication of blood
246
D.A.
Chvan
et al. / Computer
Methods
and Programs
in Biomedicine
50 (1996)
241-246
Table 3 Mean early morning and late evening blood glucose measurements of the six patients in whom unrecognised nocturnal hypoglycaemia was predicted, together with the predicted time of hypoglyca.emia Patient
Late evening blood glucose (mmol/l )
1 2 3 4 5 6
7.6 7.8 11.3 4.3 6.8 5.3
Early morning blood glucose (mmol/l)
6.1
6.1 5.2 0.3 8.4 1.6
Predicted time of hypoglycaemia
Measured values on further study for up to four consecutive nights at time hypoglycaemia predicted (mmol/l)
!Hypoglycaemia confirmed
0400 0200 0500 0200 0200 0300
t2 5.1, 4.3, 5.9, 6.3 < 2, -=z2, c: 2, 4. I <2, 3.8, <2 4.2, 12.3, 2.6, 5.0 <2
Yes No Yes Yes Yes
Yes
The patients entered a further study period, and in 5, hypoglycaemia was confirmed on at least one occasion
glucose concentration between home blood tests and this may provide useful information to aid management of such patients in the future. [5]
References [I] The Diabetes Control and Complications Trial Research Group, The effect of intensive treatment of diabetes on the development and progression of long-term complications in insulin-dependent diabetes mellitus, N. Engl. J. Med. 329 (14) (1993) 977-986. [2] 1. Cranston, J. Lomas, A. Maran, I. Macdonald and S.A. Amiel, Restoration of hypoglycaemia awareness in patients with long-duration insulin-dependent diabetes, Lancer: 344 (1994) 2833287. 131 SK. Andersen, K.G. Olessen, F.V. Jensen and F. Jensen, HUGIN - A shell for building Bayesian belief universes for expert systems, in Proceedings of IJCAI 89, pp. 1080.-10x5, 1989. [4] O.K. Hejlesen and SK. Andersen, Implementation of a learning procedure for multiple observations in a diabetes advisory system based on causal probabilistic networks,
[6]
[7]
[X]
[9]
in Artificial intelligence in medicine, eds. S Anreassen, R. Engelbrecht, and J. Wyatt, Proceedings of the Fourth Conference on Artificial Intelligence in Medicine Europe, 3-6 October 1993, pp. 63-74 (10s Press, Munich, Amsterdam, 1993). L. Aarons, J.W. Mandema and M. Da&of. A population analysis of the pharmacokinetics and pharmacodynamics of Midazolam in the rat, J. Pharmacokin. Biopharm. 19 (1991) 485-496. S. Andreassen, J.J. Benn, R. Hovorka, K.G. Olessen and E.R. Carson, A probabilistic approach to glucose prediction and insulin dose adjustment: description of metabolic model and pilot evaluation study, Comput. Methods Programs Biomed. 41 (1994) 153-165. I. Bendtson. A. Kverneland, S. Pramming and C. Binder, Incidence of nocturnal hypoglycaemia in insulin-dependent diabetic patients on intensive therapy, Acta Med. Stand. 223 (1998) 5433548. S. Pramming, B. Thorsteinsson, 1. Ben&son, B. Ronn, and C. Binder, Nocturnal hypoglycaemia in patients receiving conventional treatment with insulin, Br. Med. J. 291 (1985) 3766379. P.H. Sonksen, S.L. Judd and C. Lowy, Home monitoring of blood glucose. Methods for improving diabetic control, Lancet i (1978) 729-732.
Computer MethodsandPrograms in Biomedicine 50(1996)241-246
Use of the DIAS model to predict unrecognised hypoglycaemia in patients with insulin-dependent diabetes D.A. Cavan”“, R. Hovorkab, O.KL. Hejlesenc, S. Andreassenc, P.H. Siinksen” “Department of Endocrinology, St Thomas’ Hospital, London, UK bDepartment of Systems Science, City University, London, UK ‘Department of Medical Informatics, Aalborg University, Aalborg, Denmark
Abstract The Diabetes Advisory System (DIAS) is a model of human glucose metabolismimplementedin a causal probabilistic network. It handlesdata on insulin dose,carbohydrate intake and blood glucoseconcentration to predict hourly blood glucoseconcentrationsand thus provide an indication of blood glucosevalues betweenhome blood tests. DIAS was used to predict blood glucoseprofiles in eight patients with well-controlled insulin-dependent diabetes, who are at increasedrisk of hypoglyca.emia(abnormally low blood glucoselevels). DIAS predicted nocturral hypoglycaemia in six patients and daytime hypoglycaemiain one patient. The occurrenceof nocturnal hypoglycaemiawas not recognisedby the patient or suspectedby their doctor but was subsequentlyconfirmed by blood testing in five patients. It is known that unrecognisednocturnal hypoglycaemiais common in patients with well-controlled diabetes.The ability of DIAS to identify such periods of hypoglycaemiawith reasonableaccuracy illustrateshow the advancedtechnology it employsmay provide reliable decisionsupport to clinicians. Keywords:
Diabetes;Insulin; Computer model
1. Introduction
Insulin-dependent diabetes (IDDM) is a disease in which autoimmune destruction of insulin-secreting cells in the pancreas leads to lifelong dependency on exogenous insulin for survival. The main burden of the disease on both the patient * Corresponding author. Medical Unit, St Thomas’ Hospital, London SE1 7EH, UK. Tel: 0171 9289292; fax: 0171 9288289; e-mail: [email protected]
and national resources is attributable to the complications of the disease. These include blindness, renal failure, amputations, heart disease and stroke. The Diabetes Control and Complications Trial (DCCT) has established a clear correlation between levels of blood glucose (as reflected in glycosylated haemoglobin measurements) and development of complications, such that any improvement in blood glucose concentration towards normality carries a reduced risk [l]. The advent of home blood glucose monitoring has greatly improved the ability of patients to achieve good diabetic control. The main draw-
0169-2607/96/$15.00 0 1996 Elsevier Science Ireland Ltd. All rights reserved PZZ 0169-2607(96)01753-l
242
D.A.
Cavan
et al. i Computer
Methods
and Progrnms
back of good diabetic control, clearly demonstrated in the DCCT trial, is an increased frequency of hypoglycaemia, when blood glucose concentration falls to abnormally low levels. In its mildest form, hypoglycaemia manifests as sweating but more profound hypoglycaemia may lead to coma and rarely death. For many patients it is the most feared aspect of their condition and this fear may conflict with attempts at achieving good diabetic control. This fear is particularly pronounced in those patients who have lost the ability to recognise impending hypoglycaemia. Such ‘hypoglycaemic unawareness’ is more common in long-standing diabetes and while its precise cause is unknown, recent research has shown that avoidance of hypoglycaemia may restore awareness of any future episode in some individuals [2]. Thus the challenge in managing patients with IDDM is to maintain near normal blood glucose concentrations while minimising the risk of hypog1ycaemi.a. Experience to date suggests that this is best achieved by intensive blood glucose monitosing, regular carbohydrate intake and adjustment of insulin dose to meet changing requirements. Despite this, the increased risk of hypoglycaemia in patients with well-controlled diabetes and the loss of warning symptoms associated with recurrent hypoglycaemia is potentially hazardous; reduced alertness caused by hypoglycaemia while driving, for example, may have tragic consequences. The Diabetes Advisory System (DIAS) is a model of human glucose metabolism implemented in a causal probabilistic network (CPN). This gives it the ability to handle uncertainty, for example, in blood glucose measurements or physiological variations in glucose metabolism. The model has elements describing the blood glucose concentration and the amount of carbohydrate in the gut as functions of hepatic glucose output, carbohydrate uptake from the gut, insulin-independent glucose uptake (e.g., central nervous system), insulin-dependent glucose uptake (e.g., muscles), loss of carbohydrate from renal excretion and carbohydrate intake. DIAS includes two parameters, insulin sensitivity and time-to-peak of NPH (long acting) insulin absorption, to represent individual characteristics
in Biomedicine
50 (1996)
241-246
‘of glucose metabolism and insulin absorption from a subcutaneous depot. Standard modelling Imethodology offered by the CPN approach is employed as implemented in the HUGIN package 1131.Glucose metabolism is represented using a discrete-space (finite number of states for each variable), discrete-time (finite numlber of time instances), stochastic model. The relationships between variables are specified using conditional probability tables. Given a set of measurements, e.g., blood glucose concentrations,, carbohydrate intake and insulin doses, a posteriori (marginal) probability distributions for the remaining model variables are calculated using the 13ayes theorem. Thus, unlike in the deterministic modelling approach, unknowns are estimated as probability distributions rather than as point-estimates. The precision of parameter estimates and subsequently precision of any predictions are inherently obtained. DIAS can be run in three modes, the learning mode, the prediction mode, and the advisory mlode. In the learning mode, relevant data (blood glucose measurements, food intake represented as grams of carbohydrate, and insulin doses) from a single day are entered (instantiated) into the model and the joint probability distribution for the two parameters is calculated. This procedure is repeated for up to four days giving four daily estimates of the joint distribution. Assuming a bivariate normal form of the joint distribution, the means and the standard deviations of the two parameters are estimated 141. This hierarchical modelling approach is similar to that adopted in ph.armacokinetics [5]. In the prediction mode, the probability distribution of parameters calculated in the learning mode is used to make predictions of blood glucose given the carbohydrate intake and insulin regimen of the subject. It is this mode which is the focus of the present paper as the ability of DIAS to predict nocturnal hypoglycaemia was examined. In the advisory mode, a utility measure is minimised to find therapy which gives the least overall (predicted) risk of too low or too high blood glucose concentration [6]. DKAS was not operated in this mode for the purposes of the present study.
D.A. &van et al. / Computer Methods and Pr#ograms in Biomedicitze 50 (1996) 241-246
243
Table 1 Patient characteristics NO.
2 3 4 5 6 7 8 Mean S.D.
Sex
F F M F M M M F
Age (years)
Duration of diabetes (yr)
Weight (1%)
Body mass index (wt/ht*)
HbA, (“/) (ref 5.4-7.6X)
Insulin dose Way)
42 42 24 33 26 36 67 30 37.5 13.6
21 36 15 24 2 13 3s 14 20 11.5
73.3 68.2 80.6 56.7 82.2 74.2 70.8 70 12 7.91
26.3 24.7 27.6 24.9 21.6 23.2 23.1 27.3 24.8 2.14
9.6 7.8 9.0 9.9 8.4 8.8 7.3 7.4 8.5 1.0
42 44 56 37 54 76 48 56 52 12
The aim of this study was to determine whether DIAS can identify periods of unrecognised hypoglycaernia in subjects with well-controlled dliabetes, and whether of any predictions hypogl,ycaemia could be verified by blood testing. We therefore asked eight patients with insulin-dependent diabetes to collect data for four days to determine whether the DIAS simulations of their glucose profiles suggested unrecognised hypoglycaemia.. Patients in whom hypoglycaemia was SUSpetted underwent further study during which they measured blood glucose concentrations at the times hypoglycaemia was predicted. This paper reports the predictions of hypoglycaemia by DIAS and their subsequent confirmation in these patients.
2. Methods Eight patients with insulin-dependent diabetes mellitus (IDDM) were recruited from the diabetic clinic at St Thomas’ Hospital, London. Patient characteristics are shown in Table 1. The inclusion criteria were HbA, < 10% (representing good glycaemic control, reference range 5.47.6%) insulin-dependent from the time of diagnosis, and ability and willingness to perform home blood glucose measurements using a meter. Patients with serious concomitant illness, anorexia or other eating disorder or psychiatric illness were excluded. On recruitment they were asked to col-
lect the following data for four consecutive weekdays: four blood glucose measurements per day (before meals and before bed) using BM strips and a Reflolux meter (Boehringer Mannheim, Mannheim, Germany) and full details of insulin injections (time, type and dose). They were asked to record any hypoglycaemic attacks. Patients were seen by a dietitian and asked to keep a (detailed diary of all food intake during this period. Patients were seen within a week following the data collection period. Their records were checked for completeness and the food diary reviewed by a dietitian who assessed the carbohydrate content of each meal. Recorded blood glucose values, insulin injections and carbohydrate intake with times were entered into DIAS and a simulation of the blood glucose profile over the 4-day study period was generated. Patients for whom the model predicted hypoglycaemia (blood glucose below 3 mmol/l) which was not documented by blood glucose measurernent were asked to record four daily blood tests for a further four day period and, in addition, to test their blood at the time hypoglycaemia was predicted. In order to avoid introducing expectation of hypoglycaemia, patients were not advised of the reason for the tests at these times, but only that they would assist in verifying the model. Clinic computer records of the patients were searched for evidence of previous problems with hypoglycaemia.
244
D.A.
Cauan
rt al. I: Computer
Methods
and Programs
ill Biomedicine
50 (1996)
241-246
B meal-data g 2501 200. 150. 100. 50. O0 ins-A-data 1 it-M&data il 3224. 16. 8O-
1
8
12
16
20
24
4
8
fl bg-data f bg-sim mmol/l 16. 12. 84OIUI IO 95
Tue Jut 11 95
Fig. I. Example of simulated blood glucose profile in a patient showing episodes of nocturnal hypoglycaemia. The simulation covers :wo days (July IO-1 1). Meals expressed as grams of carbohydrate (meal-data) are shown in the upper graph. The lower panel shows insulin doses (soluble insulin; ins-A-data) as open bars and isophane (NPH) insulin (ins-N-data) as black bars. Measured blood glucose values (bg-data) are shown as black squares. The sinulated blood glucose profile (bg-sim) suggests nocturnal hypoglycaemia at around 0200 h on the 11th. Note the simulation suggests further periods of hypoglycaemia before lunch on the 10th and at 1200 h and 1800 h on the I Ith, although measured blood glucose values range from 5 to 12 mmol/l. 3. Results
DIAS predicted unrecognised hypoglycaemia in seven of the eight patients. In one (subject 8), hypoglycaemia was predicted at around midday and documented by home blood tests on one occasion during the initial study period, although no symptoms were reported. In six patients, hypoglycaemia was predicted at night, at times ranging between 0200 h and 0500 h. None had been aware of nocturnal hypoglycaemia and only four had reported any hypoglycaemia during the study period. Seben patients, however, recorded blood glucose measurements below 3 mmol/l. In only one patient was no hypoglycaemia predicted by DIAS or recorded by the patient. Study of previous records showed that four patients in whom recurrent hypoglycaemia was predicted had reported loss of symptoms and one had reported nocturnal hypoglycaemia. Fig. 1 shqws a DIAS simulation showing nocturnal hypoglycaemia and all data regarding reported and unrecognised hypoglycaemia are summarised in Table 2.
The six patients in whom nocturnal hypoglycaemia was predicted by DIAS proceeded to further study. When asked to set an alarm clock and measure their blood glucose concentration at the exact time hypoglycaemia was predicted by DIAS, all did so on at least one night. In five of the six patients, these tests confirmed hypoglycaemia on at least one occasion. Table 3 shows the mean bedtime and fasting blood glucose values for these patients during the initial study and the time at which hypoglycaemia was predicted. The number of night-time tests performed is shown together with their values.
4. Discussion
This paper describes the first clinical use of advanced technology as a management tool in insulin-dependent diabetes, The DIAS model of human glucose metabolism, implemented in causal probabilistic networks, was used to identify
D.A.
Cavan
et al. / Computer
Methods
and Programs
in Biomedicine
50 (1996)
241-246
245
Table 2 Relationship between reported hypoglycaemia, unrecognised hypoglycaemia and history of hypoglycaemia unawareness or nocturnal hypogfycaemia in each subject Patient
No. of reported hypoglycaemic episodes over 4 days
Unrecognised hypoglycaemia predicted by DIAS
Previously reported hypoglycaemia unawareness (HU) or nocturnal hypoglycaemia (NH)
I
3 0 4 4 0 1 0 0
Yes Yes Yes Yes Yes Yes No Yes
HU NH Nil Nil Nil HU
2 3 4 5 6 7 8
periods of unrecognised hypoglycaemia in patients with well-controlled insulin-dependent diabetes. The patients studied were a small sample with good overall control and thus not representative of the clinic population as a whole. A high frequency of hypoglycaemia was not unexpected: the DCCT trial suggests that such patients are prone to hypoglycaemia [l] and three had experienced hypoglycaemia of some degree in the period leading up to the study. The significance of this study is the ability of DIAS to identify periods of hypoglycaemia, which were unrecognised by the patient, yet which were subsequently confirmed in all but one of the patients studied. While previous studies have demonstrated high frequencies of nocturnal hypoglycaemia in patients with insulin-dependent diabetes, hourly measurement of blood glucose concentration in a hospital setting has usually been employed in such studies [7,8]. A number of our patients had previously reported loss of awareness of hypoglycaemia. This may result from recurrent hypoglycaemia [2] and may further exacerbate the difficulties in detecting it in an outpatient setting. This is illustrated in the case of patient 5 who had previously been asked by a specialist nurse to test his blood glucose concentration at 0300 h to check for hypoglycaemia but none was detected. The ability of DIAS to suggest a time, specific for each patient, when hypoglycaemia is likely to occur, may increase the likelihood of its being detected.
It should be emphasised that DIAS provides blood glucose simulations which are applicable only for the days for which data have been collected and it should not be assumed that hypoglycaemia would always occur at the time predicted. Day to day changes in levels of physical activity and food intake will affect blood glucose concentrations and may explain the differences between patients in the frequency with ,which hypoglycaemia was confirmed, from one (patient 5) to Ihree (patient 3) out of four measurements being under 3 mmol/l. Such factors may also explain why the prediction of hypoglycaemia could not be confirmed in patient 2. In summary, this small study concurs with previous findings that hypoglycaemia; is common in patients with well-controlled insulin-dependent diabetes [7,8] and suggests that such hypoglycaemia may be difficult to identify clinically. These preliminary data suggest that DIAS may aid detection of hypoglycaemia. DIAS is a prototype which requires much further developmental work and clinical validation; confirmation of these findings in larger studies may, however, suggest a role for DIAS as a useful clinical tool for outpatients. Fifteen years ago, the management of patients with insulin-dependent diabetes was revolutionised by the introduction of home blood glucose monitoring, which for the first time gave an indication of blood glucose levels between clinic visits [9]. Now, the advanced technology incorporated in DIAS may provide an indication of blood
246
D.A.
Chvan
et al. / Computer
Methods
and Programs
in Biomedicine
50 (1996)
241-246
Table 3 Mean early morning and late evening blood glucose measurements of the six patients in whom unrecognised nocturnal hypoglycaemia was predicted, together with the predicted time of hypoglyca.emia Patient
Late evening blood glucose (mmol/l )
1 2 3 4 5 6
7.6 7.8 11.3 4.3 6.8 5.3
Early morning blood glucose (mmol/l)
6.1
6.1 5.2 0.3 8.4 1.6
Predicted time of hypoglycaemia
Measured values on further study for up to four consecutive nights at time hypoglycaemia predicted (mmol/l)
!Hypoglycaemia confirmed
0400 0200 0500 0200 0200 0300
t2 5.1, 4.3, 5.9, 6.3 < 2, -=z2, c: 2, 4. I <2, 3.8, <2 4.2, 12.3, 2.6, 5.0 <2
Yes No Yes Yes Yes
Yes
The patients entered a further study period, and in 5, hypoglycaemia was confirmed on at least one occasion
glucose concentration between home blood tests and this may provide useful information to aid management of such patients in the future. [5]
References [I] The Diabetes Control and Complications Trial Research Group, The effect of intensive treatment of diabetes on the development and progression of long-term complications in insulin-dependent diabetes mellitus, N. Engl. J. Med. 329 (14) (1993) 977-986. [2] 1. Cranston, J. Lomas, A. Maran, I. Macdonald and S.A. Amiel, Restoration of hypoglycaemia awareness in patients with long-duration insulin-dependent diabetes, Lancer: 344 (1994) 2833287. 131 SK. Andersen, K.G. Olessen, F.V. Jensen and F. Jensen, HUGIN - A shell for building Bayesian belief universes for expert systems, in Proceedings of IJCAI 89, pp. 1080.-10x5, 1989. [4] O.K. Hejlesen and SK. Andersen, Implementation of a learning procedure for multiple observations in a diabetes advisory system based on causal probabilistic networks,
[6]
[7]
[X]
[9]
in Artificial intelligence in medicine, eds. S Anreassen, R. Engelbrecht, and J. Wyatt, Proceedings of the Fourth Conference on Artificial Intelligence in Medicine Europe, 3-6 October 1993, pp. 63-74 (10s Press, Munich, Amsterdam, 1993). L. Aarons, J.W. Mandema and M. Da&of. A population analysis of the pharmacokinetics and pharmacodynamics of Midazolam in the rat, J. Pharmacokin. Biopharm. 19 (1991) 485-496. S. Andreassen, J.J. Benn, R. Hovorka, K.G. Olessen and E.R. Carson, A probabilistic approach to glucose prediction and insulin dose adjustment: description of metabolic model and pilot evaluation study, Comput. Methods Programs Biomed. 41 (1994) 153-165. I. Bendtson. A. Kverneland, S. Pramming and C. Binder, Incidence of nocturnal hypoglycaemia in insulin-dependent diabetic patients on intensive therapy, Acta Med. Stand. 223 (1998) 5433548. S. Pramming, B. Thorsteinsson, 1. Ben&son, B. Ronn, and C. Binder, Nocturnal hypoglycaemia in patients receiving conventional treatment with insulin, Br. Med. J. 291 (1985) 3766379. P.H. Sonksen, S.L. Judd and C. Lowy, Home monitoring of blood glucose. Methods for improving diabetic control, Lancet i (1978) 729-732.