A novel fuzzy k-means latent semantic analysis (FKLSA) approach for topic modeling over medical and health text corpora

Abstract

Medical and health text documents pose a challenge for data handling and retrieving the relevant and meaningful documents. Automatically retrieval of significant knowledge with a better understanding of medical and health documents is a challenging task. One popular approach for thematically understand the medical and health text documents and finding the topics from these documents is topic modeling. In this research, we propose a novel topic modeling approach Fuzzy k-means latent semantic analysis (FKLSA) by using the fuzzy clustering. Our method generates local and global term frequencies through the bag of words (BOW) model. Principal component analysis is used for removing high dimensionality negative impact on global term weighting. Previous work shows that in medical and health documents redundancy issue has a negative impact on the quality of text mining. Therefore, the main achievement of FKLSA is the handling of the redundancy issue in medical and text documents and discover semantically more precise topics. FKLSA is socially utilized for finding the themes from medical and health text corpus. These topics are further used for text classification and clustering tasks in text mining. Experimental results show that FKLSA performs better than LDA and RedLDA for redundant corpora. FKLSA’s time performance is also stable with an increase in number of topics and thus better than LDA and LSA on a big twitter heath dataset. Quantitative evaluations of the real-world dataset for health and medical documents show that FKLSA gives a higher performance as compared to state-of-the-art topic models like Latent Dirichlet allocation and Latent semantic analysis.

Keywords

Topic modeling bag-of-words term weighting fuzzy k-means principal component analysis

1 Introduction

Nowadays, large number of medical and health data are stored as an electronic health record (EHR) and there is a need to analyze these documents. National science foundation identified that vast quantity of scientific records management and analysis is one of the challenging tasks and a new field for future study [1]. These large collections of electronic records need new tools for organizing, searching, indexing and browsing the documents. In addition, finding the most relevant documents from all available documents is a difficult task. Particularly health and medical text data are generated and stored very fast like the number of paper publications on PubMed website has been more than 6 million in 2015. US hospital average discharges are also greater than the 30 million records [2].

Therefore, by using advanced data analytics techniques for health and medical text data, companies can save $450 billion annually. There is a need to develop efficient techniques that discover the hidden themes in complicated health and medical text data.

In natural language processing to obtain representations of words distributed from the input data, the bag-of-words model is used. A popular technique for medical and health text representation is Bag-of-words (BOW) model. This model is one of the word2vec models that analyze the semantics of natural language [3]. Working of BOW has explained by an example in Fig. 1. There are three documents and an individual list is constructed for each document. In each document, words are scored according to their occurrences in their respective document. The words are converted into a vector with respect to their occurrences. Rows represent the words and columns represent documents. Such as word “cancer” represents one time in document 1 and word “of” represents two times in document 2. The rough set is a tool for conceptualizing, organizing and analyzing several types of data that deal with inaccurate and unclear information in the application associated to artificial intelligence. In accordance with the rough set theory the information which is irrelevant is eliminated from the documents for classification and decision-making process [4]. In the soft set theory, the N-soft sets are utilized in the decision-making algorithms [5]. The parameters of the soft sets are words, numbers and sentences in various documents.

Fig. 1

Example of bag-of-words.

When documents are large then the matrix is called a sparse matrix [6]. Sparsity means that the occurrence of many words in corpus but in one document a small percentage of all words. Therefore, in the BOW matrix, most elements (words) are zero in documents [7].

Topic modeling is the popular probabilistic text modeling technique that is rapidly accepted in text mining and information retrieval fields [8 –10]. Topic modeling is a standard technique to deal with high dimensionality and sparsity problems.

Topic modeling automatically classifies many documents into topics and their equivalent distribution.

Topic modeling is a text analysis technique in which the documents contain different words and these words’ term frequencies are features. Topic modeling output matrix is shown in Fig. 2.

Fig. 2

Topic modeling matrix interpretation.

In medical and health text mining, topic modeling is an efficient technique but still need improvement because the redundancy has a negative impact on topic modeling [11]. Most of the medical and health documents are redundant [12].

Therefore, in this research dissertation, we proposed a novel fuzzy k-means latent semantic analysis technique for topic modeling approach used for medical and health text corpora. There are some research questions that are solved by the proposed technique.

How semantically precise topics are discovered from medical and health text documents?

How to remove the high dimensionality effect?

How the redundancy issue is removed from the redundant corpus.

How can achieved the higher results of classification and clustering in medical text documents?

The analysis of medical and text documents is useful for documents classification and clustering tasks. It is also helpful for removing the redundancy issue in medical text corpus. Therefore, we are motivated to develop a topic modeling approach that will solve redundancy issue in medical text corpus and topics are utilized for text classification and clustering.

The proposed topic modeling approach shows higher performance in both redundant and non-redundant documents and estimates the numbers of topics within the corpus.

To measuring the performance of FKLSA, we conduct extensive experiments on four real-world health and medical text collection datasets.

Results of the experiment show that 1) FKLSA discovers semantically strong topics from medical and health text corpora. 2) Our method generates local and global term frequencies through the bag of words (BOW) model. Principal component analysis is used for removing high dimensionality negative impact on global term weighting. 3) FKLSA performance is better for redundant corpora as compare to its competitors LDA and RedLDA. 4) FKLSA’s time performance is stable for big twitter health dataset with an increase in number of topics and better than LDA and LSA. 5) FKLSA discovers more precise topics from health and medical text documents as compared to baseline topic models. Therefore, classification and clustering results of FKLSA is higher than the state-of-the-art topic models.

Rest of the paper is organized as follows. Section 2 describes previous related work. Section 3 explains the proposed technique more briefly. Section 4 describes experiments and results, Section 5 is discussion, Section 6 is example of topics and section 7 explains the conclusion.

2 Related work

In this section, we describe the previous medical and health applications for topic modeling and fuzzy clustering.

In machine learning, statistical topic modeling is a method that finds patterns and hidden information in text data [13]. Text mining has two approaches; supervised and unsupervised learning. In classification, the corpus is a train with their labels and assigning the labels to new documents [13]. In clustering, clusters are assigned to every document in the large collection of documents that is based on similarity or dissimilarity between clusters.

Topic modeling is a standard unsupervised method of text mining, which has many applications in spam detection, SMS [14] and tagging of images [15]. Topic modeling defines that documents are a probability distribution over topics and topics are a probability distribution over words.

Latent Semantic Analysis (LSA) is a topic modeling approach, particularly distributive semantics, to analyze the relationships between a set of documents and the terms they contain by producing a set of topics related to documents and terms. Unfortunately, a large collection of medical text documents is not structured and contains unfavorable patterns such as redundancies and contradictions [16]. This creates problems for precise and automatic classification for medical text documents [17]. Latent semantic analysis is an approach used for medical text classification and text mining [18 –20]. It assumes that words having a close meaning will appear in similar parts of the text. A matrix that contains word counts per document (rows represent single words and columns represent each document) is constructed from a large portion of text and a dimension reduction technique called singular value decomposition (SVD) is used to reduce the number of rows and preserve the structure of similarity between columns.

The Latent Dirichlet allocation (LDA) topic model shows better performance in medical and health text mining [21]. In topic modeling, LDA is a generative probabilistic topic model [10]. There are various applications of LDA in health and medical such as the discovery of clinical concepts and structures in patient health record [22], predicting proteins and their relationships [23], pattern identification in clinical events of brain cancer patients [24], genomic analysis of time to event outcomes [25].

Topic models used for the task of assigning an ICD-9 code for discharges summaries [26]. In [27] they apply the LDA of topic modeling to FDA of drugs side influence labels. Results of their experiments show that extracted topics are clustered the drugs by safe care and various therapeutic uses.

LDA is also used for a clinical pathway to find the treatment behavior of patients [28], predicting the clinical ordering of patterns and several treatment activities [29] are modeled with their timestamps in the clinical pathway [30].

One version of LDA for redundancy issue in medical documents is Redundancy-aware LDA (RedLDA) that shows higher performance than LDA [31].

The documents on the web are very complex and complicated links open from one web document to another web document. Effective clustering methods extract the latent meaning from documents. In [32] a technique is proposed that extract the appropriate meaning in web documents through fuzzy linguistic topological space along with the fuzzy clustering algorithm. The topics are extracted from the web documents. In [33] proposed a technique that reveal the main concepts from the text documents. The concepts are explained by word-based connections given in a semantic topological space and built by the formal model.

The two major approaches for clustering are hard and fuzzy (soft). Every object belongs to one cluster in hard clustering.

The membership is fuzzy in fuzzy clustering and objects belong to various clusters [34]. Fuzzy clustering is used to predict the citalopram in alcohol dependence in response to treatment [35], analyze the diabetic neuropathy [36], early detection of diabetic retinopathy [37], characterize the stroke and causes of ischemic stroke [38 –40], detect cancer like breast cancer [41] and improve decision making for radiation therapy [42]. Fuzzy clustering is also used for ultrasound imaging techniques improvement [43] and analyzing the data of microarray [44].

However, fuzzy clustering has many applications in the medical and health domain specifically in image processing but very less considered for topic modeling.

Therefore, in this research, a new topic modeling approach FKLSA has proposed for the analysis of medical and health text documents, which provide a link between the fuzzy clustering and topic modeling.

3 Research methodology

In this section, we describe our proposed method, fuzzy k-means latent semantic analysis (FKLSA) that discovers latent semantic features from medical and health documents. FKLSA deal with fuzzy perspective, which is a new approach for topic modeling and verified through different experiments on health and medical text corpora. FKLSA has the potential to handle the redundancy issue, discover more precise topics and gives higher performance than its competitors such as LDA and LSA.

3.1 FKLSA

Fuzzy logic is an extension of classical logic one or zero to the truth-values between one and zero. Documents and words can be fuzzy clustered through FKLSA and each cluster is a topic. For example, in given documents, FKLSA discovers the four topics as shown in Fig. 3. There are some words in the left side (part A) which relate to some topics, after applying the fuzzy process these words confirm their association with their most related topics. In this process, the words are assigned a degree of fuzzy belonging with respect to each topic (cluster). There is three colors of circles which show the membership magnitude from low (light gray) to high (dark) level. In figure’s (part B) topic 1 is “disease” which contains words germs, afflictions, infection and topic 2 is “exercise” which contains words gym, weight, body. Topic 3 is “bacteria” which consists of words toxin, bacteria, virus and topic 4 is “injury” with words pain, wounds, therapy.

Fig. 3

Example of topic modeling with the fuzzy process.

The proposed technique finds five matrices i.e. probability of words P(W), probability of documents P(D), probability of words over documents P(W|D), probability of topics over documents P(T|D) and probability of words over topics P(W|T). Figure 4 describe the overall conceptual framework for proposed FKLSA topic modeling approach.

Fig. 4

Conceptual Framework of FKLSA topic modeling approach.

Step 1. In this step preprocessing is performed on the text documents. The text data can contain a lot of noise for example words variations, words form variations, special characters, punctuations and stop words that add noise. Following steps applied for preprocessing of data that converts the raw data into clean data.

Punctuation erase from text data.

Text data converts into lower case.

Create the tokenized documents.

Stop words like “an”, “a”, “the” etc. are removed from the text data.

Short words with 2 characters are removed from the text data.

Long words with greater than 15 characters are removed from the text data.

Normalize the words using porter stemmer.

Step 2. The Bag-of-words (BOW) model is one of the most popular methods that calculates words occurrences in documents [45]. Represents the documents as a bag (multiset) of words. After preprocessing the text collection of documents, BOW model has applied. In information retrieval BOW converts unstructured text data into structured data that based on words and neglects the grammar [46]. The documents m comprising the words k and find the “relationship” of documents and words. After that, finds out the frequencies of words k in documents m. The following Equation 1 declares the words k frequencies in documents m. Where, value of summed up to n_i, j is a number of the occurrences of the considered term (t_i) in the document (d_j). The range of i and j is depending on terms occurrences in documents which are changed according to text corpus. ${tf}_{i, j} = \sum_{k^{n}}^{n_{i, j}} k, j$ (1)

After applying the BOW model, removing those words that do not appear more than two times also removed empty documents.

Step 3. In this step, Local term weighting is calculated. We use the term frequency method among different local term weighting (LTW) methods. Term frequency method [47] is the most popular method which measures that how much a term appears in the document.

The typical weight w, term k using the vector length of the normalization factor i is shown in Equation 2 for the document d. $w_{dk} / \sqrt{\sum_{vector} (w_{d} i)^{2}}$ (2)

Term weights, which are decreasing these terms, are important and assigned to the weights w_dk that will be varied constantly between zero and one. The highest weight one used for the more important term and zero for least important term. In some cases, it may be helpful to use normalize weight assignments wherever term weight individual relies to some extent of weights on different terms within the same vector. Term frequency weighting scheme is widely used in text classification and categorization [48 –52]. Let f_i,j is the occurrence of frequency of term index k_i in documents d_j then the total frequency F_i of k term is defined in Equation 3. $F_{i} = \sum_{(j = 1)}^{N} f_{i, j}$ (3)

N is the number of documents in the huge collection of data. Document frequency n_i of a term k_i is a number of documents in which k_i occurs and n_i < F_i.

Step 4. Global term weighting (GTW) has calculated in this step. It gives the “discrimination values” to every term. Many of these schemes describe that a term less frequently appears in documents collection will be more discriminative [53].

We use ten GTW methods in this research including “Normal, Unary, IDF smooth, IDF max, Entropy, Inverse document frequency (IDF), Probabilistic inverse document frequency (ProbIDF), Incremental global frequency IDF, Square root global frequency IDF and Global frequency IDF”.

Global term weighting in Table 1 are find through Equations 4 and 5. Where, tf_ij represents the occurrence of words i in j documents. N represents the amount of document and n_i is the number of documents in which i terms appear in Table 1. $b ({tf}_{ij}) = {\begin{matrix} 1 & if & {tf}_{ij} > 0 \\ 0 & if & {tf}_{ij} = 0 \end{matrix}}$ (4) $P_{ij} = \frac{{tf}_{ij}}{\sum_{j} {tf}_{ij}}$ (5)

Table 1

Global term weighting formulas

Name	Formula	Name	Formula
Normal	$\frac{1}{\sqrt{\sum_{j} {tf}^{2} i, j}}$	IDF	$log (\frac{N}{n_{i}})$
Unary	1	Entropy	$1 + \sum_{j = 1}^{N} \frac{\frac{f_{ij}}{F_{i}} log \frac{f_{ij}}{F_{i}}}{log N}$
IDF smooth	$log (1 + \frac{N}{n_{i}})$	Incremental Global Frequency IDF (IGFI)	$\frac{F_{i}}{n_{i}} + 1$
IDF max	$\log (\frac{max {t^{t} \in d}}{n_{t}^{^{'}}})$	Square Root Global Frequency IDF (IGFS)	$\sqrt{\frac{F_{i}}{n_{i}} - 0.9}$
Probabilistic IDF(ProbIDF)	${log}^{2} \frac{n - \sum_{j} b (tfij)}{\sum_{j} b (tfij)}$	Global Frequency IDF (IGFF)	$\frac{F_{i}}{n_{i}}$

Normal is used to normalize and correct discrepancies in documents [54]. In unary GTW, no global weight is used. In IDF higher weight can be assigned to infrequent terms and lower weights to common terms [55]. Entropy GTW assigns a higher weight to the terms which are the lower frequency in the documents [56]. IDF smooth and IDF max are variants of inverse document frequency. ProbIDF gives very low weight to the term that appears in each document [54]. In [57] IGFI, IGFS, and IGFF revealed and discussed.

IGFS is proposed by the combination of formulas where the square root that is good local weight assumed as a global weight. They found that subtract a large number from F_i/n_i improves the performance. IGFF performs the best and adding one to its formula and a new global term weighting formula IGFI is proposed.

The generated output of GTW step is the term frequency (TF) matrix for the documents that are TF-Normal, TF-Unary, TF- IDF smooth, TF-IDF max, TF-IDF, TF-Entropy, TF-ProbIDF, TF-IGFI, TF-IGFS, and TF-IGFF. Table 1 shows all global term weighting formulas.

Principal component analysis (PCA) [58] technique is used to avoid the high dimensionality negative impact on ten GTW methods matrices. This method used to reduce the dimensions of data. PCA is most popular multivariate technique. Its major goals are:

Extraction of most important information from the data table.

Size of the data compressed and only keeping the important information.

Dataset description is simplified.

Observations and variables structure analyzed.

To achieve these goals PCA computes the new variable that referred to the principal component, which is obtained from a linear combination of all original variables.

Step 5. Here Fuzzy k-means (FKM) clustering method is used to fuzzy clustering the documents, which represented in ten GTW methods.

The FKM algorithm partitions the data point into k clusters where S_l, (l = 1, 2, 3, … k) are associated with the clusters centered C_l. Data point and clusters relationship are fuzzy.

The membership u_i,j ∈ [0, 1] represents belonging of clusters centers C_j and data point X_i. The set of the data point is S = {X_i}. The fuzzy k means algorithm based on minimizing distortion as shown in Equation 6. Where u_I, j is membership and C_j represents clusters. $J = \sum_{j = 1}^{k} \sum_{i = 1}^{N} u_{i, j}^{q} d_{i, j}$ (6)

N represents the number of data points, fuzzifier parameter is q; numbers of clusters are represented by k and d_i,j is the squared Euclidean distance between cluster representative C_j and data point X_i.

The fuzzy k-means clustering algorithm mapped the representative vectors and improved through data partition point. Fuzzy k-means clustering algorithm start with initial clusters centered and process is repeated until stop criteria satisfied. It is supposed that no two clusters have the same representative. If d_i,j < n then, u_i,j = 1 and u_i,j = 0 for l ≠ j where n is a positive number. Now the fuzzy k-means algorithm performs following steps.

Set the initial clusters SC_o equal to (C_j (0)) and ɛ value p = 1.

Set of clusters SC_p is given and compute d_i,j for i = 1 to N, j = 1 to k and update the memberships u_i,j using Equation 7. u_i,j membership degree for each document (D) with clusters (topics). Value of u_i,j can be the probability of topics i with documents j as P (T_i|D_j). $u_{(i, j)} = {((d_{(i, j)}^{(1 / m - 1)} \sum_{l}^{k} {(\frac{1}{d_{il}})}^{\frac{1}{m} - 1}))}^{- 1}$ (7)

Where, d_i,j < n and value of n is very small and u_i,j = 1. P (T_i|D_j) can be used to find (words×topics) matrix.

Using Equation 8 centers for every cluster are computed to get new clusters which are represented by SC_p + 1. $C_{j} (p) = \frac{\sum_{j = 1}^{N} u_{i, j}^{m} X_{i}}{\sum_{j = 1}^{N} u_{i, j}^{m}}$ (8)

If the (∥ C_j (p) - C_j (p - 1) ∥) < ∈ and j = 1 to k, then stop where ∈ > zero is the small positive number. Otherwise set p + 1 → p and move to step 2 of fuzzy k-means algorithm.

In terms of numbers of calculation, the computational complexity of fuzzy-k means is O (N_kt) and t represents several iterations.

Step 6. The Documents term matrices are used with GTW methods (Words×Documents matrix) for finding the probability of the documents P (D). P (D_j) calculated by Equation 9. where j represents the various documents. $P (D_{j}) = \frac{\sum_{i = 1}^{m} (w_{i}, D_{j})}{\sum_{i = 1}^{m} \sum_{j = 1}^{n} (w_{i}, D_{j})}$ (9)

Step 7. Find the probability of documents j in topics k to find P (D_j|T_k) by using P (D_j) and P (T_k|D_j) as shown in Equation 10. $P (D_{j}, T_{k}) = P (T_{k} | D_{j}) \times P (D_{j})$ (10)

After that, normalized P (D, T) for each of the topics by Equation 11. $P (D_{j} | T_{k}) = \frac{P (D_{j}, T_{k})}{\sum_{j = 1}^{n} P (D_{j}, T_{k})}$ (11)

Step 8. In this step, the probability of words i in documents j and P (w_i|D_j) is calculated using the document-term matrix and GTW methods as shown in Equation 12. $P (w_{i} | D_{j}) = \frac{P (w_{i}, D_{j})}{\sum_{i = 1}^{m} P (w_{i}, D_{j})}$ (12)

Step 9. In this step, find the probability of words i in topics k P (w_i|T_k) using P (D_j|T_k) and P (w_i|D_j) as shown in Equation 13. $P (w_{i} | T_{k}) = \sum_{j = 1}^{n} P (W_{i}, D_{j}) \times P (D_{j} | T_{k})$ (13)

4 Experiments and results

4.1 Datasets

Four publicly available datasets are used in this research. The first dataset MuchMore Springer Bilingual Corpus 1 (M Dataset) is labeled dataset. The dataset is medical abstract of English scientific corpus from Springer website. During this research analysis, two journals are used which consists of Federal health standard sheet and Arthroskopie.

The second dataset is Ohsumed Collection 2 (O datasets) dataset (O dataset) which is labels corpus of the medical abstract from MeSH categories. In this research, categories are virus diseases, mycoses, and bacterial infections.

The third dataset is tweets of health news dataset 3 (T-datasets) which is the unlabeled dataset.

The fourth dataset is Synthetic WSJ redundant corpora based on wall street journal corpus and widely used in Natural Language Processing (NLP) [59, 60].

4.2 Documents classification

The first evaluation of classification with Bayesian optimization has performed on two datasets MuchMore Springer medical abstracts and Ohsumed datasets. The optimization is the process of locating the point that minimizes real-valued function, which is called the objective function. Bayesian optimization is a Gaussian process model for objective function and uses the evaluations of the objective function to train the model. The error in cross-validated response is minimized by using Bayesian optimization. The Fit function in Matlab is used for Bayesian optimization.

MuchMore and Ohsumed datasets are label (classes) datasets. Two classes arthroskopie and federal health sheet are selected for Springer dataset.

Classification of the document is performed on P (T|D). In this classification, documents are assigned to some class and features have extracted from text data. To achieve the high-performance ratio, we used K-nearest neighbor (KNN) with Bayesian optimization. KNN classifier is the top performing text categorization method and machine learning algorithm. KNN uses standard information retrieval technique (e.g., tfidf). KNN assigns a category to the most common neighbor documents (vote for neighbor category). KNN [61] strongly rely on the distance metric for the input data patterns. Distance metric learning is to learn a distance metric for the data entry space of a given collection of similar/different point pairs that preserves the distance relationship between the training data. K neighbors can be estimated by cross-validation [13, 62].

The FKLSA performance has checked with LDA and LSA model on a tenfold cross-validation method. By using this method data have divided into 10 subsets for 10 iterations. Each subset is selected for testing and others for training.

The KNN method has used with 50, 100 and 150 topics for input features of documents in classification. The output of KNN is presented in a confusion matrix as shown in Table 2.

Table 2
Confusion matrix

Actual Predicted

Negative Positive

Negative TN FP

Positive FN TP

Actual	Predicted
	Negative	Positive
Negative	TN	FP
Positive	FN	TP

TN: True negative (TN) is the correct predictions that instance is negative.

FP: False positive (FP) is that the incorrect predictions with the positive instance.

FN: False negative (FN) is the incorrect prediction and the instance is negative.

TP: True positive (TP) is the correct predictions with the positive instance.

We used precision, recall, accuracy, F-measures, and specificity to check the performance of FKLSA. Equations 14 –18 show the precision, recall, accuracy, f-measure, and specificity formula. $Precision = \frac{TP}{TP + FP}$ (14) $Recall = \frac{TP}{TP + FN}$ (15) $Accuracy = \frac{TP + TN}{TP + TN + FP + FN}$ (16) $F 1 Score = \frac{2 \times TP}{(2 \times TP + FP + FN)}$ (17) $Specificity = \frac{TN}{(FP + TN)}$ (18)

Table 3 shows the basic statistics of datasets used in this research with their description and numbers of documents.

Table 3

Datasets basic statistics

Dataset Names	#Documents (After Preprocess)	#Terms	#Unique terms	Description
M-Datasets	1223	16950	4302	Medical Papers
O-Datasets	3711	28630	8387	Medical papers
T-Dataset	58,927	395,635	25,309	Health news tweets
WSJ	1300	680K	36K	Redundant corpora

The classification results on two datasets have shown in Tables 4 –9. Results indicate that proposed technique with Normal, Unary, IDF smooth, IDF max, Entropy, inverse document frequency (IDF), Probabilistic inverse document frequency (ProbIDF), log global frequency IDF, incremental global frequency IDF, square root global frequency IDF and global frequency IDF methods give the highest performance as compared to LSA and LDA model on different number of topics.

Table 4

MuchMore springer datasets classification results on 50 topics

Method	Accuracy%	Precision	Recall	F1 Score	Specificity	# topics
LSA	50.00	0.4295	0.4161	0.4227	0.5659	50
LDA	64.2077	0.5949	0.5839	0.5893	0.6878	50
FKLSA (Normal)	85.52	0.8375	0.8323	0.8349	0.8732	50
FKLSA (IDF)	85.52	0.8176	0.8634	0.8399	0.8488	50
FKLSA (Unary)	80.60	0.7679	0.8012	0.7842	0.8098	50
FKLSA (IDF Smooth)	84.97	0.8155	0.8509	0.8328	0.8488	50
FKLSA (IDF Max)	80.60	0.7679	0.8012	0.7842	0.8098	50
FKLSA (ProbIDF)	84.70	0.8107	0.8509	0.8303	0.8439	50
FKLSA (Entropy)	60.66	0.5411	0.6957	0.6087	0.5366	50
FKLSA (IGFI)	67.21	0.6242	0.6398	0.6319	0.6976	50
FKLSA (IGFS)	65.85	0.6071	0.6335	0.6201	0.6780	50
FKLSA (IGFF)	69.67	0.6524	0.6646	0.6585	0.7220	50

Table 5

MuchMore springer datasets classification results on 100 topics

Method	Accuracy%	Precision	Recall	F1 Score	Specificity	# topics
LSA	50.01	0.4329	0.4410	0.4369	0.5463	100
LDA	57.65	0.5205	0.4720	0.4951	0.6585	100
FKLSA (Normal)	85.79	0.8303	0.8509	0.8405	0.8634	100
FKLSA (IDF)	84.70	0.8107	0.8509	0.8303	0.8439	100
FKLSA (Unary)	77.60	0.7005	0.8571	0.7709	0.7122	100
FKLSA (IDF Smooth)	84.70	0.7966	0.8758	0.8343	0.8244	100
FKLSA (IDF Max)	77.60	0.7005	0.8571	0.7709	0.7122	100
FKLSA (ProbIDF)	83.61	0.8024	0.8323	0.8171	0.8390	100
FKLSA (Entropy)	67.21	0.6096	0.7081	0.6552	0.6439	100
FKLSA (IGFI)	78.96	0.7692	0.7453	0.7571	0.8244	100
FKLSA (IGFS)	79.78	0.7736	0.7640	0.7688	0.8244	100
FKLSA (IGFF)	81.15	0.7875	0.7826	0.7850	0.8341	100

Table 6

MuchMore springer datasets classification results on 150 topics

Method	Accuracy%	Precision	Recall	F1 Score	Specificity	# topics
LSA	48.91	0.4261	0.4658	0.4451	0.5073	150
LDA	52.73	0.4610	0.4410	0.4508	0.5951	150
FKLSA (Normal)	81.69	0.8176	0.7516	0.7832	0.8683	150
FKLSA (IDF)	81.15	0.7584	0.8385	0.7965	0.7902	150
FKLSA (Unary)	74.59	0.6560	0.8882	0.7546	0.6341	150
FKLSA (IDF Smooth)	83.06	0.7797	0.8571	0.8166	0.8098	150
FKLSA (IDF Max)	74.59	0.6560	0.8882	0.7546	0.6341	150
FKLSA (ProbIDF)	79.51	0.7416	0.8199	0.7788	0.7756	150
FKLSA (Entropy)	68.31	0.6098	0.7764	0.6831	0.6098	150
FKLSA (IGFI)	81.15	0.8194	0.7329	0.7738	0.8732	150
FKLSA (IGFS)	78.14	0.7682	0.7205	0.7436	0.8293	150

Table 7

Ohsumed datasets classification results on 50 topics

Method	Accuracy%	Precision	Recall	F Score	Specificity	# topics
LSA	55.26	0.6701	0.6824	0.6762	0.2707	50
LDA	56.24	0.6730	0.7021	0.6872	0.2593	50
FKLSA (Normal)	74.39	0.8167	0.8071	0.8119	0.6068	50
FKLSA (IDF)	75.11	0.8195	0.8163	0.8179	0.6097	50
FKLSA (Unary)	69.54	0.7801	0.7730	0.7765	0.5271	50
FKLSA (IDF Smooth)	76.64	0.8311	0.8268	0.8289	0.6353	50
FKLSA (IDF Max)	69.27	0.7749	0.7769	0.7759	0.5100	50
FKLSA (ProbIDF)	74.21	0.8171	0.8031	0.8101	0.6097	50
FKLSA (Entropy)	62.71	0.7205	0.7441	0.7321	0.3732	50
FKLSA (IGFI)	56.33	0.6865	0.6667	0.6764	0.3390	50
FKLSA (IGFS)	57.14	0.6853	0.6916	0.6884	0.3105	50
FKLSA (IGFF)	57.59	0.6878	0.6969	0.6923	0.3134	50

Table 8

Ohsumed datasets classification results on 100 topics

Method	Accuracy%	Precision	Recall	F Score	Specificity	# topics
LSA	57.86	0.6881	0.7034	0.6957	0.3077	100
LDA	59.93	0.7047	0.7139	0.7093	0.3504	100
FKLSA (Normal)	77.09	0.8322	0.8333	0.8328	0.6353	100
FKLSA (IDF)	75.20	0.8231	0.8123	0.8177	0.6211	100
FKLSA (Unary)	70.35	0.7967	0.7612	0.7785	0.5783	100
FKLSA (IDF Smooth)	76.01	0.8349	0.8097	0.8221	0.6524	100
FKLSA (IDF Max)	69.90	0.7929	0.7585	0.7753	0.5698	100
FKLSA (ProbIDF)	74.21	0.8197	0.7992	0.8093	0.6182	100
FKLSA (Entropy)	64.33	0.7386	0.7415	0.7400	0.4302	100
FKLSA (IGFI)	64.87	0.7444	0.7415	0.74290	0.4473	100
FKLSA (IGFS)	59.30	0.7191	0.6654	0.6912	0.4359	100
FKLSA (IGFF)	58.40	0.6991	0.6890	0.6940	0.3561	100

Table 9

Ohsumed datasets classification results on 150 topics

Method	Accuracy%	Precision	Recall	F Score	Specificity	# topics
LSA	58.4007	0.7007	0.6850	0.6928	0.3647	150
LDA	60.7367	0.7124	0.7152	0.7138	0.3732	150
FKLSA (Normal)	73.41	0.8003	0.8150	0.8075	0.5584	150
FKLSA (IDF)	75.38	0.8065	0.8425	0.8241	0.5613	150
FKLSA (Unary)	70.44	0.7676	0.8150	0.7906	0.4644	150
FKLSA (IDF Smooth)	75.38	0.8096	0.8373	0.8232	0.5726	150
FKLSA (IDF Max)	69.90	0.7659	0.8071	0.7859	0.4644	150
FKLSA (ProbIDF)	73.58	0.7985	0.8215	0.8098	0.5499	150
FKLSA (Entropy)	66.22	0.7377	0.7861	0.7611	0.3932	150
FKLSA (IGFI)	64.96	0.7302	0.7743	0.7516	0.3789	150
FKLSA (IGFS)	65.86	0.7341	0.7861	0.7592	0.3818	150
FKLSA (IGFF)	62.17	0.7161	0.7415	0.7286	0.3618	150

FKLSA classification performance in terms of accuracy, precision, recall, F1-score and specificity with a different number of topics is highest and better than LDA and LSA topic models.

4.3 Documents clustering

The second performance of clustering is checked on unlabeled health news tweets dataset. Documents clustering has performed on P (T|D) and k-means clustering technique is used. There are two methods for clustering, external and internal validation methods. The internal validation method is more precise as compared to external validation one [63]. The different numbers of topics and clusters are evaluated using Calinski-Har-abasz index internal validation method. Calinsiki-Har-abasz (CH) index [64] is a popular internal validation method. CH index is the ratio type of index in which cohesion is estimated on the distance of clusters point to the centroid, as shown in Equation 19. $CH (C) = \frac{(N - K)}{K - 1} \frac{\sum_{ok} \in C | C_{k} | d_{e} | ({\bar{C}}_{k}, \bar{X})}{\sum_{ok} \in C \sum_{xi} C_{k} d_{e} (x^{i}, {\bar{C}}_{k})}$ (19)

The Calinsiki-Har-abasz index can evaluate all clusters validity on the sum of squared error clusters average. Higher Calinsiki-Har-abasz index shows the better clustering results. Calinsiki-Har-abasz index obtains best results on clusters [65].

The performance of FKLSA compared with LSA and LDA model with 2 to 10 numbers of clusters on 25, 50, 75 100,125,150 number of topics. The result of CH index indicates that the proposed technique’s clustering performance is higher than LDA and LSA topic models with different features and clusters, as shown in Figs. 5 –10.

Fig. 5

Calinski-Harabasz for 25 topics of tweets datasets.

Fig. 6

Calinski-Harabasz for 50 topics of tweets datasets.

Fig. 7

Calinski-Harabasz for 75 topics of tweets datasets.

Fig. 8

Calinski-Harabasz for 100 topics of tweets datasets.

Fig. 9

Calinski-Harabasz for 125 topics of tweets datasets.

Fig. 10

Calinski-Harabasz for 150 topics of tweets datasets.

In Figs. 5 –10 the LSA seems near to zero level because CH index value of LSA is lower than FKLSA. The clustering performance of FKLSA is better than the LSA.

4.4 Redundancy issue

This experiment analyzed the impact of redundancy issue using Synthetic WSJ redundant corpora. FKLSA has compared with LDA and RedLDA that was developed for handling redundancy issue in medical documents [31]. LDA, RedLDA, and FKLSA are trained on Synthetic WSJ redundant corpora for performance comparison of these models. Log-likelihood is used for test of evaluation of models. The higher score of log-likelihood indicates better performance of generalization and the topic model is more successes in modeling documents structure for input text corpus. Figure 11 shows Log-likelihood for Synthetic WSJ redundant corpora with a different number of topics in a range of 50 to 450. The experiment shows that FKLSA log-likelihood scores are higher than LDA and RedLDA for the redundant documents. Therefore, FKLSA performance for redundant text documents is greater than LDA and RedLDA.

Fig. 11

Likelihood comparison of WSJ corpora.

4.5 Time execution

FKLSA’s speed is compared with LDA and LSA using health news tweets big dataset in this research. The process of topic modeling based on probability distribution of topics and words with higher probability in each topic for posterior distribution. LDA approximate method is Gibbs sampling for the experiment. This algorithm needs multiple numbers of iterations, which increases computational cost with the amount of words, topics, and documents. Figure 12 shows that FKLSA’s time performance is stable even with an increasing number of topics and better than LDA and LSA.

Fig. 12

Time Execution comparison for twitter health dataset.

5 Discussion

The experimental results indicate that classification performance of FKLSA is higher than LDA and LSA with different numbers of topics. In clustering results CH-index of FKLSA with various clusters is higher than LDA and LSA. Therefore, clustering performance of FKLSA is better than others topic models LDA and LSA. The redundancy issue is also analyzed the FKLSA against LDA and RedLDA. The log-likelihood scores of FKLSA is better than LDA and RedLDA for redundant corpus. However, execution time of FKLSA is stable with different numbers for topics as compared to LDA and LSA.

6 Example

Medical and health collection of documents contains several words. FKLSA can execute on different numbers of topics to estimate the performance of the proposed technique. FKLSA discovered the more precise topics from health and medical documents. In this example, four topics are showing in Tables 10 and 11. Topics (T7 to T10) and topics (T5 to T8) are extracted from two datasets. These four topics are discovered from the Ohsumed and Twitters health news tweets dataset. Each topic contains ten different words.

Table 10
Example of four topics for Ohsumed dataset

T7 T8 T9 T10

Ulcer Injure Pylori Treatment

Diagnosis Neutrophil Gastric Therapy

Abnormal Acid Infect Antibiotic

Disease Function Biopsy Treat

Bacteria Muscle Chronic Oral

Scan Trauma Urine Drug

Common Defect Mucosa Effect

Physician Diet Cancer Penicillin

Clinic Burn Helicobacter Amphotericin

Discord Solute Urinary Intravenous

T7	T8	T9	T10
Ulcer	Injure	Pylori	Treatment
Diagnosis	Neutrophil	Gastric	Therapy
Abnormal	Acid	Infect	Antibiotic
Disease	Function	Biopsy	Treat
Bacteria	Muscle	Chronic	Oral
Scan	Trauma	Urine	Drug
Common	Defect	Mucosa	Effect
Physician	Diet	Cancer	Penicillin
Clinic	Burn	Helicobacter	Amphotericin
Discord	Solute	Urinary	Intravenous

Table 11

Example of four topics for twitter health dataset

T5	T6	T7	T8
Disorder	People	Moring	St
Workout	Think	Weekend	Rt
Reduce	Professional	Coffee	Ebola
Symptoms	Domestic	Sunshine	Health
Signs	Amazing	Headlines	Cancer
Groups	Stop	Shine	Risk
Feeling	Market	Fridays	Buck
Allergies	Hiring	Holidays	Care
Anxiety	Guy	Season	Amp
Tone	Mother	Ordering	Freak

7 Conclusions

The medical and health text collection of documents is continuously increasing nowadays, and analysis of these documents is very important for getting the valuable resource of information. The archives of medical and health like PubMed are providing valuable services in the scientific community.

Topic Modeling is a popular method that discovers the hidden theme and structure in unorganized medical and health documents. The structure has used for searching, indexing, browsing and summarizing these documents.

In machine learning, fuzzy techniques are widely used for health and medical image processing and text processing. Existing topic modeling techniques are based on linear algebra and statistical distribution approaches.

In this research, FKLSA topic modeling technique is developed and evaluated for discovering the latent semantic themes in medical and health documents. FKLSA avoids the negative effect of redundancy in medical and health documents. Experimental results show that time performance of FKLSA is stable with increasing number of topics. Furthermore, FKLSA also improves the classification accuracy for medical and health datasets and provides a new approach for text mining over medical and health datasets.

FKLSA is a new topic model for researchers and practitioners that has the flexibility to work with an extensive variety of fuzzy clustering and dimension reduction techniques. FKLSA used for discovering the themes from medical and health text documents efficiently. Additionally, FKLSA works with discrete and continuous data and estimates the number of topics in medical documents.

Quantitative evaluation of four datasets shows that FKLSA outperforms progressive baselines with vital enhancements. The experimental results suggest that FKLSA is a strong method that identifies the hidden structure in health and medical dataset. Results also show that FKLSA’s classification and clustering performance is higher than state-of-the-art baselines topic models like LDA and LSA.

Footnotes

References

E. National Academies of Sciences and Medicine, Future directions for NSF advanced computing infrastructure to support US science and engineering in 2017-2020: National Academies Press, 2016.

Karami

, Gangopadhyay

, Zhou

and Karrazi

, Flatm: A fuzzy logic approach topic model for medical documents, in Fuzzy Information Processing Society (NAFIPS) Held Jointly with 2015 5th World Conference on Soft Computing (WConSC), 2015 Annual Conference of the North American, 2015, pp. 1–6.

Tutubalina

, Miftahutdinov

Z.S.

, Nugmanov

, Madzhidov

, Nikolenko

, Alimova

, et al., Using semantic analysis of texts for the identification of drugs with similar therapeutic effects, Russian Chemical Bulletin 66 (2017), 2180–2189.

Kryszkiewicz

, Rough set approach to incomplete information systems, Information Sciences 112 (1998), 39–49.

Fatimah

, Rosadi

, Hakim

R.F.

and Alcantud

J.C.R.

, N-soft sets and their decision making algorithms, Soft Computing 22 (2018), 3829–3842.

Dhillon

I.S.

, Co-clustering documents and words using bipartite spectral graph partitioning, in Proceedings of the SeventhACMSIGKDDInternational Conference on Knowledge Discovery and Data Mining, 2001, pp. 269–274.

Aggarwal

C.C.

, Zhai

, An introduction to text mining, in Mining Text Data, ed: Springer, 2012, pp. 1–10.

Wei

and Croft

W.B.

, LDA-based document models for ad-hoc retrieval, in Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2006, pp. 178–185.

Wang

and Blei

D.M.

, Collaborative topic modeling for recommending scientific articles, in Proceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2011, pp. 448–456.

10.

Blei

D.M.

, Ng

A.Y.

and Jordan

M.I.

, Latent dirichlet allocation, Journal of Machine Learning Research 3 (2003), 993–1022.

11.

Cohen

, Elhadad

and Elhadad

, Redundancy in electronic health record corpora: Analysis, impact on text mining performance and mitigation strategies, BMC Bioinformatics 14 (2013), 10.

12.

Wrenn

J.O.

, Stein

D.M.

, Bakken

and Stetson

P.D.

, Quantifying clinical narrative redundancy in an electronic health record, Journal of the American Medical Informatics Association 17 (2010), 49–53.

13.

Hotho

, Nürnberger

and Paaß

, A brief survey of text mining, in Ldv Forum, 2005, pp. 19–62.

14.

Karami

and Zhou

, Exploiting latent content based features for the detection of static sms spams, Proceedings of the American Society for Information Science and Technology 51 (2014), 1–4.

15.

, Wang

, Hua

X.-S.

and Li

, Tag refinement by regularized LDA, in Proceedings of the 17th ACMInternational Conference on Multimedia, 2009, pp. 573–576.

16.

Uzuner

, Mailoa

, Ryan

and Sibanda

, Semantic relations for problem-oriented medical records, Artificial Intelligence in Medicine 50 (2010), 63–73.

17.

Kaur

and Wasan

S.K.

, Empirical study on applications of data mining techniques in healthcare, Journal of Computer Science 2 (2006), 194–200.

18.

Jin

, Ma

and Li

, Medical Record Text Analysis Based on Latent Semantic Analysis, in Computational Intelligence and Design (ISCID), 2015 8th International Symposium on, 2015, pp. 108–110.

19.

Griffiths

T.L.

and Steyvers

, Finding scientific topics, Proceedings of the National academy of Sciences 101 (2004), 5228–5235.

20.

Aryal

, Gallivan

and Tao

Y.Y.

, Using Latent Semantic Analysis to Identify Themes in IS Healthcare Research, 2015.

21.

Sarioglu

, Choi

H.-A.

and Yadav

, Clinical report classification using natural language processing and topic modeling, in Machine Learning and Applications (ICMLA), 2012 11th International Conference on, 2012, pp. 204–209.

22.

Arnold

C.W.

, El-Saden

S.M.

, Bui

A.A.

and Taira

, Clinical case-based retrieval using latent topic analysis, in AMIA Annual Symposium Proceedings, 2010, p. 26.

23.

Asou

and Eguchi

, Predicting protein-protein relationships from literature using collapsed variational latent dirichlet allocation, in Proceedings of the 2nd International Workshop on Data and Text Mining in Bioinformatics, 2008, pp. 77–80.

24.

Arnold

and Speier

, A topic model of clinical reports, in Proceedings of the 35th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2012, pp. 1031–1032.

25.

Dawson

J.A.

and Kendziorski

, Survival-supervised latent Dirichlet allocation models for genomic analysis of time-to-event outcomes, arXiv preprint arXiv:1202.5999, 2012.

26.

Perotte

A.J.

, Wood

, Elhadad

and Bartlett

, Hierarchically supervised latent Dirichlet allocation, in Advances in Neural Information Processing Systems, 2011, pp. 2609–2617.

27.

Bisgin

, Liu

, Fang

, Xu

and Tong

, Mining FDA drug labels using an unsupervised learning technique-topic modeling, in BMC Bioinformatics, 2011, p. S11.

28.

Huang

, Dong

, Duan

and Li

, Similarity measure between patient traces for clinical pathway analysis: Problem, method, and applications, IEEE Journal of Biomedical and Health Informatics 18 (2014), 4–14.

29.

Chen

J.H.

, Goldstein

M.K.

, Asch

S.M.

, Mackey

and Altman

R.B.

, Predicting inpatient clinical order patterns with probabilistic topic models vs conventional order sets, Journal of the American Medical Informatics Association 24 (2017), 472–480.

30.

Defossez

, Rollet

, Dameron

and Ingrand

, Temporal representation of care trajectories of cancer patients using data from a regional information system: An application in breast cancer, BMC Medical Informatics and Decision Making 14 (2014), 24.

31.

Cohen

, Aviram

, Elhadad

and Elhadad

, Redundancy-aware topic modeling for patient record notes, PloS One 9 (2014), e87555.

32.

Chiang

I.-J.

, Liu

C.C.-H.

, Tsai

Y.-H.

and Kumar

, Discovering latent semantics in web documents using fuzzy clustering, IEEE Transactions on Fuzzy Systems 23 (2015), 2122–2134.

33.

Cavaliere

, Senatore

and Loia

, Context-aware profiling of concepts from a semantic topological space, Knowledge-Based Systems 130 (2017), 102–115.

34.

Xie

X.L.

and Beni

, A validity measure for fuzzy clustering, IEEE Transactions on Pattern Analysis & Machine Intelligence (1991), 841–847.

35.

Naranjo

C.A.

, Bremner

K.E.

, Bazoon

and Turksen

I.B.

, Using fuzzy logic to predict response to citalopram in alcohol dependence, Clinical Pharmacology & Therapeutics 62 (1997), 209–224.

36.

Di Lascio

, Gisolfi

, Albunia

, Galardi

and Meschi

, A fuzzy-based methodology for the analysis of diabetic neuropathy, Fuzzy Sets and Systems 129 (2002), 203–228.

37.

Zahlmann

, Kochner

, Ugi

, Schuhmann

, Liesenfeld

, Wegner

, et al., Hybrid fuzzy image processing for situation assessment [diabetic retinopathy], IEEE Engineering in Medicine and Biology Magazine 19 (2000), 76–83.

38.

Helgason

C.M.

and Jobe

T.H.

, The fuzzy cube and causal efficacy: Representation of concomitant mechanisms in stroke, Neural Networks 11 (1998), 549–555.

39.

Helgason

C.M.

and Jobe

T.H.

, Causal interactions, fuzzy sets and cerebrovascular ‘accident’: The limits of evidence-based medicine and the advent of complexity-based medicine, Neuroepidemiology 18 (1999), 64–74.

40.

Helgason

C.M.

, Malik

, Cheng

S-C

, Jobe

T.H.

and Mordeson

J.N.

, Statistical versus fuzzy measures of variable interaction in patients with stroke, Neuroepidemiology 20 (2001), 77–84.

41.

Hassanien

A.E.

, Intelligent data analysis of breast cancer based on rough set theory, International Journal on Artificial Intelligence Tools 12 (2003), 465–479.

42.

Papageorgiou

E.I.

, Stylios

C.D.

and Groumpos

P.P.

, An integrated two-level hierarchical system for decision making in radiation therapy based on fuzzy cognitive maps, IEEE Transactions on Biomedical Engineering 50 (2003), 1326–1339.

43.

Moon

W.K.

, Chang

S.-C.

, Huang

C.-S.

and Chang

R.-F.

, Breast tumor classification using fuzzy clustering for breast elastography, Ultrasound in Medicine & Biology 37 (2011), 700–708.

44.

Gasch

A.P.

and Eisen

M.B.

, Exploring the conditional coregulation of yeast gene expression through fuzzy k-means clustering,research , Genome Biology 3 (2002), research0059. 1.

45.

Zhang

, Jin

and Zhou

Z.-H.

, Understanding bag-of-words model: A statistical framework, International Journal of Machine Learning and Cybernetics 1 (2010), 43–52.

46.

McCarthy

and Carroll

, Disambiguating nouns, verbs, and adjectives using automatically acquired selectional preferences, Computational Linguistics 29 (2003), 639–654.

47.

Salton

and Buckley

, Term-weighting approaches in automatic text retrieval, Information Processing & Management 24 (1988), 513–523.

48.

Agrawal

and Phatak

, A novel algorithm for automatic document clustering, in Advance Computing Conference (IACC), 2013 IEEE 3rd International, 2013, pp. 877–882.

49.

Choi

and Kim

, Automatic image annotation using semantic text analysis, in International Conference on Availability, Reliability, and Security, 2012, pp. 479–487.

50.

Huang

, Fu

and Chen

, Text-based video content classification for online video-sharing sites, Journal of the American Society for Information Science and Technology 61 (2010), 891–906.

51.

Iezzi

D.F.

, Centrality measures for text clustering, Communications in Statistics-Theory and Methods 41 (2012), 3179–3197.

52.

Gayathri

and Marimuthu

, Text document preprocessing with the KNN for classification using the SVM, in Intelligent Systems and Control (ISCO), 2013 7th International Conference on, 2013, pp. 453–457.

53.

Croft

W.B.

and Harper

D.J.

, Using probabilistic models of document retrieval without relevance information, Journal of Documentation 35 (1979), 285–295.

54.

Kolda

T.G.

, Limited-memory matrix methods with applications, 1998.

55.

Papineni

, Why inverse document frequency? in Proceedings of the second meeting of the North American Chapter of the Association for Computational Linguistics on Language Technologies, 2001, pp. 1–8.

56.

Dumais

, Enhancing performance in latent semantic indexing (LSI) retrieval, ed: Technical, 1992.

57.

Chisholm

and Kolda

T.G.

, New term weighting formulas for the vector space method in information retrieval, Computer Science and Mathematics Division, Oak Ridge National Laboratory 10 (1999), 5698.

58.

Abdi

and Williams

L.J.

, Principal component analysis, Wiley Interdisciplinary Reviews: Computational Statistics 2 (2010), 433–459.

59.

Gildea

, Corpus variation and parser performance, in Proceedings of the 2001 Conference on Empirical Methods in Natural Language Processing, 2001.

60.

Tsuruoka

, Tateishi

, Kim

J.-D.

, Ohta

, McNaught

, Ananiadou

, et al., Developing a robust part-of-speech tagger for biomedical text, in Panhellenic Conference on Informatics, 2005, pp. 382–392.

61.

Yang

and Jin

, Distance metric learning: A comprehensive survey, Michigan State Universiy 2 (2006), 4.

62.

Yang

, Slattery

and Ghani

, A study of approaches to hypertext categorization, Journal of Intelligent Information Systems 18 (2002), 219–241.

63.

Rendón

, Abundez

, Arizmendi

and Quiroz

E.M.

, Internal versus external cluster validation indexes, International Journal of computers and communications 5 (2011), 27–34.

64.

Caliński

and Harabasz

, A dendrite method for cluster analysis, Communications in Statistics-theory and Methods 3 (1974), 1–27.

65.

Hardy

, On the number of clusters, Computational Statistics & Data Analysis 23 (1996), 83–96.

A novel fuzzy k-means latent semantic analysis (FKLSA) approach for topic modeling over medical and health text corpora

Abstract

Keywords

1 Introduction

3 Research methodology

3.1 FKLSA

4.1 Datasets

4.2 Documents classification

Table 2 Confusion matrix Actual Predicted Negative Positive Negative TN FP Positive FN TP

6 Example

Footnotes

References

Table 2
Confusion matrix

Actual Predicted

Negative Positive

Negative TN FP

Positive FN TP