Background
Type 2 diabetes is one of the most prevalent chronic diseases worldwide, especially among older people aged ≥75 years, in whom prevalence reached 20% in 2017, and is poised to increase over the coming decades [
1]. In Europe, the cost per patient per year with diabetes mellitus was estimated at US Dollar 3,100 in 2017. Moreover, diabetes was responsible for 10% of total health care expenditure in 2010 [
2]. Diabetes in older patients has therefore a major impact on healthcare systems.
Current classification of diabetes mellitus considers 4 different categories: type 1 diabetes, type 2 diabetes, gestational diabetes and specific rare types of diabetes [
3]. In older age, type 2 diabetes is reported to represent 85–90% of all-cause diabetes, ahead of type 1 diabetes, which includes latent autoimmune diabetes in adults [
4,
5].
Type 2 diabetes induces specific acute or chronic complications (e.g. microvascular complications from chronic hyperglycaemia) and increases the incident risk of macrovascular complications from various cardiometabolic abnormalities promoting the occurrence of atherosclerosis [
6]. These vascular complications promote and intensify the development of several geriatric syndromes in older patients, such as falls, polymedication, cognitive disorders or sensorial disorders [
6,
7]. The aim of glucose lowering therapy (GLT) in these patients is to control hyperglycaemia and its associated morbidity and mortality. Nevertheless, in older patients with type 2 diabetes, GLT should be adapted according to patient’s characteristics in order to be intense enough to avoid microvascular complications but light enough to prevent potential side-effects of GLT, mainly hypoglycaemia, as it also increases morbidity and mortality [
7]. These considerations offer only a narrow frame to perform a safe and effective GLT management in patients aged 75 years or more with type 2 diabetes. Several recent guidelines provide recommendations about GLT management in older patients with diabetes, in terms of hyperglycaemia, risk factors and complications [
8,
9]. These guidelines and other reports all insist on treatments’ individualisation in order to give tailored medication for each patient [
8,
10‐
15]. At present, factors currently considered in this treatment individualisation are related to the type of diabetes [
3], but also to prevalent comorbidities, geriatric syndromes, nutrition issues, physical activity, age-specific aspects of pharmacotherapy, ethnic disparities and estimated life expectancy [
8,
11].
Indeed, type 2 diabetes is a complex condition with marked heterogeneity in pathophysiological mechanisms leading to hyperglycaemia and cardiometabolic comorbidities between patients. Ageing process enhances this heterogeneity, adding other conditions, such as nutritional deficits, sarcopenia, additional stresses on pancreatic beta-cells and micro-inflammation [
16‐
18].
Yet, current guidelines for older patients do not suggest taking into account characteristics related to the pathophysiology of diabetes or severity of residual beta-cell function (BCF) loss. Therefore, it is of interest to consider these factors in GLT individualisation in order to improve the quality, efficacy and safety of GLT management in older patients.
Therefore, the aim of the present study was to assess the heterogeneity of cardiometabolic features in patients aged 75 years or more with type 2 diabetes and to classify them into relevant cardiometabolic profiles using mixture models as Latent Profile Analysis (LPA).
Methods
Study design and patient selection
A retrospective cohort study of outpatients followed by the same investigator (MPH) between 2000 and 2017 and attending a Belgian university diabetes clinic was conducted. Among the 266 Caucasian patients followed in the diabetes clinic and aged ≥75 years at the last two visits to the endocrinologist, 147 participants had a Homeostasis Model Assessment (HOMA2) after the diagnosis of their type 2 diabetes. All 147 participants were GAD-antibodies-negative. Type 2 diabetes was defined according to the Expert Committee on the Diagnosis and Classification of Diabetes Mellitus [
3].
This study was approved by the local Ethics Committee (Commission d’Ethique Hospitalo-Facultaire, Cliniques universitaires Saint-Luc, Brussels, Belgium; ref. B403/2017/16NOV/521).
Data collection
A first part of the data was collected at the time of the HOMA2 assessment. Data included anthropometric (weight, body mass index and fat mass proportion), biochemical (HbA1c) and ongoing GLT (drug molecules and doses).
Body mass index (BMI; kg/m
2) was calculated as [Weight(kg) × Height(m)
− 2]. Body fat mass (%) was measured using a BodyFat Analyser (Omron BF 500; Omron Healthcare Europe B.V., Hoofddorp, The Netherlands). HbA1c was expressed in NGSP nomenclature (%) and was converted to IFCC nomenclature (mmol/mol) using the NGSP convertor (
www.ngsp.org/convert1.asp).
Insulin sensitivity and beta-cell function were assessed using the computer-based homeostasis model assessment (HOMA2,
http://www.dtu.ox.ac.uk
) [
19]. HOMA2 parameters were calculated from triplicates of fasting glucose and insulin level, sampled after a sufficient period of GLT washout (i.e. between 1 to 5 days, according to the molecules involved). Values of insulin secretion (
HOMA2%-β (%); normal 100%) were plotted as a function of insulin sensitivity (
HOMA2%-S (%); normal 100%), defining a
hyperbolic product area (
HOMA2%-βxS (%2); normal 100%). This product described the interaction between insulin sensitivity and insulin secretion, or more precisely, the true latent beta-cell function (BCF) indexed by insulin sensitivity. It approximates the magnitude of glucose homeostasis deficit and the required GLT intensity [
20].
GLT data corresponded to the treatment taken one week before the HOMA2 realization. Drugs were transcribed into Anatomical Therapeutic Chemical (ATC) codes and grouped by GLT classes (A10A-Insulin, A10BA-biguanides, A10BB-sulfonylureas, A10BF-alpha-glucosidase-inhibitors, A10BG-thiazolidinediones, A10BH-DPP4-inhibitors, A10BJ-GLP1-receptor agonists, A10BX-other). Sulfonylureas and repaglinide were considered as “
Oral hypoglycaemic agents (OHA)” and insulin and OHA were considered as “
Hypoglycaemic agents (HA)”. Patients with no GLT were considered as “
Lifestyle changes only”. Treatment doses were collected and converted into Defined Daily Dose (DDD), according to the ATC/DDD Index 2018 [
21]. For each patient, a sum of the GLT drugs doses, expressed in DDD, was computed and described hereafter as “
GLT’s total doses”.
The second part of the data was collected at the time of the last consultation at the diabetes clinic, at which all patients were ≥ 75 years, and included socio-demographic (age, sex) and diabetes-related data (age at diabetes diagnosis, comorbidities, vascular complications). Micro-angiopathic complications were defined as: neuropathy (clinical examination of knee and ankle reflexes; Semmes-Weinstein monofilament test, confirmed by lower-limbs electromyography) and diabetic retinopathy (based on retinal examination by an experienced ophthalmologist and/or fluorescein angiography). Diabetic nephropathy was not taken into account in this study because of its high prevalence in older age and its multiple aetiologies that cannot be attributed de facto to chronic hyperglycaemia.
Macro-angiopathic complications included coronary artery disease (CAD: myocardial infarction, angioplasty, stenting, revascularization surgery and/or significant coronary stenosis confirmed by angiography), cerebrovascular disease (CVD) or peripheral artery disease (PAD). CVD was defined according to
UK Prospective Diabetes Study criteria: any neurological deficit ≥1 month, without distinction between ischemic, embolic and haemorrhagic events [
22]. PAD was diagnosed from medical history of lower-limb claudication; clinical or imaging evidence for ischemic diabetic foot; history of angioplasty, stenting, revascularization surgery; and/or lower-limb artery stenosis at Doppler ultrasonography or angiography.
Statistical analysis
Continuous data were expressed as medians (P25, P75). Categorical data were expressed as number of people and percentages. Continuous variables were compared between 2 groups using Mann Whitney test, and between ≥3 groups using Kruskal-Wallis test. Categorical variables were compared between groups using Pearson’s χ2 test, Pearson’s χ2 test with Yates correction, Fisher’s exact test or Fisher Freeman Halton’s test, according to the conditions of validity of each test.
In order to identify profiles of patients with type 2 diabetes a latent profile analysis (LPA) was performed using the following continuous discriminant variables (indicators): insulin sensitivity (
HOMA2%-S), BCF (
HOMA2%-β), hyperbolic product
βxS (
HOMA2%-βxS) and age at diabetes diagnosis. Models with 2 to 7 profiles were ran. Evaluative information was used to select the best model, e.g. the model with the lowest Akaike information criteria, Bayesian Information Criterion (BIC) and Log Likelihood (LL) [
23]. In addition, the likelihood ratio test was used to compare a model with k-1 profiles with a model with k profiles. Finally, posterior probabilities, i.e. the probability of each patient of belonging to each profile, were computed for the final selected model. An average posterior probability per group ≥0.70 was used to consider whether profiles were sufficiently separated from each other.
Statistical analyses were performed using IBM SPSS Statistics 25® software or R software (R × 64 version 3.4.1). A p-value< 0.05 was considered statistically significant.
Discussion
The aim of the present study was to classify older patients with type 2 diabetes into profiles using a LPA methodology based on their metabolic features, in order to select more appropriate GLT in terms of their diabetes attributes and metabolic phenotype, and doing so to add another dimension to treatment individualisation [
8] based on diabetes characteristics.
The indicators used as discriminant variables input for LPA were selected on the basis of recent literature. First, as suggested in several studies, age at diabetes’ diagnosis is a major determinant of metabolic differences. Cardio-metabolic profile is usually less severe in patients with an older age at diabetes diagnosis than in those who are diagnosed younger. The former have lower HbA1c, fasting plasma glucose, fasting insulin, insulin resistance, triglyceride levels, LDL-cholesterol, BMI, obesity prevalence and family history of diabetes [
24‐
26]. Patients diagnosed with diabetes at an older age also have a lower risk of developing diabetic retinopathy, regardless of known diabetes duration [
27]. This suggests that their diabetes might have a lower propensity of generating microvascular complications.
Furthermore,
HOMA2%-S and
HOMA2%-β were used in order to distinct patients in terms of intrinsic glucose homeostasis characteristics, allowing to better select among GLT alternatives. One advantageous feature of our model is to have
HOMA2%-βxS among input variables, bringing essential information on residual BCF to better identify patients whose needs and intensity of GLT escalation are more marked [
19].
The use of these indicators allowed classifying patients into six distinct profiles. It highlights important phenotypic differences across patients sharing a common and seemingly unambiguous diagnosis of type 2 diabetes. Firstly, patients of profiles 1 and 2 had both the highest age at diabetes diagnosis combined with the highest
βxS, whereas profile 6 patients had the youngest age at diabetes diagnosis and the lowest
βxS. A link seems to exist between age at diabetes diagnosis and magnitude of glucose homeostasis’ impairment, as shown in previous studies [
24,
25]. This also implies that patients with an older age at diabetes diagnosis may need less intensive GLT, in terms of dose and drug of choice (e.g. use of hypoglycaemic agent). Inappropriate prescribing of hypoglycaemic agents in patients with late-onset type 2 diabetes may induce severe hypoglycaemic events.
Secondly, cardiometabolic risk, as shown by indices of insulin resistance, macrovascular comorbidities and BMI was very different between profiles. Some patients’ profiles had lower BMI, lower insulin resistance and few macrovascular complications (e.g. profile 1), while other profiles had higher values of these variables (e.g. profile 6). Profiling older patients with type 2 diabetes thus confirms the rationale of bringing under control modifiable risk factors taking into account the cardiometabolic risk profile for the corresponding profile of individual patients.
The LPA method used allowed for distinguishing patients based on degree of insulin resistance and/or BCF loss. The quantification of these variables provides useful information to individualise GLT (e.g. hypoglycaemic agents when impaired BCF is the major driver of hyperglycaemia or biguanides when insulin resistance is in the foreground). This is all the more relevant given the absence of phenotypic overlap of different profiles of type 2 diabetes in older patients.
The strengths of the present study are twofold. First, all patients were followed by the same endocrinologist and data were prospectively collected by one dedicated clinician. This allows for standardization of all bioclinical measurements, increases as such data’s quality and accuracy. Second, the HOMA2 was based on triplicates of fasting glucose and insulin levels sampled after a sufficient period of GLT washout. However, this sample of patients, most of whom Caucasians from a well-off Brussels suburb, was followed at a single-centre diabetes clinic, and may not de facto be representative of other populations of older patients with type 2 diabetes of various ethnicities.
Recently, Ahlqvist et al. provide a refined classification of diabetes using a data-driven cluster analysis [
28], realised on a large cohort of Swedish patients with diabetes (ANDIS cohort,
N = 8980) and replicated on three independent cohorts (
N = 5795). It classified patients into five clusters. Despite some similarities in the aims and variables chosen to classify patients, the study of Ahlqvist et al. differed from the present study in many ways.
First, the data used in ANDIS cohort were collected on incident cases at the time of the diabetes diagnosis (median time at inclusion = 40 days after diagnosis) in adult patients aged from 18 to 96 years, with a mean age at diagnosis of 60.2 years. Our study included prevalent cases of patients diagnosed with type 2 diabetes ≥75 years, with a median age at diagnosis of 62.0 years. Then, the inclusion criteria of Ahlqvist et al. were not restricted to type 2 diabetes but included all types of diabetes. The analytical method was a data-driven clustering, a classification method based on different theoretical approach as compared to latent profile analysis used in the present study. Finally, Ahlqvist et al. used six variables classifying patients into subgroups: three were identical to those used in the present study (HOMA2%-β, HOMA2%-S and age at diabetes onset), while two were not used (body mass index (BMI), GAD-antibodies and HbA1c). In the present study, BMI was not used, as it is not an optimal measure for obesity in older patients [
29]. GAD-antibodies were not used, as the present study included only patients with type 2 diabetes. Regarding HbA1c, the present study used HOMA2%-βxS instead, assessing the blood glucose control in patients taking glucose lowering therapies.
In the future, it might be of interest to assess the reproducibility of this study by increasing the number of patients, by recruiting older patients with diabetes followed by general practitioners and/or by running a study with a prospective design. It would allow predicting whether patients are ascribed to their appropriate profile and, accordingly, to propose therapeutic recommendations based on the patient’s cardiometabolic profile, keeping in mind that such recommendations could only serve as complements to existing criteria for standards of care individualisation and current guidelines [
8].
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.