Relationship between caffeine intake and autosomal dominant polycystic kidney disease progression: a retrospective analysis using the CRISP cohort

Data were taken from the Consortium for Radiologic Imaging Studies of Polycystic Kidney Disease (CRISP) study, which comprised of four clinical centers, University of Alabama, Emory University, University of Kansas and the Mayo Clinic. Patients were eligible to be enrolled in the CRISP study if they were diagnosed with ADPKD, had a creatinine clearance of at least 70 mL/min, serum creatinine level of either ≤1.6 mg/deciliter for men or ≤ 1.4 mg/deciliter for women, and were between 15 and 46 years of age at baseline. Patients were excluded from the study if they had any comorbidities that would affect kidney function besides hypertension. At each visit, TKV was determined from coronal T1- and T2-weighted MRI using a stereologic method, [12‐14] and corrected for height (htTKV, ml/m). GFR was measured by iothalamate clearance and indexed to body surface area (ml/min/1.73 m²). Lifestyle exposures, including caffeine consumption and smoking status, were collected via written surveys in the form of closed questions (Additional file 1). Patients were screened for mutations in the PKD1 and PKD2 genes. Further details about the CRISP study have been previously published [13, 15].

Two outcomes were used to assess disease progression. The primary outcome was height-adjusted total kidney volume (htTKV) [16]. Because kidney volume increases exponentially over time, htTKV was natural log transformed. The second outcome was measured glomerular filtration rate (mGFR) calculated by iothalamate clearance [17]. We chose to have the primary outcome be htTKV because htTKV has been shown to be a good proxy for disease state [15] and the secondary outcome to be mGFR because mGFR is used to estimate kidney filtering capabilities and thus kidney function.

Caffeine consumption at baseline was assessed as the number of cups of coffee/tea, and the number of glasses of other caffeinated beverages consumed per day, averaged over the prior month. For the primary analysis, the exposure variable of caffeine consumption was dichotomized (any vs. none). For further investigation of this relationship, we treated caffeine dose both as a continuous variable, and binned into daily quartiles of caffeine intake of 0 mg (n = 51), > 0-86 mg (n = 50), > 86-181 mg (n = 47), > 181-301 mg (n = 45) and > 301 mg (n = 46). To obtain a quantitative estimate of caffeine dose, each source was converted into milligrams via the following conversions: 95 mg of caffeine = 1 cup of coffee/tea (8 oz.) and 43 mg of caffeine = 1 glass of other caffeinated beverage (12 oz.) [18, 19]. When converting caffeine consumption into milligrams, if the patient was missing either (but not both) number of cups of coffee or number of other caffeinated beverages per day, it was assumed they drank none (n = 9). Patients missing both values at baseline (n = 16) were imputed to have no caffeine intake.

Additional baseline variables that were included as covariates in the multivariable adjusted models were age, race, sex, BMI, smoking, hypertension, and genetic profile. The patient’s race was categorized as either “White” or “Other” due to the small sample sizes of subsets of the other races. Based on the current understanding of the prognostic significance of ADPKD mutations, we classified the mutation data into three subgroups of gene types: truncating PKD1, non-truncating PKD1, and PKD2 + no mutation detected (NMD) [20, 21].

Descriptive analyses for continuous variables were expressed as means (± standard deviations) and for categorical variables were expressed as frequencies and relative frequencies. If a continuous variable did not appear to be normally distributed by its quantile-quantile (QQ) plot and histogram, the median and inter-quartile range (IQR) was reported as well. Two-sampled t-tests and Pearson’s chi square test for continuous and categorical variables, respectively, were used to assess differences in measures between caffeine and non-caffeine consumers among our sample. QQ plots and histogram residuals were used to check for assumption of normality of the two-sampled t-tests. If the assumptions for the t-tests were violated, the Wilcoxon rank sum test was used. Expected cell counts were used to assess assumptions for Pearson’s chi square test. When computing the median follow-up time, the last available time was used. Four patients had 4 follow-up visits during CRISP I without times or dates, so we imputed their follow-up time to be 3 years as these all corresponded to the year three visit.

The main statistical method utilized in this analysis was a linear mixed effect model (LME) with random intercepts. LMEs were utilized because the CRISP study has more than one measurement of all outcome measures for each patient collected over 14 years. Thus we used random effects that allowed each patient to have their own intercept parameter (i.e., the random effect). All other covariates were fixed effects. Single factor association models for each variable were first computed with adjustment only for time. Then, multivariable models, adjusting for age, sex, race, BMI, smoking, hypertension, gene type and time, were computed for each outcome variable (referred to as Model 1 for ln(htTKV) and Model 1 for mGFR). Lastly, models were adjusted for caffeine, time and their interaction for each outcome variable (referred to as Model 2 for ln(htTKV) and Model 2 for mGFR). The coefficient (β) for the effect of the interaction between caffeine and time on the outcome of ln(htTKV) represents the effect of caffeine consumption on the slope of ln(htTKV) over time. The effect size, as measured by the constant, annual difference in percentage growth of htTKV (without log-transformation) in caffeine-consumers compared to non-caffeine consumers, was calculated by taking (e^β)*100% [22]. Additional file 2: Table S1 summarizes the models used in this study. Model assumptions were checked using QQ plots and plots of residuals, with appropriate adjustments for model violations such as log transformations using the natural log function. Multivariable models using daily caffeine dose as continuous and multicategory variables were centered and used in sensitivity analyses, as were models which excluded subjects missing both caffeine exposure variables (cups of coffee/tea and glasses of other caffeinated beverages per day). In some cases identified as missing visits, subjects contributed some of their study measures, just not all. Linear contrasts of Model 2 for ln(htTKV) were performed to quantify the effect of caffeine over time. Kaplan Meier plots with right-censored data were used to examine the effect of caffeine on time until ESRD or death, along with corresponding log-rank tests to compare survival between caffeine groups (any vs. none). Cox proportional hazard regression analysis was performed to examine the effects of age, sex, race, BMI, smoking, hypertension, gene type and caffeine. Statistical significance was accepted if p <  0.05. R (Vienna, Austria) was used for data processing and analysis.

In 2001, 241 patients were enrolled in the CRISP study and were evaluated until 2015. Of these 241 patients, we excluded 2 because of missing genetic information, thus giving us a sample size of 239 patients. Additionally, 16 patients were missing data for both caffeine sources (coffee/tea and other caffeinated beverages) at baseline. We computed models based on the inclusion of these patients, with their caffeine intake assumed to be 0 mg, then performed additional sensitivity analyses in which these patients were excluded. The median follow-up time was 12.5 years (IQR: 8.7,13.0). The minimum number of patient visits was 4 and the maximum number of visits was 8 (25th, 50th and 75th percentiles were 6, 8 and 8, respectively).

At baseline, the average (± standard deviation [sd]) and median (IQR) age of our sample was 32.3 (±8.7) and 33.8 (25.1, 39.7) years, respectfully. Sixty percent of our patients were female and 87% were white. 61% of our patients had hypertension at baseline and 17% of our patients reported smoking. The majority of our patients had a truncating mutation in the PKD1 gene (53%) while 25% of the patients had a non-truncating mutation in PKD1 and 22% had either a mutation in PKD2 or no mutations detected (NMD) in either PKD1 or PKD2. The mGFR was relatively preserved, as the average (± sd) and median (IQR) values were 97.7 (±24.8) and 94.7 (78.8, 114.8) mL/min/1.73m², respectively. At baseline, the average (± sd) and median (IQR) of httkv in our sample was 621.6 (±373.9) and 504.4 (350.2, 773.9) ml/m, respectfully.

At baseline, 79% of the patients reported consuming caffeine. The minimum amount of caffeine consumed was 0 mg/day and the maximum was 1425 mg/day (25th, 50th and 75th percentiles = 29.3, 129, and 233 mg/day, which would be equivalent to 0.31, 1.36 and 2.45 eight-ounce cups of coffee per day, respectively). Table 1 shows the baseline characteristics of caffeine consumers compared to non-caffeine consumers. While there were no statistically significant differences between baseline characteristics, 19% of caffeine consumers reported smoking at baseline compared to 8% of non-caffeine consumers (p-value = 0.088). Caffeine consumers had a slightly lower median (IQR) baseline htTKV compared to non-caffeine consumers [479.9 (342.4, 726.9) mL/m compared to 567.5 (382.6, 872.7) mL/m; p-value = 0.22] and slightly higher mGFR [96.4 (78.7, 114.7) mL/min/1.73m² compared to 88.7 (79.1, 114.8) mL/min/1.73m²; p-value = 0.36) but neither of these were statistically significant. At the time of this analysis, 42 patients had reached end-stage renal disease (ESRD) and 3 patients died, while 194 had not reached ESRD or were censored.

Linear mixed models were first fit for each covariate separately to determine each single factor association over time (eg. baseline age and time as fixed effects in the first model, baseline hypertension and time as fixed effects in the second model, etc., see Additional file 2: Table S2). When using ln(htTKV) as the outcome variable, age, BMI, hypertension, race and gene type were statistically significant, as has been reported previously [23, 24]. The caffeine-by-time interaction term also showed a small, but statistically significant (p-value = 0.007), positive association with ln(htTKV), whereas the main effect of caffeine in this model was not statistically significant (p-value = 0.205). In models with mGFR over time as the outcome, age, BMI and hypertension were the only statistically significant measures in our single factor analyses.

Next we fit multivariable models for ln(htTKV) and mGFR adjusted for age, sex, race, BMI, smoking, hypertension, gene type and time (Model 1). As with the single factor associations over time, age and hypertension were found to be statistically significant in Model 1 for both ln(htTKV) and mGFR (Additional file 2: Table S3). Gene type was also statistically significant for Model 1 for both outcomes (Additional file 2: Table S3). BMI was not statistically significant in Model 1.

Finally, we added caffeine consumption (any vs. none) to the multivariable models (Model 2) for both outcomes. These results are presented in Table 2. For both outcomes, the variables that were significant in Model 1 remained significant after the inclusion of caffeine consumption. The effect of caffeine on ln(htTKV) varied over time as indicated by tests of interactions between caffeine and time (p-value = 0.007) but not for mGFR (p-value = 0.811). The expected difference in ln(htTKV) between caffeine consumers and non-caffeine consumers at baseline was − 0.146 (95% CI: -0.295, 0.003; p-value = 0.061). This corresponds to a 13.6% (95% CI: 0.3, 25.5%) lower baseline htTKV associated with caffeine consumption. The expected difference in the rate of change of ln(htTKV) over time between caffeine consumers and non-caffeine consumers was 0.006 (95% CI: 0.002, 0.011; p = 0.007). This corresponds to a 0.6% (95% CI: 0.2, 1.1%) greater rate of kidney growth each year associated with caffeine, so the annual rate of kidney growth for caffeine consumers was 5.3% compared to 4.6% for non-caffeine consumers. The difference in baseline mGFR associated with caffeine consumption was 1.40 mL/min/1.73m² (95% CI: -5.91, 8.71; p = 0.713) and the difference in the rate of change of mGFR was 0.069 mL/min/1.73m² (95% CI: -0.495, 0.631; p = 0.811).

Among caffeine consumers, 15.4% reached ESRD and 1.1% died during the study, while in non-caffeine consumers, 25.5% reached ESRD and 1.2% died. The Kaplan-Meier curves for survival free of ESRD or death (Fig. 1) were not statistically significantly different (log rank test p = 0.10). In a multivariable proportional hazards model adjusted for the effects of age, sex, race, BMI, smoking, hypertension and gene type on time until ESRD or death (Additional file 2: Table S11), we found that age, smoking, hypertension and genetic status were statistically significant. When caffeine consumption (any vs. none) was added to this model, age, smoking, hypertension and gene type remained statistically significant (Table 3). There were no substantive changes in hazard for any of these variables. However, caffeine consumption was not found to be a statistically significant risk factor for the time to ESRD or death (Hazard Ratio (HR) = 0.556; 95% CI: 0.279, 1.110; p = 0.096).

We performed sensitivity analyses using different measures of amount of caffeine consumed examined as continuous and multicategory variables. We also generated these results excluding patients who were missing baseline data on both sources of caffeine (cups of coffee/tea and glasses of other caffeinated beverages). Both unadjusted and adjusted models were estimated. See Additional file 2: Tables S1-S10. When modeling ln(htTKV) as the outcome, caffeine:time interactions were statistically significant in most models but with estimates near zero. When modeling mGFR as the outcome, caffeine was not statistically significant. Therefore, the sensitivity analyses were not qualitatively different than our primary results and showed that caffeine does not have a strong and consistent effect on disease progression.