Early Diabetes Prediction: A Comparative Study Using Machine Learning Techniques.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Current diabetes mellitus (DM) screening tools rely heavily on Electronic Health Record (EHR) data.
• Developing and under-developed nations face significant challenges in establishing robust EHR systems.
• This data gap limits the availability of early screening tools in resource-limited settings.

Purpose of the Study:

• To develop and validate a prediction model for early DM detection using direct questionnaires.
• To address the lack of EHR data in developing countries for diabetes screening.
• To evaluate the efficacy of various machine learning algorithms for early DM prediction via questionnaires.

Main Methods:

• A prediction model for early DM was developed using data from direct questionnaires at a tertiary hospital in Bangladesh.
• Information gain technique was employed for feature selection to reduce irrelevant variables.
• Logistic regression, support vector machine, K-nearest neighbor, Naïve Bayes, random forest (RF), and neural network models were implemented and compared.

Main Results:

• The Random Forest (RF) algorithm demonstrated superior performance among the tested machine learning models.
• The RF model achieved an accuracy of 100% in predicting diabetes at an early stage.
• Feature selection using information gain effectively identified key predictive variables from the questionnaire data.

Conclusions:

• A combination of simple questionnaires and machine learning algorithms can effectively identify undiagnosed diabetes mellitus patients.
• This approach offers a viable solution for early DM screening in regions with limited EHR infrastructure.
• The study highlights the potential of accessible data collection methods for improving global diabetes management.