Diabetes prediction system using ml & dl techniques

       

 

Nandini Gupta¹ , Shubhangi Malik ² , Hardik Chawla ³, Surinder Kaur ^{4, *}

 

¹ Bharati Vidyapeeth’s College of Engineering, GGSIPU, Delhi, INDIA; 

² Bharati Vidyapeeth’s College of Engineering, GGSIPU, Delhi, INDIA; 

³  Bharati Vidyapeeth’s College of Engineering, GGSIPU, Delhi, INDIA; 

⁴  Bharati Vidyapeeth’s College of Engineering, GGSIPU, Delhi, INDIA; 

Emails: guptanandini12345@gmail.com; shubhangimalik28@gmail.com; hardikchawla111@gmail.com; kaur.surinder@bharatividyapeeth.edu

 

 

* Correspondence: kaur.surinder@bharatividyapeeth.edu

 

Abstract

Diabetes nowadays is a familiar and long-term disease. If a prediction is made early better treatment can be provided. The data pre-processing approach is extremely useful in predicting the disease at an early stage. “A number of tools are used in determining significant characteristics such as selection, prediction, and association rule mining for diabetes. The principal component analysis method was used to select significant attributes. Our judgments denote a firm association of diabetes with body mass indicator (BMI) and with glucose degree. The study implemented logistic regression, decision trees, and ANN techniques to process Pima Indian diabetes datasets and predict whether people at risk have diabetes. It was analysed that random forest had the best accuracy of 80.52 %. Out of 500 negative records & 268 positive records our model correctly analysed 403 records & 216 records respectively. 

Keywords: Body Mass Indicator; Artificial Neural Network; Logistic Regression; Random Forest

 

1. Introduction

A sickness or condition which is continuous or whose impact can be seen in the long run is termed a persistent condition or state. “These kinds of diseases affect the quality of life, thus deteriorating it. Diabetes is one of the diseases whose presence today is worldwide”[2]. One of the major reasons for death across the world is the chronic disease of diabetes. “Diseases like these are also cost concerns. A major portion of the budget is spent on chronic diseases by governments and individuals. The worldwide statistics for diabetes within the year 2013 revealed around 382 million individuals had this disease around the world. It was the fifth major reason for death in women and the eighth-most reason for death for both sexes in 2012. It has been noted that developed nations have a high probability of diabetes. In 2017, around 451 million grown-ups were treated with diabetes around the world. It is estimated that in 2045, around 693 million patients with diabetes will exist around the globe and a large portion of the populace will be undiscovered. Likewise, in 2017, 850 million USD was spent on patients with diabetes. Research on biological data is restricted but with the passage of time, computational and statistical models are being used for analysis. A reasonable amount of knowledge is being gathered by healthcare organizations”[18]. “This can be made a reality when new models are developed to find out from the observed data using the data processing techniques. Data mining is the process of drawing out data and can also be utilized to create the choice-making process efficiently in the medical domain”[2]. A number of information handling methods are used in disease prediction from biomedical information. “Diagnosis of diabetes is itself a challenge for quantitative research. A few boundaries like A1c, fructosamine, white blood corpuscle count, fibrinogen, and haematological indices were displayed to be insufficient because of certain limitations. Diverse examinations tried to involve these boundaries for the determination of diabetes. Some of the treatments have been considered to boost A1c including chronic ingestion of liquor, salicylates, and narcotics. Ingestion of vitamin C might raise A1c when assessed by electrophoresis but levels might seem to lessen when it is assessed by chromatography. Most studies have suggested a better white blood corpuscle count, thanks to chronic inflammation during hypertension. A case history of diabetes has not been related to BMI and insulin. However, an increased BMI isn't always related to abdominal obesity”[5]. Only one boundary isn't powerful enough to precisely analyse diabetes and should be deceiving inside the dynamic interaction. “Thus different parameters are to be mixed to efficiently predict diabetes at an early stage. A few existing strategies have not given powerful outcomes when various boundaries were utilized for the Prediction of diabetes. In our review, diabetes is anticipated with the help of genuine traits, and in this way the relationship of the contrasting credits. We examined the diagnosis of diabetes.

 

 

1.1 Diabetes categories

“Diabetes is a plague disease that happens whose major reason is a decrease of insulin within the body. Different types of diabetes are distinguished at diagnosis; so determining the sort of diabetes depends on the conditions in which the disease happens. The old division was of two sorts of diabetes, that is, insulin-reliant and non-insulin reliant. The new grouping of diabetes was developed by the America Diabetes

  Association: Type I diabetes, type II, gestational diabetes, and”[13] different sorts.

 

1.1.1. Type I diabetes

“Type I diabetes (insulin-dependent diabetes mellitus) may be a chronic disease that occurs when the pancreas releases a small amount of insulin (a hormone that's required for importing sugar). A few elements, incorporating hereditary qualities and disease with certain infections can cause type I diabetes. Although type 1 diabetes usually occurs in childhood and adolescence, adults also are vulnerable to this disease”[17].

 

1.1.2. Type 2 diabetes

“Type 2 diabetes (adult diabetes or Non-insulin-dependent diabetes), is one of the common sorts of diabetes. It constitutes around 90 percent of the patients. Unlike type 1 diabetes, the body produces insulin in type 2 diabetes, but the insulin produced by the pancreas isn't enough or the body is not able to use insulin properly. When there's not enough insulin or the body does not use insulin, glucose (sugar) within the body, cannot move to the body's cells and causes an accumulation of glucose within the body therefore the body would be in trouble and deficiencies. Unfortunately, there's no cure for this disease, but a healthy diet, exercise, and keeping fit can enhance it. If diet and exercise aren't enough, you would like medication or insulin treatment. Figure 1 is an analysis of diabetes”[10].

 

Fig 1. Analysis of Diabetes

 

2. Related Work

 

Shetty et al. used KNN and “the Naïve Bayes technique for the prediction of diabetes. Their technique was implemented as a software program, where users provide input in terms of patient records and find whether the patient is diabetic”[21] or not. 

 

“Singh et al. applied different algorithms to datasets of different types. They used KNN, random forest, and Naïve Bayesian ML algorithms. The K-fold cross-validation technique was then used for evaluation. Ahmed utilized the patient's information and plan of treatment for the classification of diabetes. Three algorithms applied were Naïve Bayes, logistic, and J48 algorithms”[18].

 

“Antony et al. utilized medical data for the prediction of diabetes. Naïve Bayes, function-based multilayer perceptron (MLP), and decision tree-based random forest (RF) algorithms were applied after pre-processing of the data. A correlation-based feature selection method was employed to remove the extra features. A learning model then predicted whether the patient is diabetic or not. By using pre-processing techniques, results were improved when applying Naïve Bayes as compared to other”[11] machine learning algorithms. 

 

“Amina et al. compared different data mining techniques by using the PID dataset for the early prediction of diabetes. Sellappan Palaniappan et al. proposed a heart disease prediction system by using various algorithms like Naïve Bayes, ANN”[12], and decision trees.

 

“Shadab Adam Pattekari and Asma Parveen [14] developed a web-based application for the prediction of myocardial infarction using Naive Bayes. Anuja Kumari and R. Chitra used”[15] an SVM model to diagnose diabetes using a high-dimensional medical dataset.

 

Md. Kamrul Hasan et al. used a proposed ensemble model on the PIDD dataset to predict diabetes. It was concluded that highest accuracy was achieved using the combination of boosting type classifiers (AdaBoost and XGBoost) when the proposed preprocessing (i.e. outlier rejection + filling missing values) is applied.

 

Aishwarya Mujumdar et al. implemented a diabetes prediction system using various ML algorithms like Gradient Boost, AdaBoost, Logistic Regression, KNN, GaussianNB, Perceptron, LDA, SVC. Out of all these algorithms, application of pipeline gave AdaBoost classifier as the best model with maximum accuracy.

 

Quan Zou et al. predicted diabetes mellitus using Random forest, Decision tree, Neural Network and Decision tree algorithms. They used two datasets namely Luzhou dataset and Pima Indians Dataset. It was concluded that Random forest gave the maximum accuracy and Pima Indians dataset gave the best performance. ML algorithms can be used to predict diabetes efficiently provided that suitable attributes, classifiers & data mining methods are found properly.

 

Safial Islam Ayon et al. implemented a diabetes prediction system using deep neural networks based on several medical predictor variables. Highest accuracy was achieved for five-fold cross-validation.

 

Bala Manoj Kumar et al. implemented diabetes prediction using Deep Neural Networks classifier. For feature attribute selection the proposed model made use of feature importance. The best accuracy was achieved using DNN-FI compared with random forest & decision tree algorithms. 

 

 

3. Procedure and Entities

 

3.1.  Dataset

PIDD (Pima Indians Diabetes Dataset)

The database is known as PIDD and is extracted from the National Institute of Diabetes and Digestive and Order Conditions, a. It aims to prognosticate if a case suffers from diabetes-mellitus or not, valid for specific individual criteria of which the database consists. For this dataset, all cases of women are at least 21 years aged.

Pima Indian Diabetes (PID) data set has the following characteristics 9 = 8 1 (Class Identifier), 768 records describing womanish cases (in the case of 500 adverse events), and 268 positive cases.  A detailed description of the features is given in Table 1.

 

Table 1. Dataset description and characteristics.

Sr.	Attribute Name	Attribute Description	Mean ± S.D
1	Pregnancies	Number of times a woman got pregnant	3.8 ± 3.3
2	Glucose(mg/dl)	Glucose concentration in oral glucose tolerance test for 120 min	120.8 ± 31.9
3	Blood Pressure(mmHg)	Diastolic Blood Pressure	69.1 ± 19.3
4	Skin Thickness (mm)	Fold Thickness of Skin	20.5 ± 15.9
5	Insulin (mu U/mL)	Serum Insulin for 2 h	79.7 ± 115.2
6	BMI (kg/m2)	Body Mass Index (weight/(height)^2)	31.9 ± 7.8
7	Diabetes Pedigree Function	Diabetes pedigree Function	0.4 ± 0.3
8	Age	Age (years)	33.2 ± 11.7
9	Outcome	Class variable (class value 1 for positive 0 for Negative for diabetes)

 

Many obstacles have been placed in the selection of these conditions on a large database.

 

 

3.2 Data Preparation

Its procedure is carried out in the following way:

 

3.2.1 Data Exploration

The process is begun by locating connections and correlations between the two factors (and the difference in the result) and visualizing the connections using a heatmap (see Table 2). 

 

Table 2. Product of feature (and outcome) relationship/correlations

	Pregnancies	Glucose	BloodPressure	SkinThickness	Insulin	BMI	DPF	Age	Outcome
count	768.000000	768.000000	768.000000	768.000000	768.000000	768.000000	768.000000	768.000000	768.000000
Mean	3.845052	120.894531	69.105469	20.536458	79.799479	31.992578	0.471876	33.240885	0.348958
Std	3.369578	31.972618	19.355807	15.952218	115.244002	7.884160	0.331329	11.760232	0.476951
Min	0.000000	0.000000	0.000000	0.000000	0.000000	0.000000	0.078000	21.000000	0.0000000
25%	1.000000	99.000000	62.000000	0.000000	0.000000	27.300000	0.243750	24.000000	0.000000
50%	3.000000	117.000000	72.000000	23.000000	23.000000	32.000000	0.372500	29.000000	0.000000
75%	6.000000	140.250000	80.000000	32.000000	32.000000	36.600000	0.626250	41.000000	1.000000
max	17.000000	199.000000	122.000000	99.000000	99.000000	67.100000	2.420000	81.000000	1.000000

 

 

In the below heatmap, the bright colours show some correlations. An important correlation can be seen in the table along with the heatmap of glucose levels, age, BMI, and gestation rate with the outgrowth variability. Also, a relationship between dyads of factors, like age and gestation, or insulin and skin firmness as shown in Fig 2.

 

     Fig 2. Heatmap of feature (and effect) correlation

 

3.2.2 Data Preprocessing

 

Sometimes, data in real life can be noisy or inconsistent, it might also contain missing values. When such degraded quality data is used, the quality of results also degrades. Thus, data reprocessing becomes necessary to gain results of good quality. Drawing, incorporating, modifying, reducing, and separating data is used in pre-data processing. Thus, it’s significant to make the data correlated to mining in terms of efficiency of time taken cost of production, and standard of data.

 

3.2.3 Data Cleaning

 

Sanctification involves fulfilling missing quantities and reducing unwanted data. Data should contain excerpts to resolve inconsistencies. For this database, glucose, Blood Pressure, Skin Consistency, Insulin, and BMI have some zero or null values (0). Therefore, every null attribute’s value is replaced by the average value of that trait to remove inconsistencies. Fig 3 and Fig 4 show the outlier in the PIDD dataset and outlier junking respectively.

 

 

Fig 3.  Outliers in the dataset 

 

                                                                           Fig 4. Outlier Junking

 

 

3.2.4. Data Reduction

Data reduction reduces the representation of datasets with much less volume, but still gives the same (or nearly the same) results. Dimensionality and size are reduced to decrease the number of attributes in a database. The crucial element figuring system used, the prize’s the crucial values from the entire database. Glucose, BMI, diastolic blood pressure, and age were known to be the most important factors in the dataset after visualization.

 

3.2.5. Data Metamorphosis

Data revision includes smoothness, familiarity, and integration of data [18]. To smooth the data, a combination system was used. The age factor has helped divide the five orders, as shown in Table 3.

 

Table 3. Binning of age.

 

Age (years)	Age Bins
≤ 30	Youngest
31-40	Younger
41-50	Middle aged
51-60	Older
≥ 60	Oldest

 

Blood glucose uptake in non-diabetic cases differs from diabetic cases. Glucose values are divided into 5 orders (19) as shown in Table 4.

 

 

Table 4. Binning of glucose.

 

Glucose	Glucose Bins
≤ 60	Very Low
61-80	Low
81-140	Normal
141-180	Early Diabetes
≥ 181	Diabetes

 

 

A strong correlation was planted between non-diabetic and diabetic cases regarding their pressure situations. Blood pressure is categorized into five distinct orders as shown in Table 5.

 

 

Table 5. Types of diastolic blood pressure.

 

Blood Pressure	Diastolic Blood Pressure Bins
≤ 61	Very Low
61-75	Low
75-90	Normal
91-100	High
≥ 100	Hypertension

 

A relationship between the body mass index and diabetes is found. The original study concludes that BMI is the most dangerous factor in determining type 2 diabetes. BMI values ​​are divided into 5 classes as shown in the Table 6.

 

 

Table 6. Binning of BMI.

 

BMI	BMI Bins
≤ 19	Starvation
19-24	Normal
25-30	Overweight
31-40	Obese
≥ 40	Very Obese

 

3.2.6. Dataset Splitting and Normalization

This process is begun by unyoking the data into one training set and one test set. The database contains records of 767 cases in aggregate. To train the model only 614 (80%) records will be used, the remaining records will be used to test and estimate the model.

 

4. Experimental Result

Sample 1

Table7. Patient Data

	pregnancies	glucose	bp	skinthickness	insulin	bmi	dpf	age
0	3	120	70	20	79	20	0.4700	33

 

Visualised Patient Report

Fig 5. Pregnancy Count Graph (Others vs Yours)                   Fig 6. Glucose Value Graph (Others vs Yours)

                  Fig 7. Blood Pressure Graph (Others vs Yours)                     Fig 8. Skin Thickness Value Graph (Others vs Yours)

Fig 9. Insulin Value Graph (Others vs Yours)

Fig 10. BMI Value Graph (Others vs Yours)