Emergency attendance for acute hyper- and hypoglycaemia in the adult diabetic population of the metropolitan area of Milan: quantifying the phenomenon and studying its predictors

All subjects with DM residing in the territory of the Milan Agency for Health Protection (AHP) from 1 January 2015 to 31 December 2017 were identified using current healthcare databases and algorithms defined by the Lombardy Region (Regional Decision No. 6164 and No. 7655) (Additional file 1, Table A1) [23]. Subjects who were not registered with the regional universal coverage healthcare system could not be identified. Therefore, temporary residents in the area, who were either registered with another regional system or not registered at all in the NHS, were excluded. Gestational diabetes was also excluded from the algorithm. We included patients with a DM diagnosis, both type 1 and type 2. We excluded subjects below the age of 18 and patients who were not pharmacologically treated. Patients were followed-up for acute glycaemic events until 31 August 2018 through the ED database of the AHP, including the healthcare consumption of resident patients outside the area. Data anonymization was provided by an internal code used in every administrative database and was used for deterministic record linkage on a unique identifier.

Demographic variables, including age and gender, were extracted from the current healthcare databases of the AHP. The socioeconomic deprivation index quintile of the census section, calculated from data provided by the National Institute of Statistics (ISTAT) and normalized to the AHP, was assigned based on the address of residence [24]. The anti-diabetic treatment was determined using the drug prescription database and categorized as: insulin alone, insulin and non-insulin anti-DM drugs, non-insulin anti-DM drugs including at least one at risk for hypoglycaemia, and other non-insulin anti-DM drugs only. ATC codes included in each category are reported in Additional file 1, Table A3. From the current healthcare databases, we also determined the total number of comorbidities and the presence of chronic renal failure (CKD) and peripheral neuropathy (PN). The databases, algorithms, and codes used to define comorbidities are described in Additional file 1, Table A3. From the drug prescription database, we calculated the duration of insulin treatment, dichotomized as less vs. equal or greater than 5 years (for the algorithm, see Additional file 1, Table A3). For each individual, we also calculated two indicators of adequate glycaemic monitoring: undergoing or not undergoing glycated haemoglobin (HbA1c) level testing at least once a year, and receiving a number of glucose test strips every year from the regional healthcare service considered as adequate for the type of treatment, according to national guidelines (algorithms described in Additional file 1, Table A3).

Attendances to the ED for acute glycaemic events were identified for the cohort, through deterministic record linkage with the ED database of the AHP. We selected ED records having ICD-CM codes of acute hypo or hyperglycaemia in the diagnosis field; version 9, which is the version currently adopted in Italy, was used. To determine the codes, a validated algorithm to identify hypoglycaemia from claims was reviewed [25], and the codes were abstracted (Additional file 1, Table A2). No validated algorithm for hyperglycaemia was found, so potential codes were individuated from the ICD-9-CM codes list (Additional file 1, Table A2). Also, as it is the use of coding partially specific to a healthcare system, we validated the use of the identified ICD-9-CM codes in the ED database of the AHP, by means of reviewing the full clinical records of hospital admission following ED attendance in a random sub-sample of the cohort, as described in Additional file 1, additional methods. The codes selected after this validation were used to identify acute glycaemic events, overall and distinguished, in hypo and hyperglycaemia, for the study cohort. After identifying the events, we determined which attendances to the emergency department were followed by a hospital admission, using a specific variable present in the ED dataset.

The characteristics of the cohort were described using absolute numbers and percentages, overall, and stratifying over ED attendances for an acute glycaemic event, hypoglycaemia, and hyperglycaemia (excluding those leading to DM diagnosis). Groups were compared using the χ² test. Costs were described using mean, median, standard deviation (s.d.), and quartiles (Q).

Factors associated with the frequency of ED attendances for hypoglycaemia were first assessed using a Poisson model. Due to overdispersion and the excess of patients not experiencing the event with respect to a Poisson distribution, negative binomial (NB), zero-inflated Poisson (ZIP), and zero-inflated negative binomial (ZINB) regression models were also fitted [27, 28]. To select the best fitting model, we used the Vuong test [29]. To allow for the skewed distribution of the duration of the follow-up, we included in each model, an offset variable (natural logarithm of the individual total follow-up time) [28, 30]. Demographic and treatment variables potentially associated with hypoglycaemia risk from the literature, and available through the current healthcare databases, were included in all the multivariable regression models [3, 5, 6, 17, 19, 22, 26]. The same variables were included also in the ZIP and ZINB inflate component of the models, which predict whether a patient would be an excess zero. The same approach was adopted to evaluate the potential factors associated with ED attendances for hyperglycaemia, excluding those related to the diagnosis. The results of the models are presented as an OR of being an excess zero (i.e. not being susceptible to the event) for the inflate portion, and an IRR of the event for the count part, along with their corresponding 95% confidence intervals (CI). We also evaluated the association between demographic and treatment characteristics and the incidence of hospitalization after an ED attendance in patients having at least one acute hypoglycaemic event, and in those who experienced at least one hyperglycaemic event. In the absence of overdispersion and excess of non-events, a multivariable Poisson model was adopted for both, and the results were presented as IRRs along with their corresponding 95% CI. All analyses were performed using SAS software (v 9.4, SAS Institute Inc., Cary, NC, USA).

The ED attendances for acute glycaemic events identified with the provisional ICD-9-CM codes (Additional file 1, Table A2) were n = 2510; five of the initially selected codes were never found in the ED database (250.81, 250.83, 250.90, 250.91, 251.1). The hospital stratified random sample of ED attendances, followed by hospitalization, included 291 records. The full clinical chart, including the ED report, was available and reviewed for 261 cases (89.7% of sampled records). The characteristics of the sample and its comparability with the whole cohort are presented in Additional file 1, Table A4. From the evaluation of the clinical records, 10 hospitalizations were performed for reasons other than an acute glycaemic alteration (False discovery rate FDR = 3.8%). Of the remaining 251 episodes, 117 were hypoglycaemic (46.6%) and 134 were hyperglycaemic (53.4%). Of the latter, 27 (10.3% of examined records) were hospitalizations following the first diagnosis of diabetes during the ED attendance, and in these records, the 80th percentile of the distribution of the time interval between the date of diagnosis in the database of chronic disease and the date of the ED in the sample was 15 days. Among patients with hypoglycaemia, 2.8% were treated with glucagon and 91.6% with glucose and/or glucose solution i.v. in the emergency room. Among patients with hyperglycaemia, insulin was administered in the emergency room in 83.7% of cases. Three-point-5 % of patients with hypoglycaemia and 10.4% of those with hyperglycaemia, were admitted to an intensive care unit. Six-point-1 % of hospitalizations for hypoglycaemia and 4.8% for hyperglycaemia, led to the death of the patient.

All provisional codes had a FDR for acute glycaemic events lower than 30% (Additional file 1, Table A2); consequently, they were all retained in the final algorithm to identify ED attendances for acute glycaemic events. Three codes had a Positive Predictive Value (PPV) lower than 80%: “250.82 Diabetes with other specified manifestations, type II or unspecified type, uncontrolled”, PPV for hyperglycaemia 60%; “276.2 Acidosis”, PPV for hyperglycaemia 56%, and “790.29 Other abnormal glucose” PPV for hyperglycaemia 75. These codes had identified 38 records in the sample (14.6% of the reviewed clinical records). They were not included in the final algorithms used to identify distinctively hyper or hypoglycaemia events in the cohort (Additional file 1, Table A2).

A total of 210,425 persons with DM resided in the Milan Metropolitan area between 2015 and 2017. Age-standardized (reference population ESP 2013 [31]) prevalence of DM at 31 December 2017 was 4.9% (5.7% in men and 4.1% in women). In 2017, the age-standardized incidence of DM was 36.6 for 10,000 residents (38.3 in men and 34.9 in women). For insulin-dependent diabetes (IDD), the incidence was 4.9 for 10,000 in both genders. Excluding paediatric subjects (n = 990) and untreated adult patients (n = 41,150), the total number of subjects included in the analysis was 168,285 (Fig. 1). Almost seventy percent of the cohort was older than 64 years, 56% were male and half resided in a census section included in the two most deprived quintiles (Table 1). Nearly 6% of the cohort had CKD, and 16% had PN. Slightly over a quarter of individuals used insulin, and 43% for at least 5 years. Table 2 describes the characteristics of the subject experiencing at least one attendance to the emergency department for hypoglycemia and for hyperglycemia.

The total number of ED attendances of the cohort for any cause between 1st January 2015 and 31st August 2018 was 270,594 (mean number of events per subject: 1.61, sd: 2.64) in 99,174 subjects (58.9% of the cohort), with an incidence of 0.5 per patient-year. One-point-9 % of the cohort (n = 3164 subjects) experienced 3719 acute glycaemic events leading to an ED attendance (mean number of events per subject: 0.02, sd: 0.18), with an incidence of 7.0 per 1000 patient-years. Of those, 967 attendances were followed by hospitalization (26.0%). Excluding ED attendances within 15 days from DM diagnosis, i.e. presumably related to diagnosis (n = 263, 7.1%), the incidence was 6.5 per 1000 patient-years. The validated coding algorithms revealed that, for 348 events, a definitive characterization as hyper or hypoglycaemia was not possible (9.3% of all acute glycaemic events). Excluding not characterized and related to diagnosis ED attendances, 48 patients (1.7%) experienced hyper and hypoglycaemia events during follow-up. Those subjects were included in both analyses.

In the period between 2015 and 2018, the total patient annual cost of ED attendances for all acute glycaemic events was 174,544 euros, with a median cost per attendance of 137 euro (Q1, Q3 = 87–233 euros). The total patient annual cost for acute hypoglycaemia was 108,648 euros, with a median cost per attendance of 127 euros (Q1, Q3 = 83–227 euro). For acute hyperglycaemia, the cost was 44,697 euros with a median cost per attendance of 142 euros (Q1, Q3 = 92–226 euros). No relevant differences were detected between subjects treated with insulin and those who were not (data not shown). The median total direct annual costs for patients having at least one ED attendance between 2015 and 2017 were 2.3 times higher than those of patients not experiencing any ED attendances (median = 2533 and Q1, Q4 = 1140–5931 vs. median = 5839 and Q1, Q3 = 3052–11,757 euro, Table 3). Costs were higher, both for subjects without and with ED attendances, among elderly patients, if comorbidities were present and if they were treated with insulin. Patients with these characteristics were more represented among patients with ED attendances (Table 1), particularly 68% of patients experiencing ED attendances for an acute glycaemic event were treated with insulin vs. 25% among those who were not.

There were 2510 Hypoglycaemic events (ICD-9-CM codes: 251.0, 251.2, 962.3) leading to ED attendance, experienced by 2137 patients, with an incidence of 4.7 per 1000 patient-years. The strongest positive univariate associations between hypoglycaemia and patient characteristics were found for (Table 4): 1) having had an ED attendance for hypoglycaemia in 2014 (17.3% of subjects experiencing hypoglycaemia vs. 1.2% among subjects without ED attendances); 2) being treated with insulin (4.1% among those treated with insulin alone and 2.7% among those treated also with other anti-DM drugs vs. 0.2% for non-insulin anti-DM not at risk for hypoglycaemia); 3) having chronic renal failure (3.4% vs. 1.1% among subjects without CKD). For the hypoglycaemic events, the best fitting multivariable regression model was the zero-inflated negative binomial. Younger patients had a higher probability of not being susceptible to ED attendances for hypoglycaemia (OR = 2.21, 95% CI = 1.14–4.27 for ≤44 year-olds and OR = 2.26, 95% CI = 1.24–4.13 for 45–54 year-olds), but if susceptible, had a higher frequency of events than 55–64 year-olds (IRR = 1.93, 95% CI = 1.21–3.08 and 2.21, 95% CI = 1.41–3.46). The opposite was true for older patients (Table 4). The presence of CKD or peripheral neuropathy reduced the probability of not being susceptible to ED attendances for hypoglycaemia (OR = 0.37, 95% CI = 0.22–0.63 for CKD and OR = 0.67, 95% CI = 0.47–0.96 for PN) but did not influence the IRR. Compared to being treated only with non-insulin anti-DM drugs not at risk for hypoglycaemia, all other treatment modalities were strongly associated with a lower probability of not being susceptible to ED attendances for hypoglycaemia (OR = 0.01, 95% CI = 0.00–0.02 for insulin alone to OR = 0.09, 95% CI = 0.04–0.19 for other anti-DM drugs at risk for hypoglycaemia); while, no statistically significant effect was seen for event rate. Having had an ED attendance for hypoglycaemia in 2014 was associated both with a reduced probability of not being susceptible to ED attendances during the study period (OR = 0.20, 95% CI = 0.07–0.53) and an increased event frequency (IRR = 5.34, 95% CI = 3.93–7.26). Subjects treated with insulin for less than 5 years had an increased probability of not being susceptible to ED attendances for hypoglycaemia (OR = 1.82, 95% CI = 1.25–2.65) and a reduced event frequency (IRR = 0.62, 95% CI = 0.52–0.74), compared to those treated for 5 years or more.

The percentage of hospitalization after an ED attendance for hypoglycaemia was 15.3%. The results of the multivariable Poisson model, which evaluated the risk of hospitalization after an ED attendance for hypoglycaemia, are also reported in Table 4. The factors that increased the incidence of hospitalization after an ED attendance for acute hypoglycaemia were: having CKD (IRR = 1.42, 95% CI 1.10–1.84) and not having received an adequate number of glucose test strips (IR = 1.28, 95% CI 1.03–1.59). In contrast, being treated with insulin reduced the risk of hospitalization by more than 50% compared to being treated only with non-insulin anti-DM drugs (IRR = 0.45, 95% CI 0.30–0.68 for insulin alone).

There were 861 Hyperglycaemic events (ICD-9-CM codes: 250.1, 250.2, 250.3, 250.80, 250.92, 250.93, 276.0, 276.4, 790.21) leading to ED attendances, experienced by n = 788 patients, with an incidence of 1.63 per 1000 patient-years. Excluding patients close to diagnosis, the events were n = 689, experienced by n = 633 patients, with an incidence of 1.30 per 1000 patient-years.

The strongest positive univariate associations between hyperglycaemia and patient characteristics were found for (Table 5): 1) having had an ED attendance for hyperglycaemia in 2014 (6.6% of subjects in the cohort experienced hyperglycaemia among those with vs. 0.4% among subjects without an ED attendance in 2014 for hyperglycaemia); 2) being treated with insulin (1.1% among those treated with insulin alone vs. 0.1% for non-insulin anti-DM not at risk for hypoglycaemia); 3) having chronic renal failure: 0.9% vs. 0.3% among subjects without CKD. For the hyperglycaemic acute events, the best fitting multivariable regression model was the zero-inflated negative binomial (Table 5). Patients older than 56 years had a lower probability of not being susceptible to ED attendances for hyperglycaemia (OR = 0.22, 95% CI = 0.08–0.57), but if susceptible, had a lower frequency of events than 45–64-year-old patients (IRR = 0.24, 95% CI = 0.12–0.52). The treatment modality did not have a statistically significant effect on the probability of being susceptible to hyperglycaemia. At the same time, the use of insulin, alone or with other anti-DM drugs, increased the frequency of events compared to being treated only with non-insulin anti-DM drugs not at risk for hypoglycaemia (IRR = 7.72, 95% CI = 5.11–11.66 for insulin alone and IRR = 8.01, 95% CI = 5.21–12.32 for insulin with other anti-DM drugs). Failure to perform a glycated haemoglobin test every year reduced the probability of not being susceptible to ED attendances for hyperglycaemia (OR = 0.65, 95% CI 0.42–0.99) and increased the event rate (IRR = 1.36, 95% CI 1.07–1.72), compared to individuals having performed it. Furthermore, having had an ED attendance for hyperglycaemia in 2014 increased the event frequency (IRR = 3.68, 95% CI = 1.33–10.23). The percentage of hospitalization after an ED attendance for hypoglycaemia was 34.8%. The results of the multivariable Poisson model, which evaluated the risk of hospitalization after an ED attendance for hyperglycaemia, are also reported in Table 5. The factors that increased the incidence of hospitalization after an ED attendance for acute hyperglycaemia were being treated with insulin, doubling the risk (IR = 2.00, 95% CI 1.31–2.05), and not having performed a yearly HbA1c test (IR = 1.37, 95% CI 1.10–1.72).