Ambiguity in medical concept normalization: An analysis of types and coverage in electronic health record datasets

Question Answering for Electronic Health Records: Scoping Review of Datasets and Models

BACKGROUND

Question answering (QA) systems for patient-related data can assist both clinicians and patients. They can, for example, assist clinicians in decision-making and enable patients to have a better understanding of their medical history. Substantial amounts of patient data are stored in electronic health records (EHRs), making EHR QA an important research area. Because of the differences in data format and modality, this differs greatly from other medical QA tasks that use medical websites or scientific papers to retrieve answers, making it critical to research EHR QA.

OBJECTIVE

This study aims to provide a methodological review of existing works on QA for EHRs. The objectives of this study were to identify the existing EHR QA datasets and analyze them, study the state-of-the-art methodologies used in this task, compare the different evaluation metrics used by these state-of-the-art models, and finally elicit the various challenges and the ongoing issues in EHR QA.

METHODS

We searched for articles from January 1, 2005, to September 30, 2023, in 4 digital sources, including Google Scholar, ACL Anthology, ACM Digital Library, and PubMed, to collect relevant publications on EHR QA. Our systematic screening process followed PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. A total of 4111 papers were identified for our study, and after screening based on our inclusion criteria, we obtained 47 papers for further study. The selected studies were then classified into 2 non-mutually exclusive categories depending on their scope: "EHR QA datasets" and "EHR QA models."

RESULTS

A systematic screening process obtained 47 papers on EHR QA for final review. Out of the 47 papers, 53% (n=25) were about EHR QA datasets, and 79% (n=37) papers were about EHR QA models. It was observed that QA on EHRs is relatively new and unexplored. Most of the works are fairly recent. In addition, it was observed that emrQA is by far the most popular EHR QA dataset, both in terms of citations and usage in other papers. We have classified the EHR QA datasets based on their modality, and we have inferred that Medical Information Mart for Intensive Care (MIMIC-III) and the National Natural Language Processing Clinical Challenges datasets (ie, n2c2 datasets) are the most popular EHR databases and corpuses used in EHR QA. Furthermore, we identified the different models used in EHR QA along with the evaluation metrics used for these models.

CONCLUSIONS

EHR QA research faces multiple challenges, such as the limited availability of clinical annotations, concept normalization in EHR QA, and challenges faced in generating realistic EHR QA datasets. There are still many gaps in research that motivate further work. This study will assist future researchers in focusing on areas of EHR QA that have possible future research directions.

Health Care Language Models and Their Fine-Tuning for Information Extraction: Scoping Review

BACKGROUND

In response to the intricate language, specialized terminology outside everyday life, and the frequent presence of abbreviations and acronyms inherent in health care text data, domain adaptation techniques have emerged as crucial to transformer-based models. This refinement in the knowledge of the language models (LMs) allows for a better understanding of the medical textual data, which results in an improvement in medical downstream tasks, such as information extraction (IE). We have identified a gap in the literature regarding health care LMs. Therefore, this study presents a scoping literature review investigating domain adaptation methods for transformers in health care, differentiating between English and non-English languages, focusing on Portuguese. Most specifically, we investigated the development of health care LMs, with the aim of comparing Portuguese with other more developed languages to guide the path of a non-English-language with fewer resources.

OBJECTIVE

This study aimed to research health care IE models, regardless of language, to understand the efficacy of transformers and what are the medical entities most commonly extracted.

METHODS

This scoping review was conducted using the PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) methodology on Scopus and Web of Science Core Collection databases. Only studies that mentioned the creation of health care LMs or health care IE models were included, while large language models (LLMs) were excluded. The latest were not included since we wanted to research LMs and not LLMs, which are architecturally different and have distinct purposes.

RESULTS

Our search query retrieved 137 studies, 60 of which met the inclusion criteria, and none of them were systematic literature reviews. English and Chinese are the languages with the most health care LMs developed. These languages already have disease-specific LMs, while others only have general-health care LMs. European Portuguese does not have any public health care LM and should take examples from other languages to develop, first, general-health care LMs and then, in an advanced phase, disease-specific LMs. Regarding IE models, transformers were the most commonly used method, and named entity recognition was the most popular topic, with only a few studies mentioning Assertion Status or addressing medical lexical problems. The most extracted entities were diagnosis, posology, and symptoms.

CONCLUSIONS

The findings indicate that domain adaptation is beneficial, achieving better results in downstream tasks. Our analysis allowed us to understand that the use of transformers is more developed for the English and Chinese languages. European Portuguese lacks relevant studies and should draw examples from other non-English languages to develop these models and drive progress in AI. Health care professionals could benefit from highlighting medically relevant information and optimizing the reading of the textual data, or this information could be used to create patient medical timelines, allowing for profiling.

Early Prediction of Cardiac Arrest in the Intensive Care Unit Using Explainable Machine Learning: Retrospective Study

BACKGROUND

Cardiac arrest (CA) is one of the leading causes of death among patients in the intensive care unit (ICU). Although many CA prediction models with high sensitivity have been developed to anticipate CA, their practical application has been challenging due to a lack of generalization and validation. Additionally, the heterogeneity among patients in different ICU subtypes has not been adequately addressed.

OBJECTIVE

This study aims to propose a clinically interpretable ensemble approach for the timely and accurate prediction of CA within 24 hours, regardless of patient heterogeneity, including variations across different populations and ICU subtypes. Additionally, we conducted patient-independent evaluations to emphasize the model's generalization performance and analyzed interpretable results that can be readily adopted by clinicians in real-time.

METHODS

Patients were retrospectively analyzed using data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) and the eICU-Collaborative Research Database (eICU-CRD). To address the problem of underperformance, we constructed our framework using feature sets based on vital signs, multiresolution statistical analysis, and the Gini index, with a 12-hour window to capture the unique characteristics of CA. We extracted 3 types of features from each database to compare the performance of CA prediction between high-risk patient groups from MIMIC-IV and patients without CA from eICU-CRD. After feature extraction, we developed a tabular network (TabNet) model using feature screening with cost-sensitive learning. To assess real-time CA prediction performance, we used 10-fold leave-one-patient-out cross-validation and a cross-data set method. We evaluated MIMIC-IV and eICU-CRD across different cohort populations and subtypes of ICU within each database. Finally, external validation using the eICU-CRD and MIMIC-IV databases was conducted to assess the model's generalization ability. The decision mask of the proposed method was used to capture the interpretability of the model.

RESULTS

The proposed method outperformed conventional approaches across different cohort populations in both MIMIC-IV and eICU-CRD. Additionally, it achieved higher accuracy than baseline models for various ICU subtypes within both databases. The interpretable prediction results can enhance clinicians' understanding of CA prediction by serving as a statistical comparison between non-CA and CA groups. Next, we tested the eICU-CRD and MIMIC-IV data sets using models trained on MIMIC-IV and eICU-CRD, respectively, to evaluate generalization ability. The results demonstrated superior performance compared with baseline models.

CONCLUSIONS

Our novel framework for learning unique features provides stable predictive power across different ICU environments. Most of the interpretable global information reveals statistical differences between CA and non-CA groups, demonstrating its utility as an indicator for clinical decisions. Consequently, the proposed CA prediction system is a clinically validated algorithm that enables clinicians to intervene early based on CA prediction information and can be applied to clinical trials in digital health.

NeighBERT: Medical Entity Linking Using Relation-Induced Dense Retrieval

One of the common tasks in clinical natural language processing is medical entity linking (MEL) which involves mention detection followed by linking the mention to an entity in a knowledge base. One reason that MEL has not been solved is due to a problem that occurs in language where ambiguous texts can be resolved to several named entities. This problem is exacerbated when processing the text found in electronic health records. Recent work has shown that deep learning models based on transformers outperform previous methods on linking at higher rates of performance. We introduce NeighBERT, a custom pre-training technique which extends BERT (Devlin et al [1]) by encoding how entities are related within a knowledge graph. This technique adds relational context that has been traditionally missing in original BERT, helping resolve the ambiguity found in clinical text. In our experiments, NeighBERT improves the precision, recall, and F1-score of the state of the art by 1-3 points for named entity recognition and 10-15 points for MEL on two widely known clinical datasets.

Supplementary Information

The online version contains supplementary material available at 10.1007/s41666-023-00136-3.

An overview of biomedical entity linking throughout the years

Biomedical Entity Linking (BEL) is the task of mapping of spans of text within biomedical documents to normalized, unique identifiers within an ontology. This is an important task in natural language processing for both translational information extraction applications and providing context for downstream tasks like relationship extraction. In this paper, we will survey the progression of BEL from its inception in the late 80s to present day state of the art systems, provide a comprehensive list of datasets available for training BEL systems, reference shared tasks focused on BEL, discuss the technical components that comprise BEL systems, and discuss possible directions for the future of the field.

Silva J, Almeida JR, Matos S. Chemical identification and indexing in PubMed full-text articles using deep learning and heuristics

The identification of chemicals in articles has attracted a large interest in the biomedical scientific community, given its importance in drug development research. Most of previous research have focused on PubMed abstracts, and further investigation using full-text documents is required because these contain additional valuable information that must be explored. The manual expert task of indexing Medical Subject Headings (MeSH) terms to these articles later helps researchers find the most relevant publications for their ongoing work. The BioCreative VII NLM-Chem track fostered the development of systems for chemical identification and indexing in PubMed full-text articles. Chemical identification consisted in identifying the chemical mentions and linking these to unique MeSH identifiers. This manuscript describes our participation system and the post-challenge improvements we made. We propose a three-stage pipeline that individually performs chemical mention detection, entity normalization and indexing. Regarding chemical identification, we adopted a deep-learning solution that utilizes the PubMedBERT contextualized embeddings followed by a multilayer perceptron and a conditional random field tagging layer. For the normalization approach, we use a sieve-based dictionary filtering followed by a deep-learning similarity search strategy. Finally, for the indexing we developed rules for identifying the more relevant MeSH codes for each article. During the challenge, our system obtained the best official results in the normalization and indexing tasks despite the lower performance in the chemical mention recognition task. In a post-contest phase we boosted our results by improving our named entity recognition model with additional techniques. The final system achieved 0.8731, 0.8275 and 0.4849 in the chemical identification, normalization and indexing tasks, respectively. The code to reproduce our experiments and run the pipeline is publicly available.

Database URL

https://github.com/bioinformatics-ua/biocreativeVII_track2

Improving broad-coverage medical entity linking with semantic type prediction and large-scale datasets

OBJECTIVES

Biomedical natural language processing tools are increasingly being applied for broad-coverage information extraction-extracting medical information of all types in a scientific document or a clinical note. In such broad-coverage settings, linking mentions of medical concepts to standardized vocabularies requires choosing the best candidate concepts from large inventories covering dozens of types. This study presents a novel semantic type prediction module for biomedical NLP pipelines and two automatically-constructed, large-scale datasets with broad coverage of semantic types.

METHODS

We experiment with five off-the-shelf biomedical NLP toolkits on four benchmark datasets for medical information extraction from scientific literature and clinical notes. All toolkits adopt a staged approach of mention detection followed by two stages of medical entity linking: (1) generating a list of candidate concepts, and (2) picking the best concept among them. We introduce a semantic type prediction module to alleviate the problem of overgeneration of candidate concepts by filtering out irrelevant candidate concepts based on the predicted semantic type of a mention. We present MedType, a fully modular semantic type prediction model which we integrate into the existing NLP toolkits. To address the dearth of broad-coverage training data for medical information extraction, we further present WikiMed and PubMedDS, two large-scale datasets for medical entity linking.

RESULTS

Semantic type filtering improves medical entity linking performance across all toolkits and datasets, often by several percentage points of F-1. Further, pretraining MedType on our novel datasets achieves state-of-the-art performance for semantic type prediction in biomedical text.

CONCLUSIONS

Semantic type prediction is a key part of building accurate NLP pipelines for broad-coverage information extraction from biomedical text. We make our source code and novel datasets publicly available to foster reproducible research.

Automated Coding of Under-Studied Medical Concept Domains: Linking Physical Activity Reports to the International Classification of Functioning, Disability, and Health

Linking clinical narratives to standardized vocabularies and coding systems is a key component of unlocking the information in medical text for analysis. However, many domains of medical concepts, such as functional outcomes and social determinants of health, lack well-developed terminologies that can support effective coding of medical text. We present a framework for developing natural language processing (NLP) technologies for automated coding of medical information in under-studied domains, and demonstrate its applicability through a case study on physical mobility function. Mobility function is a component of many health measures, from post-acute care and surgical outcomes to chronic frailty and disability, and is represented as one domain of human activity in the International Classification of Functioning, Disability, and Health (ICF). However, mobility and other types of functional activity remain under-studied in the medical informatics literature, and neither the ICF nor commonly-used medical terminologies capture functional status terminology in practice. We investigated two data-driven paradigms, classification and candidate selection, to link narrative observations of mobility status to standardized ICF codes, using a dataset of clinical narratives from physical therapy encounters. Recent advances in language modeling and word embedding were used as features for established machine learning models and a novel deep learning approach, achieving a macro-averaged F-1 score of 84% on linking mobility activity reports to ICF codes. Both classification and candidate selection approaches present distinct strengths for automated coding in under-studied domains, and we highlight that the combination of (i) a small annotated data set; (ii) expert definitions of codes of interest; and (iii) a representative text corpus is sufficient to produce high-performing automated coding systems. This research has implications for continued development of language technologies to analyze functional status information, and the ongoing growth of NLP tools for a variety of specialized applications in clinical care and research.