Enhancing clinical concept extraction with contextual embeddings

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.

Criteria2Query 3.0: Leveraging generative large language models for clinical trial eligibility query generation

OBJECTIVE

Automated identification of eligible patients is a bottleneck of clinical research. We propose Criteria2Query (C2Q) 3.0, a system that leverages GPT-4 for the semi-automatic transformation of clinical trial eligibility criteria text into executable clinical database queries.

MATERIALS AND METHODS

C2Q 3.0 integrated three GPT-4 prompts for concept extraction, SQL query generation, and reasoning. Each prompt was designed and evaluated separately. The concept extraction prompt was benchmarked against manual annotations from 20 clinical trials by two evaluators, who later also measured SQL generation accuracy and identified errors in GPT-generated SQL queries from 5 clinical trials. The reasoning prompt was assessed by three evaluators on four metrics: readability, correctness, coherence, and usefulness, using corrected SQL queries and an open-ended feedback questionnaire.

RESULTS

Out of 518 concepts from 20 clinical trials, GPT-4 achieved an F1-score of 0.891 in concept extraction. For SQL generation, 29 errors spanning seven categories were detected, with logic errors being the most common (n = 10; 34.48 %). Reasoning evaluations yielded a high coherence rating, with the mean score being 4.70 but relatively lower readability, with a mean of 3.95. Mean scores of correctness and usefulness were identified as 3.97 and 4.37, respectively.

CONCLUSION

GPT-4 significantly improves the accuracy of extracting clinical trial eligibility criteria concepts in C2Q 3.0. Continued research is warranted to ensure the reliability of large language models.

Language model and its interpretability in biomedicine: A scoping review

With advancements in large language models, artificial intelligence (AI) is undergoing a paradigm shift where AI models can be repurposed with minimal effort across various downstream tasks. This provides great promise in learning generally useful representations from biomedical corpora, at scale, which would empower AI solutions in healthcare and biomedical research. Nonetheless, our understanding of how they work, when they fail, and what they are capable of remains underexplored due to their emergent properties. Consequently, there is a need to comprehensively examine the use of language models in biomedicine. This review aims to summarize existing studies of language models in biomedicine and identify topics ripe for future research, along with the technical and analytical challenges w.r.t. interpretability. We expect this review to help researchers and practitioners better understand the landscape of language models in biomedicine and what methods are available to enhance the interpretability of their models.

Generative large language models are all-purpose text analytics engines: text-to-text learning is all your need

OBJECTIVE

To solve major clinical natural language processing (NLP) tasks using a unified text-to-text learning architecture based on a generative large language model (LLM) via prompt tuning.

METHODS

We formulated 7 key clinical NLP tasks as text-to-text learning and solved them using one unified generative clinical LLM, GatorTronGPT, developed using GPT-3 architecture and trained with up to 20 billion parameters. We adopted soft prompts (ie, trainable vectors) with frozen LLM, where the LLM parameters were not updated (ie, frozen) and only the vectors of soft prompts were updated, known as prompt tuning. We added additional soft prompts as a prefix to the input layer, which were optimized during the prompt tuning. We evaluated the proposed method using 7 clinical NLP tasks and compared them with previous task-specific solutions based on Transformer models.

RESULTS AND CONCLUSION

The proposed approach achieved state-of-the-art performance for 5 out of 7 major clinical NLP tasks using one unified generative LLM. Our approach outperformed previous task-specific transformer models by ∼3% for concept extraction and 7% for relation extraction applied to social determinants of health, 3.4% for clinical concept normalization, 3.4%-10% for clinical abbreviation disambiguation, and 5.5%-9% for natural language inference. Our approach also outperformed a previously developed prompt-based machine reading comprehension (MRC) model, GatorTron-MRC, for clinical concept and relation extraction. The proposed approach can deliver the "one model for all" promise from training to deployment using a unified generative LLM.

Importance-aware adaptive dataset distillation

Herein, we propose a novel dataset distillation method for constructing small informative datasets that preserve the information of the large original datasets. The development of deep learning models is enabled by the availability of large-scale datasets. Despite unprecedented success, large-scale datasets considerably increase the storage and transmission costs, resulting in a cumbersome model training process. Moreover, using raw data for training raises privacy and copyright concerns. To address these issues, a new task named dataset distillation has been introduced, aiming to synthesize a compact dataset that retains the essential information from the large original dataset. State-of-the-art (SOTA) dataset distillation methods have been proposed by matching gradients or network parameters obtained during training on real and synthetic datasets. The contribution of different network parameters to the distillation process varies, and uniformly treating them leads to degraded distillation performance. Based on this observation, we propose an importance-aware adaptive dataset distillation (IADD) method that can improve distillation performance by automatically assigning importance weights to different network parameters during distillation, thereby synthesizing more robust distilled datasets. IADD demonstrates superior performance over other SOTA dataset distillation methods based on parameter matching on multiple benchmark datasets and outperforms them in terms of cross-architecture generalization. In addition, the analysis of self-adaptive weights demonstrates the effectiveness of IADD. Furthermore, the effectiveness of IADD is validated in a real-world medical application such as COVID-19 detection.

Complementary and Integrative Health Information in the literature: its lexicon and named entity recognition

OBJECTIVE

To construct an exhaustive Complementary and Integrative Health (CIH) Lexicon (CIHLex) to help better represent the often underrepresented physical and psychological CIH approaches in standard terminologies, and to also apply state-of-the-art natural language processing (NLP) techniques to help recognize them in the biomedical literature.

MATERIALS AND METHODS

We constructed the CIHLex by integrating various resources, compiling and integrating data from biomedical literature and relevant sources of knowledge. The Lexicon encompasses 724 unique concepts with 885 corresponding unique terms. We matched these concepts to the Unified Medical Language System (UMLS), and we developed and utilized BERT models comparing their efficiency in CIH named entity recognition to well-established models including MetaMap and CLAMP, as well as the large language model GPT3.5-turbo.

RESULTS

Of the 724 unique concepts in CIHLex, 27.2% could be matched to at least one term in the UMLS. About 74.9% of the mapped UMLS Concept Unique Identifiers were categorized as "Therapeutic or Preventive Procedure." Among the models applied to CIH named entity recognition, BLUEBERT delivered the highest macro-average F1-score of 0.91, surpassing other models.

CONCLUSION

Our CIHLex significantly augments representation of CIH approaches in biomedical literature. Demonstrating the utility of advanced NLP models, BERT notably excelled in CIH entity recognition. These results highlight promising strategies for enhancing standardization and recognition of CIH terminology in biomedical contexts.

Emergency Department Visits for Alcohol-Associated Falls Among Older Adults in the United States, 2011 to 2020

STUDY OBJECTIVE

The aim of this study was to examine the epidemiology of alcohol-associated fall injuries among older adults aged ≥65 years in the United States.

METHODS

We included emergency department (ED) visits for unintentional fall injuries by adults from the National Electronic Injury Surveillance System-All Injury Program during 2011 to 2020. We estimated the annual national rate of ED visits for alcohol-associated falls and the proportion of these falls among older adults' fall-related ED visits using demographic and clinical characteristics. Joinpoint regression was performed to examine trends in alcohol-associated ED fall visits between 2011 and 2019 among older adult age subgroups and to compare these trends with those of younger adults.

RESULTS

There were 9,657 (weighted national estimate: 618,099) ED visits for alcohol-associated falls, representing 2.2% of ED fall visits during 2011 to 2020 among older adults. The proportion of fall-related ED visits that were alcohol-associated was higher among men than among women (adjusted prevalence ratio [aPR]=3.6, 95% confidence interval [CI] 2.9 to 4.5). The head and face were the most commonly injured body parts, and internal injury was the most common diagnosis for alcohol-associated falls. From 2011 to 2019, the annual rate of ED visits for alcohol-associated falls increased (annual percent change 7.5, 95% CI 6.1 to 8.9) among older adults. Adults aged 55 to 64 years had a similar increase; a sustained increase was not detected in younger age groups.

CONCLUSION

Our findings highlight the rising rates of ED visits for alcohol-associated falls among older adults during the study period. Health care providers in the ED can screen older adults for fall risk and assess for modifiable risk factors such as alcohol use to help identify those who could benefit from interventions to reduce their risk.

A case study in applying artificial intelligence-based named entity recognition to develop an automated ophthalmic disease registry

PURPOSE

Advances in artificial intelligence (AI)-based named entity extraction (NER) have improved the ability to extract diagnostic entities from unstructured, narrative, free-text data in electronic health records. However, there is a lack of ready-to-use tools and workflows to encourage the use among clinicians who often lack experience and training in AI. We sought to demonstrate a case study for developing an automated registry of ophthalmic diseases accompanied by a ready-to-use low-code tool for clinicians.

METHODS

We extracted deidentified electronic clinical records from a single centre's adult outpatient ophthalmology clinic from November 2019 to May 2022. We used a low-code annotation software tool (Prodigy) to annotate diagnoses and train a bespoke spaCy NER model to extract diagnoses and create an ophthalmic disease registry.

RESULTS

A total of 123,194 diagnostic entities were extracted from 33,455 clinical records. After decapitalisation and removal of non-alphanumeric characters, there were 5070 distinct extracted diagnostic entities. The NER model achieved a precision of 0.8157, recall of 0.8099, and F score of 0.8128.

CONCLUSION

We presented a case study using low-code artificial intelligence-based NLP tools to produce an automated ophthalmic disease registry. The workflow created a NER model with a moderate overall ability to extract diagnoses from free-text electronic clinical records. We have produced a ready-to-use tool for clinicians to implement this low-code workflow in their institutions and encourage the uptake of artificial intelligence methods for case finding in electronic health records.

Marie and BERT-A Knowledge Graph Embedding Based Question Answering System for Chemistry

This paper presents a novel knowledge graph question answering (KGQA) system for chemistry, which is implemented on hybrid knowledge graph embeddings, aiming to provide fact-oriented information retrieval for chemistry-related research and industrial applications. Unlike other existing designs, the system operates on multiple embedding spaces, which use various embedding methods and queries the embedding spaces in parallel. With the answers returned from multiple embedding spaces, the system leverages a score alignment model to adjust the answer scores and rerank the answers. Further, the system implements an algorithm to derive implicit multihop relations to handle the complexities of deep ontologies and improve multihop question answering. The system also implements a BERT-based bidirectional entity-linking model to enhance the robustness and accuracy of the entity-linking module. The system uses a joint numerical embedding model to efficiently handle numerical filtering questions. Further, it can invoke semantic agents to perform dynamic calculations autonomously. Finally, the KGQA system handles numerous chemical reaction mechanisms using semantic parsing supported by a Linked Data Fragment server. This paper evaluates the accuracy of each module within the KGQA system with a chemistry question data set.

Modeling new trends in bone regeneration, using the BERTopic approach

Aim: Bibliometric surveys are time-consuming endeavors, which cannot be scaled up to meet the challenges of ever-expanding fields, such as bone regeneration. Artificial intelligence, however, can provide smart tools to screen massive amounts of literature, and we relied on this technology to automatically identify research topics. Materials & methods: We used the BERTopic algorithm to detect the topics in a corpus of MEDLINE manuscripts, mapping their similarities and highlighting research hotspots. Results: Using BERTopic, we identified 372 topics and were able to assess the growing importance of innovative and recent fields of investigation such as 3D printing and extracellular vescicles. Conclusion: BERTopic appears as a suitable tool to set up automatic screening routines to track the progress in bone regeneration.

Clinical named entity recognition and relation extraction using natural language processing of medical free text: A systematic review

BACKGROUND

Natural Language Processing (NLP) applications have developed over the past years in various fields including its application to clinical free text for named entity recognition and relation extraction. However, there has been rapid developments the last few years that there's currently no overview of it. Moreover, it is unclear how these models and tools have been translated into clinical practice. We aim to synthesize and review these developments.

METHODS

We reviewed literature from 2010 to date, searching PubMed, Scopus, the Association of Computational Linguistics (ACL), and Association of Computer Machinery (ACM) libraries for studies of NLP systems performing general-purpose (i.e., not disease- or treatment-specific) information extraction and relation extraction tasks in unstructured clinical text (e.g., discharge summaries).

RESULTS

We included in the review 94 studies with 30 studies published in the last three years. Machine learning methods were used in 68 studies, rule-based in 5 studies, and both in 22 studies. 63 studies focused on Named Entity Recognition, 13 on Relation Extraction and 18 performed both. The most frequently extracted entities were "problem", "test" and "treatment". 72 studies used public datasets and 22 studies used proprietary datasets alone. Only 14 studies defined clearly a clinical or information task to be addressed by the system and just three studies reported its use outside the experimental setting. Only 7 studies shared a pre-trained model and only 8 an available software tool.

DISCUSSION

Machine learning-based methods have dominated the NLP field on information extraction tasks. More recently, Transformer-based language models are taking the lead and showing the strongest performance. However, these developments are mostly based on a few datasets and generic annotations, with very few real-world use cases. This may raise questions about the generalizability of findings, translation into practice and highlights the need for robust clinical evaluation.

Hybrid approach combining deep learning and a rule based expert system for concept extraction from prescriptions

Concept extraction from prescriptions is a very important task that provides a foundation for many of the downstream healthcare applications in decision making across the areas of pharmacovigilance, medication adherence, inventory management, and other matters of value-based care. Although short, these directions can sometimes be complex. With the increase in complexity of direction, it becomes harder to extract various concepts by only rule based expert system. It identifies major concepts like frequency, dosage, duration, etc. from the natural text direction using a combination of rules and deep learning (DL) based methods on a large real world data of a pharmacy chain. The DL module includes a fine-tuned BERT transformer and Gram CNN (Convolutional Neural Network) based NER (Named Entity Recognition) architecture. The proposed method utilizes the domain heuristics along with intelligent labelling and bootstrapping to help DL models extract concepts with high evaluation scores and thus provides a way for carrying out concept extraction using targeted methods instead of one single method. To the best of our knowledge, this is the best performance reported in the literature for concept extraction from doctor's prescription.

Weakly supervised spatial relation extraction from radiology reports

Objective

Weak supervision holds significant promise to improve clinical natural language processing by leveraging domain resources and expertise instead of large manually annotated datasets alone. Here, our objective is to evaluate a weak supervision approach to extract spatial information from radiology reports.

Materials and Methods

Our weak supervision approach is based on data programming that uses rules (or labeling functions) relying on domain-specific dictionaries and radiology language characteristics to generate weak labels. The labels correspond to different spatial relations that are critical to understanding radiology reports. These weak labels are then used to fine-tune a pretrained Bidirectional Encoder Representations from Transformers (BERT) model.

Results

Our weakly supervised BERT model provided satisfactory results in extracting spatial relations without manual annotations for training (spatial trigger F1: 72.89, relation F1: 52.47). When this model is further fine-tuned on manual annotations (relation F1: 68.76), performance surpasses the fully supervised state-of-the-art.

Discussion

To our knowledge, this is the first work to automatically create detailed weak labels corresponding to radiological information of clinical significance. Our data programming approach is (1) adaptable as the labeling functions can be updated with relatively little manual effort to incorporate more variations in radiology language reporting formats and (2) generalizable as these functions can be applied across multiple radiology subdomains in most cases.

Conclusions

We demonstrate a weakly supervision model performs sufficiently well in identifying a variety of relations from radiology text without manual annotations, while exceeding state-of-the-art results when annotated data are available.

Natural Language Processing for Clinical Laboratory Data Repository Systems: Implementation and Evaluation for Respiratory Viruses

BACKGROUND

With the growing volume and complexity of laboratory repositories, it has become tedious to parse unstructured data into structured and tabulated formats for secondary uses such as decision support, quality assurance, and outcome analysis. However, advances in natural language processing (NLP) approaches have enabled efficient and automated extraction of clinically meaningful medical concepts from unstructured reports.

OBJECTIVE

In this study, we aimed to determine the feasibility of using the NLP model for information extraction as an alternative approach to a time-consuming and operationally resource-intensive handcrafted rule-based tool. Therefore, we sought to develop and evaluate a deep learning-based NLP model to derive knowledge and extract information from text-based laboratory reports sourced from a provincial laboratory repository system.

METHODS

The NLP model, a hierarchical multilabel classifier, was trained on a corpus of laboratory reports covering testing for 14 different respiratory viruses and viral subtypes. The corpus includes 87,500 unique laboratory reports annotated by 8 subject matter experts (SMEs). The classification task involved assigning the laboratory reports to labels at 2 levels: 24 fine-grained labels in level 1 and 6 coarse-grained labels in level 2. A "label" also refers to the status of a specific virus or strain being tested or detected (eg, influenza A is detected). The model's performance stability and variation were analyzed across all labels in the classification task. Additionally, the model's generalizability was evaluated internally and externally on various test sets.

RESULTS

Overall, the NLP model performed well on internal, out-of-time (pre-COVID-19), and external (different laboratories) test sets with microaveraged F1-scores >94% across all classes. Higher precision and recall scores with less variability were observed for the internal and pre-COVID-19 test sets. As expected, the model's performance varied across categories and virus types due to the imbalanced nature of the corpus and sample sizes per class. There were intrinsically fewer classes of viruses being detected than those tested; therefore, the model's performance (lowest F1-score of 57%) was noticeably lower in the detected cases.

CONCLUSIONS

We demonstrated that deep learning-based NLP models are promising solutions for information extraction from text-based laboratory reports. These approaches enable scalable, timely, and practical access to high-quality and encoded laboratory data if integrated into laboratory information system repositories.

Representing and utilizing clinical textual data for real world studies: An OHDSI approach

Clinical documentation in electronic health records contains crucial narratives and details about patients and their care. Natural language processing (NLP) can unlock the information conveyed in clinical notes and reports, and thus plays a critical role in real-world studies. The NLP Working Group at the Observational Health Data Sciences and Informatics (OHDSI) consortium was established to develop methods and tools to promote the use of textual data and NLP in real-world observational studies. In this paper, we describe a framework for representing and utilizing textual data in real-world evidence generation, including representations of information from clinical text in the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM), the workflow and tools that were developed to extract, transform and load (ETL) data from clinical notes into tables in OMOP CDM, as well as current applications and specific use cases of the proposed OHDSI NLP solution at large consortia and individual institutions with English textual data. Challenges faced and lessons learned during the process are also discussed to provide valuable insights for researchers who are planning to implement NLP solutions in real-world studies.

Clinical Concept-Based Radiology Reports Classification Pipeline for Lung Carcinoma

Rising incidence and mortality of cancer have led to an incremental amount of research in the field. To learn from preexisting data, it has become important to capture maximum information related to disease type, stage, treatment, and outcomes. Medical imaging reports are rich in this kind of information but are only present as free text. The extraction of information from such unstructured text reports is labor-intensive. The use of Natural Language Processing (NLP) tools to extract information from radiology reports can make it less time-consuming as well as more effective. In this study, we have developed and compared different models for the classification of lung carcinoma reports using clinical concepts. This study was approved by the institutional ethics committee as a retrospective study with a waiver of informed consent. A clinical concept-based classification pipeline for lung carcinoma radiology reports was developed using rule-based as well as machine learning models and compared. The machine learning models used were XGBoost and two more deep learning model architectures with bidirectional long short-term neural networks. A corpus consisting of 1700 radiology reports including computed tomography (CT) and positron emission tomography/computed tomography (PET/CT) reports were used for development and testing. Five hundred one radiology reports from MIMIC-III Clinical Database version 1.4 was used for external validation. The pipeline achieved an overall F1 score of 0.94 on the internal set and 0.74 on external validation with the rule-based algorithm using expert input giving the best performance. Among the machine learning models, the Bi-LSTM_dropout model performed better than the ML model using XGBoost and the Bi-LSTM_simple model on internal set, whereas on external validation, the Bi-LSTM_simple model performed relatively better than other 2. This pipeline can be used for clinical concept-based classification of radiology reports related to lung carcinoma from a huge corpus and also for automated annotation of these reports.

Identifying stroke-related quantified evidence from electronic health records in real-world studies

BACKGROUND

Stroke is one of the leading causes of death and disability worldwide. The National Institutes of Health Stroke Scale (NIHSS) scores in electronic health records (EHRs), which quantitatively describe patients' neurological deficits in evidence-based treatment, are crucial in stroke-related clinical investigations. However, the free-text format and lack of standardization inhibit their effective use. Automatically extracting the scale scores from the clinical free text so that its potential value in real-world studies is realized has become an important goal.

OBJECTIVE

This study aims to develop an automated method to extract scale scores from the free text of EHRs.

METHODS

We propose a two-step pipeline method to identify NIHSS items and numerical scores and validate its feasibility using a freely accessible critical care database: MIMIC-III (Medical Information Mart for Intensive Care III). First, we utilize MIMIC-III to create an annotated corpus. Then, we investigate possible machine learning methods for two subtasks, NIHSS item and score recognition and item-score relation extraction. In the evaluation, we conduct both task-specific and end-to-end evaluations and compare our method with the rule-based method using precision, recall and F1 scores as evaluation metrics.

RESULTS

We use all available discharge summaries of stroke cases in MIMIC-III. The annotated NIHSS corpus contains 312 cases, 2929 scale items, 2774 scores and 2733 relations. The results show that the best F1-score of our method was 0.9006, which was attained by combining BERT-BiLSTM-CRF and Random Forest, and it outperformed the rule-based method (F1-score = 0.8098). In the end-to-end task, our method could successfully recognize the item "1b level of consciousness questions", the score "1" and their relation "('1b level of consciousness questions', '1', 'has value')" from the sentence "1b level of consciousness questions: said name = 1", while the rule-based method could not.

CONCLUSIONS

The two-step pipeline method we propose is an effective approach to identify NIHSS items, scores and their relations. With its help, clinical investigators can easily retrieve and access structured scale data, thereby supporting stroke-related real-world studies.

Optimizing healthcare system by amalgamation of text processing and deep learning: a systematic review

The explosion of clinical textual data has drawn the attention of researchers. Owing to the abundance of clinical data, it is becoming difficult for healthcare professionals to take real-time measures. The tools and methods are lacking when compared to the amount of clinical data generated every day. This review aims to survey the text processing pipeline with deep learning methods such as CNN, RNN, LSTM, and GRU in the healthcare domain and discuss various applications such as clinical concept detection and extraction, medically aware dialogue systems, sentiment analysis of drug reviews shared online, clinical trial matching, and pharmacovigilance. In addition, we highlighted the major challenges in deploying text processing with deep learning to clinical textual data and identified the scope of research in this domain. Furthermore, we have discussed various resources that can be used in the future to optimize the healthcare domain by amalgamating text processing and deep learning.

Deep Learning to Refine the Identification of High-Quality Clinical Research Articles from the Biomedical Literature: Performance Evaluation

BACKGROUND

Identifying practice-ready evidence-based journal articles in medicine is a challenge due to the sheer volume of biomedical research publications. Newer approaches to support evidence discovery apply deep learning techniques to improve the efficiency and accuracy of classifying sound evidence.

OBJECTIVE

To determine how well deep learning models using variants of Bidirectional Encoder Representations from Transformers (BERT) identify high-quality evidence with high clinical relevance from the biomedical literature for consideration in clinical practice.

METHODS

We fine-tuned variations of BERT models (BERT_BASE, BioBERT, BlueBERT, and PubMedBERT) and compared their performance in classifying articles based on methodological quality criteria. The dataset used for fine-tuning models included titles and abstracts of >160,000 PubMed records from 2012-2020 that were of interest to human health which had been manually labeled based on meeting established critical appraisal criteria for methodological rigor. The data was randomly divided into 80:10:10 sets for training, validating, and testing. In addition to using the full unbalanced set, the training data was randomly undersampled into four balanced datasets to assess performance and select the best performing model. For each of the four sets, one model that maintained sensitivity (recall) at ≥99% was selected and were ensembled. The best performing model was evaluated in a prospective, blinded test and applied to an established reference standard, the Clinical Hedges dataset.

RESULTS

In training, three of the four selected best performing models were trained using BioBERT_BASE. The ensembled model did not boost performance compared with the best individual model. Hence a solo BioBERT-based model (named DL-PLUS) was selected for further testing as it was computationally more efficient. The model had high recall (>99%) and 60% to 77% specificity in a prospective evaluation conducted with blinded research associates and saved >60% of the work required to identify high quality articles.

CONCLUSIONS

Deep learning using pretrained language models and a large dataset of classified articles produced models with improved specificity while maintaining >99% recall. The resulting DL-PLUS model identifies high-quality, clinically relevant articles from PubMed at the time of publication. The model improves the efficiency of a literature surveillance program, which allows for faster dissemination of appraised research.

Heart disease risk factors detection from electronic health records using advanced NLP and deep learning techniques

Heart disease remains the major cause of death, despite recent improvements in prediction and prevention. Risk factor identification is the main step in diagnosing and preventing heart disease. Automatically detecting risk factors for heart disease in clinical notes can help with disease progression modeling and clinical decision-making. Many studies have attempted to detect risk factors for heart disease, but none have identified all risk factors. These studies have proposed hybrid systems that combine knowledge-driven and data-driven techniques, based on dictionaries, rules, and machine learning methods that require significant human effort. The National Center for Informatics for Integrating Biology and Beyond (i2b2) proposed a clinical natural language processing (NLP) challenge in 2014, with a track (track2) focused on detecting risk factors for heart disease risk factors in clinical notes over time. Clinical narratives provide a wealth of information that can be extracted using NLP and Deep Learning techniques. The objective of this paper is to improve on previous work in this area as part of the 2014 i2b2 challenge by identifying tags and attributes relevant to disease diagnosis, risk factors, and medications by providing advanced techniques of using stacked word embeddings. The i2b2 heart disease risk factors challenge dataset has shown significant improvement by using the approach of stacking embeddings, which combines various embeddings. Our model achieved an F1 score of 93.66% by using BERT and character embeddings (CHARACTER-BERT Embedding) stacking. The proposed model has significant results compared to all other models and systems that we developed for the 2014 i2b2 challenge.

Discerning conversational context in online health communities for personalized digital behavior change solutions using Pragmatics to Reveal Intent in Social Media (PRISM) framework

BACKGROUND

Online health communities (OHCs) have emerged as prominent platforms for behavior modification, and the digitization of online peer interactions has afforded researchers with unique opportunities to model multilevel mechanisms that drive behavior change. Existing studies, however, have been limited by a lack of methods that allow the capture of conversational context and socio-behavioral dynamics at scale, as manifested in these digital platforms.

OBJECTIVE

We develop, evaluate, and apply a novel methodological framework, Pragmatics to Reveal Intent in Social Media (PRISM), to facilitate granular characterization of peer interactions by combining multidimensional facets of human communication.

METHODS

We developed and applied PRISM to analyze peer interactions (N = 2.23 million) in QuitNet, an OHC for tobacco cessation. First, we generated a labeled set of peer interactions (n = 2,005) through manual annotation along three dimensions: communication themes (CTs), behavior change techniques (BCTs), and speech acts (SAs). Second, we used deep learning models to apply our qualitative codes at scale. Third, we applied our validated model to perform a retrospective analysis. Finally, using social network analysis (SNA), we portrayed large-scale patterns and relationships among the aforementioned communication dimensions embedded in peer interactions in QuitNet.

RESULTS

Qualitative analysis showed that the themes of social support and behavioral progress were common. The most used BCTs were feedback and monitoring and comparison of behavior, and users most commonly expressed their intentions using SAs-expressive and emotion. With additional in-domain pre-training, bidirectional encoder representations from Transformers (BERT) outperformed other deep learning models on the classification tasks. Content-specific SNA revealed that users' engagement or abstinence status is associated with the prevalence of various categories of BCTs and SAs, which also was evident from the visualization of network structures.

CONCLUSIONS

Our study describes the interplay of multilevel characteristics of online communication and their association with individual health behaviors.

Opioid death projections with AI-based forecasts using social media language

Targeting of location-specific aid for the U.S. opioid epidemic is difficult due to our inability to accurately predict changes in opioid mortality across heterogeneous communities. AI-based language analyses, having recently shown promise in cross-sectional (between-community) well-being assessments, may offer a way to more accurately longitudinally predict community-level overdose mortality. Here, we develop and evaluate, TROP (Transformer for Opiod Prediction), a model for community-specific trend projection that uses community-specific social media language along with past opioid-related mortality data to predict future changes in opioid-related deaths. TOP builds on recent advances in sequence modeling, namely transformer networks, to use changes in yearly language on Twitter and past mortality to project the following year's mortality rates by county. Trained over five years and evaluated over the next two years TROP demonstrated state-of-the-art accuracy in predicting future county-specific opioid trends. A model built using linear auto-regression and traditional socioeconomic data gave 7% error (MAPE) or within 2.93 deaths per 100,000 people on average; our proposed architecture was able to forecast yearly death rates with less than half that error: 3% MAPE and within 1.15 per 100,000 people.

Integrating structured and unstructured data for timely prediction of bloodstream infection among children

BACKGROUND

Hospitalized children with central venous lines (CVLs) are at higher risk of hospital-acquired infections. Information in electronic health records (EHRs) can be employed in training deep learning models to predict the onset of these infections. We incorporated clinical notes in addition to structured EHR data to predict serious bloodstream infections, defined as positive blood culture followed by at least 4 days of new antimicrobial agent administration, among hospitalized children with CVLs.

METHODS

Structured EHR information and clinical notes were extracted for a retrospective cohort including all hospitalized patients with CVLs at a single tertiary care pediatric health system from 2013 to 2018. Deep learning models were trained to determine the added benefit of incorporating the information embedded in clinical notes in predicting serious bloodstream infection.

RESULTS

A total of 24,351 patient encounters met inclusion criteria. The best-performing model restricted to structured EHR data had a specificity of 0.951 and positive predictive value (PPV) of 0.056 when the sensitivity was set to 0.85. The addition of contextualized word embeddings improved the specificity to 0.981 and PPV to 0.113.

CONCLUSIONS

Integrating clinical notes with structured EHR data improved the prediction of serious bloodstream infections among pediatric patients with CVLs.

IMPACT

Developed an advanced infection prediction model in pediatrics that integrates the structured and unstructured EHRs. Extracted information from clinical notes to do timely prediction in a clinical setting. Developed a deep learning model framework that can be employed in predicting rare events in a complex and dynamic environment.

Precision information extraction for rare disease epidemiology at scale

BACKGROUND

The United Nations recently made a call to address the challenges of an estimated 300 million persons worldwide living with a rare disease through the collection, analysis, and dissemination of disaggregated data. Epidemiologic Information (EI) regarding prevalence and incidence data of rare diseases is sparse and current paradigms of identifying, extracting, and curating EI rely upon time-intensive, error-prone manual processes. With these limitations, a clear understanding of the variation in epidemiology and outcomes for rare disease patients is hampered. This challenges the public health of rare diseases patients through a lack of information necessary to prioritize research, policy decisions, therapeutic development, and health system allocations.

METHODS

In this study, we developed a newly curated epidemiology corpus for Named Entity Recognition (NER), a deep learning framework, and a novel rare disease epidemiologic information pipeline named EpiPipeline4RD consisting of a web interface and Restful API. For the corpus creation, we programmatically gathered a representative sample of rare disease epidemiologic abstracts, utilized weakly-supervised machine learning techniques to label the dataset, and manually validated the labeled dataset. For the deep learning framework development, we fine-tuned our dataset and adapted the BioBERT model for NER. We measured the performance of our BioBERT model for epidemiology entity recognition quantitatively with precision, recall, and F1 and qualitatively through a comparison with Orphanet. We demonstrated the ability for our pipeline to gather, identify, and extract epidemiology information from rare disease abstracts through three case studies.

RESULTS

We developed a deep learning model to extract EI with overall F1 scores of 0.817 and 0.878, evaluated at the entity-level and token-level respectively, and which achieved comparable qualitative results to Orphanet's collection paradigm. Additionally, case studies of the rare diseases Classic homocystinuria, GRACILE syndrome, Phenylketonuria demonstrated the adequate recall of abstracts with epidemiology information, high precision of epidemiology information extraction through our deep learning model, and the increased efficiency of EpiPipeline4RD compared to a manual curation paradigm.

CONCLUSIONS

EpiPipeline4RD demonstrated high performance of EI extraction from rare disease literature to augment manual curation processes. This automated information curation paradigm will not only effectively empower development of the NIH Genetic and Rare Diseases Information Center (GARD), but also support the public health of the rare disease community.

Fine-tuning BERT for automatic ADME semantic labeling in FDA drug labeling to enhance product-specific guidance assessment

Product-specific guidances (PSGs) recommended by the United States Food and Drug Administration (FDA) are instrumental to promote and guide generic drug product development. To assess a PSG, the FDA assessor needs to take extensive time and effort to manually retrieve supportive drug information of absorption, distribution, metabolism, and excretion (ADME) from the reference listed drug labeling. In this work, we leveraged the state-of-the-art pre-trained language models to automatically label the ADME paragraphs in the pharmacokinetics section from the FDA-approved drug labeling to facilitate PSG assessment. We applied a transfer learning approach by fine-tuning the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model to develop a novel application of ADME semantic labeling, which can automatically retrieve ADME paragraphs from drug labeling instead of manual work. We demonstrate that fine-tuning the pre-trained BERT model can outperform conventional machine learning techniques, achieving up to 12.5% absolute F1 improvement. To our knowledge, we were the first to successfully apply BERT to solve the ADME semantic labeling task. We further assessed the relative contribution of pre-training and fine-tuning to the overall performance of the BERT model in the ADME semantic labeling task using a series of analysis methods, such as attention similarity and layer-based ablations. Our analysis revealed that the information learned via fine-tuning is focused on task-specific knowledge in the top layers of the BERT, whereas the benefit from the pre-trained BERT model is from the bottom layers.

Application of Deep Learning in Generating Structured Radiology Reports: A Transformer-Based Technique

Since radiology reports needed for clinical practice and research are written and stored in free-text narrations, extraction of relative information for further analysis is difficult. In these circumstances, natural language processing (NLP) techniques can facilitate automatic information extraction and transformation of free-text formats to structured data. In recent years, deep learning (DL)-based models have been adapted for NLP experiments with promising results. Despite the significant potential of DL models based on artificial neural networks (ANN) and convolutional neural networks (CNN), the models face some limitations to implement in clinical practice. Transformers, another new DL architecture, have been increasingly applied to improve the process. Therefore, in this study, we propose a transformer-based fine-grained named entity recognition (NER) architecture for clinical information extraction. We collected 88 abdominopelvic sonography reports in free-text formats and annotated them based on our developed information schema. The text-to-text transfer transformer model (T5) and Scifive, a pre-trained domain-specific adaptation of the T5 model, were applied for fine-tuning to extract entities and relations and transform the input into a structured format. Our transformer-based model in this study outperformed previously applied approaches such as ANN and CNN models based on ROUGE-1, ROUGE-2, ROUGE-L, and BLEU scores of 0.816, 0.668, 0.528, and 0.743, respectively, while providing an interpretable structured report.

The Diagnostic and Triage Accuracy of the GPT-3 Artificial Intelligence Model

Importance

Artificial intelligence (AI) applications in health care have been effective in many areas of medicine, but they are often trained for a single task using labeled data, making deployment and generalizability challenging. Whether a general-purpose AI language model can perform diagnosis and triage is unknown.

Objective

Compare the general-purpose Generative Pre-trained Transformer 3 (GPT-3) AI model's diagnostic and triage performance to attending physicians and lay adults who use the Internet.

Design

We compared the accuracy of GPT-3's diagnostic and triage ability for 48 validated case vignettes of both common (e.g., viral illness) and severe (e.g., heart attack) conditions to lay people and practicing physicians. Finally, we examined how well calibrated GPT-3's confidence was for diagnosis and triage.

Setting and Participants

The GPT-3 model, a nationally representative sample of lay people, and practicing physicians.

Exposure

Validated case vignettes (<60 words; <6th grade reading level).

Main Outcomes and Measures

Correct diagnosis, correct triage.

Results

Among all cases, GPT-3 replied with the correct diagnosis in its top 3 for 88% (95% CI, 75% to 94%) of cases, compared to 54% (95% CI, 53% to 55%) for lay individuals (p<0.001) and 96% (95% CI, 94% to 97%) for physicians (p=0.0354). GPT-3 triaged (71% correct; 95% CI, 57% to 82%) similarly to lay individuals (74%; 95% CI, 73% to 75%; p=0.73); both were significantly worse than physicians (91%; 95% CI, 89% to 93%; p<0.001). As measured by the Brier score, GPT-3 confidence in its top prediction was reasonably well-calibrated for diagnosis (Brier score = 0.18) and triage (Brier score = 0.22).

Conclusions and Relevance

A general-purpose AI language model without any content-specific training could perform diagnosis at levels close to, but below physicians and better than lay individuals. The model was performed less well on triage, where its performance was closer to that of lay individuals.

Edge Weight Updating Neural Network for Named Entity Normalization

Discriminating the matched named entity pairs or identifying the entities' canonical forms are critical in text mining tasks. More precise named entity normalization in text mining will benefit other subsequent text analytic applications. We built the named entity normalization model with a novel edge weight updating neural network. We, next, verify our model's performance on NCBI disease, BC5CDR disease, and BC5CDR chemical databases, which are widely used named entity normalization datasets in the bioinformatics field. We also tested our model with our own financial named entity normalization dataset to validate the efficacy for more general applications. Using the constructed dataset, we differentiate named entity pairs. Our model achieved the highest named entity normalization performances in terms of various evaluation metrics. Our proposed model when tested on four different datasets achieved state-of-the-art results.

Shifting machine learning for healthcare from development to deployment and from models to data

In the past decade, the application of machine learning (ML) to healthcare has helped drive the automation of physician tasks as well as enhancements in clinical capabilities and access to care. This progress has emphasized that, from model development to model deployment, data play central roles. In this Review, we provide a data-centric view of the innovations and challenges that are defining ML for healthcare. We discuss deep generative models and federated learning as strategies to augment datasets for improved model performance, as well as the use of the more recent transformer models for handling larger datasets and enhancing the modelling of clinical text. We also discuss data-focused problems in the deployment of ML, emphasizing the need to efficiently deliver data to ML models for timely clinical predictions and to account for natural data shifts that can deteriorate model performance.

Chinese Clinical Named Entity Recognition from Electronic Medical Records based on Multi-semantic Features by using RoBERTa-wwm and CNN: Model Development and Validation (Preprint)

BACKGROUND

Clinical electronic medical records (EMRs) contain important information on patients' anatomy, symptoms, examinations, diagnoses, and medications. Large-scale mining of rich medical information from EMRs will provide notable reference value for medical research. With the complexity of Chinese grammar and blurred boundaries of Chinese words, Chinese clinical named entity recognition (CNER) remains a notable challenge. Follow-up tasks such as medical entity structuring, medical entity standardization, medical entity relationship extraction, and medical knowledge graph construction largely depend on medical named entity recognition effects. A promising CNER result would provide reliable support for building domain knowledge graphs, knowledge bases, and knowledge retrieval systems. Furthermore, it would provide research ideas for scientists and medical decision-making references for doctors and even guide patients on disease and health management. Therefore, obtaining excellent CNER results is essential.

OBJECTIVE

We aimed to propose a Chinese CNER method to learn semantics-enriched representations for comprehensively enhancing machines to understand deep semantic information of EMRs by using multisemantic features, which makes medical information more readable and understandable.

METHODS

First, we used Robustly Optimized Bidirectional Encoder Representation from Transformers Pretraining Approach Whole Word Masking (RoBERTa-wwm) with dynamic fusion and Chinese character features, including 5-stroke code, Zheng code, phonological code, and stroke code, extracted by 1-dimensional convolutional neural networks (CNNs) to obtain fine-grained semantic features of Chinese characters. Subsequently, we converted Chinese characters into square images to obtain Chinese character image features from another modality by using a 2-dimensional CNN. Finally, we input multisemantic features into Bidirectional Long Short-Term Memory with Conditional Random Fields to achieve Chinese CNER. The effectiveness of our model was compared with that of the baseline and existing research models, and the features involved in the model were ablated and analyzed to verify the model's effectiveness.

RESULTS

We collected 1379 Yidu-S4K EMRs containing 23,655 entities in 6 categories and 2007 self-annotated EMRs containing 118,643 entities in 7 categories. The experiments showed that our model outperformed the comparison experiments, with F₁-scores of 89.28% and 84.61% on the Yidu-S4K and self-annotated data sets, respectively. The results of the ablation analysis demonstrated that each feature and method we used could improve the entity recognition ability.

CONCLUSIONS

Our proposed CNER method would mine the richer deep semantic information in EMRs by multisemantic embedding using RoBERTa-wwm and CNNs, enhancing the semantic recognition of characters at different granularity levels and improving the generalization capability of the method by achieving information complementarity among different semantic features, thus making the machine semantically understand EMRs and improving the CNER task accuracy.

Quality assessment of functional status documentation in EHRs across different healthcare institutions

The secondary use of electronic health records (EHRs) faces challenges in the form of varying data quality-related issues. To address that, we retrospectively assessed the quality of functional status documentation in EHRs of persons participating in Mayo Clinic Study of Aging (MCSA). We used a convergent parallel design to collect quantitative and qualitative data and independently analyzed the findings. We discovered a heterogeneous documentation process, where the care practice teams, institutions, and EHR systems all play an important role in how text data is documented and organized. Four prevalent instrument-assisted documentation (iDoc) expressions were identified based on three distinct instruments: Epic smart form, questionnaire, and occupational therapy and physical therapy templates. We found strong differences in the usage, information quality (intrinsic and contextual), and naturality of language among different type of iDoc expressions. These variations can be caused by different source instruments, information providers, practice settings, care events and institutions. In addition, iDoc expressions are context specific and thus shall not be viewed and processed uniformly. We recommend conducting data quality assessment of unstructured EHR text prior to using the information.

A Deep Learning-Based Privacy-Preserving Model for Smart Healthcare in Internet of Medical Things Using Fog Computing

With the emergence of COVID-19, smart healthcare, the Internet of Medical Things, and big data-driven medical applications have become even more important. The biomedical data produced is highly confidential and private. Unfortunately, conventional health systems cannot support such a colossal amount of biomedical data. Hence, data is typically stored and shared through the cloud. The shared data is then used for different purposes, such as research and discovery of unprecedented facts. Typically, biomedical data appear in textual form (e.g., test reports, prescriptions, and diagnosis). Unfortunately, such data is prone to several security threats and attacks, for example, privacy and confidentiality breach. Although significant progress has been made on securing biomedical data, most existing approaches yield long delays and cannot accommodate real-time responses. This paper proposes a novel fog-enabled privacy-preserving model called δ r sanitizer, which uses deep learning to improve the healthcare system. The proposed model is based on a Convolutional Neural Network with Bidirectional-LSTM and effectively performs Medical Entity Recognition. The experimental results show that δ r sanitizer outperforms the state-of-the-art models with 91.14% recall, 92.63% in precision, and 92% F1-score. The sanitization model shows 28.77% improved utility preservation as compared to the state-of-the-art.

"Note Bloat" impacts deep learning-based NLP models for clinical prediction tasks

One unintended consequence of the Electronic Health Records (EHR) implementation is the overuse of content-importing technology, such as copy-and-paste, that creates "bloated" notes containing large amounts of textual redundancy. Despite the rising interest in applying machine learning models to learn from real-patient data, it is unclear how the phenomenon of note bloat might affect the Natural Language Processing (NLP) models derived from these notes. Therefore, in this work we examine the impact of redundancy on deep learning-based NLP models, considering four clinical prediction tasks using a publicly available EHR database. We applied two deduplication methods to the hospital notes, identifying large quantities of redundancy, and found that removing the redundancy usually has little negative impact on downstream performances, and can in certain circumstances assist models to achieve significantly better results. We also showed it is possible to attack model predictions by simply adding note duplicates, causing changes of correct predictions made by trained models into wrong predictions. In conclusion, we demonstrated that EHR text redundancy substantively affects NLP models for clinical prediction tasks, showing that the awareness of clinical contexts and robust modeling methods are important to create effective and reliable NLP systems in healthcare contexts.

Evaluation of clinical named entity recognition methods for Serbian electronic health records

BACKGROUND AND OBJECTIVES

The importance of clinical natural language processing (NLP) has increased with the adoption of electronic health records (EHRs). One of the critical tasks in clinical NLP is named entity recognition (NER). Clinical NER in the Serbian language is a severely under-researched area. The few approaches that have been proposed so far are based on rules or machine-learning models with hand-crafted features, while current state-of-the-art models have not been explored. The objective of this paper is to assess the performance of state-of-the-art NER methods on clinical narratives in the Serbian language.

MATERIALS AND METHODS

We designed an experimental setup for a comprehensive evaluation of state-of-the-art NER models. The gold standard corpus we used for the evaluation is comprised of discharge summaries from the Clinic for Nephrology at the University Clinical Center of Serbia. The following models were evaluated: conditional random fields (CRF), multilingual transformers (BERT Multilingual and XLM RoBERTa), and long short-term memory (LSTM) recurrent neural networks, and their ensembles. In addition, we investigated the necessity of the pretraining task of transformer based models and the use of pretrained word embeddings with LSTM model.

RESULTS

Our results show that individually CRF had the best precision, the pretrained BERT Multilingual model had the best recall values, and the LSTM model had the best F1 score. The best performance was achieved by combining the existing models in a majority voting ensemble with an F1 score of 0.892. The presented results are similar to the inter annotator agreement on our gold standard corpus and are comparable to existing state-of-the-art results for clinical NER reported in literature.

CONCLUSION

Existing state-of-the-art models can provide viable results for clinical named entity recognition when applied to languages with the complexity of the Serbian language without major modifications.

Research on Aspect-Level Sentiment Analysis Based on Text Comments

Sentiment analysis is the processing of textual data and giving positive or negative opinions to sentences. In the ABSA dataset, most sentences contain one aspect of sentiment polarity, or sentences of one aspect have multiple identical sentiment polarities, which weakens the sentiment polarity of the ABSA dataset. Therefore, this paper uses the SemEval 14 Restaurant Review dataset, in which each document is symmetrically divided into individual sentences, and two versions of the datasets ATSA and ACSA are created. ATSA: Aspect Term Sentiment Analysis Dataset. ACSA: Aspect Category Sentiment Analysis Dataset. In order to symmetrically simulate the complex relationship between aspect contexts and accurately extract the polarity of emotional features, this paper combines the latest development trend of NLP, combines capsule network and BRET, and proposes the baseline model CapsNet-BERT. The experimental results verify the effectiveness of the model.

Improving medical term embeddings using UMLS Metathesaurus

Background

Health providers create Electronic Health Records (EHRs) to describe the conditions and procedures used to treat their patients. Medical notes entered by medical staff in the form of free text are a particularly insightful component of EHRs. There is a great interest in applying machine learning tools on medical notes in numerous medical informatics applications. Learning vector representations, or embeddings, of terms in the notes, is an important pre-processing step in such applications. However, learning good embeddings is challenging because medical notes are rich in specialized terminology, and the number of available EHRs in practical applications is often very small.

Methods

In this paper, we propose a novel algorithm to learn embeddings of medical terms from a limited set of medical notes. The algorithm, called definition2vec, exploits external information in the form of medical term definitions. It is an extension of a skip-gram algorithm that incorporates textual definitions of medical terms provided by the Unified Medical Language System (UMLS) Metathesaurus.

Results

To evaluate the proposed approach, we used a publicly available Medical Information Mart for Intensive Care (MIMIC-III) EHR data set. We performed quantitative and qualitative experiments to measure the usefulness of the learned embeddings. The experimental results show that definition2vec keeps the semantically similar medical terms together in the embedding vector space even when they are rare or unobserved in the corpus. We also demonstrate that learned vector embeddings are helpful in downstream medical informatics applications.

Conclusion

This paper shows that medical term definitions can be helpful when learning embeddings of rare or previously unseen medical terms from a small corpus of specialized documents such as medical notes.

Benchmarking for biomedical natural language processing tasks with a domain specific ALBERT

BACKGROUND

The abundance of biomedical text data coupled with advances in natural language processing (NLP) is resulting in novel biomedical NLP (BioNLP) applications. These NLP applications, or tasks, are reliant on the availability of domain-specific language models (LMs) that are trained on a massive amount of data. Most of the existing domain-specific LMs adopted bidirectional encoder representations from transformers (BERT) architecture which has limitations, and their generalizability is unproven as there is an absence of baseline results among common BioNLP tasks.

RESULTS

We present 8 variants of BioALBERT, a domain-specific adaptation of a lite bidirectional encoder representations from transformers (ALBERT), trained on biomedical (PubMed and PubMed Central) and clinical (MIMIC-III) corpora and fine-tuned for 6 different tasks across 20 benchmark datasets. Experiments show that a large variant of BioALBERT trained on PubMed outperforms the state-of-the-art on named-entity recognition (+ 11.09% BLURB score improvement), relation extraction (+ 0.80% BLURB score), sentence similarity (+ 1.05% BLURB score), document classification (+ 0.62% F1-score), and question answering (+ 2.83% BLURB score). It represents a new state-of-the-art in 5 out of 6 benchmark BioNLP tasks.

CONCLUSIONS

The large variant of BioALBERT trained on PubMed achieved a higher BLURB score than previous state-of-the-art models on 5 of the 6 benchmark BioNLP tasks. Depending on the task, 5 different variants of BioALBERT outperformed previous state-of-the-art models on 17 of the 20 benchmark datasets, showing that our model is robust and generalizable in the common BioNLP tasks. We have made BioALBERT freely available which will help the BioNLP community avoid computational cost of training and establish a new set of baselines for future efforts across a broad range of BioNLP tasks.

Can natural language processing models extract and classify instances of interpersonal violence in mental healthcare electronic records: an applied evaluative study

OBJECTIVE

This paper evaluates the application of a natural language processing (NLP) model for extracting clinical text referring to interpersonal violence using electronic health records (EHRs) from a large mental healthcare provider.

DESIGN

A multidisciplinary team iteratively developed guidelines for annotating clinical text referring to violence. Keywords were used to generate a dataset which was annotated (ie, classified as affirmed, negated or irrelevant) for: presence of violence, patient status (ie, as perpetrator, witness and/or victim of violence) and violence type (domestic, physical and/or sexual). An NLP approach using a pretrained transformer model, BioBERT (Bidirectional Encoder Representations from Transformers for Biomedical Text Mining) was fine-tuned on the annotated dataset and evaluated using 10-fold cross-validation.

SETTING

We used the Clinical Records Interactive Search (CRIS) database, comprising over 500 000 de-identified EHRs of patients within the South London and Maudsley NHS Foundation Trust, a specialist mental healthcare provider serving an urban catchment area.

PARTICIPANTS

Searches of CRIS were carried out based on 17 predefined keywords. Randomly selected text fragments were taken from the results for each keyword, amounting to 3771 text fragments from the records of 2832 patients.

OUTCOME MEASURES

We estimated precision, recall and F1 score for each NLP model. We examined sociodemographic and clinical variables in patients giving rise to the text data, and frequencies for each annotated violence characteristic.

RESULTS

Binary classification models were developed for six labels (violence presence, perpetrator, victim, domestic, physical and sexual). Among annotations affirmed for the presence of any violence, 78% (1724) referred to physical violence, 61% (1350) referred to patients as perpetrator and 33% (731) to domestic violence. NLP models' precision ranged from 89% (perpetrator) to 98% (sexual); recall ranged from 89% (victim, perpetrator) to 97% (sexual).

CONCLUSIONS

State of the art NLP models can extract and classify clinical text on violence from EHRs at acceptable levels of scale, efficiency and accuracy.

The h-ANN Model: Comprehensive Colonoscopy Concept Compilation Using Combined Contextual Embeddings

Colonoscopy is a screening and diagnostic procedure for detection of colorectal carcinomas with specific quality metrics that monitor and improve adenoma detection rates. These quality metrics are stored in disparate documents i.e., colonoscopy, pathology, and radiology reports. The lack of integrated standardized documentation is impeding colorectal cancer research. Clinical concept extraction using Natural Language Processing (NLP) and Machine Learning (ML) techniques is an alternative to manual data abstraction. Contextual word embedding models such as BERT (Bidirectional Encoder Representations from Transformers) and FLAIR have enhanced performance of NLP tasks. Combining multiple clinically-trained embeddings can improve word representations and boost the performance of the clinical NLP systems. The objective of this study is to extract comprehensive clinical concepts from the consolidated colonoscopy documents using concatenated clinical embeddings. We built high-quality annotated corpora for three report types. BERT and FLAIR embeddings were trained on unlabeled colonoscopy related documents. We built a hybrid Artificial Neural Network (h-ANN) to concatenate and fine-tune BERT and FLAIR embeddings. To extract concepts of interest from three report types, 3 models were initialized from the h-ANN and fine-tuned using the annotated corpora. The models achieved best F1-scores of 91.76%, 92.25%, and 88.55% for colonoscopy, pathology, and radiology reports respectively.

Word embeddings trained on published case reports are lightweight, effective for clinical tasks, and free of protected health information

OBJECTIVE

Quantify tradeoffs in performance, reproducibility, and resource demands across several strategies for developing clinically relevant word embeddings.

MATERIALS AND METHODS

We trained separate embeddings on all full-text manuscripts in the Pubmed Central (PMC) Open Access subset, case reports therein, the English Wikipedia corpus, the Medical Information Mart for Intensive Care (MIMIC) III dataset, and all notes in the University of Pennsylvania Health System (UPHS) electronic health record. We tested embeddings in six clinically relevant tasks including mortality prediction and de-identification, and assessed performance using the scaled Brier score (SBS) and the proportion of notes successfully de-identified, respectively.

RESULTS

Embeddings from UPHS notes best predicted mortality (SBS 0.30, 95% CI 0.15 to 0.45) while Wikipedia embeddings performed worst (SBS 0.12, 95% CI -0.05 to 0.28). Wikipedia embeddings most consistently (78% of notes) and the full PMC corpus embeddings least consistently (48%) de-identified notes. Across all six tasks, the full PMC corpus demonstrated the most consistent performance, and the Wikipedia corpus the least. Corpus size ranged from 49 million tokens (PMC case reports) to 10 billion (UPHS).

DISCUSSION

Embeddings trained on published case reports performed as least as well as embeddings trained on other corpora in most tasks, and clinical corpora consistently outperformed non-clinical corpora. No single corpus produced a strictly dominant set of embeddings across all tasks and so the optimal training corpus depends on intended use.

CONCLUSION

Embeddings trained on published case reports performed comparably on most clinical tasks to embeddings trained on larger corpora. Open access corpora allow training of clinically relevant, effective, and reproducible embeddings.

AMMU: A survey of transformer-based biomedical pretrained language models

Transformer-based pretrained language models (PLMs) have started a new era in modern natural language processing (NLP). These models combine the power of transformers, transfer learning, and self-supervised learning (SSL). Following the success of these models in the general domain, the biomedical research community has developed various in-domain PLMs starting from BioBERT to the latest BioELECTRA and BioALBERT models. We strongly believe there is a need for a survey paper that can provide a comprehensive survey of various transformer-based biomedical pretrained language models (BPLMs). In this survey, we start with a brief overview of foundational concepts like self-supervised learning, embedding layer and transformer encoder layers. We discuss core concepts of transformer-based PLMs like pretraining methods, pretraining tasks, fine-tuning methods, and various embedding types specific to biomedical domain. We introduce a taxonomy for transformer-based BPLMs and then discuss all the models. We discuss various challenges and present possible solutions. We conclude by highlighting some of the open issues which will drive the research community to further improve transformer-based BPLMs. The list of all the publicly available transformer-based BPLMs along with their links is provided at https://mr-nlp.github.io/posts/2021/05/transformer-based-biomedical-pretrained-language-models-list/.

Automatic extraction of 12 cardiovascular concepts from German discharge letters using pre-trained language models

Objective

A vast amount of medical data is still stored in unstructured text documents. We present an automated method of information extraction from German unstructured clinical routine data from the cardiology domain enabling their usage in state-of-the-art data-driven deep learning projects.

Methods

We evaluated pre-trained language models to extract a set of 12 cardiovascular concepts in German discharge letters. We compared three bidirectional encoder representations from transformers pre-trained on different corpora and fine-tuned them on the task of cardiovascular concept extraction using 204 discharge letters manually annotated by cardiologists at the University Hospital Heidelberg. We compared our results with traditional machine learning methods based on a long short-term memory network and a conditional random field.

Results

Our best performing model, based on publicly available German pre-trained bidirectional encoder representations from the transformer model, achieved a token-wise micro-average F1-score of 86% and outperformed the baseline by at least 6%. Moreover, this approach achieved the best trade-off between precision (positive predictive value) and recall (sensitivity).

Conclusion

Our results show the applicability of state-of-the-art deep learning methods using pre-trained language models for the task of cardiovascular concept extraction using limited training data. This minimizes annotation efforts, which are currently the bottleneck of any application of data-driven deep learning projects in the clinical domain for German and many other European languages.

Ensemble of Deep Masked Language Models for Effective Named Entity Recognition in Health and Life Science Corpora

The health and life science domains are well known for their wealth of named entities found in large free text corpora, such as scientific literature and electronic health records. To unlock the value of such corpora, named entity recognition (NER) methods are proposed. Inspired by the success of transformer-based pretrained models for NER, we assess how individual and ensemble of deep masked language models perform across corpora of different health and life science domains-biology, chemistry, and medicine-available in different languages-English and French. Individual deep masked language models, pretrained on external corpora, are fined-tuned on task-specific domain and language corpora and ensembled using classical majority voting strategies. Experiments show statistically significant improvement of the ensemble models over an individual BERT-based baseline model, with an overall best performance of 77% macro F1-score. We further perform a detailed analysis of the ensemble results and show how their effectiveness changes according to entity properties, such as length, corpus frequency, and annotation consistency. The results suggest that the ensembles of deep masked language models are an effective strategy for tackling NER across corpora from the health and life science domains.

A contextual multi-task neural approach to medication and adverse events identification from clinical text

Effective wide-scale pharmacovigilance calls for accurate named entity recognition (NER) of medication entities such as drugs, dosages, reasons, and adverse drug events (ADE) from clinical text. The scarcity of adverse event annotations and underlying semantic ambiguities make accurate scope identification challenging. The current research explores integrating contextualized language models and multi-task learning from diverse clinical NER datasets to mitigate this challenge. We propose a novel multi-task adaptation method to refine the embeddings generated by the Bidirectional Encoder Representations from Transformers (BERT) language model to improve inter-task knowledge sharing. We integrated the adapted BERT model into a unique hierarchical multi-task neural network comprised of the medication and auxiliary clinical NER tasks. We validated the model using two different versions of BERT on diverse well-studied clinical tasks: Medication and ADE (n2c2 2018/n2c2 2009), Clinical Concepts (n2c2 2010/n2c2 2012), Disorders (ShAReCLEF 2013). Overall medication extraction performance enhanced by up to +1.19 F1 (n2c2 2018) while generalization enhanced by +5.38 F1 (n2c2 2009) as compared to standalone BERT baselines. ADE recognition enhanced significantly (McNemar's test), out-performing prior baselines. Similar benefits were observed on the auxiliary clinical and disorder tasks. We demonstrate that combining multi-dataset BERT adaptation and multi-task learning out-performs prior medication extraction methods without requiring additional features, newer training data, or ensembling. Taken together, the study contributes an initial case study towards integrating diverse clinical datasets in an end-to-end NER model for clinical decision support.