Dataset Information

A comparative study of pretrained language models for long clinical text.

ABSTRACT:

Objective

Clinical knowledge-enriched transformer models (eg, ClinicalBERT) have state-of-the-art results on clinical natural language processing (NLP) tasks. One of the core limitations of these transformer models is the substantial memory consumption due to their full self-attention mechanism, which leads to the performance degradation in long clinical texts. To overcome this, we propose to leverage long-sequence transformer models (eg, Longformer and BigBird), which extend the maximum input sequence length from 512 to 4096, to enhance the ability to model long-term dependencies in long clinical texts.

Materials and methods

Inspired by the success of long-sequence transformer models and the fact that clinical notes are mostly long, we introduce 2 domain-enriched language models, Clinical-Longformer and Clinical-BigBird, which are pretrained on a large-scale clinical corpus. We evaluate both language models using 10 baseline tasks including named entity recognition, question answering, natural language inference, and document classification tasks.

Results

The results demonstrate that Clinical-Longformer and Clinical-BigBird consistently and significantly outperform ClinicalBERT and other short-sequence transformers in all 10 downstream tasks and achieve new state-of-the-art results.

Discussion

Our pretrained language models provide the bedrock for clinical NLP using long texts. We have made our source code available at https://github.com/luoyuanlab/Clinical-Longformer, and the pretrained models available for public download at: https://huggingface.co/yikuan8/Clinical-Longformer.

Conclusion

This study demonstrates that clinical knowledge-enriched long-sequence transformers are able to learn long-term dependencies in long clinical text. Our methods can also inspire the development of other domain-enriched long-sequence transformers.

SUBMITTER: Li Y

PROVIDER: S-EPMC9846675 | biostudies-literature | 2023 Jan

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

A comparative study of pretrained language models for long clinical text.

Li Yikuan Y Wehbe Ramsey M RM Ahmad Faraz S FS Wang Hanyin H Luo Yuan Y

Journal of the American Medical Informatics Association : JAMIA 20230101 2

<h4>Objective</h4>Clinical knowledge-enriched transformer models (eg, ClinicalBERT) have state-of-the-art results on clinical natural language processing (NLP) tasks. One of the core limitations of these transformer models is the substantial memory consumption due to their full self-attention mechanism, which leads to the performance degradation in long clinical texts. To overcome this, we propose to leverage long-sequence transformer models (eg, Longformer and BigBird), which extend the maximum ...[more]

PMID: 36451266

Similar Datasets

Project description:BackgroundIn recent years, social media has become a major channel for health-related information in Saudi Arabia. Prior health informatics studies have suggested that a large proportion of health-related posts on social media are inaccurate. Given the subject matter and the scale of dissemination of such information, it is important to be able to automatically discriminate between accurate and inaccurate health-related posts in Arabic.ObjectiveThe first aim of this study is to generate a data set of generic health-related tweets in Arabic, labeled as either accurate or inaccurate health information. The second aim is to leverage this data set to train a state-of-the-art deep learning model for detecting the accuracy of health-related tweets in Arabic. In particular, this study aims to train and compare the performance of multiple deep learning models that use pretrained word embeddings and transformer language models.MethodsWe used 900 health-related tweets from a previously published data set extracted between July 15, 2019, and August 31, 2019. Furthermore, we applied a pretrained model to extract an additional 900 health-related tweets from a second data set collected specifically for this study between March 1, 2019, and April 15, 2019. The 1800 tweets were labeled by 2 physicians as accurate, inaccurate, or unsure. The physicians agreed on 43.3% (779/1800) of tweets, which were thus labeled as accurate or inaccurate. A total of 9 variations of the pretrained transformer language models were then trained and validated on 79.9% (623/779 tweets) of the data set and tested on 20% (156/779 tweets) of the data set. For comparison, we also trained a bidirectional long short-term memory model with 7 different pretrained word embeddings as the input layer on the same data set. The models were compared in terms of their accuracy, precision, recall, F1 score, and macroaverage of the F1 score.ResultsWe constructed a data set of labeled tweets, 38% (296/779) of which were labeled as inaccurate health information, and 62% (483/779) of which were labeled as accurate health information. We suggest that this was highly efficacious as we did not include any tweets in which the physician annotators were unsure or in disagreement. Among the investigated deep learning models, the Transformer-based Model for Arabic Language Understanding version 0.2 (AraBERTv0.2)-large model was the most accurate, with an F1 score of 87%, followed by AraBERT version 2-large and AraBERTv0.2-base.ConclusionsOur results indicate that the pretrained language model AraBERTv0.2 is the best model for classifying tweets as carrying either inaccurate or accurate health information. Future studies should consider applying ensemble learning to combine the best models as it may produce better results.

Project description:The construction of hypernym taxonomic trees, a critical task in the field of natural language processing, involves extracting lexical relationships, specifically creating a tree structure that represents hypernym relationships among a given set of words within the same domain. In this work, we present a method for constructing hypernym taxonomy trees in the Chinese language domain, and we named it CHRRM (Chinese Hypernym Relationship Reasoning Model). Our method consists of two main steps: First, we utilize pre-trained models to predict hypernym relationships between pairs of words; second, we regard these relationships as edges to form a maximum spanning tree in the word graph. Our method enhances the effectiveness of constructing hypernym taxonomic trees based on pre-trained models through two key improvements: (1) We optimize the hyperparameter configuration for this task using pre-trained models from the Bert family and provide explanations for the configuration of these hyperparameters. (2) By employing generative large language models such as ChatGPT and ChatGLM to annotate words, we improve the accuracy of hypernym relationship identification and analyze the feasibility of applying generative large language models to the task of constructing taxonomy trees. We trained our model on subtrees of WORDNET and evaluated its performance on non-overlapping subtrees of WORDNET, demonstrating that our enhancements led to a significant relative improvement of 15.67%, achieving an F1 score of 67.9 on the Chinese WORDNET validation dataset compared to the previous score of 58.7. In conclusion, our study reveals the following key findings: (1) The Roberta-wwm-ext-large model consistently delivers outstanding results in constructing taxonomic trees. (2) Generative large language models, while capable of aiding pre-trained models in improving hypernym recognition accuracy, have limitations related to generation quality and computational resources. (3) Generative large language models can serve various NLP tasks either directly or indirectly; it is feasible to improve the downstream NLU task's performance through the generative content.

Project description:BackgroundDespite their growing use in health care, pretrained language models (PLMs) often lack clinical relevance due to insufficient domain expertise and poor interpretability. A key strategy to overcome these challenges is integrating external knowledge into PLMs, enhancing their adaptability and clinical usefulness. Current biomedical knowledge graphs like UMLS (Unified Medical Language System), SNOMED CT (Systematized Medical Nomenclature for Medicine-Clinical Terminology), and HPO (Human Phenotype Ontology), while comprehensive, fail to effectively connect general biomedical knowledge with physician insights. There is an equally important need for a model that integrates diverse knowledge in a way that is both unified and compartmentalized. This approach not only addresses the heterogeneous nature of domain knowledge but also recognizes the unique data and knowledge repositories of individual health care institutions, necessitating careful and respectful management of proprietary information.ObjectiveThis study aimed to enhance the clinical relevance and interpretability of PLMs by integrating external knowledge in a manner that respects the diversity and proprietary nature of health care data. We hypothesize that domain knowledge, when captured and distributed as stand-alone modules, can be effectively reintegrated into PLMs to significantly improve their adaptability and utility in clinical settings.MethodsWe demonstrate that through adapters, small and lightweight neural networks that enable the integration of extra information without full model fine-tuning, we can inject diverse sources of external domain knowledge into language models and improve the overall performance with an increased level of interpretability. As a practical application of this methodology, we introduce a novel task, structured as a case study, that endeavors to capture physician knowledge in assigning cardiovascular diagnoses from clinical narratives, where we extract diagnosis-comment pairs from electronic health records (EHRs) and cast the problem as text classification.ResultsThe study demonstrates that integrating domain knowledge into PLMs significantly improves their performance. While improvements with ClinicalBERT are more modest, likely due to its pretraining on clinical texts, BERT (bidirectional encoder representations from transformer) equipped with knowledge adapters surprisingly matches or exceeds ClinicalBERT in several metrics. This underscores the effectiveness of knowledge adapters and highlights their potential in settings with strict data privacy constraints. This approach also increases the level of interpretability of these models in a clinical context, which enhances our ability to precisely identify and apply the most relevant domain knowledge for specific tasks, thereby optimizing the model's performance and tailoring it to meet specific clinical needs.ConclusionsThis research provides a basis for creating health knowledge graphs infused with physician knowledge, marking a significant step forward for PLMs in health care. Notably, the model balances integrating knowledge both comprehensively and selectively, addressing the heterogeneous nature of medical knowledge and the privacy needs of health care institutions.

Project description:BackgroundAs advancements in artificial intelligence (AI) continue, large language models (LLMs) have emerged as promising tools for generating medical information. Their rapid adaptation and potential benefits in health care require rigorous assessment in terms of the quality, accuracy, and safety of the generated information across diverse medical specialties.ObjectiveThis study aimed to evaluate the performance of 4 prominent LLMs, namely, Claude-instant-v1.0, GPT-3.5-Turbo, Command-xlarge-nightly, and Bloomz, in generating medical content spanning the clinical specialties of ophthalmology, orthopedics, and dermatology.MethodsThree domain-specific physicians evaluated the AI-generated therapeutic recommendations for a diverse set of 60 diseases. The evaluation criteria involved the mDISCERN score, correctness, and potential harmfulness of the recommendations. ANOVA and pairwise t tests were used to explore discrepancies in content quality and safety across models and specialties. Additionally, using the capabilities of OpenAI's most advanced model, GPT-4, an automated evaluation of each model's responses to the diseases was performed using the same criteria and compared to the physicians' assessments through Pearson correlation analysis.ResultsClaude-instant-v1.0 emerged with the highest mean mDISCERN score (3.35, 95% CI 3.23-3.46). In contrast, Bloomz lagged with the lowest score (1.07, 95% CI 1.03-1.10). Our analysis revealed significant differences among the models in terms of quality (P<.001). Evaluating their reliability, the models displayed strong contrasts in their falseness ratings, with variations both across models (P<.001) and specialties (P<.001). Distinct error patterns emerged, such as confusing diagnoses; providing vague, ambiguous advice; or omitting critical treatments, such as antibiotics for infectious diseases. Regarding potential harm, GPT-3.5-Turbo was found to be the safest, with the lowest harmfulness rating. All models lagged in detailing the risks associated with treatment procedures, explaining the effects of therapies on quality of life, and offering additional sources of information. Pearson correlation analysis underscored a substantial alignment between physician assessments and GPT-4's evaluations across all established criteria (P<.01).ConclusionsThis study, while comprehensive, was limited by the involvement of a select number of specialties and physician evaluators. The straightforward prompting strategy ("How to treat…") and the assessment benchmarks, initially conceptualized for human-authored content, might have potential gaps in capturing the nuances of AI-driven information. The LLMs evaluated showed a notable capability in generating valuable medical content; however, evident lapses in content quality and potential harm signal the need for further refinements. Given the dynamic landscape of LLMs, this study's findings emphasize the need for regular and methodical assessments, oversight, and fine-tuning of these AI tools to ensure they produce consistently trustworthy and clinically safe medical advice. Notably, the introduction of an auto-evaluation mechanism using GPT-4, as detailed in this study, provides a scalable, transferable method for domain-agnostic evaluations, extending beyond therapy recommendation assessments.

Project description:BackgroundThe widespread use of electronic health records in the clinical and biomedical fields makes the removal of protected health information (PHI) essential to maintain privacy. However, a significant portion of information is recorded in unstructured textual forms, posing a challenge for deidentification. In multilingual countries, medical records could be written in a mixture of more than one language, referred to as code mixing. Most current clinical natural language processing techniques are designed for monolingual text, and there is a need to address the deidentification of code-mixed text.ObjectiveThe aim of this study was to investigate the effectiveness and underlying mechanism of fine-tuned pretrained language models (PLMs) in identifying PHI in the code-mixed context. Additionally, we aimed to evaluate the potential of prompting large language models (LLMs) for recognizing PHI in a zero-shot manner.MethodsWe compiled the first clinical code-mixed deidentification data set consisting of text written in Chinese and English. We explored the effectiveness of fine-tuned PLMs for recognizing PHI in code-mixed content, with a focus on whether PLMs exploit naming regularity and mention coverage to achieve superior performance, by probing the developed models' outputs to examine their decision-making process. Furthermore, we investigated the potential of prompt-based in-context learning of LLMs for recognizing PHI in code-mixed text.ResultsThe developed methods were evaluated on a code-mixed deidentification corpus of 1700 discharge summaries. We observed that different PHI types had preferences in their occurrences within the different types of language-mixed sentences, and PLMs could effectively recognize PHI by exploiting the learned name regularity. However, the models may exhibit suboptimal results when regularity is weak or mentions contain unknown words that the representations cannot generate well. We also found that the availability of code-mixed training instances is essential for the model's performance. Furthermore, the LLM-based deidentification method was a feasible and appealing approach that can be controlled and enhanced through natural language prompts.ConclusionsThe study contributes to understanding the underlying mechanism of PLMs in addressing the deidentification process in the code-mixed context and highlights the significance of incorporating code-mixed training instances into the model training phase. To support the advancement of research, we created a manipulated subset of the resynthesized data set available for research purposes. Based on the compiled data set, we found that the LLM-based deidentification method is a feasible approach, but carefully crafted prompts are essential to avoid unwanted output. However, the use of such methods in the hospital setting requires careful consideration of data security and privacy concerns. Further research could explore the augmentation of PLMs and LLMs with external knowledge to improve their strength in recognizing rare PHI.

Dataset Information

A comparative study of pretrained language models for long clinical text.

Objective

Materials and methods

Results

Discussion

Conclusion

Publications

A comparative study of pretrained language models for long clinical text.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets