Longitudinal changes in blood biomarkers and their ability to predict type 2 diabetes mellitus—The Tromsø study

Abstract Introduction Identification of individuals at high risk of developing type 2 diabetes mellitus (T2DM) is important for early prevention of the disease. Once T2DM is established, it is difficult to treat and is associated with cardiovascular complications and increased mortality. We aimed to describe pre‐ and post‐diagnostic changes in blood biomarker concentrations over 30 years in individuals with and without T2DM, and to determine the predictive potential of pre‐diagnostic blood biomarkers. Methods This nested case–control study included 234 participants in the Tromsø Study who gave blood samples at five time points between 1986 and 2016: 130 did not develop T2DM and were used as controls; 104 developed T2DM after the third time point and were included as cases. After stratifying by sex, we investigated changes in pre‐ and post‐diagnostic concentrations of lipids, thyroid hormones, HbA1c, glucose and gamma‐glutamyltransferase (GGT) using linear mixed models. We used logistic regression models and area under the receiver operating characteristic curve (AROC) to assess associations between blood biomarker concentrations and T2DM, as well as the predictive ability of blood biomarkers. Results Cases and controls experienced different longitudinal changes in lipids, free T3, HbA1c, glucose, and GGT. The combination of selected blood biomarker concentrations and basic clinical information displayed excellent (AROC 0.78–0.95) predictive ability at all pre‐diagnostic time points. A prediction model that included HDL (for women), HbA1c, GGT, and basic clinical information demonstrated the strongest discrimination 7 years before diagnosis (AROC 0.95 for women, 0.85 for men). Conclusion There were clear differences in blood biomarker concentrations between cases and controls throughout the study, and several blood biomarkers were associated with T2DM. Selected blood biomarkers (lipids, HbA1c, GGT) in combination with BMI, physical activity, elevated blood pressure, and family history of T2DM had excellent predictive ability 1–7 years before T2DM diagnosis and acceptable predictive ability up to 15 years before diagnosis.


| INTRODUC TI ON
The prevalence of type 2 diabetes mellitus (T2DM) has increased substantially over the past few decades and is one of the most important global health challenges of the 20th century. 1 The disease is characterized by insufficient insulin secretion and/or insulin resistance and established risk factors include among other obesity, sedentary lifestyle, excess dietary intake, and genetic factors. 2 Previous longitudinal studies of repeated pre-diagnostic measurements have demonstrated increases in lipid and glucose concentrations 1.5-20 years before T2DM diagnosis, with steeper increases closer to diagnosis. [3][4][5][6][7][8][9][10] Thus, disruption of metabolic homeostasis involving lipids, thyroid hormones, glucose, and liver enzymes is associated with T2DM. 5,8,9,[11][12][13] However, the sequence of this disruption and its relative contribution to the progression from normal to impaired glucose tolerance, and ultimately to T2DM, remains unknown. 14,15 Prediabetes (i.e., higher-than-normal blood glucose concentrations) precedes T2DM. Once T2DM has manifested, it is irreversible, difficult to treat, and associated with cardiovascular complications and increased mortality. [16][17][18] The identification of blood biomarkers and the development of risk score models for prediabetes and T2DM are therefore highly relevant, as they will enable early identification of high-risk individuals. There are currently many risk score models for diabetes (reviewed by Buijsse et al. 19 ) most are based on basic clinical information like age, body mass index (BMI), physical activity, blood pressure and genetic predisposition, but some also include blood biomarkers. For instance, the FINDRISC (including basic clinical information as well as daily consumption of vegetables, fruits or berries, and history of high glucose) and the Framingham (including basic clinical information as well as high-density lipoprotein (HDL) and triglycerides) risk scores for diabetes have been shown to successfully identify high-risk individuals 5-7 years before diagnosis. 20,21 Several studies of risk score models have shown that adding blood biomarkers to basic clinical information improves predictive ability, 4,20,22 especially biomarkers involved in glycaemic processes, uric acid, and lipids. However, most studies on prediction models are based on a single baseline blood sample. 4,23 The Tromsø Study contains blood biomarker concentrations and basic clinical information for up to five time points. Hence, we aimed to describe pre-and post-diagnostic changes in blood biomarker concentrations over 30 years in individuals with and without T2DM, and to determine the predictive potential of pre-diagnostic blood biomarkers.

| Study population
The Tromsø Study is a population-based health survey carried out in the Tromsø municipality in Northern Norway. The first survey, Tromsø1, was carried out in 1974, and six more surveys followed (Tromsø2-Tromsø7), one about every 6-7 years. During each survey, participants completed questionnaires, underwent a clinical examination and gave a blood sample. 24,25 The present, longitudinal, nested case-control study includes Initially, all participants with a T2DM diagnosis were recorded in a local diabetes registry between 2000 (T3) and 2006 (T4), and available pre-diagnostic serum samples were eligible for inclusion as cases (76 women, 69 men). We then randomly selected 76 women and 69 men who participated in the same surveys, had serum samples for T1-T3 and had no T2DM diagnosis recorded in a local diabetes registry during the surveys as controls. Of the initial 290 participants, we excluded 29 cases with glycated haemoglobin (HbA 1c ) ≥48 mmol/mol (6.5%) before or at T3, and seven controls with HbA 1c ≥48 mmol/mol (6.5%) at any time point. We also excluded participants who reported using medications that could affect glucose and thyroid hormone concentrations before T3 (

| Questionnaires, clinical examination and blood collection
The Tromsø Study questionnaire and measurements have been described in detail elsewhere. 24,25 Briefly, each survey included a K E Y W O R D S biomarkers, blood test, health service, longitudinal survey, preventive, risk factors, type 2 diabetes mellitus questionnaire that collected information on lifestyle habits, selfreported diseases such as diabetes, family history of diseases including T2DM, parity and breastfeeding. A clinical examination was also conducted at each survey and included measurements of weight, height, waist circumference and blood pressure, among others, and the collection of non-fasting blood samples. Several analyses were performed in fresh blood samples; serum samples were frozen and stored for later use. 25

| Laboratory analyses and availability of blood biomarkers
Serum samples were thawed and analysed for triglycerides, total cholesterol, low-density lipoprotein (LDL), HDL, free triiodothyronine (T 3 ), free thyroxine (T 4 ) and thyroid-stimulating hormone (TSH), but serum samples from T2 were insufficient for analyses of free T 3 , free T 4 and TSH. Data from previous analyses carried out at the time of blood collection were available for TSH (T2), HbA 1c (T2-T5), glucose (T2-T5) and gamma-glutamyltransferase (GGT; T1-T2, T4). Included blood biomarkers varied at each time point (Figure 1).  27 Quality controls are run routinely, at three different concentrations every day, and the laboratory also participates in the external quality assessment program, Lab Quality. 28 Total lipids (g/L) were calculated according to the formula 29 :

| Statistical analyses
Blood biomarker concentrations and demographic variables are reported as means with standard deviations, medians with 5 and 95 percentiles, and/or frequencies with percentages. Sample characteristics were compared between cases and controls at each time point using unpaired two-sample t-tests for continuous variables and Pearson's chi-squared for categorical variables.
Linear mixed effects models were used to explore the rate and significance of changes in blood biomarker concentrations at T1-T5, between and within cases and controls, after adjusting for the following established risk factors for T2DM 30  We assessed the associations between pre-diagnostic blood biomarker concentrations and T2DM. Logistic regression analyses were used to estimate odds ratios of T2DM for each time point separately.
We fitted two models per blood biomarker: the first included blood biomarker concentration as a continuous, independent variable; in the second model, the blood biomarker was dichotomized according to clinical guidelines and concentrations associated with an increased risk of T2DM. Both models were adjusted for established risk factors, and odds ratios were estimated either per 1-unit increase in blood biomarker concentration or above versus below the defined clinical cut-off values: triglycerides >1.70 g/L, HDL <1.30 mmol/L for women and <1.03 for men, 30 total cholesterol >5.00 mmol/L, LDL >3.00 mmol/L 31 and HbA 1c >39.0 mmol/mol (5.7%). 18 Cut-offs for blood biomarkers with no clinical guidelines were based on a receiver operating characteristics curve (ROC) analysis in pre-diagnostic samples, which yielded the highest discrimination between cases and controls, and were as follows: total lipids >7.40 g/L for women (62.7% sensitivity, 63.3% specificity) and >7.59 for men (61.5% sensitivity, 61.5% specificity), free T 3 >5.20 pmol/L for women (

| Study sample characteristics
Type 2 diabetes mellitus cases and controls were similar in age, whereas cases were heavier, had higher BMI, and larger waist circumference (except men at T5) at all time points (Table 1). At prediagnostic time points, female cases had significantly higher blood pressure than controls, except for systolic blood pressure at T2. We observed no significant differences in blood pressure for males, except at T5, when cases had significantly lower diastolic blood pressure. In general, there were no differences in alcohol consumption or physical activity between cases and controls (Table S1), and no significant differences in parity or duration of breastfeeding between female cases and controls (Table 1). Female cases reported a family history of T2DM more frequently than female controls (Table S1).
Female cases had significantly higher triglyceride, HbA 1c , and glucose concentrations, and lower HDL concentrations than controls at all time points. Female cases also had significantly higher pre-diagnostic total lipids, total cholesterol (T2), free T 3 (T3) and GGT (T1-T2) concentrations than controls ( Figure 2 and Table S2).
However, post-diagnostic total cholesterol and LDL concentra-

| Longitudinal changes in blood biomarkers
After adjusting for age, BMI, physical activity, elevated blood pressure and family history of T2DM, female cases experienced a significantly larger increase in pre-diagnostic free T 3 (T1-T3), HbA 1c (T2-T3) and GGT (T1-T2) concentrations compared to controls ( Figure 3 and Table S3). Further, there was a significantly larger increase in HbA 1c concentrations, and a larger decrease in total cholesterol, LDL and free T 3 concentrations in cases compared to controls from T3-T5.
Male cases experienced a significantly larger decrease in prediagnostic total lipid, total cholesterol, and LDL concentrations compared to controls, whereas significantly larger increases in HbA 1c and glucose concentrations were observed from T2-T3 in cases ( Figure 4 and Table S3). Further, there was a significantly larger increase in post-diagnostic HbA 1c and HDL concentrations, and a larger decrease in free T 3 concentrations in cases compared to controls from T3-T5.

| Associations between pre-diagnostic blood biomarker concentrations and T2DM
In women, pre-diagnostic concentrations above the predefined cutoffs for HDL (T1) and free T 4 (T3) were inversely associated with T2DM, while total lipids and free T 3 (T3); triglycerides, HbA 1c and glucose (T2 and T3); and GGT (T2) were positively associated with T2DM after adjusting for established risk factors (Table S4) (Tables 2 and 3). However, the combined model had increased predictive ability at every pre-diagnostic time point. The strongest discrimination between cases and controls was observed at T2 (95% for women and 85% for men), when the models for men and women were similar but not identical, as HDL was included for women only.
Excluding HDL reduced discrimination among women to 94%, with a small loss of model fit (AIC 77.1 vs. 76.4).

| DISCUSS ION
In this nested case-control study, we observed differences between cases and controls in total lipids, triglycerides, total cholesterol, HbA 1c , glucose and GGT that were present 15 years before T2DM diagnosis in cases. The model including established risk factors (age, BMI, physical activity, blood pressure and family history of T2DM) was sufficient to acceptably discriminate between cases and con-  Models were selected with backwards selection process according to best model fit.
clinical information, and sometimes different blood biomarkers, they all showed excellent discrimination (AROC: 0.78-0.90). They also displayed similar predictive abilities, although their biomarkers were different from ours, perhaps because their biomarkers were also related to prediabetic metabolic disturbances. For example, the prediction model proposed by the Framingham offspring study used personal information (age, sex, history of T2DM, BMI), blood pressure, HDL, triglycerides and fasting glucose and had excellent predictive ability (AROC: 0.85) 7 years before diagnosis. 20 Our prediction model for women at T2 (also 7 years before diagnosis) was very similar (e.g., personal information, blood pressure, total lipids, triglycerides and HDL), but we included GGT and HbA 1c , as fasting blood glucose was not available. As postprandial hyperglycaemia is more common in individuals with prediabetes, 35 In the present study, all prediction models performed better in women than in men. Specifically, we observed stronger associations between lipids (total lipids, triglycerides and HDL), free T 3 , free T 4 , HbA 1c , glucose and T2DM in women than men. Several other studies (reviewed by Kautzky-Willer et al. 42 ) demonstrated stronger associations between lipids and incident T2DM in women than men, possibly due to sex differences in fat deposition. 42 Njølstad et al. 43 also observed stronger associations between HDL, triglycerides, random glucose and T2DM in women than men in the Finnmark Study; BMI was a more important risk factor for men.
Many blood biomarkers were significant predictors of T2DM in our study; however, discrimination and model fit were not compromised even after several biomarkers were excluded from the models. This may be due to the very strong predictive abilities of We observed that cases had higher average GGT concentrations than controls and that men generally had higher concentrations than women. However, concentrations varied within the normal range of 10-75 U/l for women and 15-115 U/l for men. 31 This is in line with previous studies investigating liver biomarkers in relation to T2DM, which showed significantly higher GGT concentrations in cases than controls, and in men than women, though they remained within normal limits. 5,47,48 GGT has been identified as an independent risk factor for T2DM and is also linked to hepatic steatosis, which in turn is associated with obesity, 49  was previously reported for both men and women. 50 The authors hypothesized that this was due to changes in cholesterol-associated lifestyle factors in the Norwegian population, such as a general increase in physical activity, and decreased smoking and consumption of trans fats. In our study, the steeper post-diagnostic decrease in cholesterol concentrations among cases may be explained by targeted lifestyle changes following the diagnosis, as individuals with T2DM have been shown to improve their lipid concentrations after diagnosis. 51 The decrease may also be attributed to the use of cholesterol-lowering drugs, as cardiovascular diseases are associated with T2DM. In our study, 43%-70% of cases and 5%-24% of controls reported using lipid-lowering drugs at T4 and T5, compared to 17%-40% in the general population within similar age groups and time periods. 50 We observed different changes in free T 3 between cases and controls where cases generally had increased pre-diagnostic and decreased post-diagnostic concentrations. Free T 3 was positively associated with T2DM in men and women at T3, whereas free T 4 was inversely associated with T2DM in women at T3. This both agrees and disagrees with a recent meta-analysis including 12 prospective studies 52 that demonstrated positive associations between TSH concentrations and T2DM, and inverse associations between free T 3 and free T 4 with T2DM. We did not observe any significant associations between TSH and T2DM, possibly due to small sample size. Time of blood sampling before diagnosis as well as study design might explain the different study observations. Accordingly, we observed that concentrations of free T 3 were similar between cases and controls at T1, with a notable increase in cases to T3, followed by a post-diagnostic decline. This observation is in line with the study by Jun et al. 53 where they observed an increased T 3 con- The main strength of this study is the nested case-control design with repeated measurements which allowed us to study pre- After stratifying by sex, there were few observations at each time point among cases and controls, which limits the precision of our effect estimates. Due to a lack of serum, we were not able to analyse thyroid hormones at T2 nor glucose at T1; moreover, GGT was unavailable at T3 and T5, as was HbA 1c at T1. Waist circumference was also not available at T1, and only available for ~68% of subjects at T2. However, even though waist circumference has a stronger association with T2DM than BMI, it has not been shown to provide more accurate risk predictions of T2DM. 57 We had smaller sample sizes at post-diagnostic time points, as the inclusion criteria required an available blood sample at all prediagnostic ones. The prediction models were developed in a study sample from a northern Norwegian population, thus, the relative contribution of each predictor may vary in other populations due to genetical, environmental and lifestyle variations. Accordingly, our prediction models should be validated in different populations to verify their generalizability, and cut-offs should be re-evaluated if necessary. 19

| CON CLUS IONS
Already 15 years before diagnosis, there were clear differences in blood biomarker concentrations between T2DM cases and controls and several blood biomarkers were associated with type T2DM.
Selected blood biomarkers (lipids, HbA 1c , GGT) in combination with BMI, physical activity, elevated blood pressure, and family history of T2DM had excellent predictive ability 1-7 years before type 2 T2DM diagnosis and acceptable predictive ability up to 15 years before diagnosis.

ACK N OWLED G EM ENTS
We would like to thank all the participants of the Tromsø Study for their willingness to participate in the surveys and to donate blood. We would also like to thank the staff at the Department of Laboratory Medicine at the University Hospital of North Norway, who contributed with the laboratory analyses; especially Arnt R.
Hagen and Tom Sollid.

CO N FLI C T O F I NTE R E S T
The authors have no conflict of interest to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data set used in present study was derived from the Tromsø Study. It is not publicly available, but may be accessed through an application to the Tromsø Study (https://uit.no/resea rch/troms ostudy).