Applied Clinical Informatics

National Library Of Medicine

Investigators

Linked publications & trials

Abstract

NLP and Indexing - Developing a multi-model system for automatic classification of MEDLINE articles as per MeSH Publication Types. - Studying the growth in Author Supplied Keywords in MEDLINE and if they might improve automatic indexing of articles. A prior study showed several keywords were used to drive readers to an article even if the keywords were not relevant to the article itself. - Developing an updated Machine Learning approach to MeSH Check Tag detection in MEDLINE articles - Expanded automated indexing to 100% of NLM indexing on April 1, 2022 as part of the NLM MEDLINE 2022 effort using the NLM Medical Text Indexer (MTI). This has reduced the indexing backlog from 576,735 articles on January 4, 2021 to practically zero today. This has also reduced the Average Time to Index from 145 days in FY21 to approximately 48 hours. - Continuing to update the SPECIALIST Lexicon, SemMedDB, and MetaMap datasets in support of COVID-19. Clinical Image Processing - Developing deep learning algorithms to monitor and measure blood leakages from fluorescein angiography (FA) of uveitis patients. - Continuing development using deep learning algorithms to classify uveitis from retinal fundus images and OCT images. - Designed and implemented a segmentation algorithm, called Multiscale Average Pooling Net (MAPNet), for optic disc and cup segmentation from fundus images using deep learning models in the ImageNet, Keras, TensorFow, and Python. The MAPNet showed 0.9679 dice coefficients for optic disc and 0.8996 for cup. Clinical Data Processing - Characterizing Post-Acute Sequelae of COVID-19 (PASC) in the Philippines: Applying the WHO SMART Guidelines for Long COVID Monitoring. - Discovering long COVID patients in the Philippines using machine learning models and electronic medical records. - Developing deep learning algorithms to predict the future health status of COVID-19 patients using patients phenotype data. Initial discussions also started looking at a project to develop deep learning algorithms to the predict future status of heart-transplant patients also using their phenotype data. - Completed mapping AllofUs, Clinical Practice Research Datalink (CPRD), UK Biobank, and CCW Virtual Research Data Center (VRDC) (Medicaid/Medicare) data sets to OMOP (Observational Medical Outcomes Partnerships) and initial counts have been made across the datasets to inform future observational studies. Standards/FHIR - Designing a new API for the dbGaP clinical studies variables to support the joint LHC-NCBI dbGaP FHIR project. This will include a table-based, shopping cart-like UI as the primary interface for querying the dbGaP FHIR server. - Improving the NLM Form Builder software suite focusing on updates to the applications include removing dependencies on the now obsolete software AngularJS. - Developing a new version of the NLM Form Builder which includes a new user interface which provides more user guidance and can efficiently handle large forms. - Completed the rewrite of the LHC-Forms (FHIR Questionnaire renderer) to now be a web-component. The rewrite also addressed some of the accessibility issues and additional support for FHIR R5. Other - FISMA Moderate Cloud environment has been setup, tested, and received preliminary security approval. The new LHC FISMA Moderate Cloud environment will allow for expanded clinical dataset usage. - Continuing to move ACIB software tools into LHC CI/CD process and into the Cloud. - Developing a new version of the NLM-Scrubber designed to capture more information from external knowledge sources and represent them in the new data structures, so that we can not only better recognize PII elements in clinical text but also better recognize clinical and scientific information and other non-PII elements of clinical notes.