Inflammatory markers and RBC status exhibit concomitance with glycemic variations
Data on clinical diagnostics with a total sample size (n = 208,137) was obtained. Two sub-cohorts were derived according to HbA1c (n = 142,011) and fasting blood glucose (FBG, n = 35,362) tests. Means (± SD) for 12 different clinical parameters were measured for the glycemic states categorized under HbA1c (Table 1A) and FBG (Table 1B).
Out of the 142,011 individuals who were tested for HbA1c, 60.94% had hyperglycemia (H), while 21.18% and 17.85% had borderline hyperglycemia (BH) and Normoglycemia(N), respectively, using American Diabetes Association (ADA) thresholds15. Higher rates of Hyperglycemia (53.45% vs 46.52%) and borderline hyperglycemia (54.81% vs 45.16%) were observed amongst males as compared to females (Table 1A), with no contrasting differences in the mean ages. Of the 35,362 individuals who got tested for FBG, 38.62% had hyperglycemia, 30.10% were borderline and 31.28% had normoglycemia. Males had higher rates of hyperglycemia (58.47%) as compared to females (41.51%). A similar trend was observed amongst males and females for borderline hyperglycemia (60.45% vs 39.53%), with similar mean ages, in the FBG classified sub-group (Table 1B). For both sub-groups (HbA1c and FBG), heightened mean expression of CRP, WBC, Platelet, NLR, PLR, RBC, and HCT were observed in hyperglycemic state, whereas a negative trend was seen for RBC indices including MCV, MCH, and MCHC. Hb exhibited stable mean expression for all levels of glycemia (Table 1).
Table 1
Clinical diagnostic characteristics stratified by glycemic states using American Diabetes Association (ADA) thresholds. Mean values (± SD) for diagnostic biomarkers compared by glycemic levels under categorization by (A) HbA1c (n = 142,011) and (B) Fasting blood glucose, FBG (n = 35362).
(A) Glycemic levels categorized under HbA1c grouping |
Parameters | Normoglycemia | Borderline hyperglycemia | Hyperglycemia |
HbA1c < 5.7% | 5.7% < HbA1c < 6.4% | HbA1c > 6.4% |
(n = 25,355) | (n = 30,072) | (n = 86,584) |
Age (years) | Female | 12,123 (47.81%) | 45 ± 15 | 13,581 (45.16%) | 55 ± 12 | 40,285 (46.52%) | 55 ± 12 |
Male | 13,227 (52.18%) | 48 ± 14 | 16,484 (54.81%) | 54 ± 14 | 46,286 (53.45%) | 54 ± 13 |
HbA1C (%) | 5.3 ± 0.3 | 6.1 ± 0.2 | 8.3 ± 1.6 |
CRP (mg/dL) | 1.29 ± 3.1 | 1.73 ± 3.75 | 2.37 ± 4.26 |
WBC (x109/L) | 8.35 ± 4.4 | 8.41 ± 2.47 | 8.91 ± 3.4 |
Platelets (x109/L) | 266 ± 82 | 267 ± 86 | 271 ± 88 |
NLR | 2.22 ± 1.62 | 2.14 ± 1.47 | 2.22 ± 1.77 |
PLR | 9.59 ± 7.18 | 9.4 ± 6.26 | 9.75 ± 7.26 |
RBC (x1012/L) | 4.81 ± 0.7 | 4.86 ± 0.67 | 4.92 ± 0.68 |
MCH (pg) | 27.8 ± 3.1 | 27.5 ± 3.1 | 27.1 ± 2.9 |
MCHC (g/dl) | 32.7 ± 1.5 | 32.5 ± 1.5 | 32.5 ± 1.5 |
MCV (fl) | 84.8 ± 7.8 | 84.3 ± 7.7 | 83.0 ± 7.3 |
HCT (%) | 40.6 ± 5.9 | 40.8 ± 5.5 | 40.8 ± 5.5 |
Hb (g/dl) | 13.3 ± 2.1 | 13.3 ± 1.9 | 13.2 ± 2.0 |
(B) Glycemic levels categorized under Fasting Blood Glucose (FBG) grouping |
Parameters | Normoglycemia | Borderline hyperglycemia | Hyperglycemia |
FBG < 100 mg/dL | 100 mg/dL < FBG < 125 mg/dL | FBG > 126mg/dL |
(n = 11,061) | (n = 10,644) | (n = 13,657) |
Age (years) | Female | 4,619 (41.76%) | 49 ± 14 | 4,208 (39.53%) | 55 ± 12 | 5,670 (41.51%) | 54 ± 12 |
Male | 6,438 (58.20%) | 50 ± 14 | 6,434 (60.45%) | 53 ± 13 | 7,986 (58.47%) | 52 ± 13 |
FBG (mg/dl) | 88.9 ± 8.4 | 111.9 ± 7.1 | 178.8 ± 57.6 |
CRP (mg/dL) | 0.94 ± 1.85 | 1.63 ± 4.34 | 1.7 ± 3.56 |
WBC (x109/L) | 8.12 ± 4.82 | 8.18 ± 3.24 | 8.59 ± 2.43 |
Platelets (x109/L) | 263 ± 75 | 263 ± 76 | 266 ± 78 |
NLR | 1.93 ± 1.04 | 1.96 ± 1.23 | 2.05 ± 1.59 |
PLR | 8.67 ± 4.23 | 8.7 ± 4.4 | 9.06 ± 5.4 |
RBC (x1012/L) | 4.94 ± 0.67 | 4.97 ± 0.63 | 5.03 ± 0.665. |
MCH (pg) | 27.7 ± 2.9 | 27.6 ± 3.0 | 27.3 ± 3.0 |
MCHC (g/dl) | 32.6 ± 1.5 | 32.5 ± 1.5 | 32.5 ± 1.6 |
MCV (fl) | 84.6 ± 7.3 | 84.5 ± 7.5 | 83.7 ± 7.3 |
HCT (%) | 41.6 ± 5.4 | 42.0 ± 5.2 | 42.0 ± 5.3 |
Hb (g/dl) | 13.6 ± 1.9 | 13.6 ± 1.9 | 13.6 ± 1.9 |
HbA1c: Glycated hemoglobin type A1c, FBG: Fasting blood glucose, WBC: White blood cells, NLR: Neutrophils to lymphocyte ratio, PLR: Platelets to Lymphocyte ratio |
Together, our results show that with an increasing glycemia, CRP, WBC, Platelet, NLR, PLR, RBC, and HCT show an increasing trend whereas RBC indices tend to decrease.
CRP, WBC, NLR, and PLR are significant discriminants for differentiating glycemic control
To evaluate the inflammatory response present in different glycemic conditions, we performed a multivariate analysis of variance, MANOVA, separately for two sub-groups, HbA1c (n = 2877) and FBG (n = 616). Our results using Pillai’s trace show that glycemic states (normal, borderline, and hyperglycemia) vary significantly with the five inflammatory markers (CRP, WBC, Platelet, NLR, PLR); group effect exhibiting F-ratioHbA1c of 182.27 (p < 0.0001) and F-ratio FBG of 37.249 (p < 0.0001) indicated by Pillai's trace.
To elucidate the effect of glycemia on inflammatory markers, a univariate one-way ANOVA test was employed. The results revealed that there was a statistically significant difference (p value < 0.0001) in the mean expression at least between 2 glycemic states for HbA1c (F = 1747.63), CRP (F = 19.226), WBC (F = 19.978), NLR (F = 20.952) and PLR (F = 13.93). Platelets were found to be statistically insignificant. The results are summarized in (Table 2A) for the HbA1c cohort. For the case of FBG cohort (F = 328.067), WBC (F = 13.359) and PLR (F = 8.607) were found to be significantly different (p-value < 0.001), along with CRP (F = 4.035, p = 0.018), platelets (F = 3.382, p = 0.035) and NLR (F = 8.607, p = 0.006) (Table 2B)
Post-hoc LSD was employed for multiple pairwise comparisons while controlling the error rate at an alpha level of 0.05. LSD results for the HbA1c cohort revealed that HbA1c was significantly increased from N to BH ( 0.752 (95% CI 0.878 to 0.626) %, p < 0.0001), N to H (2.87 (95% CI, 2.97 to 2.77) %, p < 0.0001), and BH to H (2.116 (95% CI, 2.225 to 2.008) %, p < 0 .0001). NLR was statistically significantly increased from N to BH (0.3443 (95% CI 0.6207 to 0.0678), p = 0.015), N to H (0.7294 (95% CI, 0.9556 to 0.5032), p < 0.0001), and BH to H (0.3852 (95% CI, 0.6233 to 0.1470), p = 0.002). PLR was also observed to be significantly increased from N to BH (1.1573 (95% CI 2.0628 to 0.2517), p = 0.012), N to H (1.9833 (95% CI, 2.7242 to 1.2425), p < 0.0001), and BH to H (0.8261 (95% CI, 1.6063 to 0.0459), p = 0.038). Notably, CRP was only significantly increased from N to H (1.1222 (95% CI, 1.4925 to 0.7519) mg/dl, p < 0.0001), and BH to H (0.6825 (95% CI, 1.0724 to 0.2925) mg/dl, p = 0.001). WBC, on the other hand, was statistically significantly increased from N to BH (0.4797 (95% CI, 0.8407 to 0.1187) x109/L, p = 0.009), N to H (0.9372 (95% CI, 1.2326 to 0.6419), p < 0.0001) x109/L, and BH to H (0.4575 (95% CI, 0.7686 to 0.1465) x109/L, p = 0.004). Platelets were not found to significantly vary between any group. Results from the HbA1c cohort are provided in Fig. 1(A-F) and Supplementary Table S1.
The FBG cohort exhibited a significant increase in FBG from N to H (93.766 (95% CI, 101.126 to 86.407)mg/dl, p < 0.0001), and BH to H ( 70.713 (95% CI, 78.780 to 62.646) mg/dl, p < 0.0001), NLR was only statistically significantly increased from N to H ( 0.8150 (95% CI 1.3119 to 0.3181), PLR was only statistically significantly increased from N to H (2.8070 (95% CI, 4.1480 to 1.4661), p < 0.0001), and BH to H (1.8954 (95% CI, 3.3652 to 0.4255), p = 0.012). CRP was only statistically significantly increased from N to BH (0.8145 (95% CI 1.4699 to 0.1592) mg/dl, p = 0.015) mg/dl and N to H (0.7487 (95% CI, 1.3707 to 0.1266) mg/dl, p = 0.018). While, WBC was only statistically significantly increased from N to H (1.176 (95% CI, 1.6326 to 0.7193) x109/L, p < 0.0001), and BH to H (2.116 (95% CI, 1.3785 to -0.3774) x109/L, p = 0.001). Platelets were only significantly increased from N to H (17.50 (95% CI, 32.11 to 2.90) x109/L, p = 0.019), and BH to H (17.37 (95% CI, 33.38 to 1.37) x109/L, p = 0.033). Results are shown in Fig. 1(G-L) and (Supplementary Table S2).
RBC superimposed on inflammation status as an augmented discriminator of glycemic control
To investigate the effect of glycemic status on RBCs status and inflammation, MANOVA was performed for two sub-cohorts i.e., HbA1c (n = 28,577) and FBG (n = 8,376) for 11 dependent variables including inflammatory markers (WBCs, NLR, PLR, Platelets) as well as molecular markers (RBC count, Hb (hemoglobin), HCT, and RBC indices including MCV, MCH, and MCHC). Group effect estimated from multivariate test exhibits F-ratioHbA1c of 960.097 (p < 0.0001) and F-ratio FBG of 286.112 (p < 0.0001) as indicated by Pillai's trace. To further analyze the effect of glycemia on molecular and inflammatory markers, a univariate one-way ANOVA test was employed for the HbA1c cohort. ANOVA revealed that there was a statistically significant difference (p value < 0.0001) in mean expression between at least 2 states of glycemia for HbA1c (F = 16440.736), WBC (F = 87.401), RBC (F = 81.710), MCH (F = 145.581), MCHC (F = 61.533), MCV (F = 137.593). NLR (F = 7.164), PLR (F = 4.575), and Platelets (F = 3.069) were also statistically significant but with a p-value < 0.05. HCT and Hb were found to be insignificant (Table 2C). For the FBG cohort, FBG (F = 4976.068), WBC (F = 12.801), NLR (F = 15.650), PLR (F = 11.844), RBC (F = 13.102) were significantly different (p-value < 0.0001), while MCH (F = 4.785), MCV (F = 6.463), HCT (F = 5.609) and Hb (F = 4.235) at p-value < 0.05, however, platelets and MCHC showed insignificant results (Table 2D).
Post-hoc LSD for significant ANOVAs for multiple pairwise comparisons for the HbA1c cohort revealed that HbA1c was significantly increased from N to BH (0.776 (95% CI, 0.818 to 0.733) %, p < 0.0001), N to H ( 2.978 (95% CI, 3.013 to 2.943) %, p < 0.0001), and BH to H (2.203 (95% CI, 2.239 to 2.166) %, p < 0.0001). NLR statistically significantly decreased from N to BH (-0.0971 (95% CI, -0.0362 to -0.1579), p = 0.002), and statistically significantly increased from BH to H (0.0957 (95% CI, 0.1477 to 0.0437), p < 0.0001). PLR only statistically significantly decreased from N to BH (-0.2769 (95% CI, -0.0254 to -0.5284), p = 0.031), and significantly increased from BH to H (0.3288 (95% CI, 0.5437 to 0.1139) %, p = 0.003). WBC was only statistically significantly increased from N to H (0.5840 (95% CI, 0.6818 to 0.4863) x109/L, p < .0001), and BH to H (0.4796 (95% CI, 0.5811 to 0.3780) x109/L, p < 0.0001). RBC was statistically significantly increased from N to BH ( 0.0466 (95% CI, 0.0700 to 0.0231) x1012/L, p < 0.0001), N to H (0.120 (95% CI, 0.1393 to 0.1007) x1012/L, p < 0.0001), and BH to H (0.0735 (95% CI, 0.0935 to 0.0534) x1012/L, p < 0.0001). Platelets were only significantly increased from BH to H (2.83 (95% CI, 5.42 to 0.24) x109/L, p = 0.032). MCH was statistically significantly decreased from N to BH (-0.329 (95% CI, -5.42 to -0.24) pg, p < .0001), N to H (-0.724 (95% CI, -0.618 to -0.829) pg, p < 0.0001), and BH to H (-0.395 (95% CI, -0.286 to -0.504) pg, p < 0.0001). MCHC was statistically significantly increased from N to BH (0.202 (95% CI, 0.153 to 0.250) g/dl, p < 0.0001), N to H (0.222 (95% CI, 0.182 to 0.262) g/dl, p < 0.0001). MCV was only statistically significantly decreased from N to BH (-0.202 (95% CI, -0.153 to -0.250) fl, p < 0.0001) and N to H (-0.222 (95% CI, -0.182 to -0.262) fl, p < 0.0001). Results for the HbA1c cohort are given in Fig. 2 (A-F) and Supplementary Table S3. Multiple pairwise test results for the FBG cohort showed FBG to be increased significantly from N to BH ( 23.046 (95% CI, 24.944 to 21.147) mg/dl p < 0.0001), N to H (89.322 (95% CI, 91.131 to 87.512) mg/dl, p < 0.0001), and BH to H (66.276 (95% CI, 68.191 to 64.362) mg/dl, p < 0.0001). NLR was only statistically significantly increased from N to H (0.1981 (95% CI, 0.2697 to 0.1265), p < 0.0001), and BH to H (0.1476 (95% CI, 0.2233 to 0.0718), p < 0.0001). PLR was also only statistically significantly increased from N to H (0.5924 (95% CI, 0.8423 to 0.3425), p < 0.0001), and BH to H (0.4729 (95% CI, 0.7373 to 0.2085 ), p < 0.0001). WBC was also only statistically significantly increased from N to H (0.4693 (95% CI, 0.6679 to 0.2706) x109/L, p < 0.00001), and BH to H (0.4309 (95% CI, 0.6411 to 0.2207 ) x109/L, p < 0.0001). RBC was statistically significantly increased from N to BH (0.0435 (95% CI, 0.0781 to 0.0090) x1012/L, p = 0.013), N to H (0.0859 (95% CI, 0.1188 to 0.0530) x1012/L, p < 0.0001), and BH to H ( 0.0424 (95% CI, 0.0772 to 0.0076) x1012/L, p = 0.017). MCH was significantly decreased only from N to H (-0.211 (95% CI, -0.063 to -0.360) pg, p = 0.005), and BH to H (-0.202 (95% CI, -0.045 to -0.359) pg, p < 0.012). MCV was also only statistically significantly decreased from N to H (-0.568 (95% CI,- 0.190 to -0.945) fl, p < 0.0001), and BH to H (-0.658 (95% CI, -0.258 to -1.057) fl, p < 0.0001). HCT was only statistically significantly increased from N to BH (0.378 (95% CI, 0.661 to 0.095) %, p = 0.009) and N to H (0.422 (95% CI, 0.692 to 0.153) %, p = 0.002). Hb was also statistically significantly increased from N to BH (0.115 (95% CI, 0.216 to 0.014) g/dl, p = 0.026), N to H (0.133 (95% CI, 0.229 to 0.036) g/dl, p = 0.007). Results are summarized graphically for the FBG cohort in Fig. 2 (G-L) and Supplementary Table S4.
A predictive model of non-specific inflammatory markers for estimation of dysglycemia
Linear discriminant analysis (LDA) was used to develop a glycemic prediction model comprising six clinical parameters out of which five were inflammatory markers (CRP, WBC, Platelet, NLR, PLR), where CRP is an indicator of chronic inflammation and one was glycemic indicator (HbA1c or FBG) to differentiate between different states of glycemia (hyperglycemia, borderline hyperglycemia, and normoglycemia). Results are displayed in Fig. 3A and Fig. 3B for HbA1c and FBG cohorts, respectively. Using the variances from all the values, two discriminant functions were derived, which accounted for 100% of the variance. For both the HbA1c and FBG cohort, the first canonical discriminant function contributed substantially towards the total variance in the dataset with more than 99% variance with a canonical correlation of 0.7, at a significance value p < 0.001 in both cases.
The classification discriminant functions (DF0, DF1, and DF2) were therefore generated based on the estimation of corresponding β values (Table 3A) for the HbA1c cohort and (Table 3B) for the FBG cohort.
P(y = 0|x) HbA1c\(=-36.39+\left(5.996*HbA1c\right)+\left(3.305*NLR\right)-\left(1.159*PLR\right)-\)
$$\left(0.153*CRP\right)+\left(0.519*WBC\right)+\left(0.077*Platelets\right)$$
1
P(y = 1|x) HbA1c =\(-24.944+\left(4.428*HbA1c\right)+\left(3.198*NLR\right)-\left(1.128*PLR\right)-\)
$$\left(0.157*CRP\right)+\left(0.504*WBC\right)+\left(0.076*Platelets\right)$$
2
P(y = 2|x) HbA1c =\(-21.507+\left(3.877*HbA1c\right)+\left(3.173*NLR\right)-\left(1.13*PLR\right)-\)
$$\left(0.161*CRP\right)+\left(0.481*WBC\right)+\left(0.076*Platelets\right)$$
3
P(y = 0|x) FBG\(=-26.006+\left(0.118*FBG\right)+\left(3.624*NLR \right)-\left(1.545*PLR \right)-\)
$$\left(0.149*CRP \right)+\left(1.081*WBC \right)+\left(0.098*Platelets \right)$$
4
P(y = 1|x) FBG =\(-18.083+\left(0.072*FBG\right)+\left(3.769*NLR\right)-\left(1.627*PLR\right)-\)
$$\left(0.06*CRP \right)+\left(0.97*WBC\right)+\left(0.098*Platelets\right)$$
5
P(y = 2|x) FBG =\(-16.423+\left(0.057*FBG\right)+\left(3.744*NLR\right)-\left(1.625*PLR\right)-\)
$$\left(0.116*CRP\right)+\left(0.966*WBC\right)+\left(0.098*Platelets\right)$$
6
Where, y = 0 means belongingness of the hyperglycemia subset, y = 1 belongingness to borderline hyperglycemia, and y = 2, belongingness to normoglycemia.
A joint predictive model of inflammatory markers in combination with Erythrocytes status for estimation of dysglycemia
We applied LDA to develop a glycemic predictive model with enhanced accuracy by integration of the significant DVs, concluded from the outcomes of individual ANOVAs (Table 2C and 2D). Predictors of inflammation (NLR, PLR, and WBC count), predictors of erythrocytes status (RBC count, MCV, and MCH), and glycemic status indicator (HbA1c or FBG) were used to differentiate between different states of glycemia. Results are displayed in Fig. 4A and Fig. 4B for HbA1c and FBG cohorts, respectively. Two discriminant functions were derived, which accounted for 100% of the variance. For the HbA1c cohort (n = 50,116) first canonical discriminant function majorly contributed towards the total variance in the dataset with 99.9% variance and canonical correlation of 0.729, however for the FBG cohort (n = 13,861) first canonical discriminant accounted for 100% variance with the canonical correlation of 0.739, at a significance p-value < 0.001 in both cases. The classification discriminant functions (DF0, DF1, and DF2) were therefore generated based on the estimation of corresponding β values (Table 3C) for the HbA1c cohort and (Table 3D) for the FBG cohort.
P(y = 0|x) HbA1c \(=-177.133+\left(5.089*HbA1c\right)+\left(0.479*WBC\right)-\left(0.115*NLR\right)+\left(0.610*PLR\right)+\left(20.083*RBC\right)- \left(3.485*MCH\right)+ \left(3.566*MCV\right)\) (7)
P(y = 1|x) HbA1c = \(-169.510+\left(3.703*HbA1c\right)+\left(0.443*WBC\right)-\left(0.166*NLR\right)+\left(0.624*PLR\right)+\left(20.257*RBC\right)- \left(3.44*MCH\right)+ \left(3.561*MCV\right)\) (8)
P(y = 2|x) HbA1c \(=-166.396+\left(3.228*HbA1c\right)+\left(0.437*WBC\right)-\left(0.142*NLR\right)+\left(0.627*PLR\right)+\left(20.240*RBC\right)- \left(3.316*MCH\right)+ \left(3.518*MCV\right)\) (9)
P(y = 0|x) FBG\(=-193.222+\left(0.135*FBG\right)+\left(0.393*WBC\right)-\left(2.144*NLR\right)+\)
$$\left(1.596*PLR\right)+\left(23.788*RBC\right)- \left(3.172*MCH\right)+ \left(3.746*MCV\right)$$
10
P(y = 1|x) FBG\(=-186.011+\left(0.084*FBG\right)+\left(0.364*WBC\right)-\left(2.174*NLR\right)+\)
$$\left(1.601*PLR\right)+\left(23.851*RBC\right)- \left(3.152*MCH\right)+ \left(3.741*MCV\right)$$
11
P(y = 2|x) FBG\(=-183.696+\left(0.067*FBG\right)+\left(0.360*WBC\right)-\left(2.207*NLR\right)+\)
$$\left(1.609*PLR\right)+\left(23.817*RBC\right)- \left(3.104*MCH\right)+ \left(3.720*MCV\right)$$
12
Where, y = 0 means belongingness of the hyperglycemia subset, y = 1 belongingness to borderline hyperglycemia, and y = 2, belongingness to normoglycemia.
An evaluation of model accuracy to predict dysglycemia from inflammatory markers.
Discriminant classification results showed good separations of the three glycemic states for both cohorts with an accuracy of greater than 80% (Figs. 3C and D). Classification results for the HbA1c cohort (Fig. 3C) showed that the back substitution method can classify hyperglycemia with a correct discrimination proportion of 72%; for the borderline hyperglycemia subset, 95.2% and for normoglycemia cases with a correct discrimination proportion of 89.6%. Moreover, classification results for the FBG cohort (Fig. 3D) showed correct group membership of about 67% for hyperglycemia, 84.6% for borderline hyperglycemia, and for normoglycemia cases, it was 93.4%. For both cohorts, borderline hyperglycemia had the highest correct discrimination proportion results. To further evaluate the stability of the model discriminant functions, Jackknife cross-validation was employed, which showed almost similar classification accuracy for both the HbA1c (80.8%) and FBG (80.8%) LDA models. ROC (receiver-operating-characteristics curve) analysis of the model computed AUC and 95% CI values for each glycemic type for both cohorts. The model exhibited a strong diagnostic value for glycemic state with all ROCs showing AUC above 0.9 (p < 0.0001, Fig. 5).
An evaluation of joint model accuracy in the prediction of dysglycemia
Upon integration of significant candidates from RBC parameters with the inflammatory markers in the HbA1c model, the overall accuracy increased from 81.2–89.5% (Fig. 4C), thus providing evidence of a strong discriminatory value of parameter superimposition. Notably, no difference was observed for the FBG cohort (82.3% vs 82.8%) (Fig. 4D). Interestingly, the joint HbA1c cohort model showed 25% improvement in accuracy for hyperglycemia (72% vs 97.1%). The FBG cohort showed an increment of 10% in the predictive ability for borderline hyperglycemia (84% vs 94%). Jackknife cross-validation results were comparable with the classification accuracy for both HbA1c (89.5%) and FBG (82.6%) LDA models. Furthermore, ROC assessment established the diagnostic specificity and sensitivity of the joint model (p < 0.0001) with all ROC AUCs above 0.9 except for the HbA1c cohort in predicting borderline hyperglycemia (AUC = 0.87). Results are illustrated in Fig. 6.
Low cost and high accuracy risk fingerprinting of dysglycemia
The prediction accuracy for the inflammatory (“M1”: equations 1–3, “M2”: equations 4–6) and joint (“M3”: equations 7–9, “M4”: equations 10–12) model for both HbA1c and FBG are provided in Table 4. M1 provided the highest discrimination proportion of 95.2% for predicting borderline hyperglycemia but had the highest cost (~$20). M4 had 94% prediction accuracy for borderline hyperglycemia with the lowest price at just ~$5. Importantly, M3 had the highest discriminatory power for the correct classification of hyperglycemia (97.1%) and normoglycemia (94.2%) cases. In the case of the FBG cohort, the overall predictive capacity for both models M2 vs M4 (82.3% vs 82.8%) was comparable but at 2.5 times the price difference ($12 vs. $5). Overall, amongst all the reported models, M3 provided the highest cumulative predictive accuracy of 89.5% for hyperglycemia, borderline hyperglycemia, and normal cases. In conclusion, M3 and M4 models could be utilized for population-level screening programs and by clinicians for predicting hyperglycemia and borderline hyperglycemia at a lower cost.