A Domain Knowledge-Enhanced LSTM-CRF Model for Disease Named Entity Recognition.

Disease named entity recognition (NER) is a critical task for most biomedical natural language processing (NLP) applications. For example, extracting diseases from clinical trial text can be helpful for patient profiling and other downstream applications such as matching clinical trials to eligible patients. Similarly, disease annotation in biomedical articles can help information search engines to accurately index them such that clinicians can easily find relevant articles to enhance their knowledge. In this paper, we propose a domain knowledge-enhanced long short-term memory network-conditional random field (LSTM-CRF) model for disease named entity recognition, which also augments a character-level convolutional neural network (CNN) and a character-level LSTM network for input embedding. Experimental results on a scientific article dataset show the effectiveness of our proposed models compared to state-of-the-art methods in disease recognition.

Automatic Normalization of Anatomical Phrases in Radiology Reports Using Unsupervised Learning.

In today's radiology workflow, free-text reporting is established as the most common medium to capture, store, and communicate clinical information. Radiologists routinely refer to prior radiology reports of a patient to recall critical information for new diagnosis, which is quite tedious, time consuming, and prone to human error. Automatic structuring of report content is desired to facilitate such inquiry of information. In this work, we propose an unsupervised machine learning approach to automatically structure radiology reports by detecting and normalizing anatomical phrases based on the Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT) ontology. The proposed approach combines word embedding-based semantic learning with ontology-based concept mapping to derive the desired concept normalization. The word embedding model was trained using a large corpus of unlabeled radiology reports. Fifty-six anatomical labels were extracted from SNOMED CT as class labels of the whole human anatomy. The proposed framework was compared against a number of state-of-the-art supervised and unsupervised approaches. Radiology reports from three different clinical sites were manually labeled for testing. The proposed approach outperformed other techniques yielding an average precision of 82.6%. The proposed framework boosts the coverage and performance of conventional approaches for concept normalization, by applying word embedding techniques in semantic learning, while avoiding the challenge of having access to a large amount of annotated data, which is typically required for training classifiers.

Automatically correlating clinical findings and body locations in radiology reports using MedLEE.

In this paper, we describe and evaluate a system that extracts clinical findings and body locations from radiology reports and correlates them. The system uses Medical Language Extraction and Encoding System (MedLEE) to map the reports' free text to structured semantic representations of their content. A lightweight reasoning engine extracts the clinical findings and body locations from MedLEE's semantic representation and correlates them. Our study is illustrative for research in which existing natural language processing software is embedded in a larger system. We manually created a standard reference based on a corpus of neuro and breast radiology reports. The standard reference was used to evaluate the precision and recall of the proposed system and its modules. Our results indicate that the precision of our system is considerably better than its recall (82.32-91.37% vs. 35.67-45.91%). We conducted an error analysis and discuss here the practical usability of the system given its recall and precision performance.

Bridging the text-image gap: a decision support tool for real-time PACS browsing.

In this paper, we introduce an ontology-based technology that bridges the gap between MR images on the one hand and knowledge sources on the other hand. The proposed technology allows the user to express interest in a body region by selecting this region on the MR image he or she is viewing with a mouse device. The proposed technology infers the intended body structure from the manual selection and searches the external knowledge source for pertinent information. This technology can be used to bridge the gap between image data in the clinical workflow and (external) knowledge sources that help to assess the case with increased certainty, accuracy, and efficiency. We evaluate an instance of the proposed technology in the neurodomain by means of a user study in which three neuroradiologists participated. The user study shows that the technology has high recall (>95%) when it comes to inferring the intended brain region from the participant's manual selection. We are confident that this helps to increase the experience of browsing external knowledge sources.

Algorithmic and user study of an autocompletion algorithm on a large medical vocabulary.

INTRODUCTION: Autocompletion supports human-computer interaction in software applications that let users enter textual data. We will be inspired by the use case in which medical professionals enter ontology concepts, catering the ongoing demand for structured and standardized data in medicine. OBJECTIVES: Goal is to give an algorithmic analysis of one particular autocompletion algorithm, called multi-prefix matching algorithm, which suggests terms whose words' prefixes contain all words in the string typed by the user, e.g., in this sense, opt ner me matches optic nerve meningioma. Second we aim to investigate how well it supports users entering concepts from a large and comprehensive medical vocabulary (snomed ct). METHODS: We give a concise description of the multi-prefix algorithm, and sketch how it can be optimized to meet required response time. Performance will be compared to a baseline algorithm, which gives suggestions that extend the string typed by the user to the right, e.g. optic nerve m gives optic nerve meningioma, but opt ner me does not. We conduct a user experiment in which 12 participants are invited to complete 40 snomed ct terms with the baseline algorithm and another set of 40 snomed ct terms with the multi-prefix algorithm. RESULTS: Our results show that users need significantly fewer keystrokes when supported by the multi-prefix algorithm than when supported by the baseline algorithm. CONCLUSIONS: The proposed algorithm is a competitive candidate for searching and retrieving terms from a large medical ontology.