Evacluster-package {Evacluster}

Evaluation Clustering Methods for Disease Subtypes Diagnosis (Evacluster)

Description

Contains a set of clustering methods and evaluation metrics to select the best number of the clusters based on clustering stability.

Details

Package:	Evacluster
Type:	Package
Version:	0.1.0
Date:	2022-03-25
License:	LGPL (>= 2)

Purpose: The design of clustering models and evaluation metrics for finding the cluster's number via computing clustering stability. The best number of clusters is selected via consensus clustering and clustering stability.

Author(s)

Fahimeh Nezhadmoghadam, Jose Gerardo Tamez-Pena, Maintainer: <[e-mail]>

References

Nezhadmoghadam, Fahimeh, and Jose Tamez-Pena. "Risk profiles for negative and positive COVID-19 hospitalized patients.(2021) Computers in biology and medicine 136 : 104753.
Fahimeh Nezhadmoghadam, et al., Robust Discovery of Mild Cognitive impairment subtypes and their Risk of Alzheimer's Disease conversion using unsupervised machine learning and Gaussian Mixture Modeling (2021), Current Alzheimer Research, 18 (7), 595-606.

Examples

    ## Not run: 
    ### Evacluster Package Examples ####
    library(datasets)
    data(iris)

   # Split data to training set and testing set
   rndSamples <- sample(nrow(iris),100)
   trainData <- iris[rndSamples,]
   testData <- iris[-rndSamples,]

  
   ## Expectation Maximization Clustering
   # Perform Expectation Maximization Clustering on training set with 3 clusters 
   clsut <- EMCluster(trainData[,1:4],3)
   
   # Predict the labels of the cluster for new data based on cluster labels of the training set
   pre <- predict(clsut,testData[,1:4])
   
   
   ## Fuzzy C-means Clustering
   # Perform Fuzzy C-means Clustering on training set with 3 clusters 
   clsut <- FuzzyCluster(trainData[,1:4],3)
   
   # Predict the labels of the new data 
   pre <- predict(clsut,testData[,1:4])
   
   
   ## hierarchical clustering
   # Perform hierarchical clustering on training set with 3 clusters 
   clsut <- hierarchicalCluster(trainData[,1:4],distmethod="euclidean",clusters=3)
   
   # Predict the labels of the new data 
   pre <- predict(clsut,testData[,1:4])
   
   
   ## K-means Clustering
   # Perform K-means Clustering on training set with 3 clusters 
   clsut <- kmeansCluster(trainData[,1:4],3)
   
   # Predict the labels of the new data 
   pre <- predict(clsut,testData[,1:4])
   
   
   ## Partitioning Around Medoids (PAM) Clustering
   # Perform pam Clustering on training set with 3 clusters 
   clsut <- pamCluster(trainData[,1:4],3)
   
   # Predict the labels of the new data 
   pre <- predict(clsut,testData[,1:4])
   
   
   ## Non-negative matrix factorization (NMF)
   # Perform nmf Clustering on training set with 3 clusters 
   clsut <- nmfCluster(trainData[,1:4],rank=3)
   
   # Predict the labels of the new data 
   pre <- predict(clsut,testData[,1:4])
   
   
   ## t-Distributed Stochastic Neighbor Embedding (t-SNE)
   
   library(mlbench)
   data(Sonar)
 
   rndSamples <- sample(nrow(Sonar),150)
   trainData <- Sonar[rndSamples,]
   testData <- Sonar[-rndSamples,]
 
   # Perform tSNE dimensionality reduction method on training data 
   tsne_trainData <- tsneReductor(trainData[,1:60],dim = 3,perplexity = 10,max_iter = 1000)
   
   # performs an embedding of new data using an existing embedding
   tsne_testData <- predict(tsne_trainData,k=3,testData[,1:60])
   
   
   ## clustering stability function
   # Compute the stability of clustering to select the best number of clusters.
   library(mlbench)
   data(Sonar)
 
   Sonar$Class <- as.numeric(Sonar$Class)
   Sonar$Class[Sonar$Class == 1] <- 0
   Sonar$Class[Sonar$Class == 2] <- 1
   
   # Compute the stability of clustering using kmeans clustering, UMAP as 
   dimensionality reduction method, and feature selection technique
   
  ClustStab <- clusterStability(data=Sonar, clustermethod=kmeansCluster, dimenreducmethod="UMAP",
                              n_components = 3,featureselection="yes", outcome="Class",
                              fs.pvalue = 0.05,randomTests = 100,trainFraction = 0.7,center=3)
   
   
   # Get the labels of the subjects that share the same connectivity
   clusterLabels <- getConsensusCluster(ClustStab,who="training",thr=seq(0.80,0.30,-0.1))


     # Compute the stability of clustering using PAM clustering, tSNE as
     dimensionality reduction method, and feature selection technique
     
   ClustStab <- clusterStability(data=Sonar, clustermethod=pamCluster, dimenreducmethod="tSNE",
                              n_components = 3, perplexity=10,max_iter=100,k_neighbor=2,
                             featureselection="yes", outcome="Class",fs.pvalue = 0.05,
                               randomTests = 100,trainFraction = 0.7,k=3)
          
    # Get the labels of the subjects that share the same connectivity
   clusterLabels <- getConsensusCluster(ClustStab,who="training",thr=seq(0.80,0.30,-0.1))
                     
                     
    # Compute the stability of clustering using hierarchical clustering,
    PCA as dimensionality reduction method, and without applying feature selection
                                 
   ClustStab <- clusterStability(data=Sonar, clustermethod=hierarchicalCluster, 
                               dimenreducmethod="PCA", n_components = 3,featureselection="no",
                               randomTests = 100,trainFraction = 0.7,distmethod="euclidean", 
                               clusters=3)
                               
 # Get the labels of the subjects that share the same connectivity
   clusterLabels <- getConsensusCluster(ClustStab,who="training",thr=seq(0.80,0.30,-0.1))
   
   
   # Show the clustering stability resuldts
   mycolors <- c("red","green","blue","yellow")
 
   ordermatrix <- ClustStab$dataConcensus
 
   heatmapsubsample <- sample(nrow(ordermatrix),70)
 
   orderindex <- 10*clusterLabels + ClustStab$trainJaccardpoint
 
   orderindex <- orderindex[heatmapsubsample]
   orderindex <- order(orderindex)
   ordermatrix <- ordermatrix[heatmapsubsample,heatmapsubsample]
   ordermatrix <- ordermatrix[orderindex,orderindex]
   rowcolors <- mycolors[1+clusterLabels[heatmapsubsample]]
   rowcolors <- rowcolors[orderindex]
 
 
   hplot <- gplots::heatmap.2(as.matrix(ordermatrix),Rowv=FALSE,Colv=FALSE,
                            RowSideColors = rowcolors,ColSideColors = rowcolors,dendrogram = "none",
                            trace="none",main="Cluster Co-Association \n (k=3)")
                            
   
   # Compare the PAC values of clustering stability with different numbers of clusters 
   
   ClustStab2 <- clusterStability(data=Sonar, clustermethod=kmeansCluster, dimenreducmethod="UMAP",
                              n_components = 3,featureselection="yes", outcome="Class",
                              fs.pvalue = 0.05,randomTests = 100,trainFraction = 0.7,center=2)
 
   ClustStab3 <- clusterStability(data=Sonar, clustermethod=kmeansCluster, dimenreducmethod="UMAP",
                                n_components = 3,featureselection="yes", outcome="Class",
                                fs.pvalue = 0.05,randomTests = 100,trainFraction = 0.7,center=3)
 
   ClustStab4 <- clusterStability(data=Sonar, clustermethod=kmeansCluster, dimenreducmethod="UMAP",
                                n_components = 3,featureselection="yes", outcome="Class",
                                fs.pvalue = 0.05,randomTests = 100,trainFraction = 0.7,center=4)
                                
                                
   color_range<- c(black="#FDFC74", orange="#76FF7A", skyblue="#B2EC5D")
 
 
   max.temp <- c(ClustStab2$PAC,ClustStab3$PAC,ClustStab4$PAC) 
 
   barplot(max.temp,xlab = "Number of clusters",ylab = "PAC", names.arg = c( "2","3","4"), 
          ylim=c(0,0.3),col= color_range[1:length(c(1,6,2,6,1))])
                            
   
## End(Not run)