Publications by authors named "Claire Grover"
Article Synopsis
- Natural language processing (NLP) shows promise for automating the reading of radiology reports, but its adaptability and reliability for clinical use still need confirmation.
- A study compared four NLP tools (EdIE-R, ALARM+, ESPRESSO, and Sem-EHR) on their performance in identifying cerebrovascular conditions (ischaemic stroke, small vessel disease, and atrophy) using reports and imaging data from NHS Fife and Generation Scotland.
- Results indicated that EdIE-R consistently achieved the highest F1 scores across most conditions, while ALARM+ showed strong performance, especially in precision, highlighting the varied effectiveness of these tools.
- Overall, the findings suggest that while these NLP tools can classify medical conditions, their performance can differ significantly
Article Synopsis
- - The systematic review examines the use of natural language processing (NLP) in analyzing radiology reports, emphasizing the need for transparent methodologies to enable comparisons and reproducibility across studies.
- - It analyzed 164 studies published between January 2015 and October 2019, finding that most focused on disease classification (28%) and diagnostic surveillance (27.4%), primarily using English reports from various imaging modalities, with oncology being the most common disease area.
- - The review highlights issues such as inadequate reporting on essential factors like dataset preparation and validation, with only a small percentage providing details on external validation and data/code availability, suggesting a need for improved reporting standards in NLP research.
A systematic review of natural language processing applied to radiology reports.
BMC Med Inform Decis Mak
June 2021
Article Synopsis
- NLP is crucial for extracting structured information from radiology reports, but comprehensive reviews on its application are lacking.
- A systematic literature search identified 164 publications, revealing a significant increase in usage since 2015, with a shift towards deep learning despite ongoing challenges in clinical adoption and data availability.
- Enhancing the reproducibility and explainability of NLP models is essential for their integration into clinical practice, and there is a need for improved sharing of data and methodologies to facilitate comparison across studies.
Article Synopsis
- Advances in text mining technology allow for accurate extraction of structured information from unstructured Electronic Healthcare Records (EHR).
- The Edinburgh Information Extraction for Radiology reports (EdIE-R) system classifies radiologists' reports to identify occurrences and types of stroke, using a dataset of 1168 reports from the Edinburgh Stroke Study.
- High inter-annotator agreement and system accuracy on a blind test set (ranging from 92.61 to 98.27 for various annotations) suggest that EdIE-R can effectively contribute to population health monitoring and epidemiological research.
Article Synopsis
- Manual coding of brain imaging phenotypes in radiology reports is inefficient, prompting the development of an NLP algorithm for automated identification in NHS reports.
- The NLP algorithm was tested on anonymized reports from stroke/TIA patients, showing excellent agreement between expert readers and strong performance in detecting various brain conditions, including ischaemic strokes and brain tumors.
- The successful implementation of this NLP approach enables large-scale identification of patients with significant brain imaging phenotypes, which enhances clinical practice within the NHS.
Article Synopsis
- Scholarly articles are increasingly referencing dynamic web resources, such as project sites and blogs, to enhance context but this may mislead readers as the content can change over time.
- A study revisiting articles published from 1997 to 2012 found that only about 30% of the referenced web resources had stable snapshots available to compare to current content.
- Over 75% of these web resources showed significant content drift since the original reference, highlighting concerns about the reliability of web-based scholarly information.
Use of the Edinburgh geoparser for georeferencing digitized historical collections.
Philos Trans A Math Phys Eng Sci
August 2010
Article Synopsis
- The paper discusses two JISC-funded projects focused on enhancing the metadata of digitized historical collections through automatic georeferencing and information extraction.
- Understanding location is crucial in historical research, making these collections a valuable test case for automated methodologies.
- The projects, GeoDigRef and Embedding GeoCrossWalk, explored how automatic georeferencing can improve geographical searches and detailed the configurations and evaluations of the geoparser used.
Article Synopsis
- The University of Edinburgh team developed a natural language processing (NLP) system to assist in curating biomedical research papers, aiming to make the curation process easier for biologists.
- Their system excelled in interaction subtasks and performed well in gene mention tasks with minimal effort, while a string matching technique for gene normalization showed close to average results.
- Although the technology is effective for individual tasks, complex tasks requiring multiple components, like detecting and normalizing interacting protein pairs, remain difficult for current NLP systems.
Article Synopsis
- A maximum entropy-based system was developed for identifying named entities in biomedical abstracts.
- The system achieved an F-score of 83.2% in the BioCreative evaluation and 70.1% in the BioNLP evaluation.
- Key features of the system include the use of local features, attention to boundary identification, and the integration of external resources, along with a discussion of data annotation issues that affected performance.
Article Synopsis
- Text mining has potential to enhance biomedical text curation, but there’s limited evidence of its actual effectiveness.
- Three experiments were conducted to see how Natural Language Processing (NLP) can speed up the curation process for protein-protein interactions (PPIs) and to gather curator feedback on the usability of an NLP-integrated curation tool.
- The findings suggest that if NLP output were perfectly accurate, curation time could be sped up by a maximum of one third, but more research is needed to validate curators' preferences for consistent and high-recall NLP outputs.
Article Synopsis
- Good automatic information extraction tools can help process the growing amount of biomedical literature, with named entity recognition as a key feature.
- A maximum-entropy based system was developed to identify gene and protein names in biomedical abstracts, utilizing a variety of features.
- In evaluations, the system achieved high precision and recall, demonstrating effective identification of entity boundaries and innovative use of external knowledge sources.