An efficient Gene-ICD cluster based classification framework for heterogeneous micro-array databases

As microarray datasets grow larger, identifying key features in these extensive feature spaces becomes increasingly complex due to the challenges of data size and sparsity. For scientific and biomedical researchers, the major challenge lies in ranking and classifying microarray features, which are characterized by high-dimensional feature spaces and limited sample sizes. Each microarray includes numerous identical DNA molecules, which help detect gene-related diseases. Techniques such as feature transformation, feature ranking, and data classification are crucial in effectively classifying high-dimensional data with high accuracy. Feature transformation normalizes data within specified ranges, improving feature ranking in large feature spaces. Traditional feature transformation methods, like log transformation and min-max normalization, generally ignore data distribution and outliers. Machine learning allows classifiers to learn decision-making rules from expert-labeled data, reducing costs and improving scalability compared to fully manual systems. Research on medical data classification largely focuses on binary classifiers, where a classifier is built from positive and negative examples to predict class membership. For multiclass datasets, separate binary classifiers are often constructed for each class, and their results are combined. Classification can be fully automated or use a hybrid approach involving human intervention. Microarray datasets, particularly in chronic diseases, evolve gradually, from mild symptoms to severe disease and even death. Medical datasets often consist of cancer patches and their related diseases, which are difficult for doctors to detect in high-risk patients. A patient's medical history and previous knowledge also aid in disease detection. Structural changes in the airways can result in airway remodeling, reducing luminal diameter and contributing to conditions like emphysema, a chronic respiratory disease where alveolar walls are destroyed without fibrosis.

The PubMed database currently holds around 29 million articles, and approximately 1 million new articles are added each year. A substantial amount of essential information related to proteins, drugs, diseases, and chemicals is available in unstructured formats. This exponential increase in the volume of documents makes it difficult to manually collect and organize biomedical information such as protein-protein, drug-drug, and chemical-protein interactions. Biomedical information extraction is a process designed to automatically detect biomedical concepts and their relationships using advanced language processing and machine learning tools. Each year, researchers and healthcare professionals publish large numbers of biomedical research articles, most of which are accessible online. These articles are valuable not only for biomedical scientists in their research but also for healthcare professionals in their clinical work. The volume of information continues to grow across various domains due to the increase in distributed biomedical repositories. Document preprocessing is used to reduce peer documents into summaries by selecting essential information from the source. However, this has resulted in information overload. To address this, multi-document clustering and feature extraction can reduce inter-cluster variation. This research employs a feature extraction strategy and key phrase clustering and pattern discovery to minimize redundancy across multiple original documents. Text classification is a technique that helps automatically assess the significance of a document. Additionally, identifying synonyms and abbreviations is prioritized, adding complexity to biomedical literature. Hence, more efficient methods are needed to extract biomedical information from the ever-growing pool of resources. A suitable mining approach is essential for uncovering various types of knowledge from biomedical literature, where term variation is common, including numbers, capital letters, hyphens, and other special characters within terms.

Currently, the PubMed database contains roughly 29 million articles, with around 1 million new ones being added every year. A significant amount of important knowledge about proteins, drugs, diseases, and chemicals is stored in an unstructured form. The rapid growth of these documents makes manually collecting and organizing information, such as protein-protein, drug-drug, and chemical-protein interactions, highly complex. Biomedical information extraction refers to the automated process of identifying biomedical concepts and their relationships using advanced natural language processing and machine learning methods. Every year, biomedical researchers and healthcare professionals publish large quantities of biomedical research articles, the majority of which are available online. These articles aid both scientists in their research and healthcare professionals in their practice. The sheer volume of information continues to grow across various domains due to the expansion of distributed biomedical repositories. Document preprocessing is used to reduce peer-reviewed documents into summaries by selecting critical information. However, this has created the problem of information overload. To mitigate this, multi-document clustering and feature extraction are used to reduce inter-cluster variation. This research focuses on feature extraction techniques combined with key phrase clustering and pattern discovery to eliminate redundancy in multiple original documents. Text classification automatically assesses a document’s importance. Additionally, emphasis is placed on identifying synonyms and term abbreviations, adding complexity to biomedical literature. As a result, there is a growing need for more efficient approaches to extract biomedical information from large sources. A suitable mining approach must be implemented to uncover diverse types of knowledge from biomedical literature, where there is significant variation in terms, including numbers, capital letters within words, hyphens, and special characters.

In the past, the accuracy of medical disease prediction has been significantly reduced when training datasets are small, due to class imbalance and high-dimensional data. In this method, each attribute is checked for missing values. A pattern mining and classification model was introduced for disease prediction using microarray datasets, where pathways were ranked, and disease-related patterns were identified. The random forest classification model was employed to filter and classify co-related disease patterns, but this approach requires considerable computational power and memory for larger datasets. Biomedical text data are a rich source of information. This paper described the use of the MapReduce method, a parallel and distributed programming paradigm, to mine relationships among various biomedical concepts extracted from literature. Initially, biomedical concepts are extracted using text matching with the Unified Medical Language System (UMLS), the most widely adopted standard biomedical database. The MapReduce method is then implemented to evaluate specific interestingness measures. He et al. introduced a graph-based model for unstructured biomedical text, which can be applied in several real-world applications. To address these challenges, two new approaches, RT-TNG, were proposed. Topic N-Grams plays an important role in assigning these models. The researchers also examined semantic shifts in biomedical literature, analyzing word semantic changes in the biomedical domain. They identified representative words based on frequency and topic probability distributions, showing how words can cluster with their semantic neighbors and coevolve, or drift apart over time. Functional annotation of genes is a critical process, as it helps in understanding gene relationships. Multiple gene functions are described using standardized vocabularies, known as bio-ontologies. Assigning bio-ontology terms to genes is done through approaches based on data mining and machine learning methods, such as maximum entropy and support vector machines. The goal of this study is to propose an alternative method for annotating genes, including the development of efficient classification schemes, validation models, and graphical representations of outcomes. Reducing dataset dimensions is also a key concern, with classification schemes relying on linear discriminant analysis and validation models based on statistical interpretations.

2.Related works

Biswal et al. proposed an innovative approach that integrates medical entity relationships with keyword-based search, extending the personalized PageRank algorithm. In this extended keyword-based search system, user preferences act as filters, limiting relations in linked data. In their research, they presented the PRRank algorithm, which uses all relations present in linked data, with suggestions for further expansion. They also proposed a document-based graphical method, which clusters sentences based on relationships and applies ranking at the document level. A query-sensitive rating method was introduced for graph-based classification, improving traditional models that only accounted for queries to sentence nodes. This method estimates sentence-to-sentence edges and calculates responses on demand in the graph model. A variation of TextRank was provided, using the shortest path for generating summaries. Initially, the graph model was designed to represent documents and their connected sentence entities with meaningful relationships. A weighted graph system was also introduced to rate phrases and sentences for document classification. The word embedding model implemented here relies on preprocessed patterns and strict sentence identification to highlight significant subjects across various document lengths. A three-phase feature extraction method was used, consisting of preprocessing, soft clustering, and feature extraction. The clustering algorithm in this model involves four steps: vector space model initialization, similarity rank matrix determination, parameter initialization, and iterative development. The document extraction model combined document clustering with feature extraction. Preprocessing involved comparing sentences, identifying document features, feature positions, and similarity functions. Interest in word embedding has led to comparative studies in recent years. Scheepers et al. compared Word2Vec, FastText, and GloVe, but noted bias due to differences in training datasets. They also compared these models using the BLEU score without conducting individual evaluations, focusing on whether semantic relationships are preserved. Beam et al. created large word embeddings using medical data, focusing on Word2Vec and GloVe. Their benchmark involved statistical co-occurrence of concepts. Similarly, Huang et al. examined Word2Vec on three medical systems but did not focus on semantic connections. Wang et al. compared word embedding training for medical NLP tasks, focusing on models trained with the same data format. Word embeddings are used for intrinsic and extrinsic bio-NLP tasks: intrinsic tasks assess semantic similarity between words, while extrinsic tasks include relation extraction and text classification. Chiu et al. found that smaller windows benefit extrinsic tasks, while larger windows are better for intrinsic tasks, a finding confirmed in our tests. HAL generates semantic co-occurrences from phrases, while GLSA computes term and document vectors using an LSA-based method

3. Proposed Model

The overall architecture of the proposed model is represented in fig 1. Initially, each microarray gene disease dataset is processed to find the synonym of the gene feature for efficient gene-symbol to gene-name mapping.  The overall architecture of the proposed model is represented in fig 1. Initially, each microarray gene disease dataset is filtered to fill the sparsity problem and missing values of the gene featured. Here, a hybrid data transformation approach is used to transform the feature values using the gaussian transformation measure. Each value is normalized to improve the balancing property of each feature and its class. In the initial phase, each feature is transformed using the gaussian transformation process. In the second phase, essential features are extracted using the hybrid PCA approach. Finally, an optimized decision tree classifier is designed to find the essential cancer patterns for prediction process. Proposed filtered based IPCA are integrated to improve the classification rate of the ensemble classification model with weak classifiers on high dimensional feature selection as shown in figure 1. Most of the ensemble classification technique is designed and implemented using the set of weak classifiers to optimize the overall classification rate and to minimize the error rate.

Figure 1: Multi-level Gene-Disease-Chemical drug classification and Ranking Framework

In the proposed approach, biomedical documents, gene disease database and chemical drug names are taken as input for biomedical document processing. Initially, biomedical documents are filtered using the Stanford parser in order to remove the noise and tokenization. Here, each document is converted to word2vector data for data normalization. A novel  rank similarity model is used to find the essential gene to disease patterns for chemical drug mapping.  Each microarray training dataset is pre-processed using the data transformation function to remove the variation among the data distribution.

High dimensional data transformation:

Input : Training dataset  D, F(S): Feature sets.

Output: Non-linear normalization values.

Procedure:

1.                   Input dataset D with feature space F(S).

2.                   To each feature in the feature set F(S)

3.                   perform

4.                        Compute non-linear normalization to each feature values in the feature space as          

5.                        If( >0.75)

6.                        Then

7.                                   Normalize each feature using Min-max normalization [ ,1]

8.                        Else

9.                                    Normalize each feature  using Min-max normalization with lower and upper bounds as 0 and 1.

10.                     End if

11.                Done

Proposed Algorithm 1: Biomedical document filtering

Input : Chemical drugs CD, Gene-Disease pattern GDP, Documents D.

Phase  1: Data Filtering on the gene-disease patterns and Biomedical documents.

Read gene-disease patterns GDP.

Read biomedical documents BD.

Read Chemical drugs CD.

for each gene-disease pattern g[i] in GDP

Do

     To each gene-chemical document di in D

                 Do

                Tok[]=NLPOTokenizer(di)

                To each gene-chemical tokens in Tok[]

                                Do

                                Apply Stanford NLP stopword  

            removal , stemming and other text

            preprocessing.

Gt[]={RemoveStopWords(g[i]),Remove

NonspecialChars[g[i],Tokenizer(g[i])

BDt[]={RemoveStopWords(g[i]),RemoveNonspecialChars[g[i],Tokenizer(g[i])

            Mapping (Gt,BDt) to DCi

                                Done

 Done

Done

Phase 1, describes the data preprocessing of the gene tokens and biomedical documents. Stanford NLP parser is used to filter the input documents. Stemming, stop word removal and tokenization are performed on the input documents for similarity computation. Proposed similarity computation is used to find the contextual relationship among the genes and disease patterns to the chemical symbols .

Contextual  Rank Similarity for the biomedical documents:

GloVe encodes significance in embedded space as vector offsets. In this Glove vector model, word co-occurrences are taken as vectors to find the main and contextual key word vectors for biomedical gene-disease relationships.

 Proposed Glove Optimization algorithm:

2. Bio-Gene Rank similarity measure for Glove Key-phrase extraction

Input : Glove scientific corpus main vector SC, Glove scientific corpus contextual features .

Step 1: Read Optimized glove feature values.

Step 2:  Let GM(_i) ß (w_1,w₂,….w_i) represents glove key main feature i.

GC(_j) ß (w_1,w₂,…w_j) represents glove contextual feature j.

Where   i ≠ j

In this bio-gene rank similarity measure , the similarity between the each document and the optimal glove features are computed to each document. Here, each document with highest similarity value is taken as highest ranked document in the context of genes/chemicals/proteins.

Algorithm :TopKPCA (TKPCA):

Improved EM Gene-disease clustering approach :

In the expectation maximization model, two phases are implemented on the training data to predict the best clustered features for the gene-disease prediction.

Proposed Classification Algorithm

1: Read pre-processing gene-disease patterns, training cancer datasets, gene database. All these input patterns are partitioned ‘m’ clusters based on EM approach.

2: To each clustered  

3: do

4: Apply proposed ensemble decision tree model on each cluster data.

5: In the proposed classification model, a novel gene-disease feature selection measure is implemented on each cluster.

Proposed decision tree gene-disease feature selection measure

Statistical Hoeffding entropy measure

Let  D_p represents the clustered gene-disease probabilistic patterns for the decision tree classification problem. The hoeffding entropy of the gene-disease pattern analysis is given by

2. Proposed Random forest Micro-array feature selection

4. Experimental Results

Experimental results are performed on real-time medline biomedical documents and micro-array datasets. Proposed feature selection-based ensemble methods increase the efficiency of the F-measure , recall and accuracy on  high dimensional datasets. Proposed model uses the entire training data set for construction of decision patterns; therefore, the prediction accuracy of each cross validation tends to be more accurate than the traditional ensemble classification models. Simulation results represent the proposed ensemble classification improves the overall true positive and false negative rate. Also, the main advantage of using proposed model is to reduce the error rate on high dimensional features. Different types of cancer datasets and its types are presented in below table.

Experimental results are implemented in java environment with third party libraries for word embedding and NLP pre-processing. The minimum hardware configuration include amazon AWS large instance with 32GB RAM.

Table 1: Microarray  datasets used

In this experimental study, Glove word embedding model is developed to improve the efficiency of biomedical disease prediction. Glove model contains 5 billion vocabulary tokens in order to find and extract the key terms in the biomedical disease document sets. Glove provides the words in vector format.

The following table represents the different gene-disease patterns and its probabilistic scores.

Figure 2: Comparison of proposed Glove features

to the conventional feature extraction models on

ChEBI biomedical  document sets.

Figure 2, illustrates the performance of feature extraction using the proposed  Glove  approach on ChEBI document sets. From the figure 2, it is noted that the proposed feature extraction has high filtering rate as compared to the existing approaches.

Table 2: Candidate  features extraction using the proposed Glove model on PHAEDRA dataset

Table2, illustrates the performance of  feature extraction using the proposed glove approach on PHAEDRA datasets. From the table2, it is clearly shown that the present feature extraction procedure has high filtering rate as compared to the existing approaches.

Figure  3: Performance analysis of average computational runtime(ms) with different traditional feature extraction models

Figure 3 describes the performance of the various feature extraction model's average computer runtime(s) for the proposed approach on all datasets. Figure 3 shows clearly that the current function extraction process has low calculation runtime in comparison with existing approaches.

Figure 4: Proposed real-time application of gene-disease mapping and its probabilistic scores

Figure 5: Proposed real-time application of all gene-disease mapping and its probabilistic scores

Leukemia Data

Table. 3 Datasets and Its Characteristics

Table 4: Comparative analysis of proposed caner classification approach to the traditional approaches  by using accuracy on various realtime  cancer gene-disease patterns evaluation.

Table 4, describes the performance of the proposed model on cancer gene-disease realtime  datasets. Here, all the cancer datasets are evaluated using the proposed model to find the average true positive rate and precision rate on the high dimensional datasets. From the table, it is visualized that the present approach has better true positive rate and precision over the existing models.

Table 5:  Performance comparison of proposed classification accuracy  to the conventional approaches by using accuracy on various average accuracy of DLBCL, Prostate, Lymphoma, BreastCancer

gene-disease patterns.

Table 5, describes the performance of the proposed model on all gene-disease real-time  datasets. Here, all the datasets are evaluated using the proposed model to find the average true positive rate and precision rate on the high dimensional datasets. From the table, it is visualized that the present approach has better true positive rate and precision over the existing models.

Table 6:  Comparative runtime  of present approach to the traditional approaches  by using accuracy on various real-time average accuracy of DLBCL, Prostate, Lymphoma, BreastCancer  gene-disease patterns.

ICD-Gene related somatic cancer relations:

The above table describes the somatic germline related ICD codes and its description for the classification problem. Here, these ICD codes are used to find the relationships between the genes and the disease. The following table describes the somatic germline mutations for cancer prediction.

The following patterns describes the somatic mutations for caner disease prediction using the ICD codes and the gene patterns. Here, each pattern defines the somatic feature and its relationship with the cancer disease class label for decision making.

In the above results, experimental results are performed on the somatic germline cancer mutation dataset using ICD and gene sets. From the experimental results it is observed that the proposed somatic germline mutation cancer prediction has better efficiency nearly 99%  accuracy than the conventional cancer detection models with ICD and gene feature selection.

 In this paper, an advanced micro-array cancer disease based biomedical document ranking is implemented on large biomedical document sets. Most of the existing models are independent of biomedical document ranking based on micro-array gene sets. In order to overcome these issues, an advanced feature selection-based classification learning model is proposed to overcome to problem of gene based biomedical document ranking. a hybrid word embedding method and similarity metric is used to improve the efficiency of the contextual similarity between the document sets. These biomedical key entities are used to map the micro-array data classification patterns for gene to ICD mapping process. In this work, a hybrid classification model is implemented on micro-array cancer datasets to find the essential patterns for the biomedical document mapping process. Experimental results are simulated on different biomedical document sets on the microarray datasets. Proposed results proved that the gene-chemical disease clustered based classification framework has better optimization than the conventional biomedical document embedding methods, similarity metrics and classification models.

[1]B. Hosseini and K. Kiani, “A big data driven distributed density based hesitant fuzzy clustering using Apache spark with application to gene expression microarray,” Engineering Applications of Artificial Intelligence, vol. 79, pp. 100–113, Mar. 2019, doi: 10.1016/j.engappai.2019.01.006.

[2]M. Daoud and M. Mayo, “A survey of neural network-based cancer prediction models from microarray data,” Artificial Intelligence in Medicine, vol. 97, pp. 204–214, Jun. 2019, doi: 10.1016/j.artmed.2019.01.006.

[3]A. K. Shukla, P. Singh, and M. Vardhan, “A two-stage gene selection method for biomarker discovery from microarray data for cancer classification,” Chemometrics and Intelligent Laboratory Systems, vol. 183, pp. 47–58, Dec. 2018, doi: 10.1016/j.chemolab.2018.10.009.

[4]J. Lee, I. Y. Choi, and C.-H. Jun, “An efficient multivariate feature ranking method for gene selection in high-dimensional microarray data,” Expert Systems with Applications, vol. 166, p. 113971, Mar. 2021, doi: 10.1016/j.eswa.2020.113971.

[5]X. Liu, S.-C. Lee, G. Casella, and G. F. Peter, “Assessing agreement of clustering methods with gene expression microarray data,” Computational Statistics & Data Analysis, vol. 52, no. 12, pp. 5356–5366, Aug. 2008, doi: 10.1016/j.csda.2008.06.004.

[6]O. Stoss and T. Henkel, “Biomedical marker molecules for cancer – current status and perspectives,” Drug Discovery Today: TARGETS, vol. 3, no. 6, pp. 228–237, Dec. 2004, doi: 10.1016/S1741-8372(04)02459-4.

[7]F. Zhu et al., “Biomedical text mining and its applications in cancer research,” Journal of Biomedical Informatics, vol. 46, no. 2, pp. 200–211, Apr. 2013, doi: 10.1016/j.jbi.2012.10.007.

[8]Y. He, W. Pan, and J. Lin, “Cluster analysis using multivariate normal mixture models to detect differential gene expression with microarray data,” Computational Statistics & Data Analysis, vol. 51, no. 2, pp. 641–658, Nov. 2006, doi: 10.1016/j.csda.2006.02.012.

[9]R. Wang, L. Scharenbroich, C. Hart, B. Wold, and E. Mjolsness, “Clustering analysis of microarray gene expression data by splitting algorithm,” Journal of Parallel and Distributed Computing, vol. 63, no. 7, pp. 692–706, Jul. 2003, doi: 10.1016/S0743-7315(03)00085-6.

[10]S. Bruhn et al., “Combining gene expression microarray- and cluster analysis with sequence-based predictions to identify regulators of IL-13 in allergy,” Cytokine, vol. 60, no. 3, pp. 736–740, Dec. 2012, doi: 10.1016/j.cyto.2012.08.009.

[11]B. S. Biswal, A. Mohapatra, and S. Vipsita, “Ensemble Neighborhood Search (ENS) for biclustering of gene expression microarray data and single cell RNA sequencing data,” Journal of King Saud University - Computer and Information Sciences, Dec. 2019, doi: 10.1016/j.jksuci.2019.11.011.

[12]S. R. Kumaran, M. S. Othman, L. M. Yusuf, and A. Yunianta, “Estimation of Missing Values Using Hybrid Fuzzy Clustering Mean and Majority Vote for Microarray Data,” Procedia Computer Science, vol. 163, pp. 145–153, Jan. 2019, doi: 10.1016/j.procs.2019.12.096.

[13]C. L. Clayman, S. M. Srinivasan, and R. S. Sangwan, “K-means Clustering and Principal Components Analysis of Microarray Data of L1000 Landmark Genes,” Procedia Computer Science, vol. 168, pp. 97–104, Jan. 2020, doi: 10.1016/j.procs.2020.02.265.

[14]M. A. Hambali, T. O. Oladele, and K. S. Adewole, “Microarray cancer feature selection: Review, challenges and research directions,” International Journal of Cognitive Computing in Engineering, vol. 1, pp. 78–97, Jun. 2020, doi: 10.1016/j.ijcce.2020.11.001.

[15]J. Cui, N. Zhang, Y. Liu, L. Zhang, C. Gao, and S. Liu, “Microarray gene expression profiling provides insights into functions of TIPE2 in HBV-related apoptosis,” Molecular Immunology, vol. 131, pp. 137–143, Mar. 2021, doi: 10.1016/j.molimm.2020.12.031.

[16]D. R. Rhodes et al., “ONCOMINE: A Cancer Microarray Database and Integrated Data-Mining Platform,” Neoplasia, vol. 6, no. 1, pp. 1–6, Jan. 2004, doi: 10.1016/S1476-5586(04)80047-2.

[17]S. Sarbazi-Azad, M. S. Abadeh, and M. E. Mowlaei, “Using Data Complexity Measures and an Evolutionary Cultural Algorithm for Gene Selection in Microarray Data,” Soft Computing Letters, p. 100007, Oct. 2020, doi: 10.1016/j.socl.2020.100007.

[18]Y. Wang et al., “A comparison of word embeddings for the biomedical natural language processing,” Journal of Biomedical Informatics, vol. 87, pp. 12–20, Nov. 2018, doi: 10.1016/j.jbi.2018.09.008.

[19]D. Zhao, J. Wang, Y. Chu, Y. Zhang, Z. Yang, and H. Lin, “Improving Biomedical Word Representation with Locally Linear Embedding,” Neurocomputing, Mar. 2021, doi: 10.1016/j.neucom.2021.02.071.

[20]M. Moradi, M. Dashti, and M. Samwald, “Summarization of biomedical articles using domain-specific word embeddings and graph ranking,” Journal of Biomedical Informatics, vol. 107, p. 103452, Jul. 2020, doi: 10.1016/j.jbi.2020.103452.

[21]D. Dimitriadis and G. Tsoumakas, “Word embeddings and external resources for answer processing in biomedical factoid question answering,” Journal of Biomedical Informatics, vol. 92, p. 103118, Apr. 2019, doi: 10.1016/j.jbi.2019.103118.