clinical concept extraction

To address the tendency of large language models (LLMs) to overlook critical medical context and clinical decision cues in clinical text summarization, this paper proposes a context-preserving token filtering mechanism integrated with domain-specific semantic enhancement via a lightweight medical knowledge graph. The method comprises: (1) context-aware token filtering that prioritizes clinically salient entities; (2) construction and dynamic retrieval of a guideline-grounded, lightweight medical knowledge graph; and (3) LLM fine-tuning and structured prompt engineering explicitly optimized for factual consistency. Evaluated on two public clinical summarization benchmarks, our approach achieves new state-of-the-art performance—improving clinical fact accuracy by 12.7% and ROUGE-L by 4.3 points—while demonstrating strong scalability. This work establishes a novel paradigm for high-precision, interpretable clinical decision support systems.

Current clinical knowledge graph construction heavily relies on manual curation and rule-based approaches, struggling to handle the semantic complexity and contextual ambiguity inherent in clinical guidelines and biomedical literature—resulting in low automation and insufficient clinical reliability of structured, interoperable medical indicator knowledge graphs. To address this, we propose a guideline-driven, ontology-guided, retrieval-augmented generation (RAG) and large language model (LLM)-integrated framework for automated knowledge graph construction. Our method synergistically combines domain ontology modeling, dynamic multi-source guideline retrieval, structured schema generation, and expert-in-the-loop validation. It significantly improves the accuracy, scalability, and clinical consistency of knowledge extraction. We experimentally constructed a high-quality knowledge graph covering 200+ core medical indicators and empirically validated its effectiveness across downstream tasks—including intelligent diagnosis and treatment, clinical decision support, and medical question answering.

This study addresses the significant performance degradation of clinical decision extraction in narrative-rich recommendation and prevention texts, revealing a systematic blind spot. Leveraging the DICTUM classification framework and MedDec discharge summaries, the work uncovers, for the first time, associations between clinical decision categories and specific linguistic features—such as stopword density, negation, and hedging cues—and proposes a Transformer-based extraction model evaluated under both exact and overlap matching strategies. Experimental results demonstrate that traditional exact-match recall is only 48%, rising to 71% under overlap matching. Notably, segments containing high stopword ratios or negation/hedging markers exhibit recall as low as 24%, highlighting how conventional evaluation metrics systematically undervalue semantically correct but boundary-ambiguous predictions.

Electronic health record (EHR) texts are excessively long and noisy, frequently exceeding the context windows of mainstream large language models (LLMs), thereby impeding clinical reasoning. To address this, we systematically evaluate retrieval-augmented generation (RAG) against full-context input across clinical tasks. We introduce three reproducible, multi-institutional clinical tasks—key information extraction, temporal event modeling, and core diagnosis identification—and conduct experiments using three state-of-the-art LLMs, integrating targeted text retrieval with recent clinical notes as input. Results show that RAG achieves comparable or superior performance to full-context baselines across most tasks—improving average F1 by 2.3–5.1 percentage points—while consuming only ~15% of the input tokens. This yields substantial reductions in computational cost and enhances practical deployability. To our knowledge, this is the first study to empirically validate, within a unified framework, RAG’s efficiency and robustness for long-context EHR reasoning, establishing a lightweight, scalable paradigm for clinical LLM deployment.

Extracting argumentative entities and identifying support/refute relations in clinical texts—critical for diagnostic reasoning—remains challenging under data-scarce conditions. Method: We propose an end-to-end argument structure parsing framework that jointly leverages token-level sequence labeling and a pre-trained natural language inference (NLI) model (e.g., RoBERTa-MNLI). This is the first work to adapt the NLI paradigm to medical argument mining, moving beyond conventional classification-based approaches. Contribution/Results: Our method significantly improves few-shot relation classification performance by exploiting entailment semantics between argument components. Evaluated on a real-world clinical note dataset, it achieves an F1 score of 78.3%, outperforming strong baselines by 12.6 percentage points. The approach enhances the interpretability and evidential grounding of automated diagnostic conclusions, establishing a novel paradigm for low-resource medical NLP tasks.

This study addresses the challenges of accurately querying structured data and extracting information from unstructured clinical text in electronic health records (EHRs). To this end, the authors propose a unified framework that integrates large language models (LLMs) with retrieval-augmented generation (RAG): LLMs are employed to execute structured queries (e.g., Pandas operations), while RAG enhances information extraction from unstructured clinical narratives. The work introduces an innovative automatic evaluation pipeline based on synthetically generated question-answer pairs, combining exact match metrics, semantic similarity scores, and human assessments. Evaluated on a subset of MIMIC-III, the approach demonstrates improved semantic accuracy and task adaptability, offering clinical data science a flexible and reliable tool for automated reasoning and evaluation.

This study addresses the trade-off among accuracy, standardization, and computational efficiency in NLP for low-resource, non-English medical settings, using unstructured electronic health records from a Polish pediatric rehabilitation hospital. We propose a hybrid approach integrating rule-based systems—offering high precision and low computational cost—with multilingual large language models (LLMs)—providing strong generalization and adaptability. We systematically compare performance on demographic, clinical finding, and medication information extraction tasks using both original Polish text and machine-translated English text. Results show rule-based methods outperform LLMs in age and gender identification, while LLMs significantly improve drug name recognition accuracy. Critically, machine translation introduces non-negligible information loss, degrading downstream performance. This work establishes a new paradigm for resource-constrained, multilingual clinical NLP that balances accuracy, robustness, and practical deployability.

This work addresses the challenge of extracting structured symptom information from unstructured clinical notes, a task hindered by the limitations of existing approaches—either rule-based systems lacking contextual awareness that yield high false-positive rates or learning-based methods requiring extensive annotated data for fine-tuning. To overcome these issues, the authors propose Pythia, a multi-agent system that autonomously generates and refines symptom extraction prompts on locally deployed open-source large language models, eliminating the need for manual prompt engineering or model fine-tuning. The optimal prompt is selected based on sensitivity and specificity evaluated on a development set. This approach achieves state-of-the-art performance without any fine-tuning, attaining average sensitivity of 0.76 and specificity of 0.95 across 72 symptoms—substantially outperforming a fine-tuned BERT model (sensitivity: 0.23)—while maintaining high specificity and strong generalization even for low-prevalence symptoms.

This study addresses the limitations of existing post-hoc explanation methods—such as LIME and SHAP—in clinical natural language processing, particularly their inability to generate semantically coherent and clinically meaningful interpretations when applied to long, unstructured medical narratives. Focusing on the task of predicting hospital length of stay, the work presents the first systematic evaluation of these methods’ robustness and reliability on real-world clinical texts through adversarial input perturbations and attribution stability analysis. The findings reveal that current approaches frequently overemphasize irrelevant terms, produce unstable attributions, and yield high-confidence predictions even for nonsensical inputs. These shortcomings expose critical flaws for clinical deployment and underscore the necessity for explanations that simultaneously uphold clinical relevance, semantic consistency, and resilience to linguistic noise.

In clinical research, manual extraction of structured clinical features from unstructured electronic health records (EHRs) is time-consuming, inefficient, and error-prone. To address this, we propose a privacy-preserving, modular, on-premises large language model (LLM) framework that integrates retrieval-augmented generation (RAG) with structured-output prompt engineering, enabling secure, scalable, containerized deployment in HIPAA-compliant environments. Our approach uniquely synergizes RAG with deterministic structured-response mechanisms for clinical text parsing—balancing domain adaptability and strict data privacy. Evaluated across multiple medical feature extraction tasks, the framework achieves high accuracy, substantially reducing manual annotation effort and improving data consistency. Notably, its systematic evaluation uncovered previously undetected systematic errors in prior human annotations, thereby validating both its reliability and its capacity for quality assurance and error discovery.