An official website of the United States government
Abstract BackgroundDiagnostic pathology depends on complex, structured reasoning to interpret clinical, histologic, and molecular data. Replicating this cognitive process algorithmically remains a significant challenge. As large language models (LLMs) gain traction in medicine, it is critical to determine whether they have clinical utility by providing reasoning in highly specialized domains such as pathology. MethodsWe evaluated the performance of four reasoning LLMs (OpenAI o1, OpenAI o3-mini, Gemini 2.0 Flash Thinking Experimental, and DeepSeek-R1 671B) on 15 board-style open-ended pathology questions. Responses were independently reviewed by 11 pathologists using a structured framework that assessed language quality (accuracy, relevance, coherence, depth, and conciseness) and seven diagnostic reasoning strategies. Scores were normalized and aggregated for analysis. We also evaluated inter-observer agreement to assess scoring consistency. Model comparisons were conducted using one-way ANOVA and Tukey’s Honestly Significant Difference (HSD) test. ResultsGemini and DeepSeek significantly outperformed OpenAI o1 and OpenAI o3-mini in overall reasoning quality (p < 0.05), particularly in analytical depth and coherence. While all models achieved comparable accuracy, only Gemini and DeepSeek consistently applied expert-like reasoning strategies, including algorithmic, inductive, and Bayesian approaches. Performance varied by reasoning type: models performed best in algorithmic and deductive reasoning and poorest in heuristic and pattern recognition. Inter-observer agreement was highest for Gemini (p < 0.05), indicating greater consistency and interpretability. Models with more in-depth reasoning (Gemini and DeepSeek) were generally less concise. ConclusionAdvanced LLMs such as Gemini and DeepSeek can approximate aspects of expert-level diagnostic reasoning in pathology, particularly in algorithmic and structured approaches. However, limitations persist in contextual reasoning, heuristic decision-making, and consistency across questions. Addressing these gaps, along with trade-offs between depth and conciseness, will be essential for the safe and effective integration of AI tools into clinical pathology workflows.
more »
« less
This content will become publicly available on April 25, 2026
Employing Consensus-Based Reasoning with Locally Deployed LLMs for Enabling Structured Data Extraction from Surgical Pathology Reports
Surgical pathology reports contain essential diagnostic information, in free-text form, required for cancer staging, treatment planning, and cancer registry documentation. However, their unstructured nature and variability across tumor types and institutions pose challenges for automated data extraction. We present a consensus-driven, reasoning-based framework that uses multiple locally deployed large language models (LLMs) to extract six key diagnostic variables: site, laterality, histology, stage, grade, and behavior. Each LLM produces structured outputs with accompanying justifications, which are evaluated for accuracy and coherence by a separate reasoning model. Final consensus values are determined through aggregation, and expert validation is conducted by board-certified or equivalent pathologists. The framework was applied to over 4,000 pathology reports from The Cancer Genome Atlas (TCGA) and Moffitt Cancer Center. Expert review confirmed high agreement in the TCGA dataset for behavior (100.0%), histology (98.5%), site (95.2%), and grade (95.6%), with lower performance for stage (87.6%) and laterality (84.8%). In the pathology reports from Moffitt (brain, breast, and lung), accuracy remained high across variables, with histology (95.6%), behavior (98.3%), and stage (92.4%), achieving strong agreement. However, certain challenges emerged, such as inconsistent mention of sentinel lymph node details or anatomical ambiguity in biopsy site interpretations. Statistical analyses revealed significant main effects of model type, variable, and organ system, as well as model × variable × organ interactions, emphasizing the role of clinical context in model performance. These results highlight the importance of stratified, multi-organ evaluation frameworks in LLM benchmarking for clinical applications. Textual justifications enhanced interpretability and enabled human reviewers to audit model outputs. Overall, this consensus-based approach demonstrates that locally deployed LLMs can provide a transparent, accurate, and auditable solution for integrating AI-driven data extraction into real-world pathology workflows, including cancer registry abstraction and synoptic reporting.
more »
« less
- Award ID(s):
- 2234468
- PAR ID:
- 10599122
- Publisher / Repository:
- medRxiv
- Date Published:
- Format(s):
- Medium: X
- Institution:
- medRxiv
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract Objective We develop natural language processing (NLP) methods capable of accurately classifying tumor attributes from pathology reports given minimal labeled examples. Our hierarchical cancer to cancer transfer (HCTC) and zero-shot string similarity (ZSS) methods are designed to exploit shared information between cancers and auxiliary class features, respectively, to boost performance using enriched annotations which give both location-based information and document level labels for each pathology report. Materials and Methods Our data consists of 250 pathology reports each for kidney, colon, and lung cancer from 2002 to 2019 from a single institution (UCSF). For each report, we classified 5 attributes: procedure, tumor location, histology, grade, and presence of lymphovascular invasion. We develop novel NLP techniques involving transfer learning and string similarity trained on enriched annotations. We compare HCTC and ZSS methods to the state-of-the-art including conventional machine learning methods as well as deep learning methods. Results For our HCTC method, we see an improvement of up to 0.1 micro-F1 score and 0.04 macro-F1 averaged across cancer and applicable attributes. For our ZSS method, we see an improvement of up to 0.26 micro-F1 and 0.23 macro-F1 averaged across cancer and applicable attributes. These comparisons are made after adjusting training data sizes to correct for the 20% increase in annotation time for enriched annotations compared to ordinary annotations. Conclusions Methods based on transfer learning across cancers and augmenting information methods with string similarity priors can significantly reduce the amount of labeled data needed for accurate information extraction from pathology reports.more » « less
-
null (Ed.)Abstract Objective Cancer is a leading cause of death, but much of the diagnostic information is stored as unstructured data in pathology reports. We aim to improve uncertainty estimates of machine learning-based pathology parsers and evaluate performance in low data settings. Materials and methods Our data comes from the Urologic Outcomes Database at UCSF which includes 3232 annotated prostate cancer pathology reports from 2001 to 2018. We approach 17 separate information extraction tasks, involving a wide range of pathologic features. To handle the diverse range of fields, we required 2 statistical models, a document classification method for pathologic features with a small set of possible values and a token extraction method for pathologic features with a large set of values. For each model, we used isotonic calibration to improve the model’s estimates of its likelihood of being correct. Results Our best document classifier method, a convolutional neural network, achieves a weighted F1 score of 0.97 averaged over 12 fields and our best extraction method achieves an accuracy of 0.93 averaged over 5 fields. The performance saturates as a function of dataset size with as few as 128 data points. Furthermore, while our document classifier methods have reliable uncertainty estimates, our extraction-based methods do not, but after isotonic calibration, expected calibration error drops to below 0.03 for all extraction fields. Conclusions We find that when applying machine learning to pathology parsing, large datasets may not always be needed, and that calibration methods can improve the reliability of uncertainty estimates.more » « less
-
Large Language Models (LLMs) have achieved remarkable success in natural language tasks, yet understanding their reasoning processes re- mains a significant challenge. We address this by introducing XplainLLM, a dataset accom- panying an explanation framework designed to enhance LLM transparency and reliability. Our dataset comprises 24,204 instances where each instance interprets the LLM’s reasoning behavior using knowledge graphs (KGs) and graph attention networks (GAT), and includes explanations of LLMs such as the decoder- only Llama-3 and the encoder-only RoBERTa. XplainLLM also features a framework for gener- ating grounded explanations and the debugger- scores for multidimensional quality analysis. Our explanations include why-choose and why- not-choose components, reason-elements, and debugger-scores that collectively illuminate the LLM’s reasoning behavior. Our evaluations demonstrate XplainLLM’s potential to reduce hallucinations and improve grounded explana- tion generation in LLMs. XplainLLM is a re- source for researchers and practitioners to build trust and verify the reliability of LLM outputs. Our code and dataset are publicly available.more » « less
-
High-quality knowledge graphs (KGs) play a crucial role in many applications. However, KGs created by automated information extraction systems can suffer from erroneous extractions or be inconsistent with provenance/source text. It is important to identify and correct such problems. In this paper, we study leveraging the emergent reasoning capabilities of large language models (LLMs) to detect inconsistencies between extracted facts and their provenance. With a focus on ``open'' LLMs that can be run and trained locally, we find that few-shot approaches can yield an absolute performance gain of 2.5-3.4% over the state-of-the-art method with only 9% of training data. We examine the LLM architectures' effect and show that Decoder-Only models underperform Encoder-Decoder approaches. We also explore how model size impacts performance and counterintuitively find that larger models do not result in consistent performance gains. Our detailed analyses suggest that while LLMs can improve KG consistency, the different LLM models learn different aspects of KG consistency and are sensitive to the number of entities involved.more » « less
