Water T₂ as an early, global and practical biomarker for metabolic syndrome: an observational cross-sectional study

This was a biomarker discovery study with an observational, cross-sectional design. Initially, it was designed to test the hypothesis that serum water T₂ was associated with markers of insulin, glucose and lipid metabolism. Phase 1 was designed to collect data on 29 non-diabetic human subjects. The target number of subjects was derived from a power calculation with α = 0.05, β = 0.2 (statistical power of 0.8) and a correlation coefficient of 0.5 [32, 33]. The actual number of subjects analyzed in Phase 1 was 28.

Analysis of the Phase 1 data led to the observation that, in addition to insulin-, glucose- and lipid-related markers, water T₂ appeared to be correlated with inflammatory markers. So Phase 2 of the study added an expanded set of inflammatory and acute phase markers. The target number of 38 subjects for Phase 2 was derived from a power calculation with α = 0.05, β = 0.1 (statistical power of 0.9) and a correlation coefficient of 0.5 [32, 33]. The actual number of subjects analyzed in Phase 2 was 44.

Therefore, the total number of subjects analyzed and reported here was 72: 28 in Phase 1 and 44 in Phase 2. Many biomarkers were collected in both Phase 1 and 2, while some were collected only in Phase 2.

Human subject volunteers were recruited with prior written informed consent into two protocols approved by the Institutional Review Board of the University of North Texas Health Science Center in Fort Worth (UNTHSC). One protocol recruited adult subjects from the student and staff population of UNTHSC, including spouses, friends and associates. The second protocol recruited Fort Worth community members enrolled in the Health & Aging Brain Study at UNTHSC [34]. Exclusion criteria for the current study included diabetes (HbA_1C ≥ 6.5, fasting plasma glucose ≥ 125 mg/dL or prior history/diagnosis), active acute or chronic illness (C-reactive protein > 10 or history/diagnosis), history of bleeding disorders or difficulty donating blood, confirmed or suspected pregnancy from medical history, or not fasting for 12 h. Inclusion criteria were ages 18 and up. A total of 87 subjects were enrolled in the study, with 72 of the 87 subjects qualified according to the inclusion and exclusion criteria.

All subjects completed a comprehensive medical history form and a follow-up interview prior to the day of blood draw. On the morning of the blood draw, anthropometric measurements (height, weight, waist circumference, blood pressure and heart rate) were taken by the study nurse, and urine samples were screened for microalbuminuria using Chemstrip Micral (Roche Diagnostics).

Fasting blood samples were drawn at 7 a.m. by the study nurse following a standard order-of-draw protocol. For plasma preparation, blood was drawn into BD Vacutainer lavender-top tubes containing K₂EDTA as the anticoagulant. For serum used for NMR and viscosity measurements, blood was drawn into plain glass red-top tubes lacking any gel separator or clot activators (BD models 366,441 and 366,430) to avoid potential interference in the NMR and viscosity measurements. Every effort was made to collect enough blood to perform all planned measurements. However, there were instances where the amount of blood collected from a given subject was not sufficient or samples were rejected by the testing lab due to hemolysis or other reasons. That variability, along with a few laboratory errors (instrument malfunction, data not collected or accidentally overwritten), accounted for the test-to-test differences in sample size (n) for the measurements listed in the tables. No attempts were made to interpolate or fill in missing data.

The plasma and serum samples were processed immediately after each blood draw. The serum samples were allowed to clot for 30 min, while plasma samples were being centrifuged. The first spin was at 3380 rpm (1590×g) for 10 min at room temperature to pellet and remove blood cells, followed by a second spin of the supernatant at 3800 rpm (2361×g) for 15 min to remove residual cells or debris. The presence of residual platelets was ruled out by dynamic light scattering analysis of each twice-centrifuged sample using a Wyatt Mobius instrument. The water T₂ measurements were performed in triplicate on a sample of fresh plasma followed immediately by three repeats on fresh serum such that all water T₂ measurements were completed within 2 h after the blood draw. Likewise, viscosity was measured on fresh serum and plasma samples within a few hours of the blood draw using a VISCOLab3000 instrument [35]. Aliquots of fresh serum were sent on ice to Atherotech, Inc. for Vertical Autoprofile (VAP) advanced lipoprotein testing and to determine LDL-P, hs-CRP, GGT, homocysteine, and Lp(a). Other aliquots of fresh plasma and serum were temporarily stored at 4°C prior to being sent the same day to Quest or Labcorp for diagnostic testing. Plasma aliquots for amino acid analysis, glucagon, fibrinogen, free fatty acids and proinsulin were frozen immediately after preparation and stored at − 80 °C prior to shipment to Quest. Other individual aliquots of plasma and serum were frozen at − 80 °C for subsequent in-house analysis using the following assay kits: apolipoprotein E concentration (Abcam, Ab108813), ORAC antioxidant capacity (Cell Biolabs, STA-345), protein carbonyl content (Cell Biolabs, STA-307), HNE (Cell Biolabs, STA-838), phospholipids (Wako Diagnostics, Assay Kit C), α₂-macroglobulin (Abcam, ab108888), PAI-1 (Abcam, ab184863), neutrophil elastase (Abcam ab119553), soluble fibronectin (Abcam, ab181419), l-lactate (Abcam ab65331), endotoxin (Thermo Scientific, PI88282), and staphylococcus enterotoxin (Creative Diagnostics, DEIA-CL032).

Cytokines were assayed using the V-PLEX platform from Meso Scale Discovery, with a customized human cytokine plate for IL-6, IL-1β, TNF-α and IL-10 (MSD, K151AOH-1). Adiponectin and Factor VII were assayed using MSD Plate K151BXC-1, and sICAM-1, using MSD Kit K151SUD-1. All tests using − 80 °C frozen specimens were performed on samples that underwent only one freeze–thaw cycle.

Measurements of T₂, the transverse relaxation time constant, were performed at 37 °C using a Bruker Minispec mq20 benchtop time-domain NMR instrument equipped with a 10 mm variable temperature probe (Model H20-10-25-AVGX). The 10 mm-diameter sample tube included a 3 mm coaxial insert (Norell NI10CCI-B), and the insert was filled to a sample height of 1 cm, corresponding to a sample volume of ~ 50 μL.

The modified Carr–Purcell–Meiboom–Gill (CPMG) pulse sequence we employed for T₂ measurements is illustrated in Figure 1 of Ref. [31]. In our experience, a critical factor in obtaining high quality NMR relaxometry data with aqueous samples is to avoid radiation damping, particularly when higher magnetic fields and/or larger sample volumes are used (e.g., 10 mm tubes without a coaxial insert). The magnitude of the radiation damping depends on the particular instrument and probe design. We determined that the sample size in the coaxial insert was sufficiently small in this instrument to avoid radiation damping. Thus, it was not necessary to use the optional composite 180° pulse and Δ delay shown in the pulse sequence. Other experimental aspects pertinent to NMR data collection and analysis are detailed in Ref. [31].

The bivariate correlations, multiple and logistical regression analyses, categorical means comparisons, receiver operator characteristic curves and principal components analysis with variable clustering were performed using JMP Pro version 13.1 (SAS, Inc.) and GraphPad Prism v. 6.05 (GraphPad Software, Inc.). The guiding principles for the statistical analyses were derived largely from the books by Motulsky and Huber [36, 37]. Regression residuals were analyzed in GraphPad Prism using the strategy outlined by Klingenberger [38].

The bivariate correlation coefficients were calculated using three complementary methods: Pearson product-moment r, Spearman ρ, and the Huber M-value [36, 37]. These three estimators involve different assumptions about the data, and thus, have different strengths and weaknesses. Included among the assumptions for the Pearson r analysis is that both measures are sampled from a Gaussian distribution [36]. Analysis of the variables in this study revealed that more than half were not Gaussian distributed, as assessed using the D’Agostino–Pearson omnibus normality test implemented in GraphPad. However, a natural log transformation corrected the problem in nearly all cases (see Table 3, footnote c).

Another key assumption of the Pearson and Spearman correlations is that there are no outliers. Pearson is especially sensitive to outliers [36, 37, 39], which can lead to an over- or under-estimation of correlation coefficients. The Huber M-value has the distinct advantage of being robust to outliers [37]. Therefore, we chose to not eliminate any outliers, with the caveat that the Pearson and Spearman coefficients would be interpreted together with the Huber M-values and with careful inspection of the scatter plots.

In most cases, all three correlation coefficients had comparable values. Specific cases where outliers appeared to cause a significant under-estimation of the Pearson or Spearman coefficients included HbA1c, serum % globulins, VLDL-C, Rem-C, platelet count, lymphocyte count, RDW, complement C4c, and anion gap, uncorrected. Cases where outliers may have caused an overestimation of Pearson and Spearman values were asparagine, PAI-1, and haptoglobin. For these reasons, the Huber M-value was taken as the single-best estimate of the correlation, especially in cases where outliers had an influence.

The use of all three methods was particularly useful when correlations were ambiguous, i.e., when one method yielded a statistically significant result, while others did not. One notable example was body-mass-index or BMI, where a weak but statistically significant Pearson correlation was observed with both plasma and serum water T₂. However, the Spearman and Huber M-value coefficients were weaker and not statistically significant. Inspection of the scatter plots revealed that the Pearson analysis was heavily influenced by a single outlier point that fell well outside of the Huber 95% confidence ellipse. This point corresponded to the subject in this study with the highest BMI and one of the lowest water T₂ values. However, this study contained mostly non-obese individuals. Therefore, a proper assessment of the possible correlation between BMI and water T₂ values in the context of obesity will require the study of a larger number of obese subjects. In the current study population, water T₂ did not correlate with BMI.

Multiple linear regression models were built from Gaussian-distributed variables using the stepwise analysis feature in JMP v. 13.1. Potential predictor variables were chosen from the output of the bivariate analyses (Table 3, Additional file 1: Tables S1, S2), and from the principal components analyses with variable clustering, i.e., the most representative variable in each cluster. The stepwise analysis provided starting points for the exploration of different models and the reduction of possibilities. Acceptable models satisfied all three of the following criteria: (1) the p values for all predictor variables in the model were significant at α = 0.05, (2) the model avoided overfitting, as assessed using k-fold cross validation with k = 10, and (3) the adjusted R² was maximized, within the constraints of criteria (1) and (2). The highest number of predictor variables in our models was five, including the y-intercept. This number is consistent with the rule of thumb that models should contain no more than 1 predictor variable for every 8–10 observables (subjects).

For the comparison of two means, unpaired two-tailed t-tests were used to assess significance, assuming equal variances and that the data were sampled from a Gaussian distribution. Five tests to confirm equal variances (O’Brien, Brown–Forsythe, Levene, Bartlett and 2-sided F-test) were performed as implemented in JMP 13.1. In one case specified in Table 4, equal variances could not be confirmed, so significance was assessed using the Welch test instead of the t test.

Receiver operator characteristic (ROC) curves were generated and evaluated using JMP 13.1. For determining the sensitivity, specificity and cutoff values for plasma and serum water T₂, the McAuley Index value of ≤ 6.07 was chosen as the categorical reference standard for early insulin resistance [26]. The equation for the McAuley Index is provided in the abstract of Ref. [26], and the input values of fasting insulin and triglyceride are provided in Table 3 (“Method A”) of Ref. [26]. The optimal ROC cutoff points for plasma and serum water T₂ were those that fell closest to the [0,1] coordinate, i.e., those which intersected with or closely approached the gray 45° tangent line shown for each ROC curve.

The clinical characteristics of the human study population are presented in Table 1. Overall, this was a fairly diverse group of asymptomatic, non-diabetic adult volunteers spanning a wide age range. The gender distribution was approximately equal. The 72 subjects included 35 white, 23 Asian, 10 Hispanic and 4 African American individuals. The inclusion and exclusion criteria are specified in “Methods” section. As shown in Table 1, the mean values for the diagnostic markers fell near the middle of their normal reference ranges. With respect to glucose markers, 47 of the 72 subjects were normoglycemic by American Diabetes Association criteria [40], with both fasting glucose < 100 mg/dL and HbA_1c < 5.7%. The remaining 25 subjects had glucose and/or HbA_1c values consistent with prediabetes.

The plasma and serum water T₂ values were Gaussian distributed and exhibited high coefficients of variation (7–8%) across the study cohort (Table 1). This high variance did not result from imprecision in the NMR relaxometry measurements, as the coefficient of variation for multiple repeats on a single subject averaged < 1%. Rather, the high variance was caused by subject-to-subject biological variation reflecting the range of metabolic health among the subjects.

To identify the specific factors governing this variation, up to 130 diagnostic blood biomarkers were measured for each subject and correlated with plasma and serum water T₂ values. As detailed in Methods, the study was conducted in two phases, with many biomarkers measured in both phases (n = 72), and some additional biomarkers measured only in Phase 2 (n = 44). While a number of markers showed significant correlations with plasma water T₂ (Table 2, middle column), many others did not (Table 2, right column). Among those showing no correlation were albumin and sodium, markers of a subject’s hydration status. In addition, markers related to paramagnetic ions or their binding proteins (transferrin, total iron, total iron binding capacity, percent iron saturation, ferritin and ceruloplasmin) showed no correlation with plasma or serum T₂, in spite of the inherent sensitivity of T₂ to changes in paramagnetic ions. Thus, in this study population, variation in water T₂ was not associated with variation in the subject’s hydration state or iron/copper status.

Bivariate scatterplots for the correlation between plasma water T₂ and insulin C-peptide, the McAuley Index, total serum protein concentration, LDL-cholesterol, triglycerides and complement C3 (measured clinically as its stable conversion product C3c) are displayed in Fig. 1. The bivariate normal density ellipse, indicated in red, provides a visual indicator of the Huber correlation at the 95% confidence level. The Huber M-value has the advantage of being robust to outliers, as compared with the Pearson r and the Spearman ρ values [37, 41]. This enabled us to analyze the correlations without excluding any real or perceived outliers, as detailed in “Methods” section.

Table 3 lists the statistically significant bivariate Huber correlation coefficients for biomarkers that were unambiguously correlated with both plasma and serum water T₂. Strong correlations (0.5–0.7) were observed between water T₂ and fasting insulin C-peptide, insulin, and proinsulin, as well as indices derived from insulin plus glucose, or insulin plus triglycerides (Table 3, first section). Note that the correlation with fasting glucose, while statistically significant, is considerably weaker (~ 0.3). Also, strong correlations were observed with protein markers, specifically plasma and serum globulins, total serum and plasma proteins, plasma and serum viscosity, as well as alanine aminotransferase, a marker of liver function (Table 3, second section). However, no significant correlations were observed with plasma and serum albumin, which account for ~ 60% of the total plasma and serum protein mass.

Correlation coefficients of ~ 0.4 to 0.6 were observed between water T₂ and markers of cholesterol- and triglyceride-rich lipoproteins (Table 3, third section). Moreover, correlations were observed with markers of inflammation and coagulation (Table 3, fourth section), especially white blood cell count, fibrinogen, complement C3c and C4c, and C-reactive protein. Finally, statistically significant correlations were observed with electrolyte markers, namely lactate, total measured anions (Cl⁻ + HCO₃ ⁻) and the anion gap (Table 3, last section). Electrolyte abnormalities have been associated with insulin resistance, inflammation and high blood pressure in the National Health and Nutritional Examination Survey [42‐44].

The association between plasma and serum water T₂ was very strong, with a correlation coefficient of 0.8. However, some biomarkers correlated with plasma, but not serum, water T₂ or vice versa. Those that correlated only with plasma water T₂ include platelet and monocyte counts, red cell distribution width, mean corpuscular volume, mean corpuscular hemoglobin, serine, asparagine, glutamine, threonine, β-alanine, chloride, gamma glutamyl transferase, erythrocyte sedimentation rate, plasminogen activator inhibitor-1, α₁-acid glycoprotein and interleukin-6. Those that correlated only with serum water T₂ included immunoglobulin G, lipoprotein-associated phospholipase A₂, red blood cell count, tyrosine and 3-methyl histidine. Full sets of Pearson, Spearman and Huber correlation coefficients are provided in Additional file 1: Tables S1 and S2 for plasma water T₂ and serum water T₂, respectively.

The overall pattern of correlations is consistent with key elements of insulin resistance and the metabolic syndrome, namely hyperinsulinemia, dyslipidemia, pro-inflammation, pro-coagulation, and electrolyte imbalances. In all five categories, plasma and serum water T₂ values were inversely correlated with those metabolic abnormalities.

Of note, water T₂ measurements did not correlate with body-mass index or waist circumference, at least in this mostly non-obese population. Also, plasma and serum water T₂ did not correlate with free fatty acid levels or with age. The mean plasma water T₂ value was lower in women than men, but this difference was not statistically significant (752.1 vs. 775.3, p = 0.104). By contrast, the mean serum water T₂ values were nearly identical in women and men (817.0 vs. 819.8, p = 0.838). The possible gender difference seen with plasma but not serum water T₂ could be attributed to a higher level of fibrinogen observed for women vs. men, although the gender difference in fibrinogen did not reach statistical significance (289.4 vs. 255.9, p = 0.091).

The inflammatory markers showed weaker correlation coefficients with serum water T₂ (~ 0.3 to 0.5; Table 3) as compared with plasma water T₂ (~ 0.4 to 0.7, Table 3). The primary difference between plasma and serum is the absence of fibrinogen or Factor I in serum [45], providing an explanation for the weaker correlations observed for serum water T₂. Fibrinogen is a key reporter that connects plasma water T₂ with inflammation and coagulation status. However, fibrinogen is not the only such reporter, as discussed below.

The observed bivariate correlations led us to consider factors that may contribute directly to the variation in plasma and serum water T₂, as well as those that may be indirectly linked through another variable or a network of variables. Human blood plasma is a complex mixture containing hundreds of different proteins and lipoproteins as well as numerous small molecule metabolites. At first thought, de-convoluting these myriad variables would seem hopelessly complex. However, water mobility, and hence water T₂, is affected mainly by its binding to macromolecules, as the influence of small molecule metabolites such as glucose is negligible [46]. Moreover, the sixteen most abundant proteins and lipoproteins in plasma (albumin, IgG, transferrin, fibrinogen, IgA, α₂-macroglobulin, apolipoprotein AI, α₁-antitrypsin, complement C3, IgM, haptoglobin, apolipoprotein B, α₁-acid glycoprotein, apolipoprotein E, complement C4, and ceruloplasmin) account for > 99% of the total plasma protein mass. Thus, identifying the primary variables that contribute to water T₂ is not an intractable problem.

We used three approaches to reduce the complexity of the network and tease apart variables that contribute independently and/or additively to the variation in plasma and serum water T₂. The first approach utilized a principal components analysis with variable clustering [47]. This algorithm reduced the dimensionality by identifying clusters of variables that are most closely related. The results of one such analysis are presented in Additional file 1: Table S3. The statistical clusters correspond largely to the categories of markers based on physiological considerations. In the example shown in Additional file 1: Table S3, cluster 1 represents insulin and glucose markers; clusters 2–4, protein and viscosity markers; cluster 5, lipid and lipoprotein markers, clusters 6–8, inflammation markers; and cluster 9, electrolyte markers. One benefit of this analysis was to define the “most representative variable” in each category, which served as a starting point for building multiple regression models.

The second approach to reducing the complexity of the variable network used multiple regression to control for the effect of confounders and identify variables that have independent contributions to plasma water T₂. The parameters for four of the best multiple regression models for plasma water T₂ are provided in Additional file 1: Table S4. Model 1 was derived using variables collected in both phases of the study (72 subjects), while Models 2–4 included at least one variable that was measured only in Phase 2 (44 subjects). These models accounted for approximately two-thirds to three-fourths of the variation in plasma water T₂. Attempts to add more variables to the models resulted in overfitting, as assessed using k-fold cross validation and described in “Methods” section.

The primary independent contributors to plasma water T₂ were (1) insulin c-peptide, (2) total serum or plasma protein, plasma globulins, or plasma viscosity, (3) total cholesterol or apolipoprotein B, and (4) white blood cell count or fibrinogen. In general, one biomarker from each of four categories (insulin, proteins, lipids and inflammation) had independent contributions to the variation in plasma water T₂. A key observation was that plasma water T₂ was correlated with markers of insulin resistance or metabolic syndrome, even after correcting for total serum or plasma protein, serum or plasma globulins, or plasma viscosity. Attempts to correct for BMI and age did not yield statistical significance for those variables. Likewise, variables in the electrolyte category did not display a contribution to plasma water T₂ independent of the other four categories. However, lactate could be substituted for insulin c-peptide or insulin in models for serum water T₂. Multiple regression models for serum water T₂ were similar to those for plasma water T₂. Examples are presented in Additional file 1: Table S5.

The third approach used categorical, rather than continuous, variables to compare means and assess the additivity of contributions to plasma water T₂. As shown in Table 4, the subjects were categorized as having or not having hyperinsulinemia, dyslipidemia, inflammation or electrolyte abnormalities, using three measures of each condition. In addition, the subjects were categorized as having or not having clinically-defined metabolic syndrome. The differences in mean plasma water T₂ values were computed for each of the measures and conditions. The differences were greater when two or more conditions were combined. The largest difference in mean plasma water T₂ values was observed for subjects who had hyperinsulinemia plus dyslipidemia plus inflammation (Table 4, last row). Thus, the lowest plasma water T₂ values were observed in subjects who had multiple elements of early metabolic syndrome, as the effect on T₂ has both independent and additive components.

To further assess the dose–response relationship between the number of metabolic conditions and plasma water T₂, the subjects were divided into quintiles with respect to T₂ values, and the average number of conditions for each quintile was calculated on a scale of 0–4. The average number of conditions increased from 0.71 (top quintile) to 1.57, 2.36, 2.79, and 3.57 for the 4th, 3rd, 2nd and lowest quintiles of plasma water T₂, respectively. The lower the T₂ value, the greater the number of conditions associated with metabolic syndrome.

To further quantify the association of plasma water T₂ with these conditions, logistic regression models were constructed with each of the conditions in Table 4 serving as the categorical outcome. Each model was adjusted for potential confounders, specifically BMI, age and gender. For hyperinsulinemia, the unit odds ratio for plasma water T₂ was 1.034 (95% confidence limits 1.016–1.053, p < 0.0001). Thus, the observed 67.4 ms decrease in mean plasma water T₂ increased the odds of hyperinsulinemia by a factor of 1.034^67.4 = 9.5. The corresponding unit odds ratio for serum water T₂ and hyperinsulinemia was 1.026 (1.011–1.041, p < 0.0001). When dyslipidemia was the categorical outcome, the unit odds ratios were 1.028 (1.013–1.044, p < 0.0001) and 1.030 (1.015–1.047, p < 0.0001) for plasma and serum water T₂, respectively. This translated into an odds ratio of 5.8 for the decrease in mean plasma water T₂ associated with dyslipidemia (Table 4).

Even higher unit odds ratios were observed when inflammation was the categorical outcome: 1.034 (1.017–1.052, p < 0.0001) and 1.040 (1.019–1.061, p < 0.0001) for plasma and serum water T₂, respectively. Thus, the observed 69.9 ms decrease in mean plasma water T₂ translated into a 10.4-fold increase in the odds of having inflammation (Table 4). Smaller, but statistically significant odds ratios were observed when the categorical outcome was electrolyte abnormalities. Subjects who met the clinical criteria for metabolic syndrome as a whole had lower T₂ values and higher odds ratios as well (Table 4).

The large decrease in mean water T₂ values and high odds ratios provide further evidence of the strong association between water T₂ and elements of the metabolic syndrome.

As an initial attempt to assess the sensitivity and specificity of water T₂ for detecting elements of early metabolic syndrome, we performed a receiver operator characteristic curve (ROC) analysis. Insulin resistance, as defined by the McAuley Index, was used as the reference standard. The McAuley Index combines input from fasting triglycerides and fasting insulin and thus, captures two related elements of MetS. Moreover, the McAuley Index was rigorously calibrated with 178 normoglycemic subjects using a direct measure of insulin sensitivity, obtained from the euglycemic clamp, as the outcome variable [26]. Thus, it was calibrated to detect the earliest stage of insulin resistance. We categorized the 47 normoglycemic subjects in the current study using the McAuley Index cutoff of ≤ 6.07, derived using a fasting insulin of 12.2 μU/mL and a fasting triglyceride of 1.5 mM (133 mg/dL) as input variables [26]. As illustrated by the blue curve in Fig. 2, plasma water T₂ detects early insulin resistance with a sensitivity of 100% and a specificity of 87% at a cutoff value of ≤ 745.0 ms. The area under the curve (AUC) is 0.96. By contrast, HbA1c and glucose—the tools most widely used for diabetes screening and risk assessment—show lower values of area-under-the-curve (AUC) and lower combinations of sensitivity and specificity (Fig. 2, red and yellow, respectively). To detect early insulin resistance with 100% sensitivity, the HbA_1c cutoff would have to be lowered to 5.4, which would lead to poor specificity and a 42% false positive rate.

For plasma water T₂, the likelihood ratio for a positive test result (LR+) was 7.8. The likelihood ratio for a negative result (LR−) was zero.

Serum water T₂ yielded ROC curve parameters similar to those of plasma: AUC = 0.94 with a sensitivity of 100% and specificity of 80% at a cutoff value of ≤ 811.8 ms. The LR+ and LR− values were 5.0 and zero, respectively.

As a further exercise, we used the harmonized clinical criteria for metabolic syndrome [4] as the reference standard for ROC analysis, instead of the McAuley Index. For plasma water T₂, this analysis yielded sensitivity and specificity values of 73% at a cutoff value of 745 ms and an AUC = 0.75. For serum water T₂, the values were similar: 71% sensitivity and 69% specificity at a cutoff of 804.8 ms and an AUC = 0.71.

Using the plasma and serum water T₂ cutoff values of 745.0 and 811.8 ms, along with the current criteria for prediabetes [40] and metabolic syndrome [4], the subjects in this study were classified into metabolic stages, as shown in Fig. 3. Of the total of 72 subjects, 31 had normal metabolism, while the other 41 had early metabolic abnormalities, prediabetes and/or metabolic syndrome defined by clinical criteria. Of the 41, 25 had prediabetes, and 10 of those 25 met the clinical criteria for metabolic syndrome. The remaining 16 did not have prediabetes, but had plasma or serum water T₂ values below the cutoffs. Of those 16, only three met the clinical criteria for metabolic syndrome. Therefore, plasma and serum water T₂ uniquely identified 13 subjects (18% of the study population) with metabolic abnormalities that would have gone undetected by the current clinical definitions of prediabetes and metabolic syndrome.

Further examination of the 13 normoglycemic subjects with low plasma or serum water T₂ values yielded the following observations:

Thus, the 13 normoglycemic subjects with low water T₂ had a heterogeneous set of early metabolic abnormalities (hyperinsulinemia, dyslipidemia and/or inflammation) that were undetected by the clinical criteria for MetS. This observation illustrates the limitations of the clinical definition of MetS, and the unique power of water T₂ to detect this condition.

To assess the role of the 16 most abundant plasma proteins and lipoproteins in metabolic syndrome, the plasma and serum water T₂ cutoffs of ≤ 745.0 and ≤ 811.8 ms were used to classify the study subjects into those with and without the syndrome. The percent differences in the mean plasma protein concentrations for the two groups of subjects are displayed in Fig. 4. Subjects with metabolic syndrome, here defined by low water T₂, displayed statistically significant increases in the mean concentrations of fibrinogen, complement C3c, haptoglobin, apolipoprotein B, α₁-acid glycoprotein and complement C4c, as well as total plasma proteins and globulins (Fig. 4, black bars). By contrast, there were no significant changes in the concentrations of the other 10 proteins: albumin, IgG, transferrin, IgA, α₂-macroglobulin, apolipoprotein AI (HDL), α₁-antitrypsin, IgM, apolipoprotein E and ceruloplasmin (Fig. 4, grey bars). Thus, low T₂ values and metabolic syndrome are characterized by increases in the concentrations of a specific subset of acute phase proteins and lipoproteins.

This biomarker discovery study provides an initial assessment of the metabolic information content of plasma and serum water T₂ in non-diabetic human subjects. It needs to be followed up by a series of validation steps: in longitudinal cohorts, in the postprandial state, against direct measures of insulin sensitivity, in larger populations and in response to therapeutic interventions, as explained below.

This cross-sectional study does not provide direct evidence for the ability of plasma or serum water T₂ to predict future risk for type 2 diabetes or cardiovascular disease. However, the individual acute phase proteins and lipoproteins monitored by water T₂ have been shown, in prior prospective studies, to be associated with incident type 2 diabetes and cardiovascular events. Future work should assess water T₂ in longitudinal cohorts.

All blood testing was performed on subjects who underwent a 12-h overnight fast. The current results provide little or no direct insight into postprandial metabolism or glucose tolerance following an oral glucose load. Thus, T₂ will need to be correlated with results from oral glucose tolerance tests.

This study employed indirect measures of insulin resistance based on fasting insulin, insulin c-peptide, proinsulin, triglycerides and indices derived from them, most notably the McAuley Index. An important next step is to validate plasma and serum water T₂ values against direct measures of tissue insulin resistance.

The correlation of water T₂ with BMI and waist circumference in the context of obesity was inconclusive, as this study enrolled mainly non-obese subjects. Further work needs to be done to evaluate the association of water T₂ with BMI and central obesity. Likewise, the relationship of water T₂ with blood pressure warrants further evaluation.

While the 72 subjects employed in this biomarker discovery study provided sufficient statistical power to identify biologically important correlation coefficients greater than ~ 0.25, a future study with a larger number of subjects will permit a more comprehensive statistical analysis. The main benefit of a larger study would be the generation of multiple and logistical regression models that can accommodate a larger number of predictor variables. Remarkably, we were able to account for three-fourths of the variation in water T₂ with up to five predictor variables The high sensitivity and large variance in water T₂ made that possible. However, a study with a larger number of subjects may be able to account for other independent predictors of water T₂.