Diabetic retinopathy screening using artificial intelligence and its predictors among people with type 2 diabetes mellitus in an urban area of Durgapur

ABSTRACT

Introduction:

Diabetes mellitus (DM) is a metabolic disorder characterized by chronic hyperglycaemia either due to insulin resistance or due to relative or absolute insulin deficiency. Poorly controlled DM may result in both macrovascular and/or microvascular complications like diabetic retinopathy [DR]. Dilated eye examination is the most commonly employed method to diagnose DR. Nonmydriatic artificial intelligence [AI]–based technologies are the now available to screen DR.

Methods:

A cross-sectional observational study was conducted in urban field practice area of our medical college for 2 months duration. A total of 95 patients with type 2 DM were interviewed using predesigned, pretested semistructured schedule to collect data. Medical records were reviewed to collect relevant information. DR was screened using AI-based DR screening instrument, and venous blood sample was collected for glycated hemoglobin (HbA1C) testing. Data were analyzed using IBM SPSS [version 16]. Univariate and multivariate logistic regression tests were used, and P value ≤ 0.05 was taken as statistically significant.

Results:

The prevalence of DR was 17.9% in our study. Around 76.9% respondents had high fasting blood glucose [FBG: ≥126 mg/dl], and majority of the respondents [73.7%] had HbA1C value >7%. DR was significantly associated with FBG level, longer duration of diabetes, presence of hypertension, dyslipidemia, and kidney disease in univariate logistic regression and, in multivariable logistic regression, FBG level, presence of dyslipidemia and kidney disease retained their significance.

Conclusion:

This study had used AI-based DR screening instrument, to screen DR among T2DM patients. AI-based DR screening system can be encouraged in mass screening camps, especially in areas with inadequate number of ophthalmologists. This study also evaluated some important modifiable predictors of DR. Appropriate and early identification of such predictors may prevent DR-related blindness.

Introduction

Diabetes mellitus (DM) is a metabolic disorder characterized by chronic hyperglycaemia either due to insulin resistance and/or absolute or relative insulin deficiency.[1] As per International Diabetes Federation (IDF) 10^th Atlas, globally 537 million people are living with DM, and if corrective steps are not taken, it is projected to reach 643 million by 2030.[2] India is home of about 77 million diabetes patients and >90% of them are type 2 diabetes mellitus patients (T2DM).[2]

Poorly controlled DM may result in development of both macrovascular complications like ischemic heart disease, peripheral arterial disease, etc. and/or microvascular complications like diabetic retinopathy (DR), nephropathy, neuropathy, etc.[3] Almost two-thirds of people with diabetes (PwD) worldwide as well as in India are expected to develop some form of retinopathy with time.[4,5]

Gadkari et al. reported 21.7% prevalence of DR among Indian PwD.[6] DR prevalence in West Bengal was found to be 21.51% with significantly higher prevalence in rural (26.55%) areas than urban (13.89%).[7]

DR is one of the most important preventable causes of blindness among working age population.[8] Chronic retinal diseases are asymptomatic in early stages, allowing it to remain undetected until advanced stage.[8] Chronic poor glycemic control, longer duration of diabetes and associated high blood pressure are risk factors for DR.[9] Vision loss due to DR may be averted or delayed by early intervention and good glycemic control.[9] As per the American Diabetes Association (ADA) standard of care all people with T2DM must undergo retinal examination at the time of diagnosis then yearly if retina is normal.[10]

The conventional method of dilated eye examination for DR screening has many drawbacks like requirement of trained ophthalmologist, time consuming, requirement of patient’s attendant. Patient compliance is poor because of poor vision even after 2–3 hours of evaluation due to mydriatic effect.[11] To improve compliance of PwD for regular screening of DR, nonmydriatic fundus examination was used, but the availability of trained ophthalmologist remained a challenge especially in rural areas.[12] Nowadays ‘artificial intelligence (AI)’–based technologies are being used in diagnostic medicine.[13] REVELO-AI is one such commercially available, AI-based triage and notification system for fundal imaging. REVELO-AI has 90.0% sensitivity and 88.0% specificity for retinopathy.[14] Burden of DR will increase in future owing to increased number of PwD which will further need trained manpower and other resources. Providing such number of resources specially trained manpower will be difficult for resource limited countries like India. Therefore, it is important to explore the feasibility and utility of AI-based technologies in the screening and diagnosis of DR. This study aims to use REVELO-AI software for the DR screening among T2DM patients residing in the catchment area of urban health training center (UHTC) of a medical college of eastern India.

Objectives

1. To estimate the prevalence of DR among people with type-2 DM (PwD) using nonmydratic AI-based retinal scan in urban field practice area of a medical college of Eastern India.

2. To find out the predictors of DR among study participants.

Methodology

A community-based, observational, cross-sectional study was conducted from August to September 2022 among people with T2DM who were residents of the catchment area of urban health training center of a medical college of Eastern India.

Sample size and sampling technique

A sample size of 95 was calculated as per World Health Organization (WHO) guidelines[15] using formula n = {z² × p x (1-p)}/L² for cross-sectional study design. DR prevalence of 13.89% among people with T2DM was used to calculate the sample size.[7] Considering, standard normal deviate (z) = 1.96 at 95% confidence interval, prevalence (p)= 0.1389, 1 – P = 0.8611 and absolute error (L) = 7.5% minimum sample size was 82 and a nonresponse rate of 15% of the minimum sample size was added to reach final sample size of 95.

Systematic random sampling was used for the selection of study participants. Population of UHTC catchment area is about 35000 as per local municipality records.

Information regarding health camps was circulated by field workers a few weeks before the data collection in the urban field practice area of our medical college. Health-checkup camps for PwD were organized over a period of 7 days. A total of 473 diabetic patients were registered for the camps. A sampling frame was prepared with the 473 T2DM patients. It was ensured that registration occurred only once, and informed consent was taken. After finalizing the sampling frame, a sampling interval of 5 [473/95 = 4.9 ~ 5] was calculated and systematic random sampling was done. Those who were selected by systematic random sampling but did not fulfill the selection criteria were excluded from the study and the next patient from the sampling frame was included accordingly. All the PwD attending the camp received general health checkup and medical advice was given.

Inclusion and exclusion criteria

People with diagnosed T2DM of at least 6 months duration, residing in UHTC catchment area and willing to give consent were included in our study. Seriously ill, gestational diabetes mellitus (GDM) and patients on steroid and/or hormonal therapy were excluded from our study.

Study tools and technique

1. Predesigned, pretested, semistructured schedule.

2. Digital Weight Machine.

3. Nonstretchable measuring tape.

4. Fluoride and Ethylene-Diamine Tetra acetic Acid (EDTA) vial.

5. Medical records.

6. Manual blood pressure instrument and stethoscope.

7. AI-based retinopathy screening instrument: Noninvasive AI-based instrument—AI-based Radial Eye Triage and Notification [RETN] was used (14). For screening of DR, instrumental and technical support was provided by a diabetes clinic in the urban field practice area of our medical college, which is currently using nonmydriatic retinal scan using REVELO.

Study technique

Ethical clearance to conduct the study was obtained from the Institutional Ethics Committee of our Medical College [Memo No-IQMC/IEC-10/LTR/10 (02)/22].

Predesigned pretested structured schedule was used to interview study participants. After getting written informed consent data regarding clinic–social characteristics like age, sex, height, weight, etc., were collected. Medical records were reviewed to get additional clinical information. Venipuncture at ante-cubital vein was done maintaining aseptic condition. Venous blood sample was collected in “fluoride vial” and “ethylenediamine tetra acetic acid (EDTA) vial.” Fasting plasma glucose and glycated albumin (HbA1C) were estimated as per the World Health Organization (WHO) guidelines.[16] Anthropometric measurements were taken and classified as per WHO classification.[17] Type 2 diabetes mellitus and glycemic control were diagnosed as per American Diabetes Association (ADA) guidelines.[18] The retinal screening was done using an AI-based instrument using REVELO technology.

Operational definitions

1. Diabetic retinopathy (DR): Five stages of DR a) no apparent DR, b) mild nonproliferative retinopathy (NPDR), c) Moderate NPDR, d) Severe NPDR and e) proliferative DR were classified as per the “International Clinical Disease Severity Scale for classification of Diabetic-Retinopathy.”[19]

2. Diabetes mellitus: Fasting plasma glucose ≥ 126 mg/dl, post prandial plasma glucose ≥ 200 mg/dl and HbA1C ≥ 5.7%.[18]

3. Good glycemic control: HbA1C < 7.0%.[18]

4. Body Mass Index (BMI)[17]: BMI was calculated using the formula, weight (in kg)/height² (in meter) and classified as a) underweight: BMI <18.5, b) normal: 18.5–24.99, c) overweight: BMI 25.00–29.99, and d) obese: BMI ≥30.[17]

5. Hypertension: Classification of hypertension was done as per Eighth Joint National Committee (JNC-8) guidelines[20] with normal blood pressure level as <120 (and) <80 mm Hg.[20]

6. Diagnosed case of dyslipidemia and chronic kidney disease were taken after reviewing the records.

Statistical analysis

The data were entered in MS Excel and analyzed in IBM SPSS [Version 16]. The variables were categorized as applicable and presented with number and percentage. The variables were not normally distributed in the study, so the univariate binary logistic regression and multivariable logistic regression were done to find out the predictors of DR. The P value ≤ 0.05 was taken as significant difference.

Results

54.7% of study population were in the age group of 41–59 years, and 23.2% were ≥60 years old [Table 1]. About 3/5^th of the study population (61.1%) was male. 42.1% had education up to primary level followed by 22.1%, 21.1% and 14.7% had education up to secondary level, graduation and postgraduation, respectively [Table 1]. Duration of T2DM was <6 years among 44.2% of study participants followed by 33.7% and 22.1% who had T2DM for 6–10 years and ≥11 years, respectively [Table 1]. 77.9% of study participants had history of substance abuse. 47.4% of study participants were either overweight or obese and about 4/5^th (77.9%) had sedentary lifestyle. 66.3% had no dietary compliance and 73.7% had poor glycemic control, i.e., HbA1C > 7.0% [Table 1].

Table 1.

Clinico-Social Characteristics of Study Population, [n=95]

Clinico-Social characteristics	n (%)
Age group
≤40 years	21 (22.1)
41-59 years	52 (54.7)
≥60 years	22 (23.2)
Sex
Male	58 (61.1)
Female	37 (38.9)
Educational status
Primary	40 (42.1)
Secondary	21 (22.1)
Graduate	20 (21.1)
Post Graduate	14 (14.7)
Duration of Diabetes
<6 years	42 (44.2)
6-10 years	32 (33.7)
≥11 years	21 (22.1)
History of Substance abuse
Present	21 (22.1)
Absent	74 (77.9)
BMI (Kg/m2)
Underweight (<18.50)	9 (9.5)
Normal (18.5-24.99)	41 (43.2)
Overweight/Obesity (≥25.0)	45 (47.4)
Physical Activity
Sedentary	74 (77.9)
Moderate	9 (9.5)
Vigorous	12 (12.6)
Dietary Compliance
Present	32 (33.7)
Absent	63 (66.3)
Fasting Plasma Glucose
<126 mg/dl	22 (23.1)
≥126 mg/dl	73 (76.9)
HbA1C
≤7.0%	25 (26.3)
>7.0%	70 (73.7)

17.9% of the study participants had diabetic retinopathy, 15.8% respondents had moderate NPDR and 2.1% had mild NPDR [Table 2]. Among comorbidities, 32.6% study participants had diabetic peripheral neuropathy (DPN), 46.3% had hypertension, 32.6% had dyslipidaemia, 11.6% had chronic kidney disease and 25.3% had diabetic foot [Table 3].

Table 2.

Distribution of the study participants according to the prevalence of diabetes retinopathy [n=95]

Diabetes retinopathy	Number (%)
No DR	78 (82.1)
Mild NPDR	2 (2.1)
Moderate NPDR	15 (15.8)
Total	95 (100)

Table 3.

Pattern of Co-Morbidities among Study Population, [n=95]

Co-Morbidities	n (%)
Diabetic Peripheral Neuropathy (DPN)
Present	31 (32.6)
Absent	64 (67.4)
Hypertension
Present	44 (46.3)
Absent	51 (53.7)
Dyslipidemia
Present	31 (32.6)
Absent	64 (67.4)
Chronic Kidney Disease
Present	11 (11.6)
Absent	84 (88.4)
Diabetic Foot
Present	24 (25.3)
Absent	71 (74.7)

The median duration of T2DM was higher among study participants who had DR. One year increase in the duration of T2DM had 1.12 times significantly (P value: 0.012) higher odds of having DR [Table 4]. On univariate logistic regression, significant (P value: 0.033) 3.5 times higher odds of having DR was found among study participants with hypertension [Table 4]. Significantly high prevalence of DR was associated with the presence of comorbidities viz dyslipidaemia, chronic kidney disease and high fasting plasma glucose level [Table 4]. However, no significant association was found between DR and other variables like age, gender, educational status, BMI, and physical activity level [Table 4]. The presence of dyslipidemia [AOR: 5.1 (1.2, 22), P value-0.03], kidney disease [AOR: 20 (3,139), P value-0.002] and increased FBG level [AOR: 1.01 (1,1.02), P value-0.012] retained their significance in the multivariable logistic. However, the duration of diabetes [P value-0.051] and presence of hypertension [P value-0.2] lost their significance in the multivariable logistic regression model [Table 5]. The model fitting was good as the Hosmer and Lemeshow test P value was 0.463, and the Omnibus test P value was <0.001. The model could explain 31.3% to 51.3% variability of dependent variable as Cox and Snell R Square was 0.313 and Nagelkerke R Square was 0.513 [Table 5].

Table 4.

Univariate logistic regression showing association of diabetic retinopathy [DR] with different variables [n=95]

	Variables	DR Absent [n=78] Number (%)	DR Present [n=17] Number (%)	OR [CI]	P
Age [Years]	Median (IQR)	52 (43, 60.3)	55 (44.5, 58.5)	1.006 (0.96, 1.055)	0.804
Gender	Female	29 (78.4)	8 (21.6)	1.502 (0.52, 4.3)	0.451
	Male	49 (84.5)	9 (15.5)	1
Education	Primary, Secondary and Graduate	66 (81.5)	15 (18.5)	1.4 (0.3, 6.7)	0.704
	Post graduate	12 (85.7)	2 (14.3)	1
BMI [kg/m2]	Under weight (<18.5)	6 (66.7)	3 (33.3)	1
	Normal (18.5-24.9)	36 (87.8)	5 (12.2)	0.278 (0.05, 1.5)	0.133
	Overweight and obesity	36 (80)	9 (20)	0.5 (0.1, 2.4)	0.4
Duration of diabetes	Median (IQR) [Years]	5.5 (3.8, 10)	10 (6.5, 13.5)	1.12 (1.03, 1.22)	0.012
Hypertension	Present	32 (72.7)	12 (27.3)	3.5 (1.1, 10.8)	0.033
	Absent	46 (90.2)	5 (9.8)	1
Dietary compliance	Present	29 (90.6)	3 (9.4)	1
	Absent	49 (77.8)	14 (22.2)	2.8 (0.7, 10.4)	0.134
Physical activity	Sedentary	60 (81.1)	14 (18.9)	2.6 (0.3, 21.6)	0.4
	Moderate	7 (77.8)	2 (22.2)	3.1 (0.2, 41.5)	0.4
	Vigorous	11 (91.7)	1 (8.3)	1
Dyslipidemia	Present	21 (67.7)	10 (32.2)	3.9 (1.3, 11.5)	0.015
	Absent	57 (89.1)	7 (10.9)	1
Kidney disease	Present	4 (36.4)	7 (63.3)	12.9 (3.2, 52.2)	0.001
	Absent	74 (88.1)	10 (11.9)	1
Hypothyroidism	Present	13 (72.2)	5 (27.8)	2.1 (0.6, 6.9)	0.23
	Absent	65 (84.4)	12 (15.6)	1
Neuropathy	Present	24 (77.4)	7 (22.6)	1.6 (0.5, 4.6)	0.41
	Absent	54 (84.4)	10 (15.6)	1
Diabetic foot	Present	19 (79.2)	5 (20.8)	1.3 (0.4, 4.1)	0.67
	Absent	59 (83.1)	12 (16.9)	1
Substance abuse	Present	20 (95.2)	1 (4.8)	5.5 (0.7, 44.3)	0.11
	Absent	58 (78.4)	16 (21.6)	1
FBG [mg/dl]	Median (IQR)	170 (122, 217)	189 (143, 331)	1.01 (1.0, 1.02)	0.005
HBA1C	Median (IQR)	8.4 (6.8, 10.6)	10.3 (7.8, 12.2)	1.2 (0.99, 1.5)	0.054

Table 5.

Multivariable logistic regression showing association of diabetic retinopathy [DR] with different variables [n=95]#

Variable	OR (CI)	P	AOR (CI)	P
Duration of diabetes [Years]	1.12 (1.03, 1.22)	0.012	1.1 (0.9, 1.3)	0.051
Presence of hypertension	3.5 (1.1, 10.8)	0.033	2.7 (0.7, 12)	0.2
Presence of dyslipidemia	3.9 (1.3, 11.5)	0.015	5.1 (1.2, 22)	0.03
Presence of kidney disease	12.9 (3.2, 52.2)	0.001	20 (3, 139)	0.002
FBG [mg/dl]	1.01 (1.0, 1.02)	0.005	1.01 (1, 1.02)	0.012

[#Hosmer and Lemeshow Test P=0.463, Omnibus Test P<0.001, Cox & Snell R²=0.313, Nagelkerke R²=0.513.]

Discussion

This study was conducted to find out the prevalence and predictors of diabetic retinopathy among patients with T2DM. The prevalence of DR was 17.9% among the respondents, and the majority of the DR cases (15.8%) were moderate NPDR in our study. The prevalence of DR was different among different parts of the world, among patients of different age groups, among patients with different co -morbidities due to variable risk factors.[21,22,23] But almost similar i.e. 17.6% prevalence of DR was reported among urban south Indian population by Rema et al.[24] A study from West Bengal[6] reported 13.89% prevalence of DR among in urban area, which is slightly lower than our findings; however, another study by Gadkari et al.[7] reported a higher 21.7% prevalence of DR among Indian T2DM patients. This difference might be due to different study setting and different screening tool used for the detection of DR.

54.7% of study population were in the age group of 41–59 years, and 23.2% were ≥60 years old but age was not found to be significantly associated with DR. Nonsignificant association of age with DR was also reported by Lima et al.[25] However, the literature suggests increased risk of diabetes with increasing age, and there are studies which show significant association of DR with increasing age.[26] This difference in finding may be because of the younger population of the study population in our study as more than 50.0% of study participants were <60 years of age. In present study, there was a nonsignificant female preponderance of the DR similar with findings of Lima et al.[25] In disagreement with our findings, few studies[27,28,29] have reported significant male preponderance of DR and a Kajiwara et al.[30] reported a significant female preponderance.

In our study nonsignificant underweight and overweight, preponderance of DR was found but the literature shows conflicting role of BMI on DR. Many studies[31,32,33,34] have reported no association between BMI and DR, and few studies[30,35,36] have reported positive association between high BMI and DR.

Duration of diabetes is considered to be one of the most important risk factor[30,31,32] for DR, but in our study, the duration of diabetes was significant in univariate logistic regression, and the significance was lost after multivariate logistic regression. This may be due to more younger study participant with less duration of T2DM in our study.

In present study, the significance of high blood pressure on DR was lost in multivariate analysis. Many studies[37,38] have reported significant higher prevalence of DR among hypertensive T2DM patients, and few studies[31,39] are in agreement with our findings reported nonsignificant association between hypertension and DR. Nonsignificant association between hypertension and DR in our study may be due to the fact that isolated blood pressure (BP) measurement may not express the long-term effect of high BP.[39] Significant positive association between dyslipidemia and DR in our study was in agreement with many other studies[40,41,42] who reported dyslipidemia as an independent risk factor for DR.

In present study, the presence of kidney disease and high fasting plasma glucose level was found to be significant independent risk factor for DR. Chronic hyperglycemia leads to progressive changes in retinal microvasculature like increased vascular permeability, decreased perfusion, intra-ocular proliferation of retinal vessels etc., and all these could lead to DR.[8] Similar results were reported by many researchers across the globe.[43,44,45]

Strengths and limitations

Strengths of our study includes its community-based study setting, well-defined sampling technique, robust statistical analysis using univariate and multivariate analysis. and use of new technology for the screening of diabetic retinopathy.

Limitations of our study include inability to establish causality due to cross-sectional study design, use of nonmydriatic screening tool which might have been resulted in some bias, failure to compare AI-based screening tool with dilated fundus examination and nongeneralization of study results among different types of diabetes mellitus patients due to inclusion of only T2DM patients.

Conclusion

The present study successfully used AI-based screening program, REVELO-AI to screen diabetic retinopathy among patients with T2DM. Dyslipidemia, chronic kidney disease and high fasting plasma glucose are found to be significant modifiable risk factor for DR among T2DM patients. Loss of vision due to DR is preventable if corrective steps are taken at right time. Early diagnosis and timely intervention may halt the development and/or progression of DR among T2DM patients. The use of nonmydriatic AI-based program like REVELO-AI for the screening of DR may help in early diagnosis by eliminating the need of trained man power. The use of such automated screening systems can be encouraged in mass screening camps, especially in countries with inadequate number of ophthalmologists.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.