Age at diagnosis, glycemic trajectories, and responses to oral glucose-lowering drugs in type 2 diabetes in Hong Kong: A population-based observational study

Background Lifetime glycemic exposure and its relationship with age at diagnosis in type 2 diabetes (T2D) are unknown. Pharmacologic glycemic management strategies for young-onset T2D (age at diagnosis <40 years) are poorly defined. We studied how age at diagnosis affects glycemic exposure, glycemic deterioration, and responses to oral glucose-lowering drugs (OGLDs). Methods and findings In a population-based cohort (n = 328,199; 47.2% women; mean age 34.6 and 59.3 years, respectively, for young-onset and usual-onset [age at diagnosis ≥40 years] T2D; 2002–2016), we used linear mixed-effects models to estimate the association between age at diagnosis and A1C slope (glycemic deterioration) and tested for an interaction between age at diagnosis and responses to various combinations of OGLDs during the first decade after diagnosis. In a register-based cohort (n = 21,016; 47.1% women; mean age 43.8 and 58.9 years, respectively, for young- and usual-onset T2D; 2000–2015), we estimated the glycemic exposure from diagnosis until age 75 years. People with young-onset T2D had a higher mean A1C (8.0% [standard deviation 0.15%]) versus usual-onset T2D (7.6% [0.03%]) throughout the life span (p < 0.001). The cumulative glycemic exposure was >3 times higher for young-onset versus usual-onset T2D (41.0 [95% confidence interval 39.1–42.8] versus 12.1 [11.8–12.3] A1C-years [1 A1C-year = 1 year with 8% average A1C]). Younger age at diagnosis was associated with faster glycemic deterioration (A1C slope over time +0.08% [0.078–0.084%] per year for age at diagnosis 20 years versus +0.02% [0.016–0.018%] per year for age at diagnosis 50 years; p-value for interaction <0.001). Age at diagnosis ≥60 years was associated with glycemic improvement (−0.004% [−0.005 to −0.004%] and −0.02% [−0.027 to −0.0244%] per year for ages 60 and 70 years at diagnosis, respectively; p-value for interaction <0.001). Responses to OGLDs differed by age at diagnosis (p-value for interaction <0.001). Those with young-onset T2D had smaller A1C decrements for metformin-based combinations versus usual-onset T2D (metformin alone: young-onset −0.15% [−0.105 to −0.080%], usual-onset −0.17% [−0.179 to −0.169%]; metformin, sulfonylurea, and dipeptidyl peptidase-4 inhibitor: young-onset −0.44% [−0.476 to −0.405%], usual-onset −0.48% [−0.498 to −0.459%]; metformin and α-glucosidase inhibitor: young-onset −0.40% [−0.660 to −0.144%], usual-onset −0.25% [−0.420 to −0.077%]) but greater responses to other combinations containing sulfonylureas (sulfonylurea alone: young-onset −0.08% [−0.099 to −0.065%], usual-onset +0.06% [+0.059 to +0.072%]; sulfonylurea and α-glucosidase inhibitor: young-onset −0.10% [−0.266 to 0.064%], usual-onset: 0.25% [+0.196% to +0.312%]). Limitations include possible residual confounding and unknown generalizability outside Hong Kong. Conclusions In this study, we observed excess glycemic exposure and rapid glycemic deterioration in young-onset T2D, indicating that improved treatment strategies are needed in this setting. The differential responses to OGLDs between young- and usual-onset T2D suggest that better disease classification could guide personalized therapy.


Fig B.
Schematic diagram illustrating the procedure to calculate glycemic exposure, defined as the area under the curve (AUC) of the mean yearly A1C over time, for a hypothetical person with diabetes. The AUC includes only the values in excess of 7%; values <7% are excluded from the calculation. We applied this concept to the Register cohort by plotting the mean A1C in each age at diagnosis group for each year (by attained age) from the mean observed age at diagnosis until age 75 years. The observation period was defined as a constant, calculated as the duration between the mean observed age at diagnosis to age 75 years. Causal diagram depicting the relationship between the age at diagnosis (primary exposure), oral glucose-lowering drugs (secondary exposure), and hemoglobin A1c (A1C; outcome). The causal pathway of interest is the biological relationship between age at diagnosis and A1C. Sex is a confounding variable. Comorbidities (Appendix Table 2) are mediating variables, unrelated to the causal pathway of interest. The hypothesized interaction between age at diagnosis and drug effects is indicated with a dotted line.

Underlying A1C
We modeled the value of A1C (Y) for each observation i nested within each person j as shown in the basic equation below, where β is the vector of each regression coefficient, and ε is the vector of errors.
β0j is the intercept for person j (equivalent to A1C when time=0), and β1j is the rate of change of the A1C for each unit of time.
For example, consider a hypothetical person j, whose A1C is measured at baseline and annually thereafter. The table below shows the A1C values. In the plot below, each dot represents an A1C measurement, and the blue line connects all dots to display the trend in A1C over time. The blue line's intercept β0j = 6.5 and the slope β1j = 0.5:

Time-Varying Drug Prescriptions
Next, suppose person j starts taking drug A at time = 4, as shown in the table below. We define the time-varying binary variable drugA as equal to 1 when Drug A is active and 0 when Drug A is inactive. In the plot below, the blue dots are A1C measurements taken while Drug A was inactive, and the red dots are those taken while Drug A was active. The blue line connects the blue dots, representing the underlying A1C trend (no drug) and the red line represents the A1C trend on Drug A. As displayed by the downward shift in the red line from the blue line, Drug A drops the A1C by 1% from its expected value without treatment (β2j = −1): 8.0 0 4 7.5 1 5 8.0 1 6 8.5 1 7 9.0 1 8 9.5 1

Medication Adherence
The A1C still falls by 1% whenever the drug (prescription) is active, as shown in the plots below:

Switching Drugs
Suppose that Drug A is switched to Drug B. We can account for this by defining an additional time-varying binary covariate drugB, which indicates when Drug B is active. Based on the table and plot below, Drug B (green) drops the A1C by 1.5% from its expected value without treatment (β3j = −1.5):

Drug Combinations
Suppose that Drug B is added to Drug A. Clinically, the A1C-lowering effect of a two-drug combination is less than the sum of the A1C-lowering effects of each drug taken independently. For example, drug A may drop the A1C by 1%, and drug B by 1.5%, but the combination of drugs A and B may only drop A1C by 2%. We account for this by defining each drug combination as a unique covariate. In the example below, person j starts on Drug A at year 2, switches to Drug B at year 4, then switches to combination A + B (variable drugAB, black) at year 6. The combination drops the A1C by 2% (β4j = −2), while drugs A and B drop A1C by 1% and 1.5% respectively: In terms of single drugs, we defined drug terms for metformin and sulfonylureas as these were the most common first-line agents for monotherapy. We excluded other drugs for monotherapy as they were infrequently used (<0.2% of A1C observations). We additionally included the 10 most common metformin-or sulfonylurea-based combinations (2 or 3 drugs) shown in Table A.     In this sensitivity analysis, we included an interaction term for each drug combination with the baseline A1C for that combination. Results are shown for a baseline A1C level of 7.5%. This model excluded interaction terms with each drug combination and age at diagnosis due to computational limitations. The results from the original model (without baseline A1C) for people with usual-onset type 2 diabetes (age at diagnosis ≥40 years) are shown for comparison.