The average age was 68.28 ± 8.73 years, the gender ratio was 1:1.79, and the number of DR patients was 660 (34.8 %). The univariate analysis identified diabetes family history, treatment method, nephropathy, and duration, the smoking and alcohol habit, glycemic control, and the HbA1c level (P < 0.05) as DR associated factors. Gender, age, educational level, hypertension, hyperlipidemia, and diabetic foot were not associated (P > 0.05; Table 1).
Table 1
Univariate analysis of diabetic retinopathy related factors
Variable
|
DR (-)
(n = 1236)
|
DR (+)
(n = 660)
|
χ2/t
|
P
|
Sex
|
|
|
|
|
male
|
460(67.6)
|
220(32.4)
|
2.821
|
0.093
|
female
|
776(63.8)
|
440(36.2)
|
Age
|
|
|
|
|
50~
|
229(69.2)
|
102(30.8)
|
4.304
|
0.116
|
60~
|
439(62.7)
|
261(37.3)
|
70~
|
568(65.7)
|
297(34.3)
|
Education
|
|
|
|
|
Illiteracy
|
584(63.8)
|
332(36.2)
|
2.840
|
0.417
|
Primary school
|
381(67.6)
|
183(32.4)
|
Junior high school
|
204(66.2)
|
104(33.8)
|
Senior high school
|
67(62.0)
|
41(38.0)
|
Family history of diabetes
|
|
|
|
|
No
|
1050(67.5)
|
506(32.5)
|
20.067
|
< 0.001
|
Yes
|
186(54.7)
|
154(45.3)
|
Treatment method of diabetes
|
|
|
|
|
Diet control and exercise
|
477(87.7)
|
67(12.3)
|
231.801
|
< 0.001
|
Oral medicine only
|
660(61.3)
|
417(38.7)
|
Insulin
|
99(36.0)
|
176(64.0)
|
Hypertension
|
|
|
|
|
No
|
822(65.7)
|
430(34.3)
|
0.079
|
0.778
|
Yes
|
414(64.3)
|
230(35.7)
|
Hyperlipidemia
|
|
|
|
|
No
|
822(65.7)
|
430(34.3)
|
0.351
|
0.553
|
Yes
|
414(64.3)
|
230(35.7)
|
Diabetic nephropathy
|
|
|
|
|
No
|
1146(66.1)
|
588(33.9)
|
7.245
|
0.007
|
Yes
|
90(55.6)
|
72(44.4)
|
Diabetic foot
|
|
|
|
|
No
|
1196(65.2)
|
637(34.8)
|
0.083
|
0.774
|
Yes
|
40(63.5)
|
23(36.5)
|
Smoking
|
|
|
|
|
No/Quit
|
937(63.5)
|
539(36.5)
|
8.561
|
0.003
|
Yes
|
299(71.2)
|
121(28.8)
|
Drinking alcohol
|
|
|
|
|
No/Quit
|
990(63.3)
|
574(36.7)
|
14.069
|
< 0.001
|
Yes
|
246(74.1)
|
86(25.9)
|
glycemic control level
|
|
|
|
|
< 6.1
|
71(75.5)
|
23(24.5)
|
87.097
|
< 0.001
|
6.1~
|
584(76.5)
|
179(23.5)
|
≥ 8.1
|
581(55.9)
|
458(44.1)
|
Duration of diabetes(year)
|
|
|
|
|
< 5
|
619(86.8)
|
94(13.2)
|
361.013
|
< 0.001
|
5-
|
398(67.7)
|
190(32.3)
|
10-
|
124(35.0)
|
230(65.0)
|
≥ 15
|
95(39.4)
|
146(60.6)
|
BMI (kg/m2)
|
|
|
|
|
< 24
|
280(62.5)
|
168(37.5)
|
1.870
|
0.171
|
≥ 24
|
956(66.0)
|
492(34.0)
|
HbA1c (%)
|
|
|
|
|
< 7.0
|
603(76.3)
|
187(23.7)
|
88.457
|
< 0.001
|
7.0-
|
466(61.0)
|
298(39.0)
|
≥ 10
|
167(48.8)
|
175(51.2)
|
The eight statistically significant factors identified in the univariate analysis were taken as independent variables in the multivariate logistic regression analysis (the forward LR method), which determined that the family diabetes history, the treatment method, diabetes duration, and glycosylated hemoglobin should be included in the final regression model. After excluding other confounding factors (e.g., sex, age, education, hypertension, hyperlipidemia, diabetic nephropathy, diabetic foot, and BMI), the above factors in the final regression model were associated with and independent predictors of DR. Table 2 presents the partial regression coefficient (β) , odds ratio (OR), and P-value for each factor. The statistical equation model (Eq. 1) for predicting DR was:
Table 2
Multivariate analysis of diabetic retinopathy related factors
Variable
|
β
|
SE
|
Wald
|
Sig
|
OR
|
95%CI
|
Constant term
|
-3.138
|
0.191
|
268.672
|
< 0.001
|
-
|
-
|
Family history of diabetes
|
0.279
|
0.136
|
4.185
|
0.041
|
1.322
|
1.012–1.727
|
Treatment method of diabetes
|
0.730
|
0.103
|
50.451
|
< 0.001
|
2.074
|
1.696–2.537
|
Duration of diabetes (year)
|
0.107
|
0.011
|
90.912
|
< 0.001
|
1.113
|
1.089–1.138
|
HbA1c (%)
|
0.244
|
0.076
|
10.190
|
0.001
|
1.276
|
1.099–1.482
|
$$P = -3.138 + 0.279{X}_{1 }+ 0.730{X}_{2 }+ 0.107{X}_{3 }+ 0.244{X}_{4}$$
Equation 1: X1 equals diabetes family history, X2 equals the diabetes treatment method, X3 equals the diabetes duration, and X4 equals the HbA1c level.
The probability of suffering from DR per individual was calculated using the statistical equation model (Eq. 2):
$$P = 1/[1 + {e}^{-(-3.138+0.279{X}_{1}+0.730{X}_{2}+0.107{X}_{3}+0.244{X}_{4})}]$$
The complexity of the statistical equation models requires reliance on a calculator or computer, which limits the clinical applicability. Therefore, it was simplified to construct a simple scoring scale.
Based on the practical clinical significance, the diabetes duration was divided into four categories (< 5, 5–9, 10–14, and ≥ 15), and the HbA1c level was divided into three categories (< 7.0, 7.0-9.9, ≥ 10 %). Based on the scoring system and the regression Eq. 2, a simple scoring scale for DR was constructed and had a maximum score of 12 points. Table 3 presents the score assignments of various risk factors.
calculation notes: Classify the risk factors and determine the reference value of each category as:
C(i=1, ... ,n;j=1, ... ,Ci; Ci is the total number of categories of risk factors). Based on the original data distribution, the lower limit of hemoglobinA1c (HbA1c; %) was 3, the upper limit was 15, the lower limit of the diabetes duration was 1, and the upper limit was 40.
WiREF: One category for each risk is a reference and assigned 0 points, recorded as: WiREF (i=1, ...n). The HbA1c reference was <7.0 % and the diabetes duration reference was <5 years.
βi( Wij -WiREF): Calculate the difference between the reference value of each category and the reference value of the reference category and multiply it by the corresponding weight coefficient of each risk factor, recorded as: βi( Wij -WiREF), C(i=1, ... ,n;j=1, ... ,Ci; Ci is the total number of categories of risk factors)
B: Set a constant for the rating scheme system (i.e., determine a constant that makes the number of regression units for a certain category be assigned to 1 point, accordingly). In this study, the coefficient of 5-times the age was taken as a fixed constant: B= 5(0.107) = 0.535.
βi( Wij -WiREF)/B: Determine the score with each category per risk factor: Scoresij C(i=1, ... ,n;j=1, ... ,Ci; Ci is the total number of categories of risk factors).The relevant scores per risk factor in each category are calculated by: Scoresij = βi( Wij -WiREF)/B). The final score is the result rounded to the nearest integer value.
The simple scoring scale was comprehensively evaluated by the ROC curve (Fig. 1); the AUC was 0.753 (95 % confidence interval: 0.731–0.776). According to the coordinates of each point on the ROC curve, the diagnostic cut-off value at the maximum Youden Index was 3.5 points. According to the actual scoring, the cut-off value was 4 points. The sensitivity value was 0.665, the specificity value was 0.732, the positive predictive value was 0.570, the negative predictive value was 0.804, and the coincidence rate was 0.709 using a 4-point cut-off value (Table 4). For early detection and to improve the sensitivity, a 3-point cut-off value was evaluated. The sensitivity value was 0.797, specificity value was 0.571, positive predictive value was 0.498, negative predictive value was 0.840, and the coincidence rate was 0.650.
Table 4
The scale discrimination results
Scale Score
|
Clinical Diagnosis (Golden Standard)
|
Total
|
DR (+)
|
DR (-)
|
≥ 4
|
439(66.5%)
|
331(26.8%)
|
770(40.6%)
|
< 4
|
221(33.5%)
|
905(73.2%)
|
1126(59.4%)
|
Total
|
660(100.0%)
|
1236(100.0%)
|
1896(100.0%)
|