Medical Informatics
Scientific Research Project
Predictors of good glycemic control with the use of a web-based
insulin titration system
Author: Mentor: Tutor:
B. Goorden, Bsc A.C.R. Simon, MSc N.B. Peek, PhD
Predictors of good glycemic control with the use of a web-based
insulin titration system
Student B. Goorden Meibergdreef 9 1105 AZ Amsterdam Collegekaart nummer: 10464611 E-mail: b.goorden@amc.uva.nl Mentor A.C.R. Simon, MSc PhD-student
Department of Medical Informatics, J1B-109 a.c.simon@amc.uva.nl
Tutor
N.B. Peek, PhD Principal Investigator
Department of Medical Informatics, J1B-110 020 56 67872 / n.b.peek@amc.uva.nl
Location of Scientific Research Project Academic Medical Center
Department of Medical Informatics Meibergdreef 9
1105 AZ Amsterdam 020 56 69111
Practice teaching period November 2013 – August 2014
Contents
Abstract ... 1
Introduction ... 2
Methods ... 4
Study design and participants ... 4
Measures ... 6
Statistical analysis ... 9
Concordance between self-measured FPG and laboratory measured FPG ... 10
Results ... 11
Baseline characteristics ... 11
PANDIT system use ... 12
HbA1c and FPGlab during follow-up ... 13
Statistical Analysis ... 14
Concordance between self-measured FPG and laboratory measured FPG ... 18
Discussion ... 19
1
Abstract
Introduction: A lot of predictors for glycemic control in Diabetes Mellitus patients are already known. However the use of web-based assisted titration systems might change the predictors for glycemic control. These predictors provide us with insight into what sub-populations will benefit most from web-based assisted titration systems. In this study we aimed to determine the predictors of glycemic control among T2DM patients using a web-based insulin titration system.
Methods: This is a post-hoc analysis of data from a randomized controlled trial that investigated the effect of computer-assisted insulin self-titration on glycemic control (the PANDIT study). In the PANDIT study patients received either standard care (control group) or care via an internet-based titration system (PANDIT system, intervention group). For our study we only used the data of the intervention group. Patients had a minimal follow up of 12 weeks. Our primary outcome was HbA1c and our secondary outcome was laboratory FPG. All available predictors for HbA1c and FPG as well some variables related to system and computer use were considered for inclusion in our analysis. Linear mixed models were used to analyse the relationship between the predictors and the outcome variables over time. Since we did not have enough power to include all predictors in our multivariate analysis, predictors were selected based on univariate association with the outcome variables.
Results: In our study 38 patients were included. Intermediate education level (16.88 mmol/mol with 95% CI 8.79 to 25.02) (compared to low education level), the interaction of intermediate education level with time (-12.33 mmol/mol with 95% CI -19.08 to -5.46), the interaction of age with time (-0.50 mmol/mol with 95% CI -0.87 to -0.10) and the interaction of the number of logins to PANDIT per month (-1.26 mmol/mol with 95% CI -2.36 to -0.21) were found to be associated with HbA1c values. The number of logins to PANDIT per month (0.38mmol/l with 95% CI 0.08 to 0.68) and it’s interaction with time (-0.45 mmol/l with 95% CI -0.91 to -0.06) were found to be associated with FPG values.
Discussion: HbA1c values of patients with an intermediate education level, started on average higher than patients with a low education level, but their HbA1c levels decreased faster during follow up. This seems to be a regression to the mean and intermediate education level is probably an incidental finding. In line with literature, age was correlated with a greater decrease in HbA1c values during follow up. However this finding was not confirmed with FPG as outcome measure. Also as expected the number of logins to PANDIT per month was correlated with a greater decrease in HbA1c and FPG values during follow up. The other predictors we included in our analysis were not found to be associated with glycemic control. This is probably because we did not have enough power. Because of the lack of power we could not include all predictors in our multivariate model. This could have biased our results. Therefore the results of this study should be used with caution.
2
Introduction
In 2003, 600.000 people in the Netherlands were diagnosed with diabetes. It is expected that this number will rise to 1.300.000 in 2025.(1) Due to this rising incidence, diabetes puts an increasing strain on the Dutch health care system. Type 2 Diabetes Mellitus (T2DM) is associated with serious long-term complications such as cardiovascular disease, renal failure and blindness.(2) Current guidelines recommend to aim for a HbA1c goal of 53 mmol/mol or lower.(3) HbA1c values below this goal indicate good glycemic control and results in a significant reduction of diabetes related morbidity and mortality.(2;4). Treatment to reach this goal consists of diet adjustments, oral glucose-lowering medication and insulin therapy. This, however, has proven to be an elusive goal, the average HbA1c level being 7.6% for type 2 diabetes patients and 7.8% in type 1 diabetes patients in the Netherlands.(5)
Several patient factors are known to be associated with poor glycemic control. Young age, being female, a long duration of diabetes, insulin use, a high body mass index (BMI) and a high cholesterol level are associated with higher HbA1c values.(6-10) Furthermore, sociodemographic factors, such as ethnicity, a low socioeconomic status, a low income, being uninsured or being unmarried also predict higher HbA1c values.(6;8-10) Finally, severe psychological abnormalities, a poor self-efficacy and a poor understanding of HbA1c values are also associated with higher HbA1c values.(11-13) Not only patient factors, but also external factors are important for achieving good glycemic control. One study found that the individual primary care physician was found to be a significant predictor of good glycemic control.(10) Furthermore, for insulin users, frequency of titration, i.e. insulin dose adjustments, provided by the caregiver has been shown to be positively correlated with a reduction of HbA1c.(14) Such factors create possibilities for strategies to improve glycemic control.
To increase titration opportunities, several web-based self-management systems have been developed to assist patients in adjusting their insulin dose.(15-17) Studies have shown that self-management supported by E-health could improve both glycemic control and cost
3 effectiveness.(18-20) In the Netherlands, 94% of the households have an internet connection.(21) Therefore a web-based self-management system has the potential to reach a large number of patients at low costs. The use of web-based assisted titration systems to be implemented in the near future, might change the factors that predict the achievement of good glycemic control. For web-based diabetes self-management interventions it has been shown that patients with a baseline HbA1c of 53 mmol/mol or higher benefited more from the intervention.(22) However still very little is known.(20) As these factors provide us with insight into what sub-populations will benefit most from computer based diabetes self-management interventions, more research is needed on this topic.
In this study we aimed to determine the predictors of glycemic control among T2DM patients using a web-based insulin titration system. Our research group performed an open multi-center randomized controlled trial that investigated the effectiveness of a computer-assisted insulin self-titration system, called the Patient Assisting Net-based Diabetes Insulin Titration system (PANDIT), in improving glycemic control. Patients in the intervention group needed to access the PANDIT system once every three days, to enter their fasting glucose values from the last three days, their insulin dosages and whether they experienced hypoglycemia. Based on those data, the PANDIT system immediately provided an insulin dosing advice for the patient.Data from this intervention group were used to assess which factors predict good glycemic control for patients using the PANDIT web-based insulin titration system.
4
Methods
Study design and participants
This is a post-hoc analysis of data from a randomized controlled trial that investigated the effect of computer-assisted insulin self-titration on glycemic control (the PANDIT study). In brief, in this study patients received either standard care (control group) or care via an internet-based titration system (PANDIT, intervention group). The PANDIT system was tested for safety and usability in an earlier study.(23) The PANDIT study was approved by the institutional review board of the AMC and carried out in accordance with the principles of the Declaration of Helsinki and of Good Clinical Practice. The study population consisted of patients, aged between 18 and 80 years, with diabetes mellitus type 2 that were either already using basal insulin or started using basal insulin at the start of the study. Furthermore, patients needed to be familiar with the Internet and they needed to be able to read and understand the Dutch language. Patients were recruited through GP practices and hospitals in the Amsterdam region. Recruitment took place from January 2013 to January 2014 and patients were enrolled from inclusion until end of trial with a minimal follow-up time of 12 weeks. Patients in both study arms were contacted at enrollment, and approximately at each subsequent 12 weeks until their follow-up period. During contact at enrollment the patient was screened and randomized into the control or intervention group. At each contact FPG (Fasting Plasma Glucose) and HbA1c was measured. For the current study, we only used data from patients in the intervention group.
System description
Participants in the intervention group were asked to visit the PANDIT system at least once every three days. The PANDIT system contains an electronic diary, which can be seen in figure 1. In this diary patients can fill in their daily fasting plasma glucose (FPG) values, insulin dosages used, and whether they experienced symptoms of hypoglycaemic episodes since the last time they visited the system. The system provides an insulin dosing advice for the patient, based on the FPG values entered. The patients were asked to use the insulin dose advice from that moment on. If a patient did not visit the PANDIT system in time, a text
5
message was send to their mobile phone. Patients could also choose to use a different visit frequency instead of once every three days.
Fig 1: Patient diary of the PANDIT system
Insulin dose advice were based on a treat-to-target algorithm, i.e. systematically adjusting the insulin until a certain FPG range (e.g. FPG 4-7 mmol/mol) has been achieved. The default upper limit of the FPG range for patients in primary care was 7.0 mmol/l. For patients in secondary care the default upper limit was 5.5 mmol/l. The upper limit of the FPG target range could be tailored to the individual patient by their caregiver to prevent future hypoglycemic events. The lower limit of the FPG target was for all patients 4.0 mmol/L. The treat-to-target algorithm used the lowest FPG value of the preceding three days to check against the patients FPG target range. If the FPG value was within the target range, the patient was advised to use their current insulin dose. If the FPG value was lower than 4.0 mmol/L the insulin dose was decreased. If the FPG value was higher than the upper limit of
6
the FPG range the insulin dose was increased. The adjustments to the insulin dose can be found in table 1. The increase or decrease in insulin dose depends on the patient’s age and body weight. For example, if the FPG value was <2.5 mmol/l, the insulin dose was adjusted with -0.04 IU per kg body weight of the patient (with a minimal adjustment of -4UI). The insulin dose of a patient above the age of 70 years with an FPG value of >10 mmol/l, will only be increased with 2 IU (instead of 4 IU).
FPG value Insulin dose adjustment
< 2.5 mmol/l -0.04 IU/kg (min -4IU) > 2.5 and < 4.0 mmol/l -0.02 IU/kg (min -2IU) Upper target limit* – 9.9 mmol/l +0.02 IU/kg (min +2IU) > 10 mmol/l +0.04 IU/kg (min +4IU)**
Table 1: Insulin titration algorithm.
* Upper target limit can be adjusted by the caregiver
** A patient above the age of 70 years will be up titrated with only +2 IU if FPG > 10mmol/L.
Before a new insulin dose advice was given, the treat-to-target algorithm checked if there was no large variation in insulin dose usage (more than 4 units) and if a hypoglycaemic event had occurred (low blood glucose values or patient indicated he/ she experienced symptoms of hypoglycaemia). Hypoglycaemic events were graded with a questionnaire. If a severe hypoglycaemic event had occurred or two non-severe hypoglycaemic events had occurred within a month, the PANDIT system was blocked and the patient was advised to contact their caregiver. Care providers had the possibility to take over the provision of insulin dose advice after reviewing the data and to block/ unblock a patient account after a hypoglycaemic event.
Measures
All predictors of HbA1c and FPG from the literature were considered for inclusion in our analysis. Since we did not have data on each predictor, we had to use a proxy variable for some predictors. In table 2 the predictors of HbA1c and FPG and the available variables we used for them can be found. We did not have data for all known predictors.
7
Glycemic control
HbA1c and FPG measured at the laboratory (FPGlab) were our outcomes of interest. HbA1c is our primary measure of outcome. FPGlab is our secondary measure of outcome. We prefer FPGlab over the self-measured glucose values (that the patients entered in the PANDIT system) because the measurements were performed in a controlled environment (the laboratory). HbA1c and FPGlab were measured at each subsequent contact moment (every 12 weeks). For each patient we have at least 1 and maximal 4 HbA1c and FPGlab values. HbA1c is a laboratory value that indicates glycemic control over a 2 to 3 month period. In general (especially for DM patients) lower HbA1c values mean a better glycemic control. Values less than 53 mmol/mol are considered optimal.
Predictor Available variable
Age(6;8;9) Age (years)
Gender(10) Gender (male/female)
Diabetes duration(6) Diabetes duration (years)
BMI(7) BMI (kg/m2)
Ethnicity(8) Ethnicity (Surinamese or Hindu/ Other)
Socioeconomic status(9;10) Employment status (job/ no job), Education Level (Low/
Intermediate/ High)
Marital Status(8) Partner (Yes/ No), Living Alone (Yes/No)
Insulin use(6) Duration of insulin use (years), Type of insulin (NPH/ insulin analogs)
Self-efficacy(11) CIDS (0-100)
Table 2: Known predictors of HbA1c and FPG with the available variables for the analysis
Sociodemographics
Available sociodemographic variables included age, gender, ethnic group and education level. Two ethnic groups were created; Surinamese/Hindu and other (including Caucasian). This grouping was chosen because patients in the Surinamese/Hindu group have an increased risk to develop diabetes(24). Three groups were created for education level; low, intermediate and high. Patients were also asked if about their employment status (job/no job), if they currently had a partner (yes/no) and if they were currently living alone (yes/no). For our analysis, we also included whether patients were treated in primary or secondary care for their DM. Patients in primary care and secondary care started with different upper
8
limit target levels. We had no data available on the predictors income, insurance status and psychological abnormalities.
Physical characteristics
Body weight and height were measured at baseline to calculate the BMI. We had no data available on the predictor; cholesterol.
Diabetes related variables
Diabetes related variables included diabetes duration and duration of insulin use during the time of inclusion. Also the type of insulin was included. The type of insulin was divided into two groups: human insulin (NPH) and insulin analogs (insulin glargine and insulin detemir). We had no data available on the predictors; understanding of HbA1c by the patient and the primary care physician of the patient.
Behavioural variables
The CIDS (Confidence in Diabetes Self Care) questionnaire was registered at baseline.(25) The CIDS scale can be used to measure diabetes-specific self-efficacy. The CIDS questionnaire was originally designed for DM1 patients. However a lot of the questions are also applicable to DM2 patients. Only the questions applicable to DM2 patients were selected for the analysis. Questions 3, 4, 5, 7 and 9 were therefore excluded. From the scores of all other questions a total score from 0 to 100 was calculated. A higher CIDS score means the patient has a higher diabetes specific self-efficacy.
The patients also filled in a questionnaire with several questions regarding how frequently they used computers for different tasks. A total score was calculated from 0 to 100 (0; almost never use a computer, 100; daily computer users). This score for frequency of computer use was included in our analysis because an e-health intervention was used.
System use
Since the patients used a web-based insulin titration system, we expect that variables related to the use of this system will also influence glycemic control. Therefore we included the number of logins to the PANDIT system per month in our model. The logins, where the patient entered a FPG value in the PANDIT system or got an insulin dose advice, were
9
extracted from the PANDIT-database. From these data the number of logins to the PANDIT system per month was determined for each patient.
Statistical analysis
Linear mixed models were used to analyse the relationship between the predictors and the outcome variables over time. Multilevel analysis were used because of the hierarchical structure of the data: the multiple measurements per patient were nested within the patients. So our models consisted of two levels. Since our outcome variables (HbA1c and FPGlab) were continuous, multilevel linear regressions were performed. Because the observations within each patient are not independent of each other, patient IDs were included in the random part of our model. Time (since inclusion, in units of a half year), the predictors and the interaction of the predictors with time were included in the fixed part of our model.
HbA1c and FPGlab during follow-up
We will first analyse the HbA1c and FPG values measured in the lab (FPGlab) during follow-up. For this analysis we used two models, one with HbA1c as outcome variable and one with FPGlab as outcome variable. The time at which the values were measured, in units of a half year, was included as independent variable. A random intercept at patient level was included in both models.
Selection of predictors
All aforementioned variables were included in our model. For each variable an interaction between the variable and time was also included to assess the change of HbA1c and FPG over time in relation to that variable. Prior to the analyses, the distributions of HbA1c and FPGlab values were assessed to check if the normal distribution assumption was met.
For our analysis we preferred to fit a full model. However we did not have enough power to include all predictors in our model, therefore we had to make a selection which predictors to include. To select predictors for our models we first performed a univariate analysis for each predictor. Time, the predictor and the interaction between the predictor and time were included as independent variables. A random, patient-level intercept was included. If the univariate association of the interaction between the predictor and time had a p-value <
10
0.10, both the predictor and the interaction between the predictor and time were selected for the multivariate model. If the univariate association between predictor and outcome had a p-value < 0.10, only the predictor was selected for the multivariate model.
The selected predictors from the univariate analysis with HbA1c as outcome were used as independent variables in our multivariate model with HbA1c as outcome. The selected predictors from the univariate analysis with FPGlab as outcome were used as independent variables in our multivariate model with FPGlab as outcome. A random intercept at patient level was included in both multivariate models.
Concordance between self-measured FPG and laboratory measured FPG
In this study we had data on Fasting Plasma Glucose values measured in the lab (FPGlab) and self-measured FPG values (SMBG). The SMBG values were entered in the diary of the PANDIT system by the patients. Since the FPGlab measurements were performed in a controlled environment, we choose to use FPGlab as our outcome (in addition to HbA1c). To determine if SMBG values would have been a reliable alternative as an outcome for the FPGlab values we will determine the concordance between FPGlab and SMBG values.
The FPGlab was measured at each contact moment (every 12 weeks). The concordance between FPGlab and SMBG values will be analyzed with a random intercept model. In this model SMBG was the independent variable and FPGlab was the outcome. A random intercept was included at the patient level. Because of the large variation in SMBG measurements, the average of five SMBG measurements was used for each FPGlab measurement. The five SMBG values that had a measurement date closest to the measurement date of the FPGlab value were chosen to calculate the average of those SMBG values. A maximal difference of ten days between the SMBG measurement date and FPGlab measurement date was allowed.
11
Results
Baseline characteristics
In the original study the control group consisted of 36 patients and the intervention group consisted of 38 patients. Only the 38 patients of the intervention group were included in this study. The baseline characteristics of these patients can be found in Table 3. Of the 38 patients, 1 patient did not use the PANDIT system. Not all patients answered all the questions in the questionnaire. Therefore some baseline data is missing. The lab-measured FPG (FPGlab) in the table was measured at the first contact moment. The self-measured FPG (SMBG) was the computed average of the first five FPG values entered by patients into the PANDIT system.
Variables Pandit (n=37) n
Age in years (mean(SD)) 60.4 (7.2) 38
Male (n (%)) 22 (57.9%) 38
BMI in kg/m2 (mean(SD)) 29.9 (5.5) 38
Diabetes treated in primary care (n (%)) 30 (78.9%) 38
Surinamese or Hindu ethnicity (n (%)) 6 (15.8%) 38
Living alone (n (%)) 16 (43.2%) 37
No Partner (n (%)) 13 (35.1%) 37
Unemployed (n (%)) 3 (8.3%) 36
Level of education; Low/ Intermediate/ High (n (%)) 8 (22.9%)/ 18(51.4%)/9 (25.7%) 35
Diabetes duration (years, median(IQI)) 10 (6-14) 36
Duration of insulin use (median(IQI)) 0.6 (0.1-2.1) 36
Human Insulin/ Insulin Analog (n (%)) 12 (31.6%)/ 26 (68.4%) 38
HbA1c in mmol/mol (mean(SD)) 67.9 (14.2) 38
FPGlab in mmol/l (mean(SD)) 9.2 (3.3) 37
SMBG in mmol/l (mean(SD)) 8.3 (2.7) 37
CIDS (mean(SD)) 84.8 (10.2) 34
Table 3: Baseline characteristics
We had a total of 100 HbA1c measurements and 94 FPGlab measurements. We assume we need about 10 measurements for each included predictor and could therefore only include about 10 predictors in our multivariate models. Since we had 16 predictors and also 16 times
12
their interaction with time, we did not have enough power to include all predictors in our multivariate models.
PANDIT system use
Table 4 describes the frequency of using the PANDIT system by patients and caregivers during the study. Patients had to log in to the PANDIT system at least every 3 days (i.e., at least 10.1 times per month). However on average patients logged in to the PANDIT system only 3.8 times each month. On average patients logged in to the PANDIT system 4.1 times per month in the first three months, from three months to six months 3.8 times per month and from six to nine months 2.6 times per month. So patients did log in to PANDIT more in the first three months, but never close to the required amount of 10.1 times a month. They did however fill in their glucose values in PANDIT for most days (24.7 times per month, on average). The number of advice given by caregivers per patient per half year refers to advice given by caregivers through the PANDIT system. A PANDIT account was only blocked in case of a severe hypoglycemic event, therefore the number of hypoglycemic events per patient per half year is higher than the number of blockages per patient per half year. In total 37 times a patient account was blocked. The five patients that got blocked most often, were responsible for a total of 25 blockages. The upper limit of the target FPG range was adjusted for 16 patients (42.1%) by their caregiver. All frequencies in table 4 were calculated from data taken over the entire follow up period of the patients.
Variables n
Number of PANDIT visits per Month (median(IQI)) 3.8 (2.6-5.3) 37
Glucose values entered per Month (median(IQI)) 24.7 (17.2-27.8) 37
Number of advice given by PANDIT per patient per Month (median(IQI)) 3.1 (1.5-4.2) 37 Number of advice given by caregiver per patient per Half Year (median(IQI)) 0 (0-1.4) 37
Number of hypoglycemic events per patient per Half Year (median(IQI)) 0.4 (0-2.9) 37
Number of blockages per patient per Half Year (median(IQI)) 0 (0-1.4)) 37
Upper glucose target value changed 16 (42.1%) 37
Upper glucose target value primary care in mmol/l (mean(SD)) 6.6 (0.5) 30
Upper glucose target value secondary care in mmol/l (mean(SD)) 6.1 (0.7) 7
13
HbA1c and FPGlab during follow-up
The regression line of the analysis with HbA1c as outcome and time as independent variable is shown in Figure 2 (A random intercept was included at the patient level). The mean decrease in HbA1c was 4.26 mmol/mol per half year. This decrease was significant with a 95% CI -6.81 to -1.72. There was a large variation between the patients in the follow up time and therefore the number of HbA1c measurements per patient. The longest follow up time was 12 months. For most patients HbA1c decreased over time, however for a few patients the HbA1c actually increased. Furthermore HbA1c values seemed to regress to the mean over time.
Fig. 2: Blue points: HbA1c measurements, Blue lines: connect the HbA1c measurements of individual patients, Red: regression line with HbA1c as outcome and time as independent variable.
The same analysis was performed for FPGlab. The regression line of this analysis is shown in Figure 3. FPGlab decreased with 0.61 mmol/l per half year. This decrease was however not significant with a 95% CI of -1.43 to 0.19. With FPGlab there seems to be more variation in the starting values and the change in the follow up values in comparison to the HbA1c values.
14
Fig. 3: Blue points: FPG lab-measurements, Blue lines: connect the FPGlab measurements of individual patients, Red: regression line with FPG as outcome and time as independent variable.
Fig 4: Hisgrams of HbA1c and FPG
Statistical Analysis
Distributions outcome variables
The mean of HbA1c was 65.2 mmol/mol, with a standard deviation of 12.1 mmol/mol. HbA1c had a minimum of 44.0 mmol/mol and a maximum of 112.0 mmol/mol. FPGlab had a
15
mean of 8.65 mmol/l, with a standard deviation of 2.69 mmol/l. The minimum value was 4.20 mmol/l and the maximum value was 21.70 mmol/l. In figure 4 the histograms of HbA1c and FPG can be found. HbA1c and FPGlab values were both normally distributed. Therefore no transformation of the data was needed.
Univariate analysis with HbA1c as outcome
Table 5 provides the results of the univariate analyses with HbA1c as outcome, with the regression coefficient, the 95% confidence interval and the p-value for the variable and the interaction between the variable and time since the start of the trial (per half year). The strongest association for a variable with HbA1c measurements was intermediate education. Patients with an intermediate education level had on average, at the start of the trial, a 16.91 mmol/mol (95% CI 9.70 to 23.99) higher HbA1c value than patients with a low education level (the reference category for education level).
Variable Regression coefficient variable (95% CI)
p-value Regression coefficient interaction with time (95% CI)
p-value
Age (Years) -0.28 (-0.87 to 0.30) 0.344 -0.31 (-0.67 to 0.05) 0.095
Female 2.99 (-5.11 to 11.11) 0.476 -0.82 (-5.95 to 4.23) 0.754
Diabetes duration (Years) -0.44 (-1.33 to 0.44) 0.337 0.11 (-0.51 to 0.73) 0.738
BMI (kg/m2) -0.47 (-1.20 to 0.28) 0.216 0.02 (-0.42 to 0.46) 0.943
Surinamese or Hindu ethnicity 5.86 (-5.30 to 17.45) 0.305 4.47 (-1.38 to 10.11) 0.129
Unemployed -7.35 (-20.40 to 5.95) 0.279 7.06 (0.20 to 13.93) 0.048
Intermediate education 16.91 (9.70 to 23.99) <0.001 -5.24 (-11.72 to 1.22) 0.121
Higher education 5.79 (-2.55 to 14.30) 0.188 -6.43 (-13.58 to -0.66) 0.084
No Partner -3.15 (-11.72 to 5.46) 0.477 2.31 (-3.05 to 7.70) 0.403
Living Together -1.39 (-10.18 to 7.08) 0.749 -1.90 (-6.98 to 3.28) 0.468
Analogue insulin usage -5.98 (-14.75 to 2.91) 0.185 5.00 (-0.19 to 10.34) 0.066
Duration of insulin use (Years) -0.68 (-2.32 to 0.91) 0.388 0.33 (-0.83 to 1.49) 0.583
Secondary Care 9.40 (-0.22 to 19.00) 0.060 3.01 (-3.00 to 8.89) 0.323
PANDIT visits per Month 0.29 (-0.95 to 1.54) 0.646 -1.15 (-2.23 to -0.08) 0.039
CIDS Scale (0 to 100 scale) -0.43 (-0.80 to -0.06) 0.027 0.23 (-0.04 to 0.50) 0.104
Frequency of Computer use (0 to 100 scale)
-0.07 (-0.18 to 0.04) 0.213 -0.03 (-0.10 to 0.05) 0.475
Table 5: Univariate analysis with HbA1c (in mmol/mol) as outcome (time in units of half years)
The variables type of care and CIDS scale also had a p-value < 0.10. The strongest association for an interaction with time and the HbA1c measurements was, the number of logins to PANDIT per month. HbA1c values decreased on average 1.15 mmol/mol (95% CI 2.23 to -0.08) faster per half year, for each additional login per month to the PANDIT system. Other
16
interactions that had a p-value < 0.10 were, age, unemployed, high education level and insulin analog. The variables and interactions with a p-value < 0.10 were selected for the multivariate model with HbA1c as outcome. For interactions with a p-value < 0.10, the associated variable was also selected for the multivariate model.
Multivariate analysis with HbA1c as outcome
Table 6 provides the results of the multivariate analyses with HbA1c as outcome, with the regression coefficient, the 95% confidence interval and the p-value for the variable and the interaction between the variable and time since the start of the trial (per half year). Age, employment status, education level, insulin type, number of logins to PANDIT per month as well as their interactions with time were included as independent variables in the multivariate model. Type of care and CIDS scale were also included as independent variables in the model, however their interactions with time were not included. Intermediate education level, the interaction between intermediate education level and time and the interaction between logins to PANDIT per month and time were found to be significant predictors of HbA1c. Patients with an intermediate education level had, at the start of the trial, on average a HbA1c value that was 16.88 mmol/mol (95% CI 8.79 to 25.02) higher than patients with a low education level. However the HbA1c of Patients with an intermediate education level on average decreased 12.33 mmol/mol (95% CI -19.08 to -5.46) per half year faster than patients with a low education level. For each extra time a patient logged in to PANDIT per month, their HbA1c would decrease 1.26 mmol/mol (95% CI -2.36 to -0.21) faster per half year.
Variable Regression coefficient variable (95% CI)
p-value Regression coefficient interaction with time (95% CI)
p-value
Age (Years) 0.12 (-0.33 to 0.56) 0.660 -0.50 (-0.87 to -0.10) 0.017
Unemployed 0.10 (-12.84 to 13.08) 0.989 -1.00 (-13.01 to 10.70) 0.875
Intermediate Education 16.88 (8.79 to 25.02) 0.001 -12.33 (-19.08 to -5.46) 0.001
High Education 5.11 (-3.44 to 13.69) 0.314 -6.19 (-13.31 to 0.71) 0.105
Analog insulin usage 1.77 (-5.44 to 8.99) 0.678 2.44 (-2.99 to 7.80) 0.400
Secondary Care -4.79 (-4.92 to 14.46) 0.408 - -
PANDIT visits per Month 0.69 (-0.42 to 1.81) 0.295 -1.26 (-2.36 to -0.21) 0.032
CIDS Scale (1 to 100 scale) -0.30 (-0.63 to 0.02) 0.123 - -
Table 6: Multivariate analysis with HbA1c (in mmol/mol) as outcome (time in units of half years)
17
Univariate analysis with FPGlab as outcome
Table 7 provides the results of the univariate analyses with FPGlab as outcome. The variables Intermediate education, number of logins to PANDIT per month, and CIDS scale had a p-value < 0.10. Only the interaction of number of logins to PANDIT per month with time had a p-value < 0.10.
Variable Regression coefficient variable (95% CI)
p-value Regression coefficient interaction with time (95% CI)
p-value
Age (Years) 0.09 (-0.06 to 0.22) 0.189 -0.09 (-0.21 to 0.02) 0.127
Female -0.59 (-2.43 to 1.25) 0.534 0.12 (-1.48 to 1.75) 0.883
Diabetes duration (Years) -0.10 (-0.30 to 0.10) 0.354 0.07 (-0.13 to 0.26) 0.508
BMI (kg/m2) -0.06 (-0.22 to 0.11) 0.488 0.01 (-0.12 to 0.16) 0.848
Surinamese or Hindu ethnicity -1.38 (-3.96 to 1.31) 0.284 1.38 (-0.484 to 3.23) 0.151
Unemployed -0.34 (-3.62 to 3.02) 0.841 0.61 (-1.63 to 2.83) 0.591
Intermediate education 2.03 (-0.12 to 4.17) 0.079 -0.57 (-2.71 to 1.50) 0.594
Higher education 1.10 (-1.38 to 3.57) 0.398 -1.24 (-3.62 to 1.06) 0.303
No Partner -0.70 (-2.59 to 1.20) 0.474 1.03 (-0.64 to 2.74) 0.236
Living Together 1.31 (-0.63 to 3.18) 0.167 -1.31 (-2.93 to 0.27) 0.114
Analogue insulin usage 0.49 (-1.44 to 2.44) 0.628 0.62 (-1.11 to 2.31) 0.476
Duration of insulin use (Years) 0.08 (-0.27 to 0.42) 0.636 0.01 (-0.33 to 0.35) 0.951
Secondary Care 1.45 (-0.80 to 3.71) 0.218 1.25 (-0.70 to 3.11) 0.198
PANDIT visits per Month 0.30 (-0.004 to 0.61) 0.063 -0.40 (-0.79 to -0.03) 0.041
CIDS Scale (1 to 100 scale) -0.10 (-0.18 to -0.01) 0.035 0.06 (-0.04 to 0.15) 0.266
Frequency of Computer use (1 to 100 scale)
-0.02 (-0.05 to 0.01) 0.182 -0.005 (-0.03 to 0.02) 0.696
Table 7: Univariate analysis with FPGlab (in mmol/l) as outcome (time in units of half years)
Multivariate analysis with FPGlab as outcome
Table 8 provides the results of the multivariate analyses with FPGlab as outcome. Education level, number of logins to PANDIT per month, the interaction of number of logins to PANDIT per month with time and the CIDS scale were included as independent variables in the multivariate model with FPGlab as outcome. Number of logins to PANDIT per month and it’s interaction with time were both significant. For each extra time a patient logged in to PANDIT in a month, their FPGlab was on average, 0.38 mmol/l (95% CI -0.08 to -0.68) higher at the start of the trial. However for each extra time a patient logged in to PANDIT in a month, their FPGlab would decrease on average, 0.45 mmol/l (95% CI -0.91 to -0.06) faster per half year.
18
Variable Regression coefficient variable (95% CI)
p-value Regression coefficient interaction with time (95% CI)
p-value
Intermediate Education 1.35 (-0.65 to 3.34) 0.230 - -
High Education -0.07 (-2.18 to 2.04) 0.954 - -
PANDIT visits per Month 0.38 (0.08 to 0.68) 0.026 -0.45 (-0.91 to -0.06) 0.037
CIDS Scale -0.06 (-0.13 to 0.02) 0.212 - -
Table 8: Multivariate analysis with FPGlab (in mmol/l) as outcome (time in units of half years)
Concordance between self-measured FPG and laboratory measured FPG
With laboratory measured FPG (FPGlab) as outcome and self-measured FPG (SMBG) values as independent variable, the random intercept model gave a fixed intercept of 0.64 (95% CI -0.84 to 2.13) and a regression coefficient of 1.05 (95% CI 0.87 to 1.23). This indicates a high concordance. The variance for the random intercept was 1.155.
Fig 5: Red points: SMBG values plotted against FPGlab values. Blue: regression line with FPGlab as outcome and SMBG as independent variable.
19
Discussion
Principal findingsIn this study we aimed to determine the predictors of glycemic control among T2DM patients using a web-based insulin titration system. In our analysis, an intermediate education level (mavo, mbo, havo or vwo) and the interactions of intermediate education level, age and PANDIT login frequency with time were found to be associated with glycemic control. HbA1c of patients with an intermediate education level decreased 12.33 mmol/mol faster compared to patients with a low education level. However they did start with a HbA1c which was 16.88 mmol/mol higher on average, than that of patients with a low education level. Patients who were older or visited the PANDIT system more frequently benefited the most from using the PANDIT system. HbA1c decreased 0.50 mmol/mol faster for each year a patient was older and HbA1c decreased 1.26 mmol/mol faster for each additional log in to the PANDIT system per month. This suggests patients with an intermediate education level, older patients and patients that use the PANDIT system more frequently gained the most from using the web-based insulin titration system. Only PANDIT visit frequency and its interaction with time was also associated with our secondary outcome measure, FPGlab. Furthermore, we found a high concordance between self-measured FPG (SMBG) values and measurements of FPG in the lab (FPGlab), which indicates that SMBG values are a reliable source of information on glycemic control.
Findings in relation to other studies
Swinnen et al (2009) (14) found that the frequency of titration was positively correlated with a reduction of HbA1c. This is in line with our finding that a higher visit frequency to the PANDIT system, which provides more titration opportunities, was positively correlated with a reduction in HbA1c. Several studies found that age was positively correlated with a reduction of HbA1c. (6;8;9) In our analysis age was also positively correlated with a reduction of HbA1c. A low socioeconomic status was found to be correlated with poor glycemic control in several studies.(9;10) We therefore expected patients with an intermediate education level to be associated with lower HbA1c values compared to patients with a low education level. In our study, HbA1c of patients with an intermediate
20
education level did decrease faster than patients with a low education level. However, patients with an intermediate education level started on average with higher HbA1c values compared to patients with a low education level.
A long duration of diabetes, a poor self-efficacy, being female and a high BMI were all found to be associated with higher HbA1c values.(6;7;10;11) Furthermore, demographic factors such as ethnicity and not being married are also associated with HbA1c values.(8) However, these variables were not found to be predictors in our study.
Strengths and weaknesses of the study
A lot of research has already been performed on predictors of glycemic control among T2DM patients. In this study we focused on predictors of glycemic control among T2DM patients using a web-based insulin titration system. Therefore, we also included variables related to system use and computer use. To our knowledge these variables have not been analyzed before in this context.
A major limitation of this study was that we did not have enough power to include all variables in the multivariate model. In our study we selected variables that had a strong univariate association with HbA1c and FPGlab for our multivariate model. We selected variables that had a p<0.10 for the univariate association with the outcome measure. Because of this there is a chance that some predictors were not detected. Also there is a 10% chance a variable is wrongfully selected. These faults in selection of variables can lead to bias in the results of our multivariate model.
With the number of variables that we included, there is also still the risk that our multivariate model is over-parameterized. This could lead to faults in the regression coefficients and the p-values. The results of this study should therefore be used with caution.
Furthermore, the CIDS questionnaire (used to measure diabetes-specific self-efficacy) was designed for DM1 patients. We limited ourselves to the questions that were applicable for DM2 patients to measure self-efficacy. Yet, this questionnaire was not validated for DM2. However we do not expect this to have influenced our findings.
21
Meaning of the study
We found that a greater number of logins to the PANDIT system by patients was associated with a faster decrease in HbA1c during the study follow-up. This finding was confirmed by the analysis with our secondary outcome (FPGlab). A most likely explanation for this phenomenon is that more logins to PANDIT create a higher number of titration opportunities. These increased titration opportunities have probably led to the greater decrease in HbA1c values. This is in line with what Swinnen et al (2009) (14) have shown, that more titration opportunities lead to better HbA1c values.
At baseline patients with an intermediate education level had a higher mean HbA1c value compared to patients with a low level education. We did not expect this. Patients with high education level did not have a significantly higher HbA1c start value compared to patients with low education level. The mean decrease in HbA1c values of patients with an intermediate education level was higher than that of patients with a low education level. Since patients with an intermediate education level started with a higher HbA1c level, this is probably a regression to the mean. Intermediate education level was also not found to be associated with our secondary outcome (FPGlab). Therefore the statistically significant regression coefficients for intermediate education level and it’s interaction with time were probably an incidental finding.
The HbA1c of older patients decreased faster when compared to younger patients. It might be the case that older patients indeed benefit more from web-based titration than younger patients. People with higher age might lack those skills, such as calculating a new dosing advice, that are being compensated by using the web-based insulin titration. Also this finding suggests that older patients did not have trouble using the PANDIT system. However age was not found to be associated with our secondary outcome (FPGlab).
The other aforementioned predictors known from literature, were not found to be predictors in our study. For some of these predictors we did not have any data on them and could therefore not be included in the analysis. A lack of power could be the reason why known predictors, that were included in our analysis, were not statistically significant in our study.
We conclude that patients who logged in to the system more frequently and older patients benefited the most from the web-based insulin titration system. We recommend therefore, that when a web-based insulin titration system is used, to motivate the patients to use the
22
system with a high frequency. This could potentially improve glycemic control for the patients using the system. However considering the limitations of this research, the results should be used with caution.
Unanswered questions and future research
This study is a start in regard to which sub-populations of DM patients benefit the most from e-Health solutions. Overall, still very little is known. More knowledge on this could improve the targeting of certain sub-populations with e-Health solutions. We did not have enough power to include all our variables in the multivariate analysis. We therefore recommend future research on this subject to have data on all known predictors for glycemic control and enough power, so all predictors can be included in the multivariate analysis. With the results of this study in mind, it seems wise to at least include education level, age and frequency of use of the e-health solution in the analysis.
23
Reference List
(1) RIVM. Nationaal Kompas Volksgezondheid. Bilthoven; 2008 Aug 4. Report No.: versie 3.14.1.
(2) The Diabetes Control and Complications Trial Research Group. The effect of intensive treatment of diabetes on the development and pro-gression of long-term complications in insulin-dependent diabetes mellitus. New England Journal of Medicine 1993;329:977-86.
(3) Nathan DM, Buse JB, Davidson MB, Ferrannini E, Holman RR, Sherwin R, et al. Management of Hyperglycemia in Type 2 Diabetes: A Consensus Algorithm for the Initiation and Adjustment of Therapy: Update regarding thiazolidinediones: a consensus statement from the American Diabetes Association and the European Association for the Study of Diabetes. Diabetes Care 2009 Jan;(32):1-193.
(4) UK Prospective Diabetes Study (UKPDS) Group. Intensive blood-glucose control with sulphonylureas or insulin compared withconventional treatment and risk of complications in patients with type 2 diabetes . Lancet 1998;352:837-53.
(5) Inspectie voor de Gezondheidszorg. Het resultaat telt 2006; Prestatie-indicatoren als onafhankelijke graadmeter voor de kwaliteit van in ziekenhuizen verleende zorg. Den Haag; 2007 Dec 20.
(6) Benoit SR, Fleming R, Philis-Tsimikas A, Ming J. Predictors of glycemic control among patients with Type 2 diabetes: A longitudinal study. BMC Public Health 2005;5(36). (7) Yki-Järvinen H, Ryysy L, Kauppila M, Kujansuu E, Lahti J, Marjanen T, et al. Effect of
obesity on the response to insulin therapy in noninsulin-dependent diabetes mellitus. J Clin Endocrinol Metab 1997 Dec;82(12):4037-43.
(8) Ali MK, McKeever Bullard K, Imperatore G, Barker L, Gregg EW. Characteristics associated with poor glycemic control among adults with self-reported diagnosed diabetes--National Health and Nutrition Examination Survey, United States, 2007-2010. MMWR Morb Mortal Wkly Rep 2012 Jun 15;61:32-7.
(9) Kollannoor-Samuel G, Chhabra J, Fernandez ML, Vega-López S, Pérez SS, Damio G, et al. Determinants of fasting plasma glucose and glycosylated hemoglobin among low income Latinos with poorly controlled type 2 diabetes. J Immigr Minor Health 2011 Oct;13(5):809-17.
(10) Shani M, Taylor TR, Vinker S, Lustman A, Erez R, Elhayany A, et al. Characteristics of diabetics with poor glycemic control who achieve good control. J Am Board Fam Med 2008 Nov;21(6):490-6.
24
(11) Cosansu G, Erdogan S. Influence of psychosocial factors on self-care behaviors and glycemic control in Turkish patients with type 2 diabetes mellitus. J Transcult Nurs 2014 Jan;25(1):51-9.
(12) Beard E, Clark M, Hurel S, Cooke D. Do people with diabetes understand their clinical marker of long-term glycemic control (HbA1c levels) and does this predict diabetes self-care behaviours and HbA1c? Patient Educ Couns 2010 Aug;80(2):227-32.
(13) Gardner DF, Eastman BG, Mehl TD, Merimee TJ. Effect of psychosocial factors on success in a program of self-glucose monitoring. Diabetes Res 1985 Mar;2(2):89-93. (14) Swinnen SG, DeVries JH. Contact frequency determines outcome of basal
insulininitiation trials in type 2 diabetes. Diabetologia 2009;52:2324-7.
(15) Simon AC, Holleman F, Hoekstra JB, De Clercq PA, Lemkes BA, Hermanides J, et al. Development of a web-based decision support system for insulin self-titration. Stud Health Technol Inform 2011;169:103-7.
(16) Tildesley HD, Mazanderani AB, Ross SA. Effect of Internet Therapeutic Intervention on A1C Levels in Patients With Type 2 Diabetes Treated With Insulin. Diabetes Care 2010 Aug;33(8):1738-40.
(17) Kwon HS, Cho JH, Kim HS, Lee JH, Song BR, Oh JA, et al. Development of web-based diabetic patient management system using short message service (SMS). Diabetes Research and Clinical Practice 2004 Dec;66:133-7.
(18) Noh JH, Cho YJ, Nam HW, Kim JH, Kim DJ, Yoo HS, et al. Web-Based Comprehensive Information System for Self-Management of Diabetes Mellitus. Diabetes Technology & Therapeutics 2010 May;12(5):333-7.
(19) McMahon GT, Gomes HE, Hohne SH, Hu TMJ, Levine BA, Conlin PR. Web-based care management in patients with poorly controlled diabetes. Diabetes Care 2005;28(7):1624-9.
(20) Pal K, Eastwood SV, Michie S, Farmer AJ, Barnard ML, Peacock R, et al. Computer-based diabetes self-management interventions for adults with type 2 diabetes mellitus (Review) . The Cochrane Library 2013;(3).
(21)
http://www.cbs.nl/nl-NL/menu/themas/bedrijven/publicaties/digitale-economie/artikelen/2012-3636-wm.htm [online source]. CBS Webmagazine; 2012
[updated 15 June 2012; cited 15 may 2014]
(22) Lorig K, Ritter PL, Laurent DD, Plant K, Green M, Jernigan VB, et al. Online diabetes self-management program: a randomized study. Diabetes Care 2010 Jun;33(6):1275-81.
(23) Simon ACR, Holleman F, Gude WT, Hoekstra JBL, Peute LW, Jaspers MWM, et al. Safety and usability evaluation of a web-based insulin self-titration system for
25
patients with type 2 diabetes mellitus. Artificial Intelligence in Medicine 2013;59:23-31.
(24) Bindraban NR, van Valkengoed IGM, Mairuhu G, Holleman F, Hoekstra JBL, Michels BPJ, et al. Prevalence of diabetes mellitus and the performance of a risk score among Hindustani Surinamese, African Surinamese and ethnic Dutch: a cross-sectional population-based study. BMC Public Health 2008 Aug 1;8(271).
(25) Van Der Ven NC, Weinger K., Yi J, Pouwer F, Ader H, Van Der Ploeg HM, et al. The confidence in diabetes self-care scale: psychometric properties of a new measure of diabetes-specific self-efficacy in Dutch and US patients with type 1 diabetes. Diabetes Care 2003 Mar;26(3):713-8.
