clinical interpretation

Existing biomedical benchmarks largely neglect the most challenging PICO element—Outcome (clinical outcomes)—and lack high-quality, fine-grained annotated datasets. Method: We construct the first large-scale clinical outcome annotation dataset, covering 500 PubMed abstracts and an EBM-NLP subset; explicitly define and annotate fine-grained types of “clinically meaningful outcomes” through collaborative development of annotation guidelines by clinicians and NLP experts, achieving high inter-annotator agreement (Cohen’s κ = 0.76); propose an iteratively refined clinical NLP annotation protocol, fine-tune PubMedBERT, and evaluate at both entity-level and token-level granularity. Contribution/Results: Our model achieves entity-level F1 = 0.69 and token-level F1 = 0.76 on the EBM-NLP subset. The dataset is publicly released, establishing a new benchmark for automated clinical outcome extraction in evidence-based medicine.

This work addresses the critical limitation of current clinical large language models in dynamically adapting treatment decisions when patient contexts change—a capability inadequately assessed by conventional medical question-answering benchmarks. To bridge this gap, the authors propose ClinPivot, the first auditable evaluation benchmark specifically designed to measure dynamic adaptability in therapeutic decision-making. ClinPivot leverages a biomedical knowledge graph to construct interpretable context-perturbation pairs and incorporates structured decision supervision alongside a lightweight replay mechanism to enhance contextual sensitivity under constrained knowledge budgets. Experimental results reveal that state-of-the-art models, including Qwen variants, perform poorly on ClinPivot, whereas the proposed approach significantly improves dynamic decision-making without compromising general assistant capabilities, thereby exposing a notable disconnect between standard medical QA accuracy and genuine clinical reasoning proficiency.

This study addresses the limitations of manual clinical case conceptualization, which is time-consuming and subject to inter-clinician variability. The authors propose a novel approach that leverages large language models (LLMs) to automatically generate structured causal diagrams from patient–therapist dialogues, conforming to the standardized 5P framework (Presenting problem, Predisposing, Precipitating, Perpetuating, and Protective factors). By integrating LLM-based text understanding with causal graph construction, structural similarity assessment via NetSimile, semantic embedding alignment, and expert evaluation, the method bridges the gap between unstructured narrative discourse and formal clinical reasoning. Experimental results demonstrate that the generated causal graphs achieve structural similarity comparable to inter-expert agreement, exhibit strong semantic alignment, and receive above-moderate ratings from clinicians, particularly in terms of completeness, coherence, and clinical utility.

This study addresses the limited interpretability of net benefit in clinical decision curve analysis, which often stems from the absence of an intuitive comparison against “treat all” or “treat none” strategies. The authors innovatively reframe decision curves through the lens of positive predictive value (PPV) and calibration, explicitly linking net benefit to threshold-specific observed risk and PPV. For the first time, this approach establishes a direct connection between decision performance and both PPV and calibration within clinically relevant subgroups. The proposed PPV curve serves as a complementary tool to traditional decision curves, substantially enhancing the clinical interpretability of a model’s net benefit and enabling clinicians to more clearly discern at which decision thresholds model-guided interventions yield tangible clinical utility.

To address the challenge that clinical users—lacking expertise in time-series modeling—struggle to perform efficient predictive analytics, this paper proposes and implements a low-barrier time-series forecasting platform tailored for healthcare applications. The platform integrates multiple configurable forecasting models (e.g., ARIMA, Prophet, LSTM) with automated training pipelines and innovatively incorporates large language models (LLMs) to provide parameter recommendations, result interpretation, and interactive modeling guidance. It offers end-to-end support—including data upload, visualization, multi-model comparison, clinical interpretation, and semantic translation of outputs—thereby substantially lowering technical barriers. Empirical evaluation demonstrates that the platform significantly enhances clinicians’ and researchers’ understanding of and proficiency in predictive modeling. As a scalable, interpretable, and easily deployable infrastructure, it advances intelligent analytics capabilities for learning health systems.

This study addresses the clinical deployment challenges of black-box models in medical AI, which suffer from poor interpretability and reproducibility. To overcome these limitations, the authors propose a fully transparent and manually reproducible classification framework based on Bernoulli Naïve Bayes. The approach employs chi-squared test–guided supervised binarization to convert continuous clinical variables into threshold-based rules with clear clinical meaning. Probability calibration is performed using leakage-proof cross-validation, incorporating Brier score, calibration intercept and slope, and Beta posterior calibration. Evaluated on three medical datasets, the model achieves AUCs ranging from 0.800 to 0.984 while substantially improving calibration quality, enabling high-performance predictions that can be executed without software dependencies through manual inference.

Bridging the gap between general biomedical knowledge and actionable, testable hypotheses for specific experimental or clinical contexts remains a critical challenge. This work proposes SCENE, a novel framework that formalizes knowledge contextualization as an iterative search process through a dual-layer multi-agent architecture to deeply integrate knowledge-driven and data-driven reasoning. The upper-layer agent generates search directions and anchors relevant data patterns, while the lower-layer agent leverages knowledge graph guidance and multi-objective optimization to produce verifiable propositions that balance evidential strength with empirical support. Evaluated in real-world settings, SCENE successfully identified patient subgroups with heterogeneous treatment effects in clinical trials and discovered perturbation contexts with high target-response alignment in the LINCS L1000 study, significantly outperforming existing baselines. The generated hypotheses exhibit strong traceability, reproducibility, and expert verifiability.

While current large language models demonstrate accuracy in clinical diagnosis, it remains unclear whether their reasoning follows stable, structured clinical logic. This work proposes the Clinical Reasoning Graph framework—a structured graph representation grounded in a clinical ontology comprising five node types and seven edge types—and leverages natural language processing and graph similarity metrics to extract and analyze 750 diagnostic trajectories. The study reveals that graph similarity between correct and incorrect diagnoses is nearly identical (0.488 vs. 0.484), and reasoning structures show no significant consistency across similar cases, indicating a lack of schematic-level stability in cross-case reasoning. Although structured reflection prompts improve feature analysis, they do not enhance structural consistency. These findings underscore the need for process-level evaluation to complement conventional outcome-based accuracy and offer a novel paradigm for explainability in clinical AI.

Clinical practice guidelines, as lengthy multimodal documents with intricate branching structures, pose significant challenges for translation into structured, auditable, and executable clinical decision support systems. This work proposes a decomposition-first pipeline that leverages topological-aware chunking, interface-constrained block graph generation, semantic deduplication, and global provenance aggregation, all built upon a unified vision-language model to ensure cross-page control-flow continuity and logical coherence. Moving beyond single-pass generation paradigms, the approach achieves substantial performance gains on a prostate guideline benchmark: edge and triple precision/recall improve to 69.0%/87.5% (from 19.6%/16.1%), and node recall rises from 78.1% to 93.8%.

Clinical trial protocols are predominantly presented as unstructured narrative text, which hinders the explicit representation of dynamic eligibility criteria and temporal constraints, thereby limiting automated reasoning and simulation. This work proposes CT-TEL, a novel workflow that leverages large language models (LLMs) to automatically translate narrative protocols into Temporal Event Logic (TEL) formulas—enabling computable, formal modeling of trial specifications for the first time. By incorporating bidirectional translation between natural language and logical formulas alongside semantic similarity evaluation, the method successfully generates high-fidelity TEL models across 23 real-world trials. The results demonstrate the feasibility of scalable, symbolic modeling of clinical trials and open new avenues for simulation under the emerging “symbolic biomedicine” paradigm.