Triglyceride-cholesterol imbalance across lipoprotein subclasses predicts diabetic kidney disease and mortality in type 1 diabetes: the FinnDiane Study

Authors

  • V.-P. Mäkinen,

    Corresponding author
    1. Biocenter Oulu, University of Oulu, Oulu, Finland
    2. Folkhälsan Institute of Genetics, Biomedicum Helsinki, Helsinki, Finland
    3. Division of Nephrology, Helsinki University Central Hospital, Helsinki, Finland
    4. College of Chemistry and Chemical Engineering, Central South University, Changsha, Hunan Province, China
    • Computational Medicine, Institute of Clinical Medicine, Faculty of Medicine, University of Oulu, Oulu, Finland
    Search for more papers by this author
  • P. Soininen,

    1. Computational Medicine, Institute of Clinical Medicine, Faculty of Medicine, University of Oulu, Oulu, Finland
    2. NMR Metabolomics Laboratory, School of Pharmacy, University of Eastern Finland, Kuopio, Finland
    Search for more papers by this author
  • A. J. Kangas,

    1. Computational Medicine, Institute of Clinical Medicine, Faculty of Medicine, University of Oulu, Oulu, Finland
    Search for more papers by this author
  • C. Forsblom,

    1. Folkhälsan Institute of Genetics, Biomedicum Helsinki, Helsinki, Finland
    2. Division of Nephrology, Helsinki University Central Hospital, Helsinki, Finland
    Search for more papers by this author
  • N. Tolonen,

    1. Folkhälsan Institute of Genetics, Biomedicum Helsinki, Helsinki, Finland
    2. Division of Nephrology, Helsinki University Central Hospital, Helsinki, Finland
    Search for more papers by this author
  • L. M. Thorn,

    1. Folkhälsan Institute of Genetics, Biomedicum Helsinki, Helsinki, Finland
    2. Division of Nephrology, Helsinki University Central Hospital, Helsinki, Finland
    Search for more papers by this author
  • J. Viikari,

    1. Department of Medicine, University of Turku and Turku University Hospital, Turku, Finland
    Search for more papers by this author
  • O. T. Raitakari,

    1. Research Centre of Applied and Preventive Cardiovascular Medicine, University of Turku, Turku, Finland
    2. Department of Clinical Physiology, University of Turku and Turku University Hospital, Turku, Finland
    Search for more papers by this author
  • M. Savolainen,

    1. Computational Medicine, Institute of Clinical Medicine, Faculty of Medicine, University of Oulu, Oulu, Finland
    2. Biocenter Oulu, University of Oulu, Oulu, Finland
    3. Department of Internal Medicine, Clinical Research Center, University of Oulu, Oulu, Finland
    Search for more papers by this author
  • P.-H. Groop,

    1. Division of Nephrology, Helsinki University Central Hospital, Helsinki, Finland
    2. College of Chemistry and Chemical Engineering, Central South University, Changsha, Hunan Province, China
    3. Baker IDI Heart and Diabetes Institute, Melbourne, Australia
    Search for more papers by this author
  • M. Ala-Korpela,

    1. Computational Medicine, Institute of Clinical Medicine, Faculty of Medicine, University of Oulu, Oulu, Finland
    2. College of Chemistry and Chemical Engineering, Central South University, Changsha, Hunan Province, China
    3. Department of Internal Medicine, Clinical Research Center, University of Oulu, Oulu, Finland
    4. Department of Epidemiology and Biostatistics, School of Public Health, Imperial College London, London, UK
    Search for more papers by this author
  • Finnish Diabetic Nephropathy Study Group


Correspondence: Ville-Petteri Mäkinen, DSc (Tech), Folkhälsan Research Center, Biomedicum Helsinki, University of Helsinki, PO Box 63, FI-00014, Finland.

(fax: +358 9 19125452; e-mail: ville-petteri.makinen@computationalmedicine.fi).

Abstract

Background

Circulating cholesterol (C) and triglyceride (TG) levels are associated with vascular injury in type 1 diabetes (T1DM). Lipoproteins are responsible for transporting lipids, and alterations in their subclass distributions may partly explain the increased mortality in individuals with T1DM.

Design and subjects

A cohort of 3544 individuals with T1DM was recruited by the nationwide multicentre FinnDiane Study Group. At baseline, six very low-density lipoprotein VLDL, one intermediate-density lipoprotein IDL, three low-density lipoprotein LDL and four higher high-density lipoprotein HDL subclasses were quantified by proton nuclear magnetic resonance spectroscopy. At follow-up, the baseline data were analysed for incident micro- or macroalbuminuria (117 cases in 5.3 years), progression from microalbuminuria (63 cases in 6.1 years), progression from macroalbuminuria (109 cases in 5.9 years) and mortality (385 deaths in 9.4 years). Univariate associations were tested by age-matched cases and controls and multivariate lipoprotein profiles were analysed using the self-organizing map (SOM).

Results

TG and C levels in large VLDL were associated with incident albuminuria, TG and C in medium VLDL were associated with progression from microalbuminuria, and TG and C in all VLDL subclasses were associated with mortality. Large HDL-C was inversely associated with mortality. Three extreme phenotypes emerged from SOM analysis: (i) low C (<3% mortality), (ii) low TG/C ratio (6% mortality), and (iii) high TG/C ratio (40% mortality) in all subclasses.

Conclusions

TG–C imbalance is a general lipoprotein characteristic in individuals with T1DM and high vascular disease risk.

Introduction

In 1978, Nikkilä and Hormila concluded that the lipoprotein profile in uncomplicated type 1 diabetes is not atherogenic, and that patients with this disease have higher high-density lipoprotein (HDL) cholesterol levels than nondiabetic individuals [1]. More than a decade later, the diabetes control and complication trial (DCCT) reported minor lipid differences in patients with type 1 diabetes [2]. On the other hand, diabetic complications may be associated with a temporally complex deterioration of the lipid profile, which manifests as increases in the very low-density lipoprotein (VLDL)–intermediate-density lipoprotein (IDL)– low-density lipoprotein (LDL) pool and compositional changes in the HDL fraction [3-5].

In the DCCT/Epidemiology of Diabetes Intervention and Complications Cohort (EDIC) study, total triglycerides and cholesterol, VLDL and small LDL lipids were positively correlated with urinary albumin excretion rate (AER); in addition, large HDL was negatively, but small HDL positively correlated with AER [6, 7]. Adverse changes in the lipid profile included ‘an increase in VLDL subclass levels, smaller VLDL, a shift towards smaller LDL, increased LDL particle concentration, a decrease in cardioprotective large HDL and an increase in noncardioprotective small HDL’. The Pittsburgh Epidemiology of Diabetes study reported an increased risk of coronary artery disease for those patients with type 1 diabetes who had higher lipid masses and particle concentrations of VLDL, small LDL, medium LDL or medium HDL, or lower large HDL [8].

We have previously shown that diabetic patients with few or no complications have a favourable lipid profile (low triglycerides, high HDL cholesterol) and that the excess mortality in the FinnDiane cohort was strongly linked to kidney disease in combination with higher triglyceride levels and obesity [9]. However, standard clinical measures may be too insensitive to detect subtle, but biologically significant, lipoprotein abnormalities [10-12], and there is ongoing debate regarding the importance of lipoprotein subclasses, apolipoprotein species and other related markers of lipoprotein metabolism [13-16].

In this study, we investigated 14 lipoprotein subclasses by nuclear magnetic resonance (NMR) spectroscopy to increase understanding of lipoprotein risk factors. Initially, we performed a cross-sectional analysis of a large cohort of patients with type 1 diabetes to uncover the relationships between lipoproteins and clinically defined kidney disease. Due to the large number of measures, we employed advanced data modelling and visualization techniques to determine the wider systemic patterns of interactions. Then we assessed the associations between the baseline lipoprotein subclass measures, the progression of diabetic kidney disease and mortality in the same patient population to determine the lipoprotein phenotype with the highest vascular risk.

Materials and methods

At baseline, 3544 patients with type 1 diabetes were recruited by the multicentre Finnish Diabetic Nephropathy Study Group. The initial data collection was cross-sectional (serum and urine samples), but with longitudinal records of albuminuria and clinical history from local hospitals, and all-cause mortality data from the Population Registry Centre of Finland (385 deaths during 9.4 ± 2.5 years). In this study, Type 1 diabetes mellitus was defined as diagnosis of diabetes below 35 years of age and transition to insulin treatment within 1 year of onset.

The classification of renal status was made centrally according to urinary AER in at least two of three consecutive overnight or 24-h urine samples. Macroalbuminuria was defined as AER ≥200 μg min−1 or ≥300 mg per 24 h. Microalbuminuria (intermediate level of kidney disease) was defined as AER between >20 and <200 μg min−1 or >30 and < 300 mg per 24 h. Incident albuminuria was defined as the progression from normal AER to micro- or macroalbuminuria during the follow-up period (1651 nonprogressors and 117 progressors in 5.3 ± 2.0 years). Progression from microalbuminuria was defined as the transition from microalbuminuria to macroalbuminuria or end-stage renal disease (290 nonprogressors and 63 progressors in 5.8 ± 2.0 years). Progression from macroalbuminuria was defined as the transition from macroalbuminuria to end-stage renal disease (295 nonprogressors and 109 progressors in 6.4 ± 2.2 years). Matched subsets were used in some analyses and they are specified in the relevant figure legends. In addition to the clinical classification of albuminuria, continuous AER was centrally measured from a 24-h urine collection at baseline.

The lipoprotein subclasses were defined by particle size and calibrated according to high-performance liquid chromatography [17]: chylomicrons and extremely large VLDL particles (average diameter >75 nm), very large VLDL (64.0 nm), large VLDL (53.6 nm), medium VLDL (44.5 nm), small VLDL (36.8 nm), very small VLDL (31.3 nm), IDL (28.6 nm), large LDL (25.5 nm), medium LDL (23.0 nm), small LDL (18.7 nm), very large HDL (14.3 nm), large HDL (12.1 nm), medium HDL (10.9 nm) and small HDL (8.7 nm). Concentrations of subclass lipid constituents were estimated by regression modelling of the 1H NMR spectra [18, 19]. Other clinical and biochemical variables measured in the FinnDiane Study, including citations for laboratory assays have been described previously [9].

Statistical analyses

Statistical significance and 95% confidence intervals were empirically estimated by bootstrapping. Patients with kidney disease were older than those without diabetic complications, which may induce observational bias particularly in the prospective analyses. We therefore eliminated age, diabetes duration and gender effects by matching the cases and controls for each kidney disease trait separately, instead of including numerical adjustment for covariates. The means ± standard deviations for confounders in the disease groups are described in the figure legends.

The self-organizing map (SOM) is an unsupervised pattern recognition method and was used here for automated comparisons between multivariate biochemical profiles. Examples and technical details have been previously reported [20, 21]. The training set for the SOM included 14 NMR-determined lipoprotein subclasses and their lipid constituents, haemoglobin A1c, five urinary measures (albumin, creatinine, sodium, potassium and urea), serum albumin, creatinine, sodium, potassium, cystatin-C, total cholesterol, triglycerides, total HDL cholesterol, HDL3 cholesterol, apolipoproteins A-I, A-II and B-100, C-reactive protein, adiponectin, mannan-binding lectin and soluble receptor for advanced glycation end-products. Prospective or baseline clinical traits were not used for the model training to prevent over-fitting. The training data were adjusted for gender prior to analysis, but not for age or diabetes duration. Observational bias due to age stratification is not a problem for the SOM, because the map will explicitly reveal the characteristics of different age groups, if such effects exist.

Results

Baseline characteristics of the study cohort are shown in Table 1. Women had less kidney disease at baseline, as well as lower triglyceride and higher HDL cholesterol levels. There were minor differences in age and diabetes duration between the sexes. Further gender comparisons were performed using the SOM (see below).

Table 1. Characteristics of patients with type 1 diabetes. Median values and standard deviations are shown for continuous measures. P-values for gender differences were estimated using the Kolmogorov–Smirnov statistic for continuous measures. C-peptide was below the detection limit for most patients
 Men = 1817Women = 1727P-value
Age (years)37.6 ± 12.237.2 ± 11.70.13
Diabetes duration (years)21.9 ± 12.622.9 ± 12.1<10−5
C-peptide (nmol L−1)0.04 ± 0.130.03 ± 0.11-
Microalbuminuria (%)15110.00083
Macroalbuminuria (%)16120.00046
End-stage renal disease (%)860.0018
Body mass index (kg m−2)25.1 ± 3.424.9 ± 3.7<10−4
Systolic blood pressure (mmHg)137 ± 18131 ± 19<10−30
Diastolic blood pressure (mmHg)81 ± 1078 ± 10<10−38
Total triglycerides (mmol L−1)1.30 ± 0.731.11 ± 0.57<10−16
Total cholesterol (mmol L−1)4.7 ± 0.94.9 ± 0.8<10−7
HDL cholesterol (mmol L−1)1.28 ± 0.321.49 ± 0.36<10−61

Figure 1a shows the comparisons of lipoprotein triglycerides and cholesterol between 454 patients with microalbuminuria and 1362 control subjects with normal AER matched for age, diabetes duration and sex (see figure legend for confounder statistics). Both cholesterol and triglycerides in very large, large and medium VLDL subclasses were significantly increased in patients with microalbuminuria (< 0.0001). IDL or LDL subclasses were not significantly different between the two groups (> 0.01). Figure 1b shows the difference in lipoprotein triglycerides and cholesterol between patients with macroalbuminuria and normal AER (508 patients in each group). Both lipids are increased in all VLDL subclasses (< 0.0001), and triglycerides are also increased in IDL, large and medium LDL and very large and small HDL subclasses (< 0.0001). Large and medium HDL cholesterol are decreased in patients with macroalbuminuria (< 0.0001). Fold differences in lipoprotein subclasses are shown in Table 2. Figure 1c shows the subclass differences between end-stage renal disease (243 cases) and normal AER (486 controls). The overall profile was similar to that of patients with macroalbuminuria, except for very large HDL cholesterol which was significantly decreased only in end-stage renal disease (< 0.0001).

Table 2. Lipoprotein subclasses in patients with normal AER or macroalbuminuria at baseline, and alive or dead at follow-up. Matching criteria between the groups are given in the legends to Figs 1 and 2. Mean ± SD concentrations are shown in mmol L−1 unless otherwise indicated. Fold-change was calculated by dividing the difference between cases (macroalbuminuria or dead) and controls (normal AER or alive) by the mean value in the respective control group. P-values were estimated by bootstrapping
 Normal AER = 508Macroalbuminuria = 508Difference (%)Alive = 1155Dead = 385Difference (%)
  1. a

    < 0.01.

  2. b

    < 0.0001.

Age (years)42 ± 1042 ± 10 45 ± 1045 ± 10 
Diabetes duration (years)30 ± 830 ± 8 32 ± 1032 ± 10 
Largest VLDL-TG0.017 ± 0.0150.028 ± 0.026+68b0.019 ± 0.0220.028 ± 0.023+43b
Very large VLDL-TG0.040 ± 0.0270.062 ± 0.044+69b0.037 ± 0.0340.054 ± 0.040+48b
Large VLDL-TG0.11 ± 0.070.17 ± 0.11+61b0.11 ± 0.090.15 ± 0.10+39b
Medium VLDL-TG0.18 ± 0.100.27 ± 0.15+48b0.22 ± 0.130.28 ± 0.15+29b
Small VLDL-TG0.21 ± 0.070.28 ± 0.12+34b0.25 ± 0.100.31 ± 0.12+23b
Very small VLDL-TG0.11 ± 0.040.14 ± 0.06+32b0.12 ± 0.050.15 ± 0.06+23b
IDL-TG0.11 ± 0.040.13 ± 0.06+28b0.12 ± 0.050.14 ± 0.06+22b
Large LDL-TG0.09 ± 0.040.12 ± 0.05+25b0.10 ± 0.050.12 ± 0.05+23b
Medium LDL-TG0.05 ± 0.030.07 ± 0.04+35b0.052 ± 0.0300.067 ± 0.034+30b
Very large HDL-TG0.010 ± 0.0060.013 ± 0.008+31b0.011 ± 0.0070.014 ± 0.008+23b
Large HDL-TG (μmol L−1)1.8 ± 4.34.2 ± 8.6+127a4.7 ± 9.97.5 ± 11.8+59a
Medium HDL-TG0.012 ± 0.0110.015 ± 0.015+22a0.017 ± 0.0130.020 ± 0.017+21a
Small HDL-TG0.037 ± 0.0140.047 ± 0.020+22b0.041 ± 0.0170.048 ± 0.021+16b
Largest VLDL-C (μmol L−1)9.0 ± 7.313.1 ± 10.6+45b9.0 ± 9.412.0 ± 9.9+34b
Very large VLDL-C0.014 ± 0.0110.023 ± 0.016+68b0.014 ± 0.0130.020 ± 0.015+45b
Large VLDL-C0.064 ± 0.0360.097 ± 0.053+51b0.063 ± 0.0430.084 ± 0.049+33b
Medium VLDL-C0.18 ± 0.050.22 ± 0.08+27b0.19 ± 0.070.22 ± 0.08+18b
Small VLDL-C0.28 ± 0.070.32 ± 0.10+14b0.28 ± 0.080.31 ± 0.09+12b
Very small VLDL-C0.23 ± 0.070.25 ± 0.08+11b0.22 ± 0.060.24 ± 0.08+10
IDL-C0.70 ± 0.170.69 ± 0.21−10.67 ± 0.180.67 ± 0.200
Large LDL-C0.95 ± 0.220.95 ± 0.2700.93 ± 0.230.94 ± 0.26+1
Medium LDL-C0.56 ± 0.130.57 ± 0.17+30.56 ± 0.140.58 ± 0.16+3
Small LDL-C0.36 ± 0.090.38 ± 0.11+6a0.36 ± 0.090.38 ± 0.11+5a
Very large HDL-C0.26 ± 0.100.25 ± 0.12−30.26 ± 0.110.25 ± 0.13−2
Large HDL-C0.24 ± 0.130.18 ± 0.12−25b0.24 ± 0.130.20 ± 0.14−19b
Medium HDL-C0.25 ± 0.070.23 ± 0.09−9b0.22 ± 0.080.21 ± 0.09−8a
Small HDL-C0.52 ± 0.090.51 ± 0.12−20.52 ± 0.100.51 ± 0.11−2
Figure 1.

Comparisons of lipid concentrations within lipoprotein subclasses for patients with type 1 diabetes with or without diabetic kidney disease. (a) shows the differences in group means with 95% confidence intervals between 454 patients with microalbuminuria (age 39 ± 13 years, diabetes duration 27 ± 11 years, 59% male) and 1362 matched controls with normal AER (age 39 ± 13 years, diabetes duration 25 ± 11 years, 51% male). (b) shows the differences between 508 patients with macroalbuminuria (age 42 ± 10 years, diabetes duration 30 ± 8 years, 59% male) and 508 matched controls with normal AER (age 42 ± 10 years, diabetes duration 30 ± 8 years, 53% male). (c) shows the differences between 243 patients with end-stage renal disease (age 45 ± 8 years, diabetes duration 33 ± 8 years, 44% male) and 486 matched controls with normal AER (age 45 ± 9 years, diabetes duration 32 ± 8 years, 61% male). Age and diabetes duration were summarized as mean values ± SD. *< 0.01; **< 0.0001.

Glycaemic control is essential in diabetes care and therefore we evaluated whether it explained the increase in VLDL lipids. Haemoglobin A1c was weakly correlated with triglycerides (6.3% variance explained), medium VLDL cholesterol (5.1% variance), small LDL cholesterol (2.7% variance) and large HDL cholesterol (2.8% variance), when adjusted for diabetes duration and gender. In a linear regression model of centrally measured 24-h AER with haemoglobin A1c and these four lipid measures as regressors, triglyceride concentration was borderline significant (= 0.012), whereas the other three lipid measures and haemoglobin A1c were highly significant (< 10−6).

Figure 2 is similar to Fig. 1, but shows the results from prospective comparisons between patients who progressed or died during follow-up, and those who maintained their baseline classification (see figure legend for confounder statistics). Incident albuminuria (Fig. 2a, 117 cases and 1404 controls) was associated with increased cholesterol and triglycerides in the medium VLDL subclass (< 0.01). Progression from microalbuminuria (Fig. 2b, 63 cases and 252 controls) was associated with medium VLDL lipids (< 0.01), and similar differences were seen for IDL, LDL and HDL triglycerides (< 0.01, except for small LDL and large HDL). Medium and small LDL cholesterol also predicted progression from microalbuminuria (< 0.01). Patients who progressed from macroalbuminuria to end-stage renal disease (Fig. 2c) had higher triglycerides in VLDL, IDL and LDL subclasses (except the largest VLDL) and the small HDL subclass (< 0.01). Cholesterol was higher in the four largest VLDL subclasses and lower in the medium HDL subclass (< 0.01).

Figure 2.

Comparisons of lipid concentrations within lipoprotein subclasses for patients with type 1 diabetes with or without progressive kidney disease. (a) shows the differences in group means with 95% confidence intervals between 117 patients with normal AER at baseline, but micro- or macroalbuminuria at follow-up (age 34 ± 12 years, diabetes duration 19 ± 12 years, 62% male, 5.7 ± 1.7 years of follow-up), and 1404 patients with stable normal AER (age 35 ± 12 years, diabetes duration 19 ± 12 years, 53% male, 5.1 ± 2.0 years of follow-up). (b) shows the differences between 63 patients with microalbuminuria at baseline, but macroalbuminuria or end-stage renal disease at follow-up (age 37 ± 12 years, diabetes duration 26 ± 11 years, 78% male, 6.1 ± 1.7 years of follow-up), and 252 patients with stable microalbuminuria (age 38 ± 12 years, diabetes duration 26 ± 12 years, 61% male, 5.4 ± 2.0 years of follow-up). (c) shows the differences between 109 patients with macroalbuminuria at baseline who developed end-stage renal disease during follow-up (age 41 ± 9 years, diabetes duration 30 ± 8 years, 61% male, 5.9 ± 2.2 years of follow-up) and 218 patients with stable macroalbuminuria (age 42 ± 10 years, diabetes duration 30 ± 8 years, 64% male, 6.5 ± 2.2 years of follow-up). Age, diabetes duration and follow-up period were summarized as mean values ± SD. *< 0.01; **< 0.0001.

All-cause mortality (Fig. 3, 385 deaths) was associated with a general increase in triglycerides across subclasses (< 0.01). Both triglycerides and cholesterol were increased in medium VLDL (< 0.0001), and large HDL cholesterol was decreased (< 0.0001) in patients who died during follow-up. Fold differences in lipid subclasses between those who had died and those who were alive at the end of follow-up are shown in Table 2.

Figure 3.

Differences between 385 patients who died during the study (age 46 ± 10 years, diabetes duration 32 ± 10 years, 61% male, 6.0 ± 3.1 years of follow-up until death) and 1155 patients who were alive at the end of follow-up (age 45 ± 10 years, diabetes duration 32 ± 10 years, 53% male, 9.7 ± 2.2 years of follow-up). Age, diabetes duration and follow-up period were summarized as mean values ± SD. *< 0.01; **< 0.0001.

Multivariate analyses of the biochemical data were performed with the SOM. We chose this data-driven bottom-up approach to determine those lipoprotein characteristics that are associated with the spectrum of vascular injuries, rather than reduce the statistical power into top-down subanalyses of specific clinical categories. The variables that were included in the SOM are given in the Methods, and a summary of the technique has been previously published [21]. Finally, we focused on extreme phenotypes, as in a classical univariate case–control study.

Three extreme phenotype models were selected to summarize the SOM results: (i) a low cholesterol, (ii) a cholesterol-rich, and (iii) a triglyceride-rich phenotype. The specific map units are shown in Fig. 4a, with map colourings for selected subclasses in Fig. 4b, and phenotypic lipid concentrations for all subclasses in Fig. 4c, d. Additional colourings and sex-specific analyses are given in Online Supplements 1–3. Despite some differences in mean concentrations of some of the lipoprotein measures, no differences in the shapes of SOM patterns were detected between men and women.

Figure 4.

Phenotype models from self-organizing map analysis to summarize the principal lipoprotein features in patients with type 1 diabetes. The map was constructed from all biochemical data including the conventional lipid profile and the NMR lipoprotein subclasses. (a) shows the criteria for the selection of extreme phenotypes: we focused on the main clinical measures of total triglycerides, cholesterol and HDL cholesterol, then examined the extreme values on the map to determine a maximally diverse, but simple characterization. (b) shows a selected set of subclass lipids that indicated or predicted clinical end-points in Figs 1 and 2. c and d show the full subclass profiles for the three phenotypes. Additional map colourings are given in Online Supplements 1–3.

Phenotype I represents patients with overall low serum lipids. This was manifested by reduced triglycerides in large and medium VLDL, and large and medium HDL subclasses, and reduced cholesterol in IDL, LDL and large and medium HDL. Phenotype II closely resembles the typical profile in the background population in terms of conventional lipids [22]. Compared with Phenotype I, increased cholesterol can be seen particularly in the IDL and LDL subclasses. Phenotype III shows increased triglycerides in all apolipoprotein B-containing subclasses (VLDL, IDL and LDL), and in small HDL. In addition, large and medium HDL cholesterol were substantially depleted, whereas increases were seen in large, medium and small VLDL cholesterol. IDL and LDL cholesterol levels were similar between Phenotypes II and III.

Biological and statistical significance of the phenotype models was assessed using the clinical data that were not used for model training. Table 3 presents the estimated percentages of clinical categories, and concentrations of selected metabolites and map colourings for prospective traits are available in Online Supplement 2Q–T. There was a low prevalence of complications in Phenotypes I and II, whereas Phenotype III was characterized by a longer diabetes duration and advanced kidney disease. Furthermore, Phenotype III was overweight (body mass index >25 kg m−2) with poor glycaemic control (haemoglobin A1c >9%, despite a higher insulin dose) and chronic inflammation (elevated C-reactive protein). The progression rate from a normal AER was the highest for Phenotype III (10% in 5.6 years). Similarly, progression from microalbuminuria (38% in 5.4 years) and from macroalbuminuria (42% in 6.2 years) occurred more frequently for Phenotype III. Finally, Phenotype III was associated with poor outcomes during follow-up (40% mortality in 9.4 years).

Table 3. Clinical characteristics of phenotype models from the self-organizing map of biochemical data (see also Fig. 3). Global P-values and 95% confidence intervals are reported for variables that were not part of the map training set. The P-values reflect the statistical connection between the global map structure and the clinical trait (denoted as ‘map-wide’), but do not measure the differences between the phenotype models. P-values were estimated by permutation analysis and confidence intervals by bootstrapping (see [21] for technical details)
 Phenotype I lowest cholesterolPhenotype II lowest TG:C ratioPhenotype III highest TG:C ratio–log10 P (map-wide values)
Normal AER at baseline (%)93 (91–95)77 (74–82)8 (6–11)33
Microalbuminuria at baseline (%)5 (4–7)12 (9–15)7 (5–10)5
Macroalbuminuria at baseline (%)<35 (3–7)57 (48–62)35
End-stage renal disease at baseline (%)<14 (2–6)27 (22–34)23
Type 1 diabetes duration (years)12 (11–13)19 (21–23)30 (28–31)21
Body mass index (kg m−2)23.4 (23.1–23.8)23.8 (23.6–24.2)25.6 (25.0–26.4)16
Systolic blood pressure (mmHg)124 (122–126)131 (129–133)145 (141–149)19
Diastolic blood pressure (mmHg)76 (75–77)80 (79–81)84 (82–85)11
Cholesterol (mmol L−1)3.95.05.3Not available
Triglycerides (mmol L−1)0.760.812.11Not available
HDL cholesterol (mmol L−1)1.221.821.06Not available
Haemoglobin A1c (%)7.58.19.1Not available
Insulin dose (IU kg−1)0.67 (0.63–0.71)0.66 (0.64–0.68)0.69 (0.66–0.72)9
Urinary albumin excretion (mg per 24 h)8.613596Not available
Estimated GFR (mL min−1 per 1.73 m2)1059448Not available
C-reactive protein (mg L−1)1.151.443.86Not available
Serum adiponectin (mg L−1)9.31618Not available
Progression from normal AER (%)4 (2–6)2 (1–4)10 (4–20)3
Progression from microalbuminuria (%)<115 (6–29)38 (18–57)5
Progression from macroalbuminuria (%)1 (0–4)13 (3–26)42 (31–51)4
Dead at follow-up (%)<36 (4–9)40 (32–47)28

Discussion

Chronic kidney disease per se is associated with multiple lipoprotein abnormalities [23, 24], but dysfunctional lipoprotein formation and clearance may be present already in patients with type 1 diabetes and albuminuria, before kidney function is affected. Lipotoxicity may also be a causative factor in the pathogenesis of kidney disease [25]. Here, we present a data-driven analysis of the lipoprotein diversity in type 1 diabetes, and show a consistent pattern of triglyceride-enrichment at all stages of diabetic kidney disease and across all subclasses, both cross-sectionally and prospectively.

In the DCCT/EDIC study, lipoprotein subclasses were measured by proton NMR spectroscopy in 958 patients with type 1 diabetes [6]. The strongest signals for albuminuria were obtained for VLDL and HDL subclasses, whereas the LDL subclasses were weaker indicators of kidney disease. In this study, we were able to further dissect the lipoprotein phenotypes by separating cholesterol and triglycerides, and link the changes in these measures to the progression of kidney disease. VLDL cholesterol and VLDL triglycerides were positively associated with kidney disease, and the absolute increases in triglycerides were greater compared with cholesterol in every VLDL subclass (Figs 1 and 2). Furthermore, the ratio of cholesterol was decreased (Fig. 4 and Online Supplement 2) and we know from a previous study that apolipoprotein B-100 is also increased [9]. A high degree of proteinuria is associated with increased synthesis and decreased catabolism of the VLDL–IDL–LDL cascade [26], which may explain some of the associations between VLDL, IDL and LDL subclasses and macroalbuminuria. Nevertheless, our findings suggest that VLDL particles increase in number and become enriched with triglycerides across all stages of diabetic kidney disease. Therefore, it is plausible that the causes of dyslipidaemia may vary during the course of the disease.

Glycaemic control is the primary modifiable factor in diabetes care [27, 28] and hyperglycaemia may simultaneously explain the risk of complications and dyslipidaemia. Accordingly, we saw increased VLDL lipids in patients with high levels of haemoglobin A1c (Phenotype III). On the other hand, intensive insulin treatment did not lead to a major improvement in serum lipid levels in the DCCT [2]. In this study, insulin deprivation is unlikely to cause the high haemoglobin A1c, because weight-adjusted insulin doses were comparable in all phenotypes (Table 3). Therefore, dyslipidaemia in type 1 diabetes should not be automatically attributed to suboptimal insulin dosage, but other causes such as insulin resistance and obesity from a sedentary lifestyle and energy-rich diet should also be considered.

Persistent microalbuminuria is the earliest clinical sign of kidney stress, and lipotoxicity may have a (preventable) role in the pathway that leads to the escalation of glomerular injury, but the precise mechanisms remain unclear. It is interesting that atherosclerosis and glomerulosclerosis have similar features [29] and it has been hypothesized that an increase in the hepatic output of circulatory lipids (with VLDL as the main transporter) initiates a vicious cycle of glomerular and tubular events that ultimately cause a progressive decline in kidney function [25]. The role of cholesterol, in particular, has been investigated in a number of animal models. When guinea-pigs and rats were fed cholesterol-rich food, they developed various forms of glomerular and other injuries and the effects could be modulated by partial or unilateral nephrectomy and hypertension [30-32]. However, cholesterol alone may not be sufficient to initiate the disease process as not all hyperlipidaemic animals develop glomerular lesions. Moreover, nondiabetic human individuals with elevated cholesterol or triglyceride levels rarely develop kidney disease. Therefore, it is plausible that the combination of hyperglycaemia and hypertension are necessary to observe lipotoxic effects.

Ceramide accumulation from excess availability of saturated fatty acids (via the sphingolipid pathway) has been implicated in insulin resistance and glomerular injury [33, 34]. We have previously investigated fatty acid species in a subset of the FinnDiane cohort, and found that serum sphingomyelin was positively correlated with 24-h AER [35] and, together with saturated fatty acid levels, was higher in patients at high risk of progression [21]. Of interest, altered sphingomyelinase activity may partly explain how rituximab inhibits the recurrence of (nondiabetic) focal segmental glomerulosclerosis [36], although the direction of the effect suggests increased production rather than clearance of ceramides. These findings nevertheless point to a lipotoxic effect on the kidneys, presumably from the excess saturated fatty acids in the circulation (much of which is contained in VLDL).

Intermediate-density lipoprotein and LDL particles contain a large proportion of total cholesterol and are considered atherogenic in the general population; however, their relevance to diabetic kidney disease was found to be limited in this study. Progression from microalbuminuria was weakly associated with elevated absolute concentrations of cholesterol in every LDL subclass (Fig. 2b), and we have previously observed similar patterns in a selected subset of the FinnDiane cohort [21]. Other studies have also identified calculated LDL or total cholesterol as significant predictors of progression [37-39], but the results could be confounded by VLDL or IDL cholesterol. In this study, much of the increased total cholesterol seemed to originate from the VLDL rather than LDL subclasses, which supports the concept of hepatic overproduction as an early lipidaemic insult on the path from microalbuminuria to progressive injury.

The weak link between mortality and LDL subclasses is noteworthy in this study. Patients with type 1 diabetes with microvascular complications are highly susceptible to cardio- and cerebrovascular events, but it seems that the dyslipidaemia that is associated with kidney disease is more important than the traditional population-based macrovascular disease risk factors. For example, statins reduce LDL cholesterol considerably, but they are less effective in protecting from kidney disease [40, 41]. Therefore, lowering total cholesterol or calculated LDL may not be sufficiently specific goals for drug development, but clinical trials involving patients with type 1 diabetes should verify the pharmacological effects at the subclass level, especially now that there are cost-effective technologies [11, 42-45]. Furthermore, because there is an association between the severity of kidney disease and lipoprotein profile, these trials should address carefully the additional modulating effect of complications.

The HDL fraction contains a complex set of multi-functional particles at different stages of maturation. In general, HDL particles (particularly the large mature ones) are considered protective against vascular diseases: they are able to remove excess cholesterol from peripheral tissues and attenuate oxidative stress, and they also have anti-inflammatory properties [46-48]. Depletion of HDL cholesterol is typically observed in connection with increased triglycerides (a feature of the metabolic syndrome), and an abnormal cholesteryl ester transfer protein-mediated exchange of lipids between VLDL and HDL particles may explain some of this association [49, 50]. In this study, the large HDL subclass exhibited both absolute and relative depletion of cholesterol for the triglyceride-rich phenotype (Fig. 4 and Online Supplement 1), and large HDL cholesterol was inversely associated with kidney disease and mortality (Figs 1-3); these findings are consistent with the above-mentioned anti-atherogenic properties of HDL particles. When kidney function declines, HDL fails to mature properly to its cholesterol-rich form and remains as a small lipid-poor particle [51]. The NMR method cannot detect this type of particle-specific effect, but our findings nevertheless suggest that end-stage renal disease causes secondary disruption of HDL metabolism. In particular, an insulin-resistant phenotype with depleted large HDL cholesterol is the typical feature of the albuminuric phase and a secondary loss of the largest HDL subclass (Figure 1c) may occur when kidney function declines.

Our biochemical study was limited to a single time-point. Hence, it was not possible to determine whether there was a change from one phenotype to another in patients during their lifetime. The study population consisted of ethnically homogeneous Finns, and the heritability of subclass lipids is substantial in the Finnish population [52]. Therefore, our results should not be directly translated to other populations without further studies. Diabetes duration is another important factor to be considered: the FinnDiane Study is not an inception cohort, and the prospective results, although carefully matched, are only applicable to patients with long-standing type 1 diabetes. Furthermore, the visual SOM patterns and the selected extreme phenotypes are always subject to interpretation. To minimize the possibility of false-positive findings due to over-fitting of the statistical model, we excluded the clinical classifications and prospective data from the model training phase.

The large number of samples, biochemical measures and longitudinal clinical diagnoses gives strength to our study. For instance, we intentionally focused on the NMR lipoprotein data for consistency, but also analysed apolipoproteins A-I, A-II and B-100 to verify dataset integrity. Of note, the apolipoprotein B/A-I ratio resembled the triglyceride/cholesterol ratio observed on the SOM (Fig. 4a and Online Supplement 2J).

The combination of low serum cholesterol (bottom tertile <4.4 mmol L−1) and low triglycerides (bottom tertile <0.87 mmol L−1) indicates the most favourable lipoprotein phenotype in Finnish individuals with type 1 diabetes. If serum lipids are elevated, a high triglycerides/cholesterol ratio (top tertile >0.26) denotes the high-risk phenotype with poor glycaemic control, excess body weight and hypertension. Our detailed analyses revealed that the imbalance between cholesterol and triglycerides affects the whole of lipoprotein metabolism simultaneously. Therefore, it is not sufficient to focus on the established risk factors such as LDL cholesterol, but anti-atherogenic interventions in type 1 diabetes should target all the lipoprotein subclasses to negate the effects of any structural and compositional abnormalities due to the diabetic milieu. It is also noteworthy that a safe triglyceride concentration may be much lower in type 1 diabetes than in the general population.

Author contributions

V-PM, P-HG and MA-K designed the study. V-PM analysed the data and wrote the manuscript. V-PM, CF, NT, LMT and P-HG collected the clinical data and samples. JV, OTR and MS provided clinical interpretation. PS, AJK and MA-K designed the NMR platform. PS performed the NMR experiments and V-PM, PS and AJK analysed the NMR data. All authors reviewed, commented on, and accepted the final manuscript.

No outside editorial assistance was used in writing this manuscript.

Conflict of interest statement

No conflict of interest was declared.

Acknowledgements

The authors are grateful for the contribution of healthcare professionals who collected data on the type 1 diabetic patients in the local hospitals (the full list of the FinnDiane Study Group is available in Online Supplement 4).The FinnDiane Study was supported by grants from the Folkhälsan Research Foundation, the Wilhelm and Else Stockmann Foundation, the Liv och Hälsa Foundation and the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement 223211 (Collaborative European Effort to Develop Diabetes Diagnostics). The work was also supported by grants from the Orion-Farmos Research Foundation (to V-PM), Finnish Foundation for Cardiovascular Research (to MS and MA-K), Jenny and Antti Wihuri Foundation (to AJK), the Academy of Finland (137870; to PS), the Responding to Public Health Challenges Research Programme of the Academy of Finland (129269 and 129429; to MS and MA-K respectively) and the Strategic Research Funding from the University of Oulu (to MA-K).

Ancillary