The Proposed Medical Recommender System

Here, the RS model has become more valid research in developing artificial intelligence (AI) methods. In contrast, several numbers of RSs aim at e-business, book, and movie recommendation, which offers a virtual experience to apply the proper predictions. As there are higher accuracy and efficiency, it becomes more crucial for online medicine RSs, thus it is estimated with the help of data mining (DM) approaches to attain optimal trade-off from accuracy, efficiency, and reliability. As depicted in Fig. 6.2, the RS framework is comprised of five steps:

  • 1. Database system
  • 2. Data preparation
  • 3. Recommendation system
  • 4. Model evaluation
  • 5. Data visualization

Database System Module

This module offers data links that are composed with the diagnosis case, drug database, as well as a professional knowledge database. The initial database saves the details of the diagnosis case and gives access to alternate modules. Secondly, the drug database gathers every drug and an index is developed. Finally, experts’ knowledge is attained by consolidating professional knowledge.

Overall process of proposed method

FIGURE 6.2 Overall process of proposed method.

Data Preparation Module

It is treated as a data-cleaner in this approach. A practical data is original, which might be partial and noisy. Thus, data preparation has been created for producing clear information. It is comprised of missing value computation, correlation determining, as well as data optimization.

Recommendation Model Module

This method is deployed under the application of the b-ELM model. Here, it is majorly deployed for building RS on the basis of three techniques. Also, it has been added with novel DM methods to resolve the problems. The visualization module offers visualization methods to project little valid knowledge in diagnosis case data.

Model Evaluation Module

This system estimates diverse RS models in the concrete data set. In the case of a diverse data set, it is required to compute the techniques to attain optimal trade-off between model accuracy, model efficiency, and model scalability. It is assumed to be the overall medicine RS, which uses the DM approaches to medical analysis that complete exploitation of diagnosis case details as well as experts’ knowledge. Few RS methods are developed according to diagnosis case data and obtain the drug for an individual integrated with professional knowledge.

Proposed Recommendation Model

Extreme Learning Machine (ELM)

ELM is defined as a unique learning model to a single hidden-layer feed-forward neural network (SLFN). The flow diagram, as well as SLFN ELM, is stated in Fig. 6.3 and Fig. 6.4, correspondingly. It is obvious that a diverse gradient-based learning model is applied for normal ANN, the biases and input weights are resolved, and finally, the simulation weights are applied for easy matrix estimations in ELM, and

Single-layer feed forward ELM

FIGURE 6.3 Single-layer feed forward ELM.

Structure of ELM

FIGURE 6.4 Structure of ELM.

reducing training duration. Also, ELM is an improvising mechanism from regression applications and massive data set classifying approaches.

Given a collection of A diverse instance {(x j, t j)} Nj = 1 with inputs Xj= [Xj,

Xj2,.... Х,„ Te _ b" and outputs = l{, t)2, ..., tjm]Те__, the ELM with N

hidden neurons with activation function § (•) is denoted numerically as

where Wt = [ , [ W)2, ...,[Wjn]T is the weight vector that connects the input neurons andyth secret neurons; pj = [pjl, Pj2,..., Pj/и] Trepresents a weight vector from yth secret neurons to resultant neurons; and bt implies a threshold of y'th secret node. The relation among a target input and output layers of a network is described as

The formula can be presented effectively as H P = T, where P = [pi, .... p" N] T,

T = [t 1.....t N]T and H signify a secret layer of ELM.

The hidden neurons undergo a conversion of input data to the diverse presentation as two procedural ways. Initially, the data is detected as a hidden layer by weights with biases of the input layer and implies the result of nonlinear activation functions.

ELMs are resolved as typical neural networks (NN) in matrix form as demonstrated in Fig. 6.4. The matrix form is provided as

Providing that T shows the target, an exclusive outcome of a system with minimum squared error is identified w'ith the help of the Moore-Penrose generalized inverse. Hence, it is computed in an individual process of values in weights of a hidden layer, which tends in a solution w'ith reduced error to predict a destination T

b-ELM Classifier

Here, it is established with ELM ensemble classification models such as b-ELM. The main aim of using this method is to reach a reliable, effective means of classification in large-scale data. b-ELM applies a Bag of Little Bootstraps (BLB) approach in processing gains as well as productive scalability. BLB is capable of capturing diversities of fundamental classifiers from small subsets. In ELM, training dataset are produced using BLB small subsets for generating the larger original training data set. It makes use of ensemble predictors on training sets, and making decisions and assumed as a detected label. The w'orkflow of b-ELM is provided in Fig. 6.5.

It is composed of two pairs of nested loops in a method: the initial pair of nested loops are employed to find the optimal parameters for base classifiers relied on k-fold cross-validation (CV); the second pair is applied for training base classifiers and obtains predictions for the testing data set. The aggregation stage has a major number of votes to get the final outcome. The parameter b = nY in technology is a size of subsamples tested with no replacement from the whole actual training data set. Also, b = n'< for 0.5 < y< 0.9. In fact, b-ELM has a maximum favorable space profile when

Workflow of b-ELM

FIGURE 6.5 Workflow of b-ELM.


Attributes of the Diagnosis Case Data

Attribute Name

Attribute Type

Attribute Values


Numeric attribute



Binary attribute


BP (Blood Pressure)

Discrete attribute



Discrete attribute



Numeric attribute

0.5005 - 0.899S


Numeric attribute

0.0202 - 0.0799


Discrete attribute

Drug A. B,C, X.Y

compared to Bagging ELM. The parameter s implies a count of subsamples, whereas r signifies re-samples bootstrapped. Therefore, s and r compute the entire number of predictions for the testing data set.

< Prev   CONTENTS   Source   Next >