Improved biomedical document retrieval system with PubMed term statistics and expansions

2010

IJDMB

Self Cite

We developed a new paradigm with the ultimate goal of enabling disease-specific drug candidate discovery with molecular-level evidences generated from literature and prior knowledge. We showed how to implement the paradigm by building a prototype literature-mining framework and performing drug-protein association mining for breast cancer drug discovery. In a molecular pharmacology study of breast cancer, 79.2% of 729 enriched drugs in 'Organic Chemicals' category were validated to be disease-related, and the remaining 20.8% were also investigated as potential for future molecular therapeutics studies. 'Doxorubicin', 'Etoposide' and 'Paclitaxel' were identified as having similar pharmacological profiles to treat breast cancer.

Section: Select Significant Breast Cancer Drugsmentioning

confidence: 99%

“…In this work, we first calculate a p-value for each drug term t j in T BR using methods described in Li and Chen (2009) and later derive its false discovery rate (FDR). Let the null hypothesis Ho be that document frequency of drug term t j in T BR comes from a random distribution T Random .…”

Section: Select Significant Breast Cancer Drugsmentioning

confidence: 99%

Discovering breast cancer drug candidates from biomedical literature

Zhu

2010

IJDMB

Self Cite

“…The system involving three major components in text mining performed the task of text classification. Li and Chen (2009) also described a new framework to improve the performance of medical document retrieval. Other systems as reported in Mamlin et al (2003) included MedLEE and LifeCode®.…”

Section: Expert Systems For Medical Data Analysismentioning

confidence: 99%

Indexing ICD-9 codes for free-textual clinical diagnosis records by a new ensemble classifier

Hsu

2009

IJCIBSB

An expert system which can support in indexing automatically an ICD-9 code for a clinical diagnosis record is necessary and in great demand for hospitals. The ICD-9 code determines reimbursement amount and an incorrect code may result in fining. In this paper we present a new expert system to index automatically an ICD-9 code with respect to two perspectives. First, the system analyses the free-textual medical documents as would be necessary to activate the uses of natural language process and text mining. A free-textual document has to be represented by a large number of vocabulary words as analogy with a high dimensional data vector. Second, we drive a new ensemble classifier which combines the uses of the majority voting approach with multiple learning algorithms and the boosting approach at the same time. The motivation is stimulated in that when a predicted ICD-9 code of a majority voting of a clinical diagnosis record is incorrect, the record needs to be trained more often. The experimental results show that the proposed ensemble technique is able to achieve simultaneously stability and performance in terms of classification accuracy.

“…We use a term frequency statistical method described by Li and Chen [16]. This method makes use of term statistical distribution from the entire PubMed abstracts to calculate p-value of each term's significance in being observed in any collection of retrieved PubMed abstracts.…”

Section: Build Biogists For Breast Cancer Related Drug Compoundsmentioning

confidence: 99%

“…In this work, we first calculate a p-value for each term in using methods described in [16] and later derive its false discovery rate. Let the null hypothesis Ho be that document frequency of term in come from a random distribution .…”

Section: Build Biogists For Breast Cancer Related Drug Compoundsmentioning

confidence: 99%

Mining disease-specific molecular association profiles from biomedical literature

Zhu

Proceedings of the 2008 ACM Symposium on Applied Computing

2008

Self Cite

We developed a new literature mining paradigm with the ultimate goal of enabling knowledge discovery in molecular association profiles generated from literature and prior knowledge. We show how to implement the paradigm by building a prototype literature mining framework and performing molecule-bioGist association mining. The framework consists of two modules. The first module, Textual Data Mining, takes the synonym-expanded disease-related molecule names and outputs a list of bioGist list. The second module, Structured Data Mining, takes two inputs, initial disease-related molecular query terms and extracted bioGist list from the first module, and outputs a molecule-bioGist association matrix. Our approach is novel because biomedical literature mining is used here not only as an "information retrieval" tool, but also as a "hypothesis generation and validation" platform. We applied the framework to a molecular pharmacology study of breast cancer. Based on 214 breast cancerrelated proteins, 429,067 MEDLINE abstracts were retrieved, and 4,491 drug compounds were identified as bioGists. We evaluated 172 hydrocarbons in the above bioGist list, and found that more than 82.5% hydrocarbons were verified to be related to breast cancer. BRCA1 and BRCA2 were found to have similar profiles in drug compound studies, whereas "doxorubicin", "etoposide", and "paclitaxel" were identified to have similar pharmacological profiles to treat breast cancer.