Predicting glycated hemoglobin levels in the non-diabetic general population: Development and validation of the DIRECT-DETECT prediction model - a DIRECT study

Scientific Advances in Diabetes: The Impact of the Innovative Medicines Initiative

Diabetes Mellitus is one of the World Health Organization's priority diseases under research by the first and second programmes of Innovative Medicines Initiative, with the acronyms IMI1 and IMI2, respectively. Up to October of 2019, 13 projects were funded by IMI for Diabetes & Metabolic disorders, namely SUMMIT, IMIDIA, DIRECT, StemBANCC, EMIF, EBiSC, INNODIA, RHAPSODY, BEAT-DKD, LITMUS, Hypo-RESOLVE, IM2PACT, and CARDIATEAM. In general, a total of €447 249 438 was spent by IMI in the area of Diabetes. In order to prompt a better integration of achievements between the different projects, we perform a literature review and used three data sources, namely the official project's websites, the contact with the project's coordinators and co-coordinator, and the CORDIS database. From the 662 citations identified, 185 were included. The data collected were integrated into the objectives proposed for the four IMI2 program research axes: (1) target and biomarker identification, (2) innovative clinical trials paradigms, (3) innovative medicines, and (4) patient-tailored adherence programmes. The IMI funded projects identified new biomarkers, medical and research tools, determinants of inter-individual variability, relevant pathways, clinical trial designs, clinical endpoints, therapeutic targets and concepts, pharmacologic agents, large-scale production strategies, and patient-centered predictive models for diabetes and its complications. Taking into account the scientific data produced, we provided a joint vision with strategies for integrating personalized medicine into healthcare practice. The major limitations of this article were the large gap of data in the libraries on the official project websites and even the Cordis database was not complete and up to date.

Improving Current Glycated Hemoglobin Prediction in Adults: Use of Machine Learning Algorithms With Electronic Health Records

Background

Predicting the risk of glycated hemoglobin (HbA_1c) elevation can help identify patients with the potential for developing serious chronic health problems, such as diabetes. Early preventive interventions based upon advanced predictive models using electronic health records data for identifying such patients can ultimately help provide better health outcomes.

Objective

Our study investigated the performance of predictive models to forecast HbA_1c elevation levels by employing several machine learning models. We also examined the use of patient electronic health record longitudinal data in the performance of the predictive models. Explainable methods were employed to interpret the decisions made by the black box models.

Methods

This study employed multiple logistic regression, random forest, support vector machine, and logistic regression models, as well as a deep learning model (multilayer perceptron) to classify patients with normal (<5.7%) and elevated (≥5.7%) levels of HbA_1c. We also integrated current visit data with historical (longitudinal) data from previous visits. Explainable machine learning methods were used to interrogate the models and provide an understanding of the reasons behind the decisions made by the models. All models were trained and tested using a large data set from Saudi Arabia with 18,844 unique patient records.

Results

The machine learning models achieved promising results for predicting current HbA_1c elevation risk. When coupled with longitudinal data, the machine learning models outperformed the multiple logistic regression model used in the comparative study. The multilayer perceptron model achieved an accuracy of 83.22% for the area under receiver operating characteristic curve when used with historical data. All models showed a close level of agreement on the contribution of random blood sugar and age variables with and without longitudinal data.

Conclusions

This study shows that machine learning models can provide promising results for the task of predicting current HbA_1c levels (≥5.7% or less). Using patients’ longitudinal data improved the performance and affected the relative importance for the predictors used. The models showed results that are consistent with comparable studies.

Next-generation epidemiology: the role of high-resolution molecular phenotyping in diabetes research

Epidemiologists have for many decades reported on the patterns and distributions of diabetes within and between populations and have helped to elucidate the aetiology of the disease. This has helped raise awareness of the tremendous burden the disease places on individuals and societies; it has also identified key risk factors that have become the focus of diabetes prevention trials and helped shape public health recommendations. Recent developments in affordable high-throughput genetic and molecular phenotyping technologies have driven the emergence of a new type of epidemiology with a more mechanistic focus than ever before. Studies employing these technologies have identified gene variants or causal loci, and linked these to other omics data that help define the molecular processes mediating the effects of genetic variation in the expression of clinical phenotypes. The scale of these epidemiological studies is rapidly growing; a trend that is set to continue as the public and private sectors invest heavily in omics data generation. Many are banking on this massive volume of diverse molecular data for breakthroughs in drug discovery and predicting sensitivity to risk factors, response to therapies and susceptibility to diabetes complications, as well as the development of disease-monitoring tools and surrogate outcomes. To realise these possibilities, it is essential that omics technologies are applied to well-designed epidemiological studies and that the emerging data are carefully analysed and interpreted. One might view this as next-generation epidemiology, where complex high-dimensionality data analysis approaches will need to be blended with many of the core principles of epidemiological research. In this article, we review the literature on omics in diabetes epidemiology and discuss how this field is evolving. Graphical abstract.

The role of physical activity in metabolic homeostasis before and after the onset of type 2 diabetes: an IMI DIRECT study

AIMS/HYPOTHESIS

It is well established that physical activity, abdominal ectopic fat and glycaemic regulation are related but the underlying structure of these relationships is unclear. The previously proposed twin-cycle hypothesis (TC) provides a mechanistic basis for impairment in glycaemic control through the interactions of substrate availability, substrate metabolism and abdominal ectopic fat accumulation. Here, we hypothesise that the effect of physical activity in glucose regulation is mediated by the twin-cycle. We aimed to examine this notion in the Innovative Medicines Initiative Diabetes Research on Patient Stratification (IMI DIRECT) Consortium cohorts comprised of participants with normal or impaired glucose regulation (cohort 1: N ≤ 920) or with recently diagnosed type 2 diabetes (cohort 2: N ≤ 435).

METHODS

We defined a structural equation model that describes the TC and fitted this within the IMI DIRECT dataset. A second model, twin-cycle plus physical activity (TC-PA), to assess the extent to which the effects of physical activity in glycaemic regulation are mediated by components in the twin-cycle, was also fitted. Beta cell function, insulin sensitivity and glycaemic control were modelled from frequently sampled 75 g OGTTs (fsOGTTs) and mixed-meal tolerance tests (MMTTs) in participants without and with diabetes, respectively. Abdominal fat distribution was assessed using MRI, and physical activity through wrist-worn triaxial accelerometry. Results are presented as standardised beta coefficients, SE and p values, respectively.

RESULTS

The TC and TC-PA models showed better fit than null models (TC: χ² = 242, p = 0.004 and χ² = 63, p = 0.001 in cohort 1 and 2, respectively; TC-PA: χ² = 180, p = 0.041 and χ² = 60, p = 0.008 in cohort 1 and 2, respectively). The association of physical activity with glycaemic control was primarily mediated by variables in the liver fat cycle.

CONCLUSIONS/INTERPRETATION

These analyses partially support the mechanisms proposed in the twin-cycle model and highlight mechanistic pathways through which insulin sensitivity and liver fat mediate the association between physical activity and glycaemic control.

Discovery of biomarkers for glycaemic deterioration before and after the onset of type 2 diabetes: descriptive characteristics of the epidemiological studies within the IMI DIRECT Consortium

AIMS/HYPOTHESIS

Here, we describe the characteristics of the Innovative Medicines Initiative (IMI) Diabetes Research on Patient Stratification (DIRECT) epidemiological cohorts at baseline and follow-up examinations (18, 36 and 48 months of follow-up).

METHODS

From a sampling frame of 24,682 adults of European ancestry enrolled in population-based cohorts across Europe, participants at varying risk of glycaemic deterioration were identified using a risk prediction algorithm (based on age, BMI, waist circumference, use of antihypertensive medication, smoking status and parental history of type 2 diabetes) and enrolled into a prospective cohort study (n = 2127) (cohort 1, prediabetes risk). We also recruited people from clinical registries with type 2 diabetes diagnosed 6-24 months previously (n = 789) into a second cohort study (cohort 2, diabetes). Follow-up examinations took place at ~18 months (both cohorts) and at ~48 months (cohort 1) or ~36 months (cohort 2) after baseline examinations. The cohorts were studied in parallel using matched protocols across seven clinical centres in northern Europe.

RESULTS

Using ADA 2011 glycaemic categories, 33% (n = 693) of cohort 1 (prediabetes risk) had normal glucose regulation and 67% (n = 1419) had impaired glucose regulation. Seventy-six per cent of participants in cohort 1 was male. Cohort 1 participants had the following characteristics (mean ± SD) at baseline: age 62 (6.2) years; BMI 27.9 (4.0) kg/m²; fasting glucose 5.7 (0.6) mmol/l; 2 h glucose 5.9 (1.6) mmol/l. At the final follow-up examination the participants' clinical characteristics were as follows: fasting glucose 6.0 (0.6) mmol/l; 2 h OGTT glucose 6.5 (2.0) mmol/l. In cohort 2 (diabetes), 66% (n = 517) were treated by lifestyle modification and 34% (n = 272) were treated with metformin plus lifestyle modification at enrolment. Fifty-eight per cent of participants in cohort 2 was male. Cohort 2 participants had the following characteristics at baseline: age 62 (8.1) years; BMI 30.5 (5.0) kg/m²; fasting glucose 7.2 (1.4) mmol/l; 2 h glucose 8.6 (2.8) mmol/l. At the final follow-up examination, the participants' clinical characteristics were as follows: fasting glucose 7.9 (2.0) mmol/l; 2 h mixed-meal tolerance test glucose 9.9 (3.4) mmol/l.

CONCLUSIONS/INTERPRETATION

The IMI DIRECT cohorts are intensely characterised, with a wide-variety of metabolically relevant measures assessed prospectively. We anticipate that the cohorts, made available through managed access, will provide a powerful resource for biomarker discovery, multivariate aetiological analyses and reclassification of patients for the prevention and treatment of type 2 diabetes.

Development of type 2 diabetes mellitus in people with intermediate hyperglycaemia

BACKGROUND

Intermediate hyperglycaemia (IH) is characterised by one or more measurements of elevated blood glucose concentrations, such as impaired fasting glucose (IFG), impaired glucose tolerance (IGT) and elevated glycosylated haemoglobin A1c (HbA1c). These levels are higher than normal but below the diagnostic threshold for type 2 diabetes mellitus (T2DM). The reduced threshold of 5.6 mmol/L (100 mg/dL) fasting plasma glucose (FPG) for defining IFG, introduced by the American Diabetes Association (ADA) in 2003, substantially increased the prevalence of IFG. Likewise, the lowering of the HbA1c threshold from 6.0% to 5.7% by the ADA in 2010 could potentially have significant medical, public health and socioeconomic impacts.

OBJECTIVES

To assess the overall prognosis of people with IH for developing T2DM, regression from IH to normoglycaemia and the difference in T2DM incidence in people with IH versus people with normoglycaemia.

SEARCH METHODS

We searched MEDLINE, Embase, ClincialTrials.gov and the International Clinical Trials Registry Platform (ICTRP) Search Portal up to December 2016 and updated the MEDLINE search in February 2018. We used several complementary search methods in addition to a Boolean search based on analytical text mining.

SELECTION CRITERIA

We included prospective cohort studies investigating the development of T2DM in people with IH. We used standard definitions of IH as described by the ADA or World Health Organization (WHO). We excluded intervention trials and studies on cohorts with additional comorbidities at baseline, studies with missing data on the transition from IH to T2DM, and studies where T2DM incidence was evaluated by documents or self-report only.

DATA COLLECTION AND ANALYSIS

One review author extracted study characteristics, and a second author checked the extracted data. We used a tailored version of the Quality In Prognosis Studies (QUIPS) tool for assessing risk of bias. We pooled incidence and incidence rate ratios (IRR) using a random-effects model to account for between-study heterogeneity. To meta-analyse incidence data, we used a method for pooling proportions. For hazard ratios (HR) and odds ratios (OR) of IH versus normoglycaemia, reported with 95% confidence intervals (CI), we obtained standard errors from these CIs and performed random-effects meta-analyses using the generic inverse-variance method. We used multivariable HRs and the model with the greatest number of covariates. We evaluated the certainty of the evidence with an adapted version of the GRADE framework.

MAIN RESULTS

We included 103 prospective cohort studies. The studies mainly defined IH by IFG5.6 (FPG mmol/L 5.6 to 6.9 mmol/L or 100 mg/dL to 125 mg/dL), IFG6.1 (FPG 6.1 mmol/L to 6.9 mmol/L or 110 mg/dL to 125 mg/dL), IGT (plasma glucose 7.8 mmol/L to 11.1 mmol/L or 140 mg/dL to 199 mg/dL two hours after a 75 g glucose load on the oral glucose tolerance test, combined IFG and IGT (IFG/IGT), and elevated HbA1c (HbA1c5.7: HbA1c 5.7% to 6.4% or 39 mmol/mol to 46 mmol/mol; HbA1c6.0: HbA1c 6.0% to 6.4% or 42 mmol/mol to 46 mmol/mol). The follow-up period ranged from 1 to 24 years. Ninety-three studies evaluated the overall prognosis of people with IH measured by cumulative T2DM incidence, and 52 studies evaluated glycaemic status as a prognostic factor for T2DM by comparing a cohort with IH to a cohort with normoglycaemia. Participants were of Australian, European or North American origin in 41 studies; Latin American in 7; Asian or Middle Eastern in 50; and Islanders or American Indians in 5. Six studies included children and/or adolescents.Cumulative incidence of T2DM associated with IFG5.6, IFG6.1, IGT and the combination of IFG/IGT increased with length of follow-up. Cumulative incidence was highest with IFG/IGT, followed by IGT, IFG6.1 and IFG5.6. Limited data showed a higher T2DM incidence associated with HbA1c6.0 compared to HbA1c5.7. We rated the evidence for overall prognosis as of moderate certainty because of imprecision (wide CIs in most studies). In the 47 studies reporting restitution of normoglycaemia, regression ranged from 33% to 59% within one to five years follow-up, and from 17% to 42% for 6 to 11 years of follow-up (moderate-certainty evidence).Studies evaluating the prognostic effect of IH versus normoglycaemia reported different effect measures (HRs, IRRs and ORs). Overall, the effect measures all indicated an elevated risk of T2DM at 1 to 24 years of follow-up. Taking into account the long-term follow-up of cohort studies, estimation of HRs for time-dependent events like T2DM incidence appeared most reliable. The pooled HR and the number of studies and participants for different IH definitions as compared to normoglycaemia were: IFG5.6: HR 4.32 (95% CI 2.61 to 7.12), 8 studies, 9017 participants; IFG6.1: HR 5.47 (95% CI 3.50 to 8.54), 9 studies, 2818 participants; IGT: HR 3.61 (95% CI 2.31 to 5.64), 5 studies, 4010 participants; IFG and IGT: HR 6.90 (95% CI 4.15 to 11.45), 5 studies, 1038 participants; HbA1c5.7: HR 5.55 (95% CI 2.77 to 11.12), 4 studies, 5223 participants; HbA1c6.0: HR 10.10 (95% CI 3.59 to 28.43), 6 studies, 4532 participants. In subgroup analyses, there was no clear pattern of differences between geographic regions. We downgraded the evidence for the prognostic effect of IH versus normoglycaemia to low-certainty evidence due to study limitations because many studies did not adequately adjust for confounders. Imprecision and inconsistency required further downgrading due to wide 95% CIs and wide 95% prediction intervals (sometimes ranging from negative to positive prognostic factor to outcome associations), respectively.This evidence is up to date as of 26 February 2018.

AUTHORS' CONCLUSIONS

Overall prognosis of people with IH worsened over time. T2DM cumulative incidence generally increased over the course of follow-up but varied with IH definition. Regression from IH to normoglycaemia decreased over time but was observed even after 11 years of follow-up. The risk of developing T2DM when comparing IH with normoglycaemia at baseline varied by IH definition. Taking into consideration the uncertainty of the available evidence, as well as the fluctuating stages of normoglycaemia, IH and T2DM, which may transition from one stage to another in both directions even after years of follow-up, practitioners should be careful about the potential implications of any active intervention for people 'diagnosed' with IH.

Predicting Current Glycated Hemoglobin Values in Adults: Development of an Algorithm From the Electronic Health Record

Background

Electronic, personalized clinical decision support tools to optimize glycated hemoglobin (HbA_1c) screening are lacking. Current screening guidelines are based on simple, categorical rules developed for populations of patients. Although personalized diabetes risk calculators have been created, none are designed to predict current glycemic status using structured data commonly available in electronic health records (EHRs).

Objective

The goal of this project was to create a mathematical equation for predicting the probability of current elevations in HbA_1c (≥5.7%) among patients with no history of hyperglycemia using readily available variables that will allow integration with EHR systems.

Methods

The reduced model was compared head-to-head with calculators created by Baan and Griffin. Ten-fold cross-validation was used to calculate the bias-adjusted prediction accuracy of the new model. Statistical analyses were performed in R version 3.2.5 (The R Foundation for Statistical Computing) using the rms (Regression Modeling Strategies) package.

Results

The final model to predict an elevated HbA_1c based on 22,635 patient records contained the following variables in order from most to least importance according to their impact on the discriminating accuracy of the model: age, body mass index, random glucose, race, serum non–high-density lipoprotein, serum total cholesterol, estimated glomerular filtration rate, and smoking status. The new model achieved a concordance statistic of 0.77 which was statistically significantly better than prior models. The model appeared to be well calibrated according to a plot of the predicted probabilities versus the prevalence of the outcome at different probabilities.

Conclusions

The calculator created for predicting the probability of having an elevated HbA_1c significantly outperformed the existing calculators. The personalized prediction model presented in this paper could improve the efficiency of HbA_1c screening initiatives.

The Association between Social Jetlag, the Metabolic Syndrome, and Type 2 Diabetes Mellitus in the General Population: The New Hoorn Study

Only a few studies have investigated the metabolic consequences of social jetlag. Therefore, we examined the association of social jetlag with the metabolic syndrome and type 2 diabetes mellitus in a population-based cohort. We used cross-sectional data from the New Hoorn Study cohort (n = 1585, 47% men, age 60.8 ± 6 years). Social jetlag was calculated as the difference in midpoint sleep (in hours) between weekdays and weekend days. Poisson and linear regression models were used to study the associations, and age was regarded as a possible effect modifier. We adjusted for sex, employment status, education, smoking, physical activity, sleep duration, and body mass index. In the total population, we only observed an association between social jetlag and the metabolic syndrome, with prevalence ratios adjusted for sex, employment status, and educational levels of 1.64 (95% CI 1.1-2.4), for participants with >2 h social jetlag, compared with participants with <1 h social jetlag. However, we observed an interaction effect of median age (<61 years). In older participants (≥61 years), no significant associations were observed between social jetlag status, the metabolic syndrome, and diabetes or prediabetes. In the younger group (<61 years), the adjusted prevalence ratios were 1.29 (95% CI 0.9-1.9) and 2.13 (95% CI 1.3-3.4) for the metabolic syndrome and 1.39 (95% CI 1.1-1.9) and 1.75 (95% CI 1.2-2.5) for diabetes/prediabetes, for participants with 1-2 h and >2 h social jetlag, compared with participants with <1 h social jetlag. In conclusion, in our population-based cohort, social jetlag was associated with a 2-fold increased risk of the metabolic syndrome and diabetes/prediabetes, especially in younger (<61 years) participants.