R Data Mining Blueprints (2016)

Chapter 9. Applying Neural Network to Healthcare Data

Neural-network-based models are gradually becoming the backbone of artificial intelligence and machine learning implementations. The future of data mining will be governed by the usage of artificial neural-network-based advanced modeling techniques. One obvious question: why is neural network gaining so much importance recently though it was invented in 1950s? Borrowed from the computer science domain, a neural network can be defined as a parallel information processing system where the inputs are connected with each other like neurons in the human brain to transmit information so that activities such as face recognition, image recognition, and so on can be performed. In this chapter, we are going to learn about application of neural-network-based methods in various data mining tasks such as classification, regression, time series forecasting, and feature reduction. Artificial Neural Network (ANN) functions in a way that is similar to the human brain, where billions of neurons link to each other for information processing and insight generation.

In this chapter, you will learn about various types of neural networks, methods, and variants of neural networks with different functions to control the training of artificial neural networks in performing standard data mining tasks such as:

· Prediction of real valued output using regression-based methods

· Prediction of output levels in a classification-based task

· Forecasting future values of a numerical attribute based on historical data

· Compressing features to recognize important ones in order to perform prediction or classification

Introduction to neural networks

The brain's biological network provides the basis for connecting elements in a real-life scenario for information processing and insight generation. It's a hierarchy of neurons connected through layers, where the output of one layer becomes the input for another layer; information passes from one layer to another layer as weights. The weights associated with each neuron contain insights so that the recognition and reasoning become easier for the next level. Artificial neural network is a very popular and effective method that consists of layers associated with weights. The association between different layers is governed by a mathematical equation that passes information from one layer to the other. In fact, a bunch of mathematical equations are at work inside one artificial neural network model. The following graph shows the general architecture for a neural-network-based model:

Introduction to neural networks

Figure 1

In the preceding graph, there are three layers—Input, Hidden and Output layer—which are the core of any neural network-based architecture. ANNs are a powerful technique used to solve many real-world problems such as classification, regression, and feature selection. ANNs have the ability to learn from new experiences in the form of new input data in order to improve the performance of classification- or regression-based tasks and to adapt themselves to changes in the input environment. Each circle in the preceding figure represents a neuron.

There are different variants of neural networks that are used in multiple different scenarios; we are going to explain a few of them conceptually in this chapter, and also their usage in practical applications:

· Single hidden layer neural network: This is the simplest form of neural network, as shown in the preceding figure. In it, there is only one hidden layer.

· Multiple hidden layer neural networks: In this form, more than one hidden layer will connect the input data to the output data. The complexity of calculation increases in this form as it requires more computational power in the system to process information.

· Feed forward neural networks: In this form of neural network architecture, the information passed is one-directional from one layer to another layer; there is no iteration from the first level of learning.

· Back propagation neural networks: In this form of neural network, there are two important steps. Feed forward works by passing information from the input to the hidden and from the hidden to the output layer; secondly, it calculates the error and propagates it back to the previous layers.

The feed-forward neural network model architecture is shown in the following figure, and backpropagation method is explained in Figure 3:

Introduction to neural networks

Figure 2

In the following figure, the red-colored arrows indicate information that has not passed through the output layer, and is again fed back to the input layer in terms of errors:

Introduction to neural networks

Figure 3

Having displayed the general architecture for different types of neural networks, let's visit the underlying math behind them.

Understanding the math behind the neural network

The neurons present in different layers--input, hidden, and output--are interconnected through a mathematical function called activation function, as displayed in Figure 1. There are different variants of the activation function, which are explained as follows. Understanding the activation function will help in implementation of the neural network model for better accuracy:

· Sigmoid function: This is frequently used by professionals in data mining and analytics, as it is easier to explain and implement too. The equation is mentioned here: