Consensus-Based Reasoning with Locally Deployed LLMs for Structured Data Extraction from Surgical Pathology Reports DOI Creative Commons
Aakash Tripathi, Asim Waqas, Ehsan Ullah

и другие.

medRxiv (Cold Spring Harbor Laboratory), Год журнала: 2025, Номер unknown

Опубликована: Апрель 25, 2025

Surgical pathology reports contain essential diagnostic information, in free-text form, required for cancer staging, treatment planning, and registry documentation. However, their unstructured nature variability across tumor types institutions pose challenges automated data extraction. We present a consensus-driven, reasoning-based framework that uses multiple locally deployed large language models (LLMs) to extract six key variables: site, laterality, histology, stage, grade, behavior. Each LLM produces structured outputs with accompanying justifications, which are evaluated accuracy coherence by separate reasoning model. Final consensus values determined through aggregation, expert validation is conducted board-certified or equivalent pathologists. The was applied over 4,000 from Cancer Genome Atlas (TCGA) Moffitt Center. Expert review confirmed high agreement the TCGA dataset behavior (100.0%), histology (98.5%), site (95.2%), grade (95.6%), lower performance stage (87.6%) laterality (84.8%). In (brain, breast, lung), remained variables, (98.3%), (92.4%), achieving strong agreement. certain emerged, such as inconsistent mention of sentinel lymph node details anatomical ambiguity biopsy interpretations. Statistical analyses revealed significant main effects model type, variable, organ system, well × variable interactions, emphasizing role clinical context performance. These results highlight importance stratified, multi-organ evaluation frameworks benchmarking applications. Textual justifications enhanced interpretability enabled human reviewers audit outputs. Overall, this consensus-based approach demonstrates LLMs can provide transparent, accurate, auditable solution integrating AI-driven extraction into real-world workflows, including abstraction synoptic reporting.

Язык: Английский

Anonymizing medical documents with local, privacy preserving large language models: The LLM-Anonymizer DOI Creative Commons
Isabella C. Wiest, Marie-Elisabeth Leßmann, F M Wolf

и другие.

medRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Июнь 13, 2024

Abstract Background Medical research with real-world clinical data can be challenging due to privacy requirements. Ideally, patient are handled in a fully pseudonymised or anonymised way. However, this make it difficult for medical researchers access and analyze large datasets exchange between hospitals. De-identifying free text is particularly the diverse documentation styles unstructured nature of data. recent advancements natural language processing (NLP), driven by development models (LLMs), have revolutionized ability extract information from text. Methods We hypothesize that LLMs highly effective tools extracting patient-related information, which subsequently used de-identify reports. To test hypothesis, we conduct benchmark study using eight locally deployable (Llama-3 8B, Llama-3 70B, Llama-2 7B, 7B “Sauerkraut”, 70B Mistral Phi-3-mini) dataset 100 letters. then remove identified our newly developed LLM-Anonymizer pipeline. Results Our results demonstrate LLM-Anonymizer, when achieved success rate 98.05% removing characters carrying personal identifying information. When evaluating performance relation number manually as containing identifiable characteristics, system missed only 1.95% erroneously redacted 0.85% characters. Conclusion provide full LLM-based Anonymizer pipeline under an open source license user-friendly web interface operates on local hardware requires no programming skills. This powerful tool has potential significantly facilitate enabling secure efficient de-identification premise, thereby addressing key challenges sharing.

Язык: Английский

Процитировано

6
Generalizable and automated classification of TNM stage from pathology reports with external validation DOI Creative Commons

Jenna Kefeli,

Jacob Berkowitz, Jose Miguel Acitores Cortina

и другие.

Nature Communications, Год журнала: 2024, Номер 15(1)

Опубликована: Окт. 16, 2024

Cancer staging is an essential clinical attribute informing patient prognosis and trial eligibility. However, it not routinely recorded in structured electronic health records. Here, we present BB-TEN: Big Bird - TNM Extracted from Notes, a generalizable method for the automated classification of stage directly pathology report text. We train BERT-based model using publicly available reports across approximately 7000 patients 23 cancer types. explore use different types, with differing input sizes, parameters, architectures. Our final goes beyond term-extraction, inferring context when included text explicitly. As external validation, test our on almost 8000 Columbia University Medical Center, finding that trained achieved AU-ROC 0.815-0.942. This suggests can be applied broadly to other institutions without additional institution-specific fine-tuning.

Язык: Английский

Процитировано

6
Use of Artificial Intelligence for Liver Diseases: A Survey from the EASL Congress 2024 DOI Creative Commons
Laura Žigutytė, Thomas Sorz, Jan Clusmann

и другие.

JHEP Reports, Год журнала: 2024, Номер 6(12), С. 101209 - 101209

Опубликована: Сен. 6, 2024

Язык: Английский

Процитировано

5
Fine-tuning language model embeddings to reveal domain knowledge: An explainable artificial intelligence perspective on medical decision making DOI Creative Commons
Ceca Kraišniković, Robert Harb, Markus Plass

и другие.

Engineering Applications of Artificial Intelligence, Год журнала: 2024, Номер 139, С. 109561 - 109561

Опубликована: Ноя. 15, 2024

Язык: Английский

Процитировано

4
An Automated Information Extraction Model For Unstructured Discharge Letters Using Large Language Models and GPT-4 DOI Creative Commons
Robert Siepmann, Giulia Baldini, Cynthia S. Schmidt

и другие.

Healthcare Analytics, Год журнала: 2025, Номер unknown, С. 100378 - 100378

Опубликована: Янв. 1, 2025

Язык: Английский

Процитировано

0
Natural Language Processing in Gastroenterology DOI
Sravanthi Parasa, Arun Raghav Mahankali Sridhar

Gastrointestinal Endoscopy Clinics of North America, Год журнала: 2025, Номер 35(2), С. 309 - 317

Опубликована: Янв. 18, 2025

Язык: Английский

Процитировано

0
Using Generative AI to Extract Structured Information from Free Text Pathology Reports DOI Creative Commons
Farah Shahid, Min-Huei Hsu, Yung‐Chun Chang

и другие.

Journal of Medical Systems, Год журнала: 2025, Номер 49(1)

Опубликована: Март 13, 2025

Abstract Manually converting unstructured text pathology reports into structured is very time-consuming and prone to errors. This study demonstrates the transformative potential of generative AI in automating analysis free-text reports. Employing ChatGPT Large Language Model within a Streamlit web application, we automated extraction structuring information from 33 breast cancer Taipei Medical University Hospital. Achieving 99.61% accuracy rate, system notably reduced processing time compared traditional methods. not only underscores efficacy medical data but also highlights its enhance efficiency reliability analysis. However, this limited was conducted using obtained hospitals associated with single institution. In future, plan expand scope research include for other types incrementally conduct external validation further substantiate robustness generalizability proposed system. Through technological integration, aimed capabilities improving both speed processing. The outcomes affirm that can significantly transform handling reports, promising substantial advancements biomedical by facilitating complex data.

Язык: Английский

Процитировано

0
Synoptic reporting by summarizing cancer pathology reports using large language models DOI Creative Commons
Sivaraman Rajaganapathy,

Shaika Chowdhury,

Xiaodi Li

и другие.

Опубликована: Апрель 1, 2025

Abstract Synoptic reporting, the documenting of clinical information in a structured manner, enhances patient care by improving accuracy, readability, and report completeness, but imposes significant administrative burdens on physicians. The potential Large Language Models (LLMs) for automating synoptic reporting remains underexplored. In this study, we explore state-of-the-art LLMs automatic using 7774 pathology reports from 8 cancer types, paired with physician annotated Mayo Clinic EHR. We developed comprehensive automation framework, combining LLMs, incorporating parameter-efficient optimization, scalable prompt templates, robust evaluation strategies. validate our results both internal external data, ensuring alignment pathologist responses. Using fine-tuned LLAMA-2 achieved BERT F1 scores above 0.86 across all data elements exceeding 0.94 over 50% (11 22) elements, translating to manually assessed mean semantic accuracies 77% up 81% short reports.

Язык: Английский

Процитировано

0
A Systematic Review of Large Language Models in Medical Specialties: Applications, Challenges and Future Directions DOI
Asma Musabah Alkalbani, Ahmed Salim Alrawahi, Ahmad Salah

и другие.

Research Square (Research Square), Год журнала: 2025, Номер unknown

Опубликована: Апрель 16, 2025

Abstract Background: Large Language Models (LLMs) are one of the artificial intelligence (AI) technologies used to understand and generate text, summarize information, comprehend contextual cues. LLMs have been increasingly by researchers in various medical applications, but their effectiveness limitations still uncertain, especially across specialties. Objective: This review evaluates recent literature on how utilized research studies 19 It also explores challenges involved suggests areas for future focus. Methods: Two performed searches PubMed, Web Science Scopus identify published from January 2021 March 2024. The included usage LLM performing tasks. Data was extracted analyzed five reviewers. To assess risk bias, quality assessment using revised tool intelligence-centered diagnostic accuracy (QUADAS-AI). Results: Results were synthesized through categorical analysis evaluation metrics, impact types, validation approaches A total 84 this mainly originated two countries; USA (35/84) China (16/84). Although reviewed applications spread specialties, multi-specialty demonstrated 22 studies. Various aims include clinical natural language processing (31/84), supporting decision (20/84), education (15/84), diagnoses patient management engagement (3/84). GPT-based BERT-based most (83/84) Despite reported positive impacts such as improved efficiency accuracy, related reliability, ethics remain. overall bias low 72 studies, high 11 not clear 3 Conclusion: dominate specialty with over 98.8% these models. potential benefits process diagnostics, a key finding regarding substantial variability performance among LLMs. For instance, LLMs' ranged 3% support 90% some NLP Heterogeneity utilization diverse tasks contexts prevented meaningful meta-analysis, lacked standardized methodologies, outcome measures, implementation approaches. Therefore, room improvement remains wide developing domain-specific data establishing standards ensure reliability effectiveness.

Язык: Английский

Процитировано

0
Deidentifying Medical Documents with Local, Privacy-Preserving Large Language Models: The LLM-Anonymizer DOI Open Access
Isabella C. Wiest, Marie-Elisabeth Leßmann, F M Wolf

и другие.

NEJM AI, Год журнала: 2025, Номер 2(4)

Опубликована: Март 27, 2025

Язык: Английский

Процитировано

0