Automated deidentification of radiology reports combining transformer and "hide in plain sight" rule-based methods

Automated anonymization of radiology reports: comparison of publicly available natural language processing and large language models

PURPOSE

Medical reports, governed by HIPAA regulations, contain personal health information (PHI), restricting secondary data use. Utilizing natural language processing (NLP) and large language models (LLM), we sought to employ publicly available methods to automatically anonymize PHI in free-text radiology reports.

MATERIALS AND METHODS

We compared two publicly available rule-based NLP models (spaCy; NLPac, accuracy-optimized; NLPsp, speed-optimized; iteratively improved on 400 free-text CT-reports (test set)) and one offline LLM approach (LLM-model, LLaMa-2, Meta-AI) for PHI-anonymization. The three models were tested on 100 randomly selected chest CT reports. Two investigators assessed the anonymization of occurring PHI entities and whether clinical information was removed. Subsequently, precision, recall, and F1 scores were calculated.

RESULTS

NLPac and NLPsp successfully removed all instances of dates (n = 333), medical record numbers (MRN) (n = 6), and accession numbers (ACC) (n = 92). The LLM model removed all MRNs, 96% of ACCs, and 32% of dates. NLPac was most consistent with a perfect F1-score of 1.00, followed by NLPsp with lower precision (0.86) and F1-score (0.92) for dates. The LLM model had perfect precision for MRNs, ACCs, and dates but the lowest recall for ACC (0.96) and dates (0.52), corresponding F1 scores of 0.98 and 0.68, respectively. Names were removed completely or majorly (i.e., one first or family name non-anonymized) in 100% (NLPac), 72% (NLPsp), and 90% (LLM-model). Importantly, NLPac and NLPsp did not remove medical information, while the LLM model did in 10% (n = 10).

CONCLUSION

Pre-trained NLP models can effectively anonymize free-text radiology reports, while anonymization with the LLM model is more prone to deleting medical information.

KEY POINTS

Question This study compares NLP and locally hosted LLM techniques to ensure PHI anonymization without losing clinical information. Findings Pre-trained NLP models effectively anonymized radiology reports without removing clinical data, while a locally hosted LLM was less reliable, risking the loss of important information. Clinical relevance Fast, reliable, automated anonymization of PHI from radiology reports enables HIPAA-compliant secondary use, facilitating advanced applications like LLM-driven radiology analysis while ensuring ethical handling of sensitive patient data.

Visualizing Clinical Data Retrieval and Curation in Multimodal Healthcare AI Research: A Technical Note on RIL-workflow

Curating and integrating data from sources are bottlenecks to procuring robust training datasets for artificial intelligence (AI) models in healthcare. While numerous applications can process discrete types of clinical data, it is still time-consuming to integrate heterogenous data types. Therefore, there exists a need for more efficient retrieval and storage of curated patient data from dissimilar sources, such as biobanks, health records, and sensors. We describe a customizable, modular data retrieval application (RIL-workflow), which integrates clinical notes, images, and prescription data, and show its feasibility applied to research at our institution. It uses the workflow automation platform Camunda (Camunda Services GmbH, Berlin, Germany) to collect internal data from Fast Healthcare Interoperability Resources (FHIR) and Digital Imaging and Communications in Medicine (DICOM) sources. Using the web-based graphical user interface (GUI), the workflow runs tasks to completion according to visual representation, retrieving and storing results for patients meeting study inclusion criteria while segregating errors for human review. We showcase RIL-workflow with its library of ready-to-use modules, enabling researchers to specify human input or automation at fixed steps. We validated our workflow by demonstrating its capability to aggregate, curate, and handle errors related to data from multiple sources to generate a multimodal database for clinical AI research. Further, we solicited user feedback to highlight the pros and cons associated with RIL-workflow. The source code is available at github.com/magnooj/RIL-workflow.

Pragmatic De-Identification of Cross-Domain Unstructured Documents: A Utility-Preserving Approach with Relation Extraction Filtering

The volume of information, and in particular personal information, generated each day is increasing at a staggering rate. The ability to leverage such information depends greatly on being able to satisfy the many compliance and privacy regulations that are appearing all over the world. We present READI, a utility preserving framework for the unstructured document de-identification. READI leverages Named Entity Recognition and Relation Extraction technology to improve the quality of the entity detection, thus improving the overall quality of the data de-identification process. In this proof of concept study, we evaluate the proposed approach on two different datasets and compare with the existing state-of-the-art approaches. We show that Relation Extraction-based Approach for De-Identification (READI) notably reduces the number of false positives and improves the utility of the de-identified text.

Leveraging large language models for generating responses to patient messages-a subjective analysis

OBJECTIVE

This study aimed to develop and assess the performance of fine-tuned large language models for generating responses to patient messages sent via an electronic health record patient portal.

MATERIALS AND METHODS

Utilizing a dataset of messages and responses extracted from the patient portal at a large academic medical center, we developed a model (CLAIR-Short) based on a pre-trained large language model (LLaMA-65B). In addition, we used the OpenAI API to update physician responses from an open-source dataset into a format with informative paragraphs that offered patient education while emphasizing empathy and professionalism. By combining with this dataset, we further fine-tuned our model (CLAIR-Long). To evaluate fine-tuned models, we used 10 representative patient portal questions in primary care to generate responses. We asked primary care physicians to review generated responses from our models and ChatGPT and rated them for empathy, responsiveness, accuracy, and usefulness.

RESULTS

The dataset consisted of 499 794 pairs of patient messages and corresponding responses from the patient portal, with 5000 patient messages and ChatGPT-updated responses from an online platform. Four primary care physicians participated in the survey. CLAIR-Short exhibited the ability to generate concise responses similar to provider's responses. CLAIR-Long responses provided increased patient educational content compared to CLAIR-Short and were rated similarly to ChatGPT's responses, receiving positive evaluations for responsiveness, empathy, and accuracy, while receiving a neutral rating for usefulness.

CONCLUSION

This subjective analysis suggests that leveraging large language models to generate responses to patient messages demonstrates significant potential in facilitating communication between patients and healthcare providers.

De-identification of clinical free text using natural language processing: A systematic review of current approaches

BACKGROUND

Electronic health records (EHRs) are a valuable resource for data-driven medical research. However, the presence of protected health information (PHI) makes EHRs unsuitable to be shared for research purposes. De-identification, i.e. the process of removing PHI is a critical step in making EHR data accessible. Natural language processing has repeatedly demonstrated its feasibility in automating the de-identification process.

OBJECTIVES

Our study aims to provide systematic evidence on how the de-identification of clinical free text written in English has evolved in the last thirteen years, and to report on the performances and limitations of the current state-of-the-art systems for the English language. In addition, we aim to identify challenges and potential research opportunities in this field.

METHODS

A systematic search in PubMed, Web of Science, and the DBLP was conducted for studies published between January 2010 and February 2023. Titles and abstracts were examined to identify the relevant studies. Selected studies were then analysed in-depth, and information was collected on de-identification methodologies, data sources, and measured performance.

RESULTS

A total of 2125 publications were identified for the title and abstract screening. 69 studies were found to be relevant. Machine learning (37 studies) and hybrid (26 studies) approaches are predominant, while six studies relied only on rules. The majority of the approaches were trained and evaluated on public corpora. The 2014 i2b2/UTHealth corpus is the most frequently used (36 studies), followed by the 2006 i2b2 (18 studies) and 2016 CEGS N-GRID (10 studies) corpora.

CONCLUSION

Earlier de-identification approaches aimed at English were mainly rule and machine learning hybrids with extensive feature engineering and post-processing, while more recent performance improvements are due to feature-inferring recurrent neural networks. Current leading performance is achieved using attention-based neural models. Recent studies report state-of-the-art F1-scores (over 98 %) when evaluated in the manner usually adopted by the clinical natural language processing community. However, their performance needs to be more thoroughly assessed with different measures to judge their reliability to safely de-identify data in a real-world setting. Without additional manually labeled training data, state-of-the-art systems fail to generalise well across a wide range of clinical sub-domains.

A practical guide to the development and deployment of deep learning models for the orthopaedic surgeon: Part III, focus on registry creation, diagnosis, and data privacy

Deep learning is a subset of artificial intelligence (AI) with enormous potential to transform orthopaedic surgery. As has already become evident with the deployment of Large Language Models (LLMs) like ChatGPT (OpenAI Inc.), deep learning can rapidly enter clinical and surgical practices. As such, it is imperative that orthopaedic surgeons acquire a deeper understanding of the technical terminology, capabilities and limitations associated with deep learning models. The focus of this series thus far has been providing surgeons with an overview of the steps needed to implement a deep learning-based pipeline, emphasizing some of the important technical details for surgeons to understand as they encounter, evaluate or lead deep learning projects. However, this series would be remiss without providing practical examples of how deep learning models have begun to be deployed and highlighting the areas where the authors feel deep learning may have the most profound potential. While computer vision applications of deep learning were the focus of Parts I and II, due to the enormous impact that natural language processing (NLP) has had in recent months, NLP-based deep learning models are also discussed in this final part of the series. In this review, three applications that the authors believe can be impacted the most by deep learning but with which many surgeons may not be familiar are discussed: (1) registry construction, (2) diagnostic AI and (3) data privacy. Deep learning-based registry construction will be essential for the development of more impactful clinical applications, with diagnostic AI being one of those applications likely to augment clinical decision-making in the near future. As the applications of deep learning continue to grow, the protection of patient information will become increasingly essential; as such, applications of deep learning to enhance data privacy are likely to become more important than ever before. Level of Evidence: Level IV.

Examining the Generalizability of Pretrained De-identification Transformer Models on Narrative Nursing Notes

BACKGROUND

 Narrative nursing notes are a valuable resource in informatics research with unique predictive signals about patient care. The open sharing of these data, however, is appropriately constrained by rigorous regulations set by the Health Insurance Portability and Accountability Act (HIPAA) for the protection of privacy. Several models have been developed and evaluated on the open-source i2b2 dataset. A focus on the generalizability of these models with respect to nursing notes remains understudied.

OBJECTIVES

 The study aims to understand the generalizability of pretrained transformer models and investigate the variability of personal protected health information (PHI) distribution patterns between discharge summaries and nursing notes with a goal to inform the future design for model evaluation schema.

METHODS

 Two pretrained transformer models (RoBERTa, ClinicalBERT) fine-tuned on i2b2 2014 discharge summaries were evaluated on our data inpatient nursing notes and compared with the baseline performance. Statistical testing was deployed to assess differences in PHI distribution across discharge summaries and nursing notes.

RESULTS

 RoBERTa achieved the optimal performance when tested on an external source of data, with an F1 score of 0.887 across PHI categories and 0.932 in the PHI binary task. Overall, discharge summaries contained a higher number of PHI instances and categories of PHI compared with inpatient nursing notes.

CONCLUSION

 The study investigated the applicability of two pretrained transformers on inpatient nursing notes and examined the distinctions between nursing notes and discharge summaries concerning the utilization of personal PHI. Discharge summaries presented a greater quantity of PHI instances and types when compared with narrative nursing notes, but narrative nursing notes exhibited more diversity in the types of PHI present, with some pertaining to patient's personal life. The insights obtained from the research help improve the design and selection of algorithms, as well as contribute to the development of suitable performance thresholds for PHI.

Generative Large Language Models for Detection of Speech Recognition Errors in Radiology Reports

This study evaluated the ability of generative large language models (LLMs) to detect speech recognition errors in radiology reports. A dataset of 3233 CT and MRI reports was assessed by radiologists for speech recognition errors. Errors were categorized as clinically significant or not clinically significant. Performances of five generative LLMs-GPT-3.5-turbo, GPT-4, text-davinci-003, Llama-v2-70B-chat, and Bard-were compared in detecting these errors, using manual error detection as the reference standard. Prompt engineering was used to optimize model performance. GPT-4 demonstrated high accuracy in detecting clinically significant errors (precision, 76.9%; recall, 100%; F1 score, 86.9%) and not clinically significant errors (precision, 93.9%; recall, 94.7%; F1 score, 94.3%). Text-davinci-003 achieved F1 scores of 72% and 46.6% for clinically significant and not clinically significant errors, respectively. GPT-3.5-turbo obtained 59.1% and 32.2% F1 scores, while Llama-v2-70B-chat scored 72.8% and 47.7%. Bard showed the lowest accuracy, with F1 scores of 47.5% and 20.9%. GPT-4 effectively identified challenging errors of nonsense phrases and internally inconsistent statements. Longer reports, resident dictation, and overnight shifts were associated with higher error rates. In conclusion, advanced generative LLMs show potential for automatic detection of speech recognition errors in radiology reports. Keywords: CT, Large Language Model, Machine Learning, MRI, Natural Language Processing, Radiology Reports, Speech, Unsupervised Learning Supplemental material is available for this article.

Enhancing phenotype recognition in clinical notes using large language models: PhenoBCBERT and PhenoGPT

To enhance phenotype recognition in clinical notes of genetic diseases, we developed two models-PhenoBCBERT and PhenoGPT-for expanding the vocabularies of Human Phenotype Ontology (HPO) terms. While HPO offers a standardized vocabulary for phenotypes, existing tools often fail to capture the full scope of phenotypes due to limitations from traditional heuristic or rule-based approaches. Our models leverage large language models to automate the detection of phenotype terms, including those not in the current HPO. We compare these models with PhenoTagger, another HPO recognition tool, and found that our models identify a wider range of phenotype concepts, including previously uncharacterized ones. Our models also show strong performance in case studies on biomedical literature. We evaluate the strengths and weaknesses of BERT- and GPT-based models in aspects such as architecture and accuracy. Overall, our models enhance automated phenotype detection from clinical texts, improving downstream analyses on human diseases.

Enhancing Phenotype Recognition in Clinical Notes Using Large Language Models: PhenoBCBERT and PhenoGPT

To enhance phenotype recognition in clinical notes of genetic diseases, we developed two models - PhenoBCBERT and PhenoGPT - for expanding the vocabularies of Human Phenotype Ontology (HPO) terms. While HPO offers a standardized vocabulary for phenotypes, existing tools often fail to capture the full scope of phenotypes, due to limitations from traditional heuristic or rule-based approaches. Our models leverage large language models (LLMs) to automate the detection of phenotype terms, including those not in the current HPO. We compared these models to PhenoTagger, another HPO recognition tool, and found that our models identify a wider range of phenotype concepts, including previously uncharacterized ones. Our models also showed strong performance in case studies on biomedical literature. We evaluated the strengths and weaknesses of BERT-based and GPT-based models in aspects such as architecture and accuracy. Overall, our models enhance automated phenotype detection from clinical texts, improving downstream analyses on human diseases.

Leveraging Large Language Models for Generating Responses to Patient Messages

Objective

This study aimed to develop and assess the performance of fine-tuned large language models for generating responses to patient messages sent via an electronic health record patient portal.

Methods

Utilizing a dataset of messages and responses extracted from the patient portal at a large academic medical center, we developed a model (CLAIR-Short) based on a pre-trained large language model (LLaMA-65B). In addition, we used the OpenAI API to update physician responses from an open-source dataset into a format with informative paragraphs that offered patient education while emphasizing empathy and professionalism. By combining with this dataset, we further fine-tuned our model (CLAIR-Long). To evaluate the fine-tuned models, we used ten representative patient portal questions in primary care to generate responses. We asked primary care physicians to review generated responses from our models and ChatGPT and rated them for empathy, responsiveness, accuracy, and usefulness.

Results

The dataset consisted of a total of 499,794 pairs of patient messages and corresponding responses from the patient portal, with 5,000 patient messages and ChatGPT-updated responses from an online platform. Four primary care physicians participated in the survey. CLAIR-Short exhibited the ability to generate concise responses similar to provider's responses. CLAIR-Long responses provided increased patient educational content compared to CLAIR-Short and were rated similarly to ChatGPT's responses, receiving positive evaluations for responsiveness, empathy, and accuracy, while receiving a neutral rating for usefulness.

Conclusion

Leveraging large language models to generate responses to patient messages demonstrates significant potential in facilitating communication between patients and primary care providers.