Application of survival tree analysis for exploration of potential interactions between predictors of incident chronic kidney disease: a 15-year follow-up study

The TLGS is an ongoing community-based prospective study performed on a representative sample of residents of Tehran, the capital of the Islamic Republic of Iran. This city composed of 22 urban districts. Study samples were selected from the district 13. The rationales for choosing district 13 were: (1) high stability of the population residing in that district compared to other districts of Tehran, and (2) the age distribution of the population of district 13 was representative of the overall population of Tehran. At first phase of study (1999–2002), a multistage stratified cluster random sampling technique was used to select more than 15,000 people aged ≥ 3 year. The participants then were followed in four consecutive phases, i.e. 2002–2005 (phase 2), 2005–2008 (phase 3), 2008–2011 (phase 4) and the final survey in 2011–2014 (phase 5). Another 3551 residents were invited for the second phase and were followed in the next 3 phases. The details of the study have been published elsewhere [20]. The present study was a post hoc analysis of the TLGS. A total of 12,808 adult participants aged ≥ 20 year from the first and second phases were selected as baseline population and the cohort was re-examined at next phases until the phase 5. Among this participants, we excluded subjects with prevalent CKD (n = 2061) at baseline, and those with missing data on creatinine level (n = 414). Finally, after excluding those without any follow-up data after recruitment (n = 2095), 8238 participants (3795 men and 4443 women) (80% of eligible subjects) were followed until the end of study (Fig. 1). This study was approved by the ethical committee of the Research Institute for Endocrine Sciences of Shahid Beheshti University of Medical Sciences, Tehran, Iran, and conducted in accordance with the Declaration of Helsinki. All participants signed informed consent forms.

At baseline, information on demographics, education, smoking status, medical and drug history was collected using interview. Anthropometric measures including body weight and height, hip, wrist and waist circumferences (WC) were measured according to a standard protocol. Body mass index (BMI) was calculated as weight in kilogram divided by the square of height in meter. Blood pressure was measured twice and mean of two measurements was considered as the participant’s blood pressure. Blood samples were drawn after an overnight fast of ≥ 12 h to assess the following parameters: fasting plasma glucose (FPG) (measured by the enzymatic colorimetric glucose oxidase); total cholesterol (TC) and triglycerides (TGs) (were assayed using the enzymatic colorimetric method) and the Serum creatinine (cr) level which was assayed by kinetic colorimetric Jaffe using commercial kits (Pars Azmoon Inc., Tehran, Iran) by a Selectra 2 auto analyzer (Vital Scientific, Spankeren, The Netherlands) [20, 21]. The sensitivity of the assay was 0.2 mg/dl. Reference intervals according to manufacturer’s recommendation were 53–97 mmol/l (0.6–1.1 mg/dl) and 80–115 mmol/l (0.9–1.3 mg/dl) in women and men respectively. Both intra-assay and inter-assay CVs were less than 3.1% in both baseline and follow-up phases. Assay performance was monitored after every 25 tests using lyophilized serum controls in normal and abnormal ranges and all samples were assayed when internal quality control met the standard satisfactory criteria [17, 18]. In the first phase, physical activity level (PAL) was assessed by the lipid research clinic questionnaire (LRC) [22], which was validated for this study [23]. LRC is one of the most common questionnaires to assess PAL in large-scale epidemiological studies [24]. It was replaced by the Modifiable Activity Questionnaire (MAQ) from the 2nd phase for obtaining the quantitative measure of PAL. This questionnaire measures all three forms of activities including leisure time, job, and household activities in the past year [25].

Kidney function was estimated by the estimated glomerular filtration rate (eGFR) that was calculated using the modification of diet in renal disease (MDRD) equation [26]. The definition of incident CKD was an eGFR of < 60 ml/min per 1.73 m² (CKD Stage 3–5) occurring at any time during the follow-up period. Educational status was categorized to 4 levels as illiterate, 1–9 years, 10–12 years and over 12 years of schooling. Marital status was categorized as single, married, and widowed/divorced. Different categories of smoking status were defined according to WHO guidelines [27]. A current smoker was defined as a person who smokes cigarettes or other smoking implements (water-pipes, or pipes) at least once a day or occasionally. Ex-smoker was formerly daily or occasional smoker who currently does not smoke and non-smoker defined as people who never smoked before. A family history of premature cardio-vascular diseases (CVDs) was considered as any experience of fatal or non-fatal myocardial infarction, stroke or sudden cardiac arrest in first-degree relatives, if it occurred before 55 year of age in male relatives (father, brother and son) and before 65 year of age in female relatives (mother, sister and daughter). Family history of diabetes (FHD) was defined as having type 2 diabetes in first-degree relatives. In first phase, physically active participants were identified as those who were participating in a vigorous physical activity at least 3 days per week [22], and in the second phase, it was defined as achieving a minimum of at least 600 MET (metabolic equivalent task)-minutes per week [28].

Descriptive statistics (mean, SD, proportion) were used to describe baseline characteristics. Comparison of baseline characteristics between men vs. women, between participants with and without incident CKD and between followed and non-followed participants was done using student’s t test, Chi square and Mann–Whitney test, as appropriate. Incidence density rate of CKD and respective 95% confidence interval (CI) were calculated for each gender, by dividing the number of events to person-years at risk. Time to event was defined as time of censoring or date of incidence of CKD, whichever occurred first. The event date for the incident cases of CKD was defined as midtime between the date of follow-up visit at which the CKD was diagnosed for the first time, and the most recent follow-up visit prior to the diagnosis. For censored subjects, the time was the interval between the first and the last observation dates. Study participants were censored due to death, loss to follow-up, or the end of the observation period. The Stepwise Cox PH regression model using Akaike information criteria (AIC), as model selection approach, was used to identify significant predictors of incident CKD by calculating multivariable hazard ratios (HRs) with 95% CI (after confirming the proportionality in the model). Due to significant effect modification of gender on FPG, SBP, diastolic blood pressure (DBP), eGFR and WC (all p < 0.05), we stratified our analysis by gender. Moreover, for examination of the gender effect in the multivariable model we fitted the Cox PH model on whole population. In further analysis, conditional survival tree was implemented with the same outcome and covariates as for Cox method [29]. In step 1, among m covariate, survival tree algorithm selects the variable X _i with the highest ability to separate survivors and non-survivors using p-values from permutation distributions. In step 2, in order to split X _i into two disjoint sets (dichotomizing), the algorithm first sorts the values of X _i in increasing order. Typically, the midpoint between each pair of adjacent values is considered as a possible cut-point. Therefore, given v values of X _i , then v-1 possible splits are evaluated. For each possible cut-point for X _i , the value of the log-rank statistic associated with the considered cut-point is computed; then, the cut-point associated with the smallest p-value is selected. This procedure is applied recursively until the tree grew to an optimal number of terminal nodes. A survival tree can naturally group subjects according to their survival time and based on their covariates; hence, it automatically detects complex interactions between covariates without the need to specify them beforehand. In the present study, all baseline variables were entered for constructing survival trees. The minimum criterion for node split was defined as p < 0.05 and the minimum final nodes size were defined as 100 persons. Kaplan–Meier (KM) curves were constructed for each subgroup identified by the survival tree. Also, according the number of subgroups were found by survival tree, we specified one categorical variables with k level (k is the number of terminal nodes or subgroups in survival trees) and estimated the HR of CKD events among the identified subgroups by considering the low risk group (highest survival probability) as the reference category. We used two packages of Party and Survival from the R software (http://​www.​r-project.​org/​) for the analysis. Two-tailed p-values < 0.05 were considered statistically significant.

The study sample consisted of 3795 men, aged 20–84 year (mean, 41.4 ± 14.1) and 4443 women, aged 20–83 year (mean, 37.5 ± 12.1). Table 1 shows the baseline characteristics of the participants. There was a gender difference in all baseline variables except for TC, family history of premature CVD in male relatives, PAL and use of blood glucose lowering drugs. Women were younger and had lower eGFR, WC, wrist circumference, FPG, SBP and DBP, but had higher BMI, hip circumference and heart rate, compared with men (Table 1).

A comparison of baseline characteristics between followed and non-followed subjects is shown in Table 2. Followed subjects had a significantly higher BMI (26.5 vs. 26.1 kg/m²), WC (87.4 vs. 86.3 cm), hip circumference (100.2 vs. 99.6 cm) and TC (5.2 vs. 5.1 mmol/l), but lower FPG (5.2 vs. 5.3 mmol/l) and eGFR (75.5 vs. 76.1) compared to non-followed subjects, respectively.

Table 3 shows baseline characteristics of the study participants with and without incident CKD. In both genders, the participants who developed incident CKD were older, had higher BMI, FPG, TG, TC, SBP, DBP, WC and wrist circumferences but lower heart rate and were more likely to be married compared with participants free of CKD. Moreover, frequency of taking lipid, blood glucose and blood pressure lowering drugs as well as aspirin was higher in incident cases of CKD compared to non CKD participants (p < 0.01 for all of these measures).

Overall, 2308 new cases of CKD (men = 793 and women = 1515) were identified after a median follow-up of 12.4 [interquartile range (IQR), 9.8–13.6] years. Among incident cases of CKD, 2294 subjects developed stage 3 CKD (eGFR 30–59 ml/min/1.73 m²), 12 cases developed stage 4 CKD (eGFR 15–29 ml/min/1.73 m²), and 2 cases developed stage 5 CKD (eGFR < 15 ml/min/1.73 m²). The incidence density rates of CKD were 20.2 (95% CI 18.8–21.7) and 35.2 (33.5–37.1) per 1000 person-years in men and women, respectively. Figure 2 depicts the survival curve for men and women. Survival of women was lower than that of men during study period (Log-Rank χ² 176, p < 0.001). From the 8238 baseline population, 269 (men = 196) cases of mortality were documented during the study period; of these, 119 (men = 93) cases died from CVD.

Among men, age, FPG and DBP were positively associated with the risk of CKD (Table 4). In contrast, WC and TC were negatively associated with the risk of CKD. Also, an inverse association between eGFR and CKD was found in men; every 1-unit increase in eGFR was associated with a 10% reduction in the risk of CKD (HR 0.90; 95% CI 0.89–0.91). The HR for CKD among men with higher education (> 12 years) was 1.56 (1.06–2.28) compared to illiterate group. Being divorced/widower increased the risk of CKD by 204% (1.09–3.84) compared with being married. Interestingly, active men had 21% (1.03–1.42) higher risk of CKD compared to the inactive group. Current smoking was associated with a 24% (1.03–1.48) increase in the risk of CKD. The use of blood glucose lowering drugs at study entry was associated with 59% (1.08–2.35) higher risk of CKD.

The associations of age, FPG and WC with incidence of CKD in women were similar to those observed for men (Table 4). Every 1-unit increase in eGFR was associated with a 9% reduction in the risk of CKD (HR 0.91; 0.91–0.92). Additionally, SBP and heart rate were significantly associated with CKD incidence. Prevalent CVD at baseline was associated with a 35% (0.99–1.82) increase in the risk of CKD.

In the whole population, women had more than double risk (HR 2.09; 1.82–2.39) of CKD compared with men. Moreover, age, FPG, and DBP were positively associated with CKD incidence. In contrast, eGFR, WC, TC and heart rate were negatively associated with the risk of CKD. Every 1-unit increase in eGFR was associated with a 9% reduction in the risk of CKD (HR 0.91; 0.90–0.92). Prevalent CVD at baseline was associated with a 25% (1.01–1.54) increase in the risk of CKD. The HR for CKD among individuals with higher education (> 12 years) was 1.41 (1.14–1.72) compared to illiterate individuals (Table 5).

The results of the survival tree in men are presented in Fig. 3. The eGFR, age, FPG, PAL and WC were selected by the tree to group the male population. The best cutoff values for these predictors were found by the survival tree algorithm. The eGFR was found as the most important predictor for CKD with a cutoff value of 69.8 ml/min/1.73 m². The group with eGFR ≤ 69.8 ml/min/1.73 m² (left side of the tree) was further split by age; so that, the group with eGFR < 63.4 ml/min/1.73 m² and age > 50 year had the highest risk for CKD. The lowest risk was found among younger men aged ≤ 46 year and eGFR > 74 ml/min/1.73 m². A moderate risk was explored among the groups with eGFR > 69.85 ml/min/1.73 m² and age > 46 year; survival probability among these groups depended on the SBP and FPG. So that, the group with SBP > 119 mmHg and FPG > 5.8 mmol/l had the lowest survival. Ultimately, 12 groups were identified by the survival tree algorithm. The Kaplan–Meier curves for these groups are shown in Fig. 4a. The curves of nodes 4, 5, 7, 8, 22 and 23 were lower than the whole population’s curve, showing that they covered high risk groups for the CKD events among men. The symmetric remark can be made for nodes (11, 15, 16, 17, 18 and 20) which had a survival probability above the whole population’s curve which covered low risk groups for the CKD events in men. Using node 15 with the lowest risk as the reference group, HRs of CKD was estimated for all other groups (nodes). Accordingly, node 5 had the highest risk for CKD incidence, compared to node 15 (HR 70.6; 34.6–144.5; p < 0.001) (Table 6).

The survival tree for women is shown in Fig. 5. Three variables (age, eGFR and SBP) were found as the most significant predictors associated with CKD. Initially, the participants were split by age with the cutoff value of 45 year (p < 0.001). On the right side of the tree, the group aged > 45 year was split by eGFR with a cutoff value of 69.35 ml/min/1.73 m²; then, the group with eGFR > 69.35 ml/min/1.73 m² was split by SBP with a cutoff value of 133 mmHg. Further inspection of the tree showed that in age > 45 year, a SBP > 133 mmHg can significantly increase the risk of CKD. Among women in younger age (< 45 year), those with eGFR > 83.5 ml/min/1.73 m² (node 12) and eGFR < 65 ml/min/1.73 m² (node 4) had the lowest and highest risk, respectively, for incidence of CKD. The Kaplan–Meier curves for the ten groups are shown in Fig. 4b. The survival curves of nodes 16, 4, 19, 15, and 7 were lower than the whole population’s curve, and survival curves of nodes 12, 11, 6, and 9 had a survival probability above the whole population’s curve, showing the high risk and low risk groups for the CKD events among women, respectively. As Table 7 shows, node 16 had the highest risk for CKD incidence, compared to node 12 as the reference group (HR 27.25; 19.88–37.34; p < 0.001).

According to survival tree analysis, eGFR and age were found as the most important predictors for CKD incidence among men and women, respectively. Current evidences from cross sectional and longitudinal studies suggest that decline in GFR is a universal phenomenon which happens after age 30 in both men and women, and the rate of decline accelerates after age 50–60 years [44]. This decline appears to be a part of the normal physiologic process and is associated with structural changes in the kidney [45].

Survival tree in woman (Fig. 5) shows that not only aging is the most important predictor for CKD progression, but suggesting that the critical threshold of age for developing CKD is much lower than 30. As Fig. 5 shows, a group of women with age ≤ 29 year and eGFR between 65 and 71 ml/min/1.73 m² had a moderate risk of developing CKD during the follow-up period (node 6).

According to our findings, the lowest risk was found among women aged ≤ 45 year and with eGFR > 83 ml/min/1.73 m²; suggesting that the threshold of ≥ 83 ml/min/1.73 m² for eGFR is a safe value for decreasing the risk of CKD among women aged ≤ 45 year.

Figure 5 also shows that among women aged > 45 year, a SBP > 133 mmHg (high-range pre-hypertension and hypertension) was significantly associated with increased risk of CKD even with eGFR > 69 ml/min/1.73 m². According to a recent study, the overall prevalence of prehypertension was approximately 46% among adult Iranian population [46]; therefore, the effective strategies for the prevention and treatment of prehypertension is required to be considered among Iranian women.

Overall, the survival tree in women found 10 prognostic groups, defined only by the age, eGFR and SBP; it does not mean that other variables have no predictive ability in the CKD incidence. If the tree is growing in full depth, other variables may be involved in the prediction of CKD; however, the number of subjects in the terminal nodes will be so small for significant results.

Survival tree in male population identified 12 subgroups with different patterns (Fig. 3). The highest risk was found among men with eGFR < 63 ml/min/1.73 m² (nodes 4 and 5).

As we previously mentioned, the Cox PH model found negative association between WC and PAL in relation to risk of CKD among male population (Table 4). These unexpected results may be explained by the interactions found by the survival tree; as the Fig. 3 shows, low PAL and WC > 100 cm was significantly associated with increased risk of CKD (node 17); while, WC < 100 cm and FPG ≤ 4.7 mmol/l improved survival rate among inactive men (node 15). These findings suggest that there is a particularly strong interaction between these factors in relation to CKD incidence among male population. Survival tree also found a group of men aged > 46 year with a SBP > 119 mmHg (prehypertension and hypertension) and FPG > 5.83 mmol/l (pre-diabetes and diabetes), who had moderate risk of CKD incidence. In line with our findings, Yamagata et al. [37] showed that hypertension and diabetes were independent predictor for developing CKD. Because of the high prevalence and incidence of pre-hypertension, hypertension, pre-diabetes and diabetes among Iranian population [47], more attention should be paid to early intervention and risk factor modification for this high risk group.

In summary, survival trees are the helpful exploratory tools that have been applied in recent years in different medical studies [48‐51]. However, investigators should be aware of their drawbacks and limitations. A major issue with survival trees is that they can be built to fit the training data too perfectly (over fit); while it may not perform well on other unknown data [52]. We used the conditional inference survival tree developed by Hothorn et al. [53], which is shown to be more reliable and less likely to over fit compared to the other survival tree algorithms.

This study is subject to some limitations. First, our study was a post hoc analysis of a prospective cohort that was not able to control for some covariates related to CKD such as family history of kidney disease that is related to decline in kidney function beginning at early ages [54]. Additionally, no quantitative urinary indexes were available at the baseline or follow-up visit, because, the cohort was not originally designed for kidney diseases. Second, this study has been conducted on a sample of Iranian population and further studies should be conducted to determine whether our findings could be applicable to other populations. Third, as inherent to any prospective cohort study, the level of risk factors at the baseline examination might change during the follow up, e.g. some degree of misclassification may have occurred in this study, leading to biased estimated hazard ratios towards the null. Fourth, about 20% of participants were lost to follow-up in our study; we found statistically but not clinically important differences between the followed vs. non-followed population in some baseline variables. In total population, mean values of FPG and eGFR were lower, and TC and WC were higher in followed compared to non-followed population. Since FPG and eGFR were reported to be the most important predictors for CKD among Iranian population [32], we may underestimate the incidence of CKD in our study. To evaluate the extent to which lost to follow-up may have influenced measures of association in final models, we calculated propensity scores in both genders and in the total population. The PS is the conditional probability that an individual participate in the follow-up, given a set of covariates [55]. To obtain PS, we applied multivariable logistic regression with participation in the follow-up as the outcome and all baseline variables as the predictors. The estimated probability of participation, or the PS, derived from the multivariable logistic regression model described above was added to the final Cox PH models. In general, in both genders and total population, there were no changes in estimated HRs and CIs for predictors in the Cox PH models. We also added PS as a covariate to develop survival tree models along with all baseline covariates; interestingly, adding the PS did not alter the structure of survival tree in both genders and total population (data not shown). Therefore, lost to follow-up did not cause selection bias in our study. And last but not least, our diagnosis of CKD was based on a single estimate of GFR, which we acknowledge tends to overestimate the incidence of kidney disease [37].

As strengths, this is the first study that identified interactions between CKD risk factors in a well-characterized cohort of adults using the survival tree approach. We used data from a large, long-term, population-based cohort of men and women. In fact, in men, the statistical power of our study to detect a HR of 20, 30 and 50% for incident CKD was 72, 95 and 99%; the corresponding values in women were 94, 99 and 100%, respectively. We assessed the actual measurement of blood pressure and laboratory parameters, rather than self-reported data. In addition, we investigated a wide range of variables as potential predictors.

As, there will never be enough resources to implement every programme for all population, health policy makers prefer interventions that target high-risk groups [56]. Our study, identified not only risk factors of CKD in total population by Cox model but also identified high-risk groups by survival tree analysis, which can be used for implement specific interventions for each group according their level of risk.