Friday, June 9, 2023

Application of supervised machine learning algorithms for classification and prediction of type-2 diabetes disease status in Afar regional state, Northeastern Ethiopia 2021 – Scientific Reports

Study area and period

This study was conducted at public hospitals of Afar regional state Northeastern Ethiopia from February to June of 2021. Afar regional state is one of the nine federal states of Ethiopia located in the Northeastern part of the country 588 kms from Addis Ababa. The altitude of the region ranges from 1500 m above mean sea level (m.a.s.l) in the western highlands to − 120 m.a.s.l in the Danakil/Dallol depression. The total geographical area of the region is about 270,000 km2 and is geographically located between 39o34′ and 42o28′ East Longitude and 8o49′ and 14o30′ North Latitude (CSA, 2008). It has an estimated population of 1.2 million, of which, 90% are pastoralists (56% male and 44% female) and 10% are agro-pastoralists.

Study design

A retrospective cross-sectional study design using medical database and medical chart record review was used.

Target population

All hospital clients who ever diagnosed or will be diagnosed and/or suspected for type-2 diabetes in public hospitals of Afar regional state.

Study population

All clients who ever diagnosed for diabetes disease status, and confirmed as free from type-2 diabetes (normal) and type-2 diabetic patient in public hospitals of Afar region starting from the year when the database for electronic health information record was fully functional up to the date of sample collection (2012 GC–April 22/2020) were considered as part of the study population.

Eligibility criteria

Inclusion criteria

All clients who ever diagnosed for diabetes disease status and confirmed as free from type-2 diabetes (normal) and type-2 diabetic patients in public hospitals of Afar region starting from 2012 GC to April 22/2020 GC.

Exclusion criteria

Patients who can unable to obtain required information due to incomplete or total absence of their record status; which cannot be found in their registration book, diabetic patient under follow-up with unknown start date and diabetic patient referred from other hospitals were excluded from the study population to avoid misleading of the machine learning algorithms.

Sample size determination, sampling method and procedure

All patients who have been diagnosed for diabetes and confirmed as a type-2 diabetic patient and normal after standard diagnostic activities from 2012 GC up to April 22/2020 GC were used as a sample. The whole dataset used as a sample because data mining needs considerable amount of data for effective prediction and classification by reducing the probability of error to be committed9. Based on this, the study has been conducted on a total of 2239 population.

From this record, the clients who ever diagnosed for diabetes disease status were extracted and collected with their required variables and some variables which are not available in the database and the parameters for the normal clients were searched from the medical record book by its medical registration number (MRN) because in DHIS database of public hospitals there is no recording site for normal individuals after every diagnosis.

Data collection tools, techniques and procedures

Data related to that have been confirmed as a type-2 diabetic patient and normal at public hospitals of Afar regional state from the date of database being fully functional (2012) up to data obtaining date (April 22/2020) GC were collected from DHIS database public hospitals. Clients who have been positive for type-2 diabetes were obtained from the database and those who were normal were collected from the medical registration book by medical chart review and used for comparative purposes. The parameters of these samples were collected from their first date of their diagnosis by cross-checking with their start date of their diagnosis to avoid the misleading of machine learning algorithms. From this, the essential variables were collected, therefore, in this study was used to classify and predict type-2 diabete disease status among clients of public hospitals for all ages and both sexes who were diagnosed for diabetes in the region.

Study variable

Dependent variable

The dependent variable is type-2 diabetes disease status with dichotomous response to the question “tested and confirmed as type-2 diabetic patients” if yes = 1″ and “if no = 0”.

Independent variables

The above dependent variable was then being modeled to historical predictor variables and reasons that were selected based on existing evidences. The independent variables that have been assessed in this research were:-

  • Diastolic Blood Pressure (DBP)

  • Systolic Blood Pressure (SBP)

  • Fasting Blood Sugar Level (FBS)

  • Random Blood Sugar Level(RBS)

  • Body Mass Index (BMI)

  • Age

  • Sex

Operational definitions

Accuracy Accuracy of classifier refers to the ability of classifier to predict the class label correctly, and it also refers to how well a given predictor can guess the value of the predicted attribute for a new data10.

Classification Is a data mining function that assigns items in a collection to target categories or classes. The goal of classification is to accurately predict (for categorical independent variables) the target class for each case in the data11.

Confusion Matrix Is a simple performance analysis tool typically used in supervised machine learning. It is used to represent the test result of a prediction model. Each column of the matrix represents the instances in a predicted class, while each row represents the instances in an actual class12.

Kappa Statistics Are a metric that compares Observed Accuracy with Expected Accuracy (Random chance) that the samples randomly would be expected to reveal13.

Receiver Operating Characteristic test (ROC) Are a plot of the true positive rate against the false positive rate for the different possible cut points of a diagnostic test. The area under the curve (AUC) is a measure of diagnostic accuracy8.

Prediction Are a data mining approach that aims at building a prediction (for continuous independent variables) model for classifying new instances into one of the two class values (yes or no)14.

Sensitivity Is the proportion of positive cases that were correctly identified, which is the most important thing in disease diagnosis15.

Specificity Is defined as the proportion of negative cases that were classified correctly16.

Data management and quality control

Data related to the outcome variable of type-2 diabetic disease status were selected and extracted from the dataset of DHIS database and from the medical registration book of public hospitals of the region. Furthermore, data cleaning, labeling, coding were done for all selected variables. On the data preprocessing phase, data manipulation and data transformation for incomplete data handling and missing value management were conducted to achieve the best data quality for prediction and classification of type-2 diabetes disease status.

Data labeling and processing

The raw (preprocessed) data were contained 13 independent; one dependent variable which is categorical variable with 2,239 samples (instances).

Moreover, in this phase, the variable set which had few records (only 2 records) due to its expensiveness, i.e., glycated hemoglobin level, and the other variables which had no role on prediction of type-2 diabetes disease like (MRN and start date), the variables which have similarity among each other (height and weight with BMI) were removed from the final predictor list of variables.

The processed variables were proceeding to data mining activity for prediction and classification of type-2 diabetes disease status and their missing values of each record were replaced with mean value of each variable. Finally, after the data were processed, a total of 2239 clients with their 7 major explanatory variables (attributes) based on variable selection method that were diagnosed for diabetes disease were included for classification and prediction of type-2 diabetic disease status using supervised machine learning algorithms (Table 1).

Table 1 Processed diabetes disease status data format from DHIS of public hospitals of Afar in 2021.

Methods of data analysis

Data of clients for diabetes disease status were collected from DHIS database of public hospitals of the region, and then it was checked for its completeness prior to analysis to increase data quality. Data preprocessing was carried out before final analysis to treat missing values and incomplete records. This processed data was changed in to comma separated value (CSV) and attribute relation file format (ARFF) to be loaded in to WEKA-3.7 tool which was used for data analysis purposes. Then the major classification and prediction supervised machine learning algorithms were applied.

On the descriptive part, appropriate descriptive statistical methods such as frequencies, percentages, tables, and graphs were used to summarize and present the findings. For the inferential part, the major supervised machine learning techniques of data mining algorithm for prediction of type-2 diabetes (Logistic regression, ANN, RF, K-NN, SVM, DT pruned J48 and Naïve Bayes) were used. The performance evaluation based on their output for effective prediction and classification of type-2 diabetes disease status among these algorithms was assessed.

Experimental setup

Since the main objective of this study is applying supervised machine learning algorithms for classifying and predicting new clients’ whether they have type-2 diabetes or not using the information extracted from the diabetes’ dataset. The model building phase in the data mining process of this investigation was carried out under classification and prediction data mining approach. This classification and prediction tasks was conducted using the major supervised machine learning algorithms (DT, pruned J48, artificial neural network, k-nearest neighbor, random forest, Naïve Bayes, support vector machine and Logistic regression) for classifying and predicting the presence or absence of type-2 diabetes disease up on the 70% of the full dataset which was the training data set. To increase classification accuracy, it is better to use many of the dataset for training and few dataset for testing on percentage split of 70 : 30 division12.

After that, this task was repeated on the testing dataset, 30% of the full dataset which doesn’t have the target class of type-2 diabetes disease status. Finally, the performance of these supervised machine learning algorithms was evaluated based on their capacity of classification and prediction on the testing dataset15. All of these tasks were implemented using WEKA 3.7 data mining tool (Fig. 1).

Figure 1

Schematic representation of data mining approach applied using supervised machine learning algorithms for classification and prediction of type-2 diabetic disease status of public hospitals of Afar regional state in 2021.

Descriptive statistics

Frequency and percentage were used to report categorical variables and mean followed by standard deviation for continuous explanatory variables were used. In addition, confusion matrices were done to show the proportion of different categories of each characteristic with respect to the outcome variable (type-2 diabetes disease status).

Variable selection method for model building

Commonly used variable selection methods available in commercial software packages, i.e., WEKA tool are Wrapper Subset Forward Selection, Best First -D 1 -N 5 variable selection, and linear forward selection methods. For this study, the Best First -D 1 -N 5 variable selection method was used at tenfold cross-validation and seed number 50 methods which is better for binary outcome studies17.

Model fit statistics

Receiver operating characteristics (ROC) curve was used to assess the general accuracy of the model to the dataset using the area under receiver operating characteristics (AUC). ROC curve is a commonly used measure for summarizing the discriminatory ability of a binary prediction model.

The ROC curve describes the relationship between sensitivity and specificity of the classifier. Since the ROC curve cannot quantitatively evaluate the classifiers, AUC is usually adopted as the evaluation index. AUC value refers to the area under the ROC curve. An ideal classification model has an AUC value of 1, with a value between 0.5 and 1.0, and the larger AUC represents that the classification model has better performance18.

Model diagnostics

The performance evaluation (model diagnosis) of different supervised machine learning techniques of data mining algorithms for prediction and classification of type-2 diabetes were carried out. This was being done using cross-validation and different confusion matrices.

Cross-validation is a technique used to evaluate the performance of classification algorithms. It is used to evaluate error rate for learning techniques. The dataset is portioned in to n-folds; each fold is used for testing and training purposes. The procedure repeats for n times in testing and training dataset. In a tenfold cross validation the data is divided in to 10 parts where each part is approximately the same to form the full dataset. Each term is held out and during the learning scheme which trained on the remaining nine-tenths, the error rate is calculated in the holdout set. Learning procedure executes 10 times on training sets and finally the error rates for 10 sets are averages to yield an overall error rate10. A confusion matrix is used to present the accuracy of classifiers obtained through classification. It is used to show the relationship between outcomes and predicted classes (Table 2).

Table 2 Different outcome of two-class prediction used for performance evaluation.

In addition to the confusion matrices, there are also different parameters used to compare the performance of supervised machine learning algorithms for their classification and prediction capacity. The table below contains performance comparison matrices with their respective formulas (Table 3).

Table 3 Performance evaluation matrix for the classification and prediction model with their formula.

Each model which was used for classification and prediction algorithm was diagnosed based on their accuracy. Since this study was a medical database record review design which is having known class (confirmed as type-2 diabetic patient or not), then we were used for the supervised data mining techniques. The machine learning algorithms and model specifications for prediction and classification of diabetes disease were focused on high performance algorithm and from high dimensional medical dataset. These model comparisons were conducted according to the following table format (Table 4).

Table 4 Criterion’s used for Performance comparison of applied supervised machine learning algorithms used for classification and prediction of type-2 diabetes disease status in Afar regional state.

Model specification

Classification models have been used to determine categorical class labels, meanwhile prediction models were used to predict continuous functions, this was done in the following steps:- Data cleaning, Relevance analysis, and Data Transformation. The obtained datasets was preprocessed and split into 2 sets, training (70% of the total dataset) and test data (30% of the dataset) (13, 15). The following models were specified for data mining in classification and prediction of type-2 diabetes disease status accordingly.

Ethics approval and consent to participate

The study was approved by the Institutional Review Board of Samara University. A letter of support was obtained from Samara University. All results of this research were based on the use of secondary data and the data collection was performed prospectively. Therefore, an informed written consent form from the public hospital DHIS Database coordinator was obtained and the study was conducted in accordance with the ethical standards of the institutional and national research committee.

Source link

Related Articles

Leave a Reply

Stay Connected

- Advertisement -spot_img

Latest Articles

%d bloggers like this: