Baseline characteristics
Among 10,999 study participants, 9,221 (83.8%) were followed up during 2011–2016. Of these participants, 1,058 patients were diagnosed with diabetes. We excluded 277 patients with one or more glycaemic measures missing at baseline. Additionally, the highest and lowest 0.5% for ANGPTL8 was trimmed, leaving 781 patients for analysis. After the propensity matching, we acquired 769 matched pairs of patients from the diabetes group and control group (Supplementary Figure S1). We found that serum ANGPTL8 levels were elevated in patients with diabetes compared to subjects in the control group (618.82 318.08 vs. 581.20 299.54, p = 0.03).
Baseline characteristics of participants with and without diabetes are presented in Table 1 according to quartiles of ANGPTL8. Age, sex, aspartate aminotransferase (AST), creatinine, eGFR and smoking status changed with ANGPTL8 levels in both the diabetes and control groups (all values < 0.05). Furthermore, ANGPTL8 was also associated with waist-hip ratio (WHR), systolic blood pressure (SBP), HbA1c and history of hypertension in the control group (all p values < 0.05). ANGPTL8 was associated with BMI and smoking status in the diabetes group (all p values < 0.05).
Next, we studied correlations between ANGPTL8 levels and related variables in the control and diabetes groups by using Pearson correlation analysis. After controlling for multiple variables, ANGPTL8 levels positively correlated with age (r = 0.42), BMI (r = 0.10), and creatinine (r = 0.11) but inversely correlated with eGFR (r = -0.12) in the control group (all p values < 0.05) (Supplementary Table S1, Model 3). Moreover, ANGPTL8 levels positively correlated with age (r = 0.18), duration of diabetes (r = 0.08), 2h PG (r = 0.08), alanine transaminase (ALT) (r = 0.07), AST (r = 0.13) and creatinine (r = 0.10) but inversely correlated with BMI (r = -0.07), high-density lipoprotein (HDL) (r = -0.09) and eGFR (r = -0.13) in diabetic patients (all values < 0.05) (Supplementary Table S1, Model 3). However, the positive correlation of ANGPTL8 with TG (p = 0.001, Model 2) was diminished after adjusting for other lipid profiles (p = 0.41, Model 3).
ANGPTL8 correlates with all-cause mortality
During up to 5 years of follow-up, there were 19 participants (2.5%) who died and 44 (5.7%) incident cases for the secondary outcome among 769 participants in the control group (Table 2). The incidence of death (N = 56, 7.3%) and the secondary outcome (N = 91, 11.8%) was increased in participants with diabetes. Increasing quartiles of ANGPTL8 were associated with an elevated incidence of death and renal dysfunction in the diabetes group (all p values < 0.05, Table 2) but not in the control group. Furthermore, the CVD mortality in diabetic patients also increased numerically in the highest quartile of ANGPTL8 levels (p = 0.06, Table 2). Binary logistic regression analyses showed that compared with the first quartile, non-adjusted RRs (96% CIs) (Model 1) for the primary outcome (all-cause mortality) were 4.67 (1.00-21.92) and 4.64 (1.86-11.59) for the fourth ANGPTL8 quartile in the control group and diabetes group, respectively (Table 3). The non-adjusted RR (96% CIs) (Model 1) for the secondary outcome was 2.57 (1.04-6.34) for the fourth ANGPTL8 quartile compared with the first quartile (Table 3) in the control group. However, the associations persisted in the diabetes group, although they were slightly attenuated after additional adjustment for covariables, including age, sex and BMI (Model 2) and further adjustment for lipids (Model 3). Then, we further analysed the association of ANGPTL8 with any single component of the secondary outcome and found that elevated ANGPTL8 was associated with an increased risk for renal dysfunction in the diabetes group (RR in quartile 4 vs quartile 1, 10.50; 95% CI 1.32–83.60; Table S2) after adjusting for covariables.
Multivariable-adjusted restricted cubic spline analyses suggested linear associations of ANGPTL8 with all-cause mortality in all participants ( for nonlinear trend = 0.16, p for linear trend = 0.01; Figure 1A). Further analysis indicated a significant linear relationship between ANGPTL8 and all-cause mortality in diabetic patients (p for nonlinear trend = 0.99, p for linear trend = 0.01; Figure 1B) but not in controls (p for nonlinear trend = 0.26, p for linear trend = 0.80; Figure 1C) after adjusting for age, sex, BMI and lipid profiles.
Predictive values of ANGPTL8 for death
We observed that the model combined with ANGPTL8 was more highly predictive for death than the QFrailty score alone but that the AUC for ANGPTL8 + QMortality increased numerically only in comparison with that of QMortality alone (Figure 2A-2F, Supplementary Table S3; comparison P values were calculated using Medcalc ROC analysis). The AUC for the ANGPTL8 + QFrailty model was 0.71 versus 0.59 for the QFrailty score alone (p < 0.001; Figure 2B) in all participants. The AUC for the ANGPTL8 + QMortality model was 0.80 versus 0.79 for QMortality score alone (p = 0.78; Figure 2A) in all participants. Consistently, the addition of ANGPTL8 improved the prediction accuracy of QFrailty alone (AUC 0.70 vs. 0.59; p < 0.001) and numerically increased the prediction accuracy compared with QMortality alone in diabetic patients (AUC 0.77 vs. 0.79; p = 0.51) (Figure 2C-2D). Furthermore, the addition of ANGPTL8 was more highly predictive of death than the QFrailty score alone (AUC 0.80 vs. 0.84; p = 0.01) and increased prediction accuracy better numerically than QMortality alone (AUC 0.71 vs. 0.57; p = 0.21) in the control group (Figure 2E-2F).