Neural Network Diabetes Prediction

Students of Computational Intelligence (CSC3034) at Sunway University were tasked with finding a dataset from Kaggle and implement a Neural Network System based on the dataset chosen.

---

Chosen dataset

The dataset chosen for this assignment was the Pima Indians Diabetes Database. The neural network will utilize this dataset to predict whether a patient is diabetic based on:

• Pregnancies – Number of pregnancies the patient has had
• Glucose – Plasma glucose concentration from 2 hours in oral glucose tolerance test
• BloodPressure – Diastolic blood pressure (mm Hg)
• SkinThickness – Thickness of the patient’s triceps skin folds (mm)
• Insulin – 2 hours serum insulin (mu U/ml)
• BMI – Body Mass Index measured from weight(kg) / height(m)2
• Age – Age of the patient in years
• DiabetesPedigreeFunction – Risk of diabetes based on family history

---

Implementation

The assignment was completed by implemented Multi-Layer Perceptron (MLP) Neural Network using Python and the following libraries:

• Pandas
• Matplotlib
• NumPy
• Scikit-learn

Test Cases for Optimization

1. Activation Function and Solver combination
2. Number of Hidden Layers and Neurons
3. Number of Training Iteration

---

Results

Read the full report here

The final setting for the MLP Neural Network are as such:

• Activation function - Identity Function
• Solver - Limited-memory BFGS solver
• Hidden layers and Neurons - 4 Hidden Layer with 6 Neurons in each layer

After running 200 randomized test cases, the results of the prediction system are as follow:

Mean Accuracy = 77.45%

	Actual Non-Diabetic	Actual Diabetic
Predicted Non-Diabetic	17715	2379
Predicted Diabetic	4567	6139

Possible Reasoning for Low Accuracy

Dataset chosen has biasness towards non-diabetic patients which consisted of 65% of the dataset.