Identifying patients at high risk of diabetic kidney disease (DKD) helps improve clinical outcome.
To establish a model for predicting DKD.
The derivation cohort was from a meta-analysis. The validation cohort was from a Chinese cohort.
Cohort studies that reported risk factors of DKD with their corresponding risk ratios (RRs) in patients with type 2 diabetes were selected. All patients had estimated glomerular filtration rate (eGFR) ≥60 mL/min/1.73 m2 and urinary albumin-to-creatinine ratio (UACR) <30 mg/g at baseline.
Risk factors and their corresponding RRs were extracted. Only risk factors with statistical significance were included in our DKD risk prediction model.
Twenty cohorts including 41,271 patients with type 2 diabetes were included in our meta-analysis. Age, BMI, smoking, diabetic retinopathy, hemoglobin A1c, systolic blood pressure, HDL cholesterol, triglycerides, UACR, and eGFR were statistically significant. All these risk factors were included in the model except eGFR because of the significant heterogeneity among studies. All risk factors were scored according to their weightings, and the highest score was 37.0. The model was validated in an external cohort with a median follow-up of 2.9 years. A cutoff value of 16 was selected with a sensitivity of 0.847 and a specificity of 0.677.
There was huge heterogeneity among studies involving eGFR. More evidence is needed to power it as a risk factor of DKD.
The DKD risk prediction model consisting of nine risk factors established in this study is a simple tool for detecting patients at high risk of DKD.
Introduction
Diabetic kidney disease (DKD) is a major microvascular complication of diabetes with high prevalence, mortality, and treatment cost, yet there is low awareness of it and it has a poor effective prevention and treatment rate (1). The lack of early diagnosis and effective therapies targeting DKD is mainly responsible for the problem. The usefulness of microalbuminuria as the traditional marker of DKD and the optimal opportunity for intervention (2) has recently been challenged. As a dynamic, fluctuating condition rather than a linearly progressive process, microalbuminuria is observed to follow a remission/regression trajectory of less certain prognostic value, lacking both sensitivity and specificity in evaluating early DKD (3). Furthermore, the relationship between renal function and albuminuria is weak (4). In the Intensified Multifactorial Intervention in Patients with Type 2 Diabetes and Microalbuminuria (Steno-2) study, even with positive intervention 31% of participants with microalbuminuria progressed to macroalbuminuria during a 7.8-year follow-up (5). This pattern is seen because the presence of microalbuminuria in patients with diabetes often implies the kidneys have undergone different degrees of structural injuries.
Therefore, researchers have explored early predictive markers of DKD, including through proteomics and genomics. These novel biomarkers, however, are difficult to promote in clinical practice because of their instability and high cost. Type 2 diabetes coexists with multiple metabolic disorders. Besides hyperglycemia, other factors such as hypertension, dyslipidemia, and unhealthy lifestyle are all involved in DKD onset. A comprehensive assessment of these risk factors, along with early detection and individualized intervention in high-risk individuals, maybe the most effective strategy for preventing DKD.
Methods
Study Populations
Derivation Cohort
The derivation cohort came from a systematic review and meta-analysis of 14 prospective cohorts and 6 retrospective cohorts. These 20 cohorts were identified by searching the electronic databases of MEDLINE, Embase, and the Cochrane Library from the time of their inception to October 2017, using a combined text and MeSH heading search strategy with the terms: “diabetes mellitus, type 2,” “kidney disease,” “risk factor,” and “cohort study.” In total, 41,271 patients with type 2 diabetes from Europe (including Italy, Denmark, Sweden, and Finland), Asia (including China, China Taiwan, Japan, India, and Singapore), and the Americas (including the U.S. and Brazil) with an estimated glomerular filtration rate (eGFR) ≥60 mL/min/1.73 m2 and urinary albumin-to-creatinine ratio (UACR) <30 mg/g at baseline were included in our derivation cohort. Of the cohort patients, 80% were white and 20% were Asian. All the studies reported risk ratios (RRs) and corresponding 95% CIs for risk factors and were assessed using the Newcastle–Ottawa scales. A flowchart of the study selection methodology is shown in Fig. 1A, and search strategy, selection criteria, data extraction, and quality assessment are detailed in the Supplementary Data.
Validation Cohort
A total of 2,089 patients with diabetes who were hospitalized at least twice in Tianjin Medical University Metabolic Diseases Hospital (baseline from 27 July 2012 to 24 August 2017) were considered for this study. To be included in the validation cohort, patients needed to be between 39 and 75 years old, show normoalbuminuria (UACR <30 mg/g or albumin excretion rate [AER] <30 mg) and eGFR ≥60 mL/min/1.73 m2 at baseline, and be with a follow-up for more than 12 months. Given that other kidney diseases may affect kidney function and albuminuria and increase the chance of hospitalization, we excluded those patients whose last hospitalization were due to acute kidney injury, primary glomerulonephritis, urinary tract infection, urinary calculi, etc. Ultimately, we excluded 120 patients with type 1 diabetes, 46 patients aged <39 or >75 years, 589 patients with baseline UACR ≥30 mg/g or AER ≥30 mg or eGFR <60 mL/min/1.73 m2, 213 patients with incomplete baseline data, 199 patients who lacked the last UACR/AER/eGFR measurements, 219 patients with a follow-up for less than 12 months, and 323 patients with acute diabetic complications or serious infection. The remaining 380 patients with type 2 diabetes were included in the validation cohort. A flowchart outlining the selection of study patients is shown in Fig. 1B.
Outcome
The outcome was the occurrence of DKD defined as eGFR <60 mL/min/1.73 m2 and/or UACR ≥30 mg/g (or AER ≥30 mg) for ≥3 months caused by diabetes (6).
Ethics Statement
This study was approved by the Institutional Review Board of Tianjin Medical University Metabolic Diseases Hospital and Tianjin Institute of Endocrinology. It was agreed to waive the requirement for informed consent, as this study was designed to retrospectively collect available data from articles published in peer-reviewed journals and databases.
Statistical Analysis
Meta-analysis
RRs with 95% CIs for each initiating risk factor for DKD were extracted and then generated using a random effects model or a fixed effects model according to the heterogeneity. The inverse of the variance of the RR was used to weight studies on the basis of an estimate of statistical size (7). Heterogeneity across studies was assessed using the Cochrane Q test, and measured by I2. An I2 value of more than 50% or a P value in the Cochrane Q test of less than 0.10 indicated statistically significant heterogeneity, in which case the random effects model was applied. Otherwise, the fixed effects model was selected (8). Subgroup analyses were conducted according to the magnitude of the change in the continuous variables. Continuous variables included age (1-year increments vs. 5–10 year increments), BMI (1 kg/m2 increments vs. 5 kg/m2 increments), and systolic blood pressure (SBP) (1 mmHg increments vs. 5 mmHg increments vs. 10–20 mmHg increments). Sensitivity analyses were conducted by omitting a single study in turn to test the robustness of the results (9). Publication bias was assessed by Egger’s linear regression test at a P < 0.10 significance level. All tests were two-sided and statistical significance was defined as P < 0.05, with the exception of the heterogeneity and publication assessment, which were considered statistically significant at P < 0.10. All analyses were performed using Stata software, version 14.0 (StataCorp, College Station, TX).
Model Development
We developed a categorization point system according to the methods suggested by Sullivan et al. (10). First, all risk factors included in the prediction model were selected from the systematic review and meta-analysis described above. Second, the pooled RR of each risk factor was extracted. To make our prediction model more practical, we chose RR from subgroup or sensitivity analyses. For example, in considering age as a factor, we selected the RR of age at 5–10 year increments. β-Coefficients of each risk factor were calculated according to the pooled RR and its corresponding 95% CI. Then, scores were calculated by multiplying the β-coefficients by 10 and rounding off one decimal place. Finally, all risk factors in the prediction model were categorized based on the meta-analysis and clinical practice guidelines, and each category was assigned a score. The gross score was calculated by summing the scores of all components in the prediction model. The difference was considered nonsignificant at P > 0.05.
Model Validation
External data from the retrospective cohort study described above were used to examine the generalizability of the risk prediction model, identifying the best implementation strategy. Baseline variables were used to calculate the total score according to the risk prediction model. However, follow-up DKD incidence was used to calculate the area under the receiver operating characteristic curve. Sensitivity, specificity, and the area under the curve (AUC) were calculated at different cutoff values, and the calculations were used to identify the optimal cutoff point (11). According to the optimal cutoff point, patients were categorized into four risk levels, including relatively low, moderate, high, and very high risk. To validate the predictive power of the prediction model, the actual incidence of DKD was calculated and a Kaplan-Meier curve was generated in each risk group. All analyses were performed using SPSS software, version 22 (IBM Corp.).
Results
Description of the Cohorts
Derivation Cohort
Patients in the derivation cohort were between 39 and 75 years of age and their diabetes duration ranged from 1 to 20 years. They were followed for 1 to 20 years, equivalent to 41,271 to 825,420 person-years. Baseline characteristics of the derivation cohort from the meta-analysis were shown in Supplementary Table 1. According to Newcastle-Ottawa scales, all 20 studies included in the systematic review and meta-analysis were of high quality with scores over 8 (shown in Supplementary Table 2). We performed a crude analysis of the baseline data for the participants in the included cohort studies. In the cohort of white patients, mean BMI ranged from 27.0 to 33.0 kg/m2, mean hemoglobin A1c (HbA1c) ranged from 8.0% to 10.0% (64 to 86 mmol/mol), mean triglycerides (TG) ranged from 1.46 to 2.30 mmol/L, mean HDL cholesterol (HDL-C) ranged from 1.03 to 1.38 mmol/L, and mean SBP ranged from 127 to 153 mmHg. Across the studies, 38.9–86.8% of white patients received antidiabetes treatment, 10.6–17.98% received insulin treatment, 37.5–59.2% received antihypertensive treatment, and 10.3–81.7% received ACE inhibitors or angiotensin receptor blockers. In the cohort of Asian patients, mean BMI ranged from 23.4 to 26.2 kg/m2, mean HbA1c ranged from 8.6% to 10.0% (70 to 86 mmol/mol), mean TG ranged from 1.5 to 2.3 mmol/L, mean HDL-C ranged from 1.01 to 1.48 mmol/L, and mean SBP ranged from 125 to 144 mmHg. Across the studies, 36.2–90.0% of Asian patients received antidiabetes treatment, 13.5–31.4% received insulin treatment, 10.0–90.6% received antihypertensive treatment, and 7.4–80.0% received ACE inhibitors or angiotensin receptor blockers. Except for BMI, which was significantly lower in Asian patients than white patients, there were no significant differences between Asian and white patients.
During the patient follow-up, 11,991 incident DKD cases were observed. The estimated incidence of DKD was approximately 29.1%. There were 19 risk factors available from the studies of these cohorts, including age, sex, diabetes duration, smoking, BMI, diabetic retinopathy (DR), HbA1c, fasting plasma glucose, hypertension, SBP, pulse pressure, total cholesterol, HDL-C, TG, LDL cholesterol, uric acid, UACR, serum creatinine, and eGFR. Detailed information on these studies and respective cohorts are shown in the Supplementary Data.
Validation Cohort
Of 380 patients with type 2 diabetes included in the validation cohort, 44.7% were female. The mean age ± SD at baseline was 55 ± 9 years, and the median follow-up was 2.9 years (interquartile range [IQR] 2.0–3.9). The proportion of smokers was 41.3%, and 24.7% of the patients had DR. The baseline data for the validation cohort indicated poor control of weight (mean ± SD BMI 26.57 ± 3.36 kg/m2), HbA1c (median 8.5% [69 mmol/mol]; IQR 7.3–9.8% [56–84 mmol/mol]), and serum lipid levels (median TG 1.74 mmol/L [IQR 1.18–2.70]; median HDL-C 1.20 mmol/L [IQR 1.10–1.40]). The median SBP was 130 mmHg (IQR 120–140). All patients showed normoalbuminuria and maintained renal function at baseline with median UACR of 14.7 mg/g (IQR 11.9–18.5) and mean ± SD eGFR of 99.4 ± 17.9 mL/min/1.73 m2. Most patients (375 [98.7%]) received oral antihyperglycemic treatment and 105 (27.6%) received antihypertensive treatment. Baseline characteristics of patients in the validation cohort are shown in Supplementary Table 4. By the end of the observation, 98 (25.8%) patients developed DKD.
Model Derivation
Of the 19 risk factors identified in the systematic review and meta-analysis, 10 risk factors were associated with the onset of DKD. These factors, followed by the respective pooled RR, were age (1.09), BMI (1.07), DR (1.72), smoking (1.49), HbA1c (1.17), SBP (1.03), HDL-C (0.75), TG (1.15), UACR (1.25), and eGFR (2.20). A forest plot of these factors is shown in Fig. 2A (detailed information and forest plots of these risk factors were shown in Supplementary Table 3 and Supplementary Figs. 1–10). We selected the results from subgroup or sensitivity analyses which were most reasonable considering the feasibility of clinical practice. There were different definitions and large heterogeneities (I2 = 100%) among studies involving eGFR, so eGFR was not included in our risk prediction model. Risk factors included in the final risk prediction model were age incremented by 5–10 years (RR 1.38, 95% CI 1.20–1.59; P < 0.001), BMI incremented by 5 kg/m2 (RR 1.16, 95% CI 1.09–1.23; P < 0.001), DR (RR 1.31, 95% CI 1.00–1.73; P = 0.05), smoking (RR 1.49, 95% CI 1.30–1.71; P < 0.001), HbA1c incremented by 1% (11 mmol/mol) (RR 1.17, 95% CI 1.09–1.26; P < 0.001), SBP incremented by 10–20 mmHg (RR 1.21, 95% CI 1.15–1.27; P < 0.001), HDL-C incremented by 1 mmol/L (RR 0.78, 95% CI 0.61–0.99; P < 0.001), TG incremented by 1 mmol/L (RR 1.42, 95% CI 1.16–1.74; P < 0.001), and UACR incremented by 1 mg/g (RR 1.13, 95% CI 1.10–1.17; P < 0.001). The forest plot of the subgroup and sensitivity analyses are shown in Fig. 2B. Risk factors with the number of involved studies, sample size, pooled RRs (95% CIs), β-coefficients, and risk scores of risk factors included in the DKD risk prediction model are shown in Supplementary Table 5.
Conclusively, a simple DKD risk prediction model was developed as follows: age (years; 39–49 scores 0, 50–59 scores 3.0, and 60–75 scores 6.0), BMI (kg/m2; <25.00 scores 0, 25.00–29.99 scores 1.5, and ≥30.00 scores 3.0), smoker (nonsmoker scores 0 and smoker scores 4.0), DR (0 if no and 3.0 if yes), HbA1c (<7.0% [<53 mmol/mol] scores 0, 7.0–7.9% [53–63 mmol/mol] scores 1.5, 8.0–8.9% [64–74 mmol/mol] scores 3.0, and ≥9.0% [≥75 mmol/mol] scores 4.5), SBP (mmHg; <130 scores 0, 130–139 scores 2.0, 140–149 scores 4.0, and ≥150 scores 6.0), HDL-C (mmol/L; ≥1.3 scores 0 and <1.3 scores 2.5), TG (mmol/L; <1.7 scores 0 and ≥1.7 scores 4.0), and UACR (mg/g; <10.00 scores 0, 10.00–19.99 scores 2.0, and 20.00–29.99 scores 4.0) (shown in Table 1).
Risk factor of DKD . | Category . | Point . |
---|---|---|
Age** (years) | 39–49 | 0 |
50–59 | 3 | |
60–75 | 6 | |
BMI***(kg/m2) | <25.00 | 0 |
25.00–29.99 | 1.5 | |
≥30.00 | 3 | |
Smoker**** | Nonsmoker | 0 |
Smoker | 4 | |
DR | No | 0 |
Yes | 3 | |
HbA1c (% [mmol/mol]) | <7.0 [<53] | 0 |
7.0–7.9 [53–63] | 1.5 | |
8.0–8.9 [64–74] | 3 | |
≥9.0 [≥75] | 4.5 | |
SBP (mmHg) | <130 | 0 |
130–139 | 2 | |
140–149 | 4 | |
≥150 | 6 | |
HDL-C (mmol/L) | ≥1.30 | 0 |
<1.30 | 2.5 | |
TG (mmol/L) | <1.70 | 0 |
≥1.70 | 4 | |
UACR (mg/g) | <10.00 | 0 |
10.00–19.99 | 2 | |
20.00–29.99 | 4 |
Risk factor of DKD . | Category . | Point . |
---|---|---|
Age** (years) | 39–49 | 0 |
50–59 | 3 | |
60–75 | 6 | |
BMI***(kg/m2) | <25.00 | 0 |
25.00–29.99 | 1.5 | |
≥30.00 | 3 | |
Smoker**** | Nonsmoker | 0 |
Smoker | 4 | |
DR | No | 0 |
Yes | 3 | |
HbA1c (% [mmol/mol]) | <7.0 [<53] | 0 |
7.0–7.9 [53–63] | 1.5 | |
8.0–8.9 [64–74] | 3 | |
≥9.0 [≥75] | 4.5 | |
SBP (mmHg) | <130 | 0 |
130–139 | 2 | |
140–149 | 4 | |
≥150 | 6 | |
HDL-C (mmol/L) | ≥1.30 | 0 |
<1.30 | 2.5 | |
TG (mmol/L) | <1.70 | 0 |
≥1.70 | 4 | |
UACR (mg/g) | <10.00 | 0 |
10.00–19.99 | 2 | |
20.00–29.99 | 4 |
DKD risk prediction model applied to patients with type 2 diabetes without DKD.
Patients in derivation and validation cohorts aged 39–75 years old.
BMI was categorized as <25.00, 25.00–29.99, and ≥30.00 kg/m2 in white patients but <24.00, 24.00–27.99, and ≥28.00 kg/m2 in Asians.
Smoker was defined as having smoked more than 100 cigarettes in their lifetime.
This prediction model is recommended to be applied to patients with type 2 diabetes aged 39–75 years old and mainly for white and Asians. It is vital to note that because of the ethnic differences, BMI was classified according to two different sets of criteria: 1) World Health Organization criteria for white patients (12) were normal (18.50–24.99 kg/m2), overweight (25.00–29.99 kg/m2), and obese (≥30.00 kg/m2); and 2) Chinese criteria for Chinese patients (13) were normal (18.50–23.99 kg/m2), overweight (24.00–27.99 kg/m2), and obese (≥28.00 kg/m2) according to the criteria established in 2003 by the Working Group on Obesity in China.
Model Validation
As the validation cohort was from Chinese patients, we applied the Chinese criteria of BMI in our DKD risk prediction model. The receiver operating characteristic curve of the DKD risk prediction model is shown in Fig. 3A. The AUC was 0.765 (95% CI 0.710–0.821) in the validation cohort. Given that the aim of the model development was an early detection of patients with diabetes at high-risk for DKD, 16.0 was selected as the optimal cutoff risk score with a higher sensitivity of 0.847 and a specificity of 0.677. Sensitivity and specificity at different cutoff risk score are shown in Supplementary Table 6. Based on the obtained frequencies of DKD using different risk scores, the 380 patients with diabetes were further categorized into four risk-level groups: relatively low (n = 89), moderate (n = 113), high (n = 165), and very high (n = 13) risk, corresponding to risk scores of <12.0, 12.0–15.5, 16.0–26.5, and 27.0–37.0, respectively. The numbers of patients who developed DKD at the end of the follow-up were 6 (6.7%), 16 (14.2%), 66 (40.0%), and 10 (76.9%) in these four groups, respectively (Fig. 3B). Figure 3C shows the Kaplan-Meier curves for DKD grouped by risk scores. Compared with the low-risk group, high- and very high-risk groups had significantly higher hazard ratios. Compared with the low-risk group, the RRs of developing DKD in high- and very high-risk groups were 9.22 (3.81–22.35) and 46.11 (9.95–213.66), respectively (P < 0.001).
Discussion
Diabetic kidney disease (DKD) is a complex and multifactorial disease, involving both genetic and environmental factors (14). Current knowledge of DKD risk factors is derived mainly from cohort studies with sample sizes ranging from a few hundred to a few thousand subjects (15–22) (23–34). Considering the results of these studies were not completely consistent, systematic review and meta-analysis was chosen as a method to integrate these studies. High-quality systematic review and meta-analysis is a proven method that sits at the top of the evidence pyramid, which could improve statistical efficiency (35).
Using 20 high-quality cohort studies of 41,271 individuals, the following independent risk factors for DKD development were identified: age, BMI, smoking status, DR, HbA1c, SBP, HDL-C, TG, UACR, and eGFR. These factors are essentially consistent with the risk factors traditionally associated with DKD, further verifying that age and harmful lifestyle, together with poor glycemia, blood pressure, and plasma lipid control, initiate DKD. Except for age, most of these risk factors are modifiable, which is particularly important for DKD prevention. Comprehensive, targeted, and timely treatment including medication and lifestyle interventions would delay or even prevent the development of DKD and relieve personal and national financial burden.
Age, smoking, and TG were the most powerful baseline risk factors for detecting DKD in our systematic review and meta-analysis. With age incremented by 5–10 years, the risk for DKD was increased by 38% in our study. Smoking has been confirmed as an independent risk factor for DKD onset and progression (36). In our study, the risk for DKD increased by 49% in patients with diabetes who smoked. Dyslipidemia plays a crucial role in the development and progression of DKD, especially high TG and/or low HDL-C levels (37–39). In our study, as TG increased by 1 mmol/L, the risk for DKD increased by 42%. With HDL-C increased by 1 mmol/L, the risk for DKD was reduced by 22%. Hyperglycemia and hypertension are the main risk factors for DKD development (40). Many randomized controlled trials have shown that intensive blood glucose control (HbA1c 6.5–7.0% [48–53 mmol/mol]) can reduce the risk of DKD (41–44). The earlier the initiation of antihyperglycemic therapy, the greater the prognosis benefits (45). In our study, with HbA1c increased by 1% (11 mmol/mol), the risk for DKD increased by 17%. However, there was no statistical difference in risk for DKD with increased FPG (RR with its corresponding 95% CI was 1.05 [0.96, 1.14], P = 0.30). With SBP increased by 10–20 mmHg, the risk of DKD increased by 21% in our study. BMI is the most common measure to assess obesity. We found that the risk of DKD increased by 16% when BMI increased by 5 kg/m2. Considering the ethnic and dietary differences, BMI was classified according to different criteria in white and Asian patients. We selected Chinese criteria in our validation cohort. DR and DKD are both microvascular complications of diabetes, which are often concurrent. Our meta-analysis showed that patients with DR had their risk of DKD increased by 31%. Considering the significant heterogeneity among the relevant studies involving DR, more prospective cohorts are needed to further verify our result. Babazono et al. (46) reported that an increase in baseline UACR within the normal range predicted a faster decline in eGFR. The Kidney Disease Outcomes Quality Initiative (KDOQI) recommended a stratification of normal albuminuria according to UACR <10 mg/g (optimal) and 10 ≤ UACR < 30 mg/g (normal high limit) (47). In our study, with UACR increased by 1 mg/g, the risk for DKD increased by 13%. Considering the weight of UACR is too high, we converted it into categorical variable defined as <10.00 mg/g, 10.00–19.99 mg/g, and 20.00–29.99 mg/g. In our meta-analysis, decreased eGFR will increase the risk for DKD development by 120%. There were several reasons for excluding eGFR out of our prediction model. First, there was significant heterogeneity among the four studies involving eGFR (I2 = 100%), and the heterogeneity cannot be reduced or eliminated by statistical methods such as subgroup analysis or sensitivity analysis. We carefully searched for the sources of this huge heterogeneity and found it was mainly from the different definitions of eGFR as a risk factor; it was defined as each 5 mL/min/1.73 m2 drop in the study by Low et al. (18), as a continuous variable in the study by Takagi et al. (19), as a categorical variable of 60 ≤ eGFR < 90 mL/min/1.73 m2 and eGFR ≥ 120 mL/min/1.73 m2 in the study by Hu et al. (20), and as each 10 mL/min/1.73 m2 drop in the context of eGFR <90 mL/min/1.73 m2 in the study by De Cosmo et al. (21). Second, we did establish a DKD risk prediction model including eGFR. We defined 90 ≤ eGFR < 120 mL/min/1.73 m2 as the normal filtration group with a 0 score, eGFR ≥120 mL/min/1.73 m2 as the high filtration group with a score of 4, and 60 ≤eGFR <90 mL/min/1.73 m2 as the low filtration group with a score of 4. A score of 15.5 was selected as the cutoff value with a sensitivity of 0.755 and specificity of 0.603. Compared with our established model without eGFR (16 was selected as the cutoff value with a sensitivity of 0.847 and specificity of 0.677), eGFR did not improve the predictive performance. We think this may be related to the characteristic of eGFR: the rate of eGFR decline may have a stronger predictive value for DKD compared with a static value at a certain point in time in patients with type 1 diabetes (48). However, we still need stronger evidence to prove it in a larger population. This provides a direction for our future research. Last but not least, our DKD risk prediction model will only be applied in those with eGFR >60 mL/min/1.73 m2 and/or UACR <30 mg/g.
Any marker of microalbuminuria, proteomics, or genomics alone is not enough to accurately predict DKD. Instead, integrating the traditional risk factors mentioned above to assess the risk of DKD will be more reliable and feasible in moving forward with prevention and risk management. Several investigators have developed prediction models for diabetes-related kidney disease (48–54). Most of those prediction models, however, were developed based on small cross-sectional studies or post hoc analyses of randomized controlled trials. Thus, we established a DKD risk prediction model based on a systematic review and meta-analysis of 20 high-quality cohort studies including 41,271 patients with type 2 diabetes, which greatly improved statistical performance. A simple prediction model including common clinical data of age, BMI, smoking, DR, HbA1c, SBP, HDL-C, TG, and UACR was developed. A prior model developed by Ravizza et al. (55) consisted of data that are unconventional in clinical practice and used complex machine learning–based embedded feature selection methods. Our DKD risk prediction model, on the other hand, requires no complex calculations and is more convenient and intuitive for both patients and clinicians. Furthermore, most of the established models evaluate a relatively later stage of DKD, such as doubling of creatinine or end stage renal disease. For early detection and intervention, our prediction model predicts the early stage of DKD using a combination of UACR ≥30 mg/g and/or eGFR <60 mL/min/1.73 m2. This early detection will be helpful for DKD prevention.
Moreover, we validated our DKD risk prediction model in an external cohort of 380 patients with type 2 diabetes. The maximum score of our model is 37.0. Considering that early detection and prevention will bring huge benefits to patients, and there will be no harm even if they are misclassified as high-risk groups, we conservatively selected 16.0 as the cutoff value, with the sensitivity and specificity of 84.7% and 67.7%, respectively. The prediction model had good discriminative power with an AUC of 0.765. By calculating risk score, it was easy to distinguish high-risk patients from other populations. We further categorized patients into relatively low, moderate, high and very high risk groups according to their risk scores. Compared with participants in the low-risk group, those in high- and very high-risk groups had 9.22- and 46.11-fold increases in the odds of developing DKD, respectively. Our prediction model can dynamically assess the risk of future DKD in patients with type 2 diabetes, allowing them to constantly adjust treatment options. Through early detection and reasonable intervention, individuals at high risk of DKD could be converted to lower risk, which is the main purpose of our model development.
Nevertheless, there are several limitations to this study. First, the method of systematic review and meta-analysis was inevitably heterogeneous due to differences in research design and method as well as different race and sex compositions of the cohorts in included studies. But the heterogeneity could be reduced by further investigating the source and performing subgroup analysis and sensitivity analysis. Second, our derivation cohort consisted of about 80% white patients and 20% Asian patients, leaving out some of the ethnicities with higher risk for DKD incidence (for instance, black and Hispanic patients). This may affect the universality of the application of our DKD risk prediction model. We hope there will be more cohorts consisting of black and Hispanic patients to be included in our meta-analysis in the future. Beyond that, 80% of the derivation cohort were white patients, but we only validated the prediction model with a monocentric retrospective cohort from Chinese individuals. Therefore, we need to verify the model in a prospective multicentric cohort consisting of white patients in the future. Third, a median follow-up period of 2.9 years may be insufficient for our validation cohort, but considering that most of the patients were not newly diagnosed with diabetes at baseline but had a median diabetes duration of 7 years, it took about 10 years for these patients to develop DKD. This may partly compensate for our inadequate follow-up time. Finally, baseline risk may vary across study populations, so a model’s implementation may need to be tailored to each population (often referred to as recalibration) to improve performance in a new populations (56).
Conclusion
Based on a systematic review and meta-analysis, we developed a simple risk prediction model for early DKD that integrates lifestyle and clinical risk factors, including age, BMI, smoking, DR, HbA1c, SBP, HDL-C, TG, and UACR. We validated the model in an external cohort, showing that the model is important for early detection, diagnosis, and intervention of DKD.
W.J. and J.W. contributed equally to this study. Bao.C. and J.Y. contributed equally to this study.
Article Information
Funding. This work was supported by the National Key R&D Program of China (2018YFC1314000), National Natural Science Foundation of China (81603461, 81774043), Science and Technology Program of Tianjin (17ZXMFSY00140, 18ZXZNSY00280), Tianjin Health and Family Planning Commission (16KG167), Natural Science Foundation of Tianjin (17JCZDJC34700), and Scientific Research Funding of Tianjin Medical University Chu Hsien-I Memorial Hospital (2015DX05).
Duality of Interest. No potential conflicts of interest relevant to this article were reported.
Author Contributions. W.J. and J.W. performed the research, acquired data, and wrote the manuscript. X.S., W. Lu, Y.W., W. Li, Z.G., J.X., X.L., R.L., M.Z., Bai C., and J.L. contributed to the discussion and reviewed the manuscript. J.Y. and Bao.C. contributed to concept and design, study quality assessment, statistical analysis, and manuscript revision.