DTMiner: identification of potential disease targets through biomedical literature mining.

MOTIVATION: Biomedical researchers often search through massive catalogues of literature to look for potential relationships between genes and diseases. Given the rapid growth of biomedical literature, automatic relation extraction, a crucial technology in biomedical literature mining, has shown great potential to support research of gene-related diseases. Existing work in this field has produced datasets that are limited both in scale and accuracy. RESULTS: In this study, we propose a reliable and efficient framework that takes large biomedical literature repositories as inputs, identifies credible relationships between diseases and genes, and presents possible genes related to a given disease and possible diseases related to a given gene. The framework incorporates name entity recognition (NER), which identifies occurrences of genes and diseases in texts, association detection whereby we extract and evaluate features from gene-disease pairs, and ranking algorithms that estimate how closely the pairs are related. The F1-score of the NER phase is 0.87, which is higher than existing studies. The association detection phase takes drastically less time than previous work while maintaining a comparable F1-score of 0.86. The end-to-end result achieves a 0.259 F1-score for the top 50 genes associated with a disease, which performs better than previous work. In addition, we released a web service for public use of the dataset. AVAILABILITY AND IMPLEMENTATION: The implementation of the proposed algorithms is publicly available at http://gdr-web.rwebox.com/public_html/index.php?page=download.php The web service is available at http://gdr-web.rwebox.com/public_html/index.php CONTACT: jenny.wei@astrazeneca.com or kzhu@cs.sjtu.edu.cn Supplementary information: Supplementary data are available at Bioinformatics online.

Layout-aware information extraction from semi-structured medical images.

Textual information embedded in the medical image contains rich structured information about the medical condition of a patient. This paper aims at extracting structured textual information from semi-structured medical images. Given the recognized text spans of an image preprocessed by optical character recognition (OCR), due to the spatial discontinuity of texts spans as well as potential errors brought by OCR, the structured information extraction becomes more challenging. In this paper, we propose a domain-specific language, called ODL, which allows users to describe the value and layout of text data contained in the images. Based on the value and spatial constraints described in ODL, the ODL parser associates values found in the image with the data structure in the ODL description, while conforming to the aforementioned constraints. We conduct experiments on a dataset consisting of real medical images, our ODL parser consistently outperforms existing approaches in terms of extraction accuracy, which shows the better tolerance of incorrectly recognized texts, and positional variances between images. This accuracy can be further improved by learning from a few manual corrections.

Data-Driven Information Extraction from Chinese Electronic Medical Records.

OBJECTIVE: This study aims to propose a data-driven framework that takes unstructured free text narratives in Chinese Electronic Medical Records (EMRs) as input and converts them into structured time-event-description triples, where the description is either an elaboration or an outcome of the medical event. MATERIALS AND METHODS: Our framework uses a hybrid approach. It consists of constructing cross-domain core medical lexica, an unsupervised, iterative algorithm to accrue more accurate terms into the lexica, rules to address Chinese writing conventions and temporal descriptors, and a Support Vector Machine (SVM) algorithm that innovatively utilizes Normalized Google Distance (NGD) to estimate the correlation between medical events and their descriptions. RESULTS: The effectiveness of the framework was demonstrated with a dataset of 24,817 de-identified Chinese EMRs. The cross-domain medical lexica were capable of recognizing terms with an F1-score of 0.896. 98.5% of recorded medical events were linked to temporal descriptors. The NGD SVM description-event matching achieved an F1-score of 0.874. The end-to-end time-event-description extraction of our framework achieved an F1-score of 0.846. DISCUSSION: In terms of named entity recognition, the proposed framework outperforms state-of-the-art supervised learning algorithms (F1-score: 0.896 vs. 0.886). In event-description association, the NGD SVM is superior to SVM using only local context and semantic features (F1-score: 0.874 vs. 0.838). CONCLUSIONS: The framework is data-driven, weakly supervised, and robust against the variations and noises that tend to occur in a large corpus. It addresses Chinese medical writing conventions and variations in writing styles through patterns used for discovering new terms and rules for updating the lexica.