DISEASE PREDICTION USING ARTIFICIAL NEURAL NETWORKS

Abstract

In recent times, the medical world has been dealing with a very serious challenge  detecting diseases early enough before they become life-threatening. Many patients, especially in developing countries like Nigeria, only come to know they have a chronic disease when it has already progressed to an advanced and harder-to-treat stage. This is mostly because traditional methods of diagnosis rely heavily on doctors and laboratory technicians manually reviewing test results, a process that can be slow, expensive, and sometimes prone to human error, especially in environments with limited medical personnel or equipment.

This work investigates how Artificial Neural Networks (ANN), which is a core technique in the field of machine learning and artificial intelligence, can be trained to automatically predict whether a patient has diabetes based on certain routine medical examination inputs. The work makes use of the widely accepted Pima Indians Diabetes Dataset, which contains the medical records of female patients of Pima Indian heritage, including features such as glucose concentration, blood pressure, body mass index (BMI), age, insulin level, skin thickness, number of pregnancies, and a diabetes pedigree function score. The ANN model was designed with three hidden layers, batch normalization, and dropout regularization to ensure it generalized well on unseen data. After complete training and evaluation on the test portion of the dataset, the model achieved a Test Accuracy of 73.38%, Precision of 59.70%, Recall (Sensitivity) of 74.07%, an F1-Score of 66.12%, and an AUC-ROC score of 0.798. The relatively high Recall score is especially important in a medical context because it means the model successfully identified the majority of actual diabetic patients. These results confirm that Artificial Neural Networks are a highly promising tool for disease prediction and early diagnosis in clinical settings. The simulation was implemented entirely in Python using TensorFlow and Keras, and all results are verifiable from the charts included in this report.