Artificial Neural Networks for Detection of Diabetes Mellitus by Selecting Significant Risk Factors Using Backward Stepwise Feature Selection Method

Diabetes Mellitus (DM) is a chronic disorder affecting more than 300 million people world-wide and it needs urgent attention. Selecting of significant risk factors (SRFs) and their contributions to the risk of the disease is the key for early detection of the disease. The aim of this paper is to use Backward Stepwise Feature Selection Method (BSFSM) to select the SRFs and their contribution to the risk of DM and Kappa statistic value (KSV) to evaluate the model performance. Dataset consists of 400 patients with demographic, clinical, lifestyle and dietary risk factors collected from General Hospital Kaura Namoda, Zamfara State, Nigeria from 2019 to 2023 by checking the file of patients suffering from DM. The results obtained revealed that BSFSM retained twelve (12) SRFs, namely Blood Glucose level (BGL), High Body Mass Index (BMI), Family History of DM (FHDM), Preference for Sweet Food (PSWF), Age, Lack of Physical Activity (LPA), Blood Pressure (BP), Red Meat (RM), Refined Carbs (RC), Energy Drink (ED), White Rice (WR) and Processed Meat (PM), and removed two (2) Non-SRFs: Sex and Preference for Salty Food (PSF). The SRFs contributed 85.40%, 51.34%, 55.72%, 68.23%, 57.50%, 29.96%, 66.18%, 41.42%, 12.20%, 18.65%, 29.76% and 10.11% respectively to the risk of DM. Similarly, the Non-SRFs contributed 0.98% and 1.16% respectively to the risk of DM. The MLP model detected 98.6% DM patients in the training set, 96.3% in the validation set and 92.9% in test set. There were 97.8% Non-DM patients in the training set, 93.9% in the validation set and 93.8% for the test set. The KSV of the model was 0.94 and it was capable of distinguishing between DM and Non-DM patients. This paper demonstrated that BSFSM was capable of selecting the SRFs and their contributions, and KSV adequately evaluate the performance of the model.