Application of Artificial Neural Network model for Diagnosis of Type 2 Diabetes Mellitus (T2DM) and Determine the Importance Degree of Risk Factors
Authors:
Shiva Borzouei1, Ali Reza Soltanian2,3,*
Affiliation and address: 1. Assistant Professor (M.D.), Department of Internal Medicine, School of Medicine, Hamadan University of Medical Sciences, Hamadan, Iran. Tel +98 (81) 38380208, Postal code 6517838736, Email
[email protected]. 2. Associate Professor (Ph.D.), Department of Biostatistics, School of Public Health, Hamadan University of Medical Sciences, Hamadan, Iran. Tel +98 (81) 38380208, Postal code 6517838736, Email
[email protected]. 3. Associate Professor (Ph.D.), Modeling of Noncommunicable Diseases Research Center, School of Public Health, Hamadan University of Medical Sciences, Hamadan, Iran.
* Corresponding author: Ali Reza Soltanian (PhD), Blvd. Ghaem, Lona Park, Hamadan University of Medical Sciences, Tel +98(81)38380208, Fax +98(81)38380509, Postal code 6517838736, Email
[email protected].
Running Title: Identify the predictors of type 2 diabetes
Application of Artificial Neural Network model for Diagnosis of Type 2 Diabetes Mellitus (T2DM) and Determine the Importance Degree of Risk Factors Running Title: Identify the predictors of type 2 diabetes Abstract Objectives: Identify the most important of demographic risk factors to the diagnosis of T2DM using the neural network model Methods: In this study was conducted on 234 samples. Diagnosis of normal people and diabetics was performed by HbA1c measures. Multilayer perceptron artificial neural network used to identify demographic risk factors on T2DM and their importance rate. DeLong's method was used to compare the models fitting in sequential steps. Results: Using univariate logistic regression risk factors age, hypertension, waist, BMI, sedentary, smoke, vegetables, family history of T2DM, stress, walking, fruit and sex which had a significant level of less than 0.2, entered the model at first. After seven stages of neural network modeling, only age (78.5%), BMI (78.2%), hypertension (69.4%), stress (54.2%), smoke (49.3%) and family history of T2DM (37.2%), respectively, were identified to the diagnosis of T2DM. Conclusions: In this study waist and age were the most important predictors of T2DM. Due to the sensitivity, specificity and accuracy of the final model, it is suggested that they are used as a proper tool for risk assessment of type 2 diabetes in screening tests. Key wordy: Statistical model, Glycated Hemoglobin A, Epidemiology
Introduction Type 2 diabetes mellitus (T2DM), which known as non-insulin dependent, is a noncontagious and chronic disease (1). T2DM can because of many other diseases, such as cardiovascular disease (2), stroke (3), blindness (4) and kidney loss (5). The prevalence of diabetes is increasing. Worldwide, 285 million people have diabetes in 2010, compared to 422 million in 2014 (6), 438 million in 2030 (7) and 592 million in 2035 (8). The prevalence of diabetes in low- or moderate-income countries is higher than in highincome countries (7), and it has a high share in the mortality and disability rate among communities (6). One of the reasons for the high prevalence of diabetes in low-income countries may be due to low knowledge and awareness of people about diabetes disease (9). In 2010 and 2012, the number of undiagnosed diabetics was reported 7 and 1.8 million, respectively, which is approximately a quarter of diabetic patients diagnosed (8) and the cost of their treatment is greater than prevention.
|Page1
Therefore, the prevention of diabetes mellitus is high importance in all communities. The first step in the prevention of T2DM is the identification of its risk factors. Our literature reviews showed that factors such as age (10, 11), gender (10, 12), history of diabetes in the family (11, 13), hypertension (14), obesity (10, 15), abdominal obesity (16), stress in the workplace or home (17, 18), sedentary (19, 20), smoking (21), fruit and vegetables (22) and physical activity (23, 24) are the risk factors associated with type 2 diabetes. Many previous studies have predicted T2DM by types of lipids (e.g., LDL, HDL, FBS and etc.) (1, 8, 25). Because, all of these variables are costly, hence we used variables that do not require much cost to measure them (e.g., sex, age, BMI and etc.). Inadequate healthcare facilities in many countries, especially low-income countries, as well as a complete failure to prevent type 2 diabetes, forced us to identify the importance of demographic risk factors for Type 2 diabetes mellitus. One of the advanced methods for estimating the outcome and prioritizing risk factors is the artificial neural network. Also, there are two medical criteria (i.e., Fast blood sugar and HbA1c) for diagnosing T2DM, but they are not cost-effective in the screening of T2DM in communities. An artificial neural network technique (ANN) is one of the advanced modeling techniques that is based on brain neurons and has been widely used in recent years, and can be helpful in diagnosing, estimating, and predicting diseases (1). Our aim is present a diagnostic model that can predict and determine the importance of risk factors affecting on T2DM using the ANN model.
Methods and Materials Setting and participants In this descriptive-analytical study was conducted on 234 samples, and data were collected from individuals referring to diabetes center in Hamadan city (west of Iran) from 27 November to 15 March 2016. In the Hamadan Diabetes Risk Score study (HDRS), 130 normal and 130 diabetic volunteers were invited that referring to Hamadan diabetes center as a companion patient with aged 18 or more. Of the all invited normal volunteers (n=130), only 106 volunteers refer to the laboratory to measure their hemoglobin A1c (HbA1c), while all diabetic volunteers have compliance. The criteria for choosing normal people were age≥18, no mental disability, people without type 1 and 2 or gestational diabetes and women should not be pregnant, do not currently use metformin or other glucose control drugs”. Inclusion criteria for diabetic volunteers were age≥18, no mental disability, people with type 2 diabetes and without type 1 or gestational diabetes and women should not be pregnant. After obtaining informed consent, people were referred to the lab for testing HbA1c, then the diagnosis of normal people and diabetics was performed based on the results of HBA1C by the endocrinologist. We applied the American Diabetes Association criteria to the HbA1c results with cut-off points of less than 5.8% (
