-hop retrieval on large KGs by building on symbolic KG formulations and executing traversal as hardware-efficient operations over decomposed subject, object, and relation representations. To scale to billion-edge graphs, LogosKG integrates degree-aware partitioning, cross-graph routing, and on-demand caching. Experiments show substantial efficiency gains over CPU and GPU baselines without loss of retrieval fidelity. With proven performance in KG retrieval, a downstream two-round KG-LLM interaction demonstrates how LogosKG enables large-scale, evidence-grounded analysis of how KG topology, such as hop distribution and connectivity, shapes the alignment between structured biomedical knowledge and LLM diagnostic reasoning, thereby opening the door for next-generation KG-LLM integration. The source code is publicly available at https://github.com/LARK-NLP-Lab/LogosKG, and an online demo is available at https://lark-nlp-lab-logoskg.hf.space/.

Large language models (LLMs) are increasingly proposed for detecting and responding to violent content online, yet their ability to reason about morally ambiguous, real-world scenarios remains underexamined. We present the first study to evaluate LLMs using a validated social science instrument designed to measure human response to everyday conflict, namely the Violent Behavior Vignette Questionnaire (VBVQ). To assess potential bias, we introduce persona-based prompting that varies race, age, and geographic identity within the United States. Six LLMs developed across different geopolitical and organizational contexts are evaluated under a unified zero-shot setting. Our study reveals two key findings: (1) LLMs surface-level text generation often diverges from their internal preference for violent responses; (2) their violent tendencies vary across demographics, frequently contradicting established findings in criminology, social science, and psychology.

Large language models (LLMs) often behave inconsistently across inputs, indicating uncertainty and motivating the need for its quantification in high-stakes settings. Prior work on calibration and uncertainty quantification often focuses on individual models, overlooking the potential of model diversity. We hypothesize that LLMs make complementary predictions due to differences in training and the Zipfian nature of language, and that aggregating their outputs leads to more reliable uncertainty estimates. To leverage this, we propose MUSE (Multi-LLM Uncertainty via Subset Ensembles), a simple information-theoretic method that uses Jensen-Shannon Divergence to identify and aggregate well-calibrated subsets of LLMs. Experiments on binary prediction tasks demonstrate improved calibration and predictive performance compared to single-model and naïve ensemble baselines. In addition, we explore using MUSE as guided signals with chain-of-thought distillation to fine-tune LLMs for calibration. MUSE is available at:https://github.com/LARK-NLP-Lab/MUSE.

OBJECTIVES: As large language models (LLMs) are integrated into electronic health record (EHR) workflows, validated instruments are essential to evaluate their performance before implementation and as models and documentation practices evolve. Existing instruments for provider documentation quality are often unsuitable for the complexities of LLM-generated text and lack validation on real-world data. The Provider Documentation Summarization Quality Instrument (PDSQI-9) was developed to evaluate LLM-generated clinical summaries. This study aimed to validate the PDSQI-9 across key aspects of construct validity. MATERIALS AND METHODS: Multi-document summaries were generated from real-world EHR data across multiple specialties using several LLMs (GPT-4o, Mixtral 8x7b, and Llama 3-8b). Validation included Pearson correlation analyses for substantive validity, factor analysis and Cronbach's α for structural validity, inter-rater reliability (ICC and Krippendorff's α) for generalizability, a semi-Delphi process for content validity, and comparisons of high- versus low-quality summaries for discriminant validity. Raters underwent standardized training to ensure consistent application of the instrument. RESULTS: Seven physician raters evaluated 779 summaries and answered 8329 questions, achieving over 80% power for inter-rater reliability. The PDSQI-9 demonstrated strong internal consistency (Cronbach's α = 0.879; 95% CI, 0.867-0.891) and high inter-rater reliability (ICC = 0.867; 95% CI, 0.867-0.868), supporting structural validity and generalizability. Factor analysis identified a 4-factor model explaining 58% of the variance, representing organization, clarity, accuracy, and utility. Substantive validity was supported by correlations between note length and scores for Succinct (ρ = -0.200, P = .029) and Organized (ρ = -0.190, P = .037). The semi-Delphi process ensured clinically relevant attributes, and discriminant validity distinguished high- from low-quality summaries (P<.001). DISCUSSION: The PDSQI-9 showed high inter-rater reliability, internal consistency, and a meaningful factor structure that reliably captured key dimensions of documentation quality. It distinguished between high- and low-quality summaries, supporting its practical utility for health systems needing an evaluation instrument for LLMs. CONCLUSIONS: The PDSQI-9 demonstrates robust construct validity, supporting its use in clinical practice to evaluate LLM-generated summaries and facilitate safer, more effective integration of LLMs into healthcare workflows.

OBJECTIVE: To evaluate the efficacy of digital twins developed using a large language model (LLaMA-3), fine-tuned with Low-Rank Adapters (LoRA) on intensive care units (ICU) physician notes, and to determine whether specialty-specific training enhances treatment recommendation accuracy compared to other ICU specialties or zero-shot baselines. MATERIALS AND METHODS: Digital twins were created using LLaMA-3 fine-tuned on discharge summaries from the Medical Information Mart for Intensive Care III dataset, where medications were masked to construct training and testing datasets. The medical ICU dataset (1000 notes) was used for evaluation, and performance was assessed using Bidirectional Encoder Representations from Transformers Score (BERTScore) and ROUGE-L. A zero-shot baseline model, relying solely on contextual instructions without training, was also evaluated. While our approach moves toward digital twin capabilities, it does not incorporate real-time, patient-specific electronic health records data and can be viewed as an ICU specialty-level language model adaptation. RESULTS: Models fine-tuned on medical ICU notes achieved the highest BERTScore (0.842), outperforming models trained on other specialties or mixed datasets. Zero-shot models showed the lowest performance, highlighting the importance of training. DISCUSSION: The findings demonstrate that specialty-specific training significantly improves treatment recommendation accuracy in digital twins compared to generalized or zero-shot approaches. Tailoring models to specific ICU domains strengthens their clinical decision-support capabilities. CONCLUSION: Context-specific fine-tuning of LLMs is crucial for developing effective digital twins, offering foundational insights for personalized clinical decision support.

OBJECTIVES: Applying large language models (LLMs) to the clinical domain is challenging due to the context-heavy nature of processing medical records. Retrieval-augmented generation (RAG) offers a solution by facilitating reasoning over large text sources. However, there are many parameters to optimize in just the retrieval system alone. This paper presents an ablation study exploring how different embedding models and pooling methods affect information retrieval for the clinical domain. MATERIALS AND METHODS: Evaluating on 3 retrieval tasks on 2 electronic health record (EHR) data sources, we compared 7 models, including medical- and general-domain models, specialized encoder embedding models, and off-the-shelf decoder LLMs. We also examine the choice of embedding pooling strategy for each model, independently on the query and the text to retrieve. RESULTS: We found that the choice of embedding model significantly impacts retrieval performance, with BGE, a comparatively small general-domain model, consistently outperforming all others, including medical-specific models. However, our findings also revealed substantial variability across datasets and query text phrasings. We also determined the best pooling methods for each of these models to guide future design of retrieval systems. DISCUSSION: The choice of embedding model, pooling strategy, and query formulation can significantly impact retrieval performance and the performance of these models on other public benchmarks does not necessarily transfer to new domains. The high variability in performance across different query phrasings suggests that the choice of query may need to be tuned and validated for each task, or even for each institution's EHR. CONCLUSION: This study provides empirical evidence to guide the selection of models and pooling strategies for RAG frameworks in healthcare applications. Further studies such as this one are vital for guiding empirically-grounded development of retrieval frameworks, such as in the context of RAG, for the clinical domain.

OBJECTIVE: Risk prediction models are used in hospitals to identify pediatric patients at risk of clinical deterioration, enabling timely interventions and rescue. The objective of this study was to develop a new explainer algorithm that uses a patient's clinical notes to generate text-based explanations for risk prediction alerts. MATERIALS AND METHODS: We conducted a retrospective study of 39 406 patient admissions to the American Family Children's Hospital at the University of Wisconsin-Madison (2009-2020). The pediatric Calculated Assessment of Risk and Triage (pCART) validated risk prediction model was used to identify children at risk for deterioration. A transformer model was trained to use clinical notes from the 12-hour period preceding each pCART score to predict whether a patient was flagged as at risk. Then, label-aware attention highlighted text phrases most important to an at-risk alert. The study cohort was randomly split into derivation (60%) and validation (20%) data, and a separate test (20%) was used to evaluate the explainer's performance. RESULTS: Our pCART Explainer algorithm performed well in discriminating at-risk pCART alert vs no alert (c-statistic 0.805). Sample explanations from pCART Explainer revealed clinically important phrases such as "rapid breathing," "fall risk," "distension," and "grunting," thereby demonstrating excellent face validity. DISCUSSION: The pCART Explainer could quickly orient clinicians to the patient's condition by drawing attention to key phrases in notes, potentially enhancing situational awareness and guiding decision-making. CONCLUSION: We developed pCART Explainer, a novel algorithm that highlights text within clinical notes to provide medically relevant context about deterioration alerts, thereby improving the explainability of the pCART model.

OBJECTIVES: The application of natural language processing (NLP) in the clinical domain is important due to the rich unstructured information in clinical documents, which often remains inaccessible in structured data. When applying NLP methods to a certain domain, the role of benchmark datasets is crucial as benchmark datasets not only guide the selection of best-performing models but also enable the assessment of the reliability of the generated outputs. Despite the recent availability of language models capable of longer context, benchmark datasets targeting long clinical document classification tasks are absent. MATERIALS AND METHODS: To address this issue, we propose Long Clinical Document (LCD) benchmark, a benchmark for the task of predicting 30-day out-of-hospital mortality using discharge notes of Medical Information Mart for Intensive Care IV and statewide death data. We evaluated this benchmark dataset using baseline models, from bag-of-words and convolutional neural network to instruction-tuned large language models. Additionally, we provide a comprehensive analysis of the model outputs, including manual review and visualization of model weights, to offer insights into their predictive capabilities and limitations. RESULTS: Baseline models showed 28.9% for best-performing supervised models and 32.2% for GPT-4 in F1 metrics. Notes in our dataset have a median word count of 1687. DISCUSSION: Our analysis of the model outputs showed that our dataset is challenging for both models and human experts, but the models can find meaningful signals from the text. CONCLUSION: We expect our LCD benchmark to be a resource for the development of advanced supervised models, or prompting methods, tailored for clinical text.

Deep learning-based methods have demonstrated encouraging results in tackling the task of panoramic image inpainting. However, it is challenging for existing methods to distinguish valid pixels from invalid pixels and find suitable references for corrupted areas, thus leading to artifacts in the inpainted results. In response to these challenges, we propose a panoramic image inpainting framework that consists of a Face Generator, a Cube Generator, a side branch, and two discriminators. We use the Cubemap Projection (CMP) format as network input. The generator employs gated convolutions to distinguish valid pixels from invalid ones, while a side branch is designed utilizing contextual reconstruction (CR) loss to guide the generators to find the most suitable reference patch for inpainting the missing region. The proposed method is compared with state-of-the-art (SOTA) methods on SUN360 Street View dataset in terms of PSNR and SSIM. Experimental results and ablation study demonstrate that the proposed method outperforms SOTA both quantitatively and qualitatively.

OBJECTIVE: The classification of clinical note sections is a critical step before doing more fine-grained natural language processing tasks such as social determinants of health extraction and temporal information extraction. Often, clinical note section classification models that achieve high accuracy for 1 institution experience a large drop of accuracy when transferred to another institution. The objective of this study is to develop methods that classify clinical note sections under the SOAP ("Subjective," "Object," "Assessment," and "Plan") framework with improved transferability. MATERIALS AND METHODS: We trained the baseline models by fine-tuning BERT-based models, and enhanced their transferability with continued pretraining, including domain-adaptive pretraining and task-adaptive pretraining. We added in-domain annotated samples during fine-tuning and observed model performance over a varying number of annotated sample size. Finally, we quantified the impact of continued pretraining in equivalence of the number of in-domain annotated samples added. RESULTS: We found continued pretraining improved models only when combined with in-domain annotated samples, improving the F1 score from 0.756 to 0.808, averaged across 3 datasets. This improvement was equivalent to adding 35 in-domain annotated samples. DISCUSSION: Although considered a straightforward task when performing in-domain, section classification is still a considerably difficult task when performing cross-domain, even using highly sophisticated neural network-based methods. CONCLUSION: Continued pretraining improved model transferability for cross-domain clinical note section classification in the presence of a small amount of in-domain labeled samples.

Automatically summarizing patients’ main problems from daily progress notes using natural language processing methods helps to battle against information and cognitive overload in hospital settings and potentially assists providers with computerized diagnostic decision support. Problem list summarization requires a model to understand, abstract, and generate clinical documentation. In this work, we propose a new NLP task that aims to generate a list of problems in a patient’s daily care plan using input from the provider’s progress notes during hospitalization. We investigate the performance of T5 and BART, two state-of-the-art seq2seq transformer architectures, in solving this problem. We provide a corpus built on top of progress notes from publicly available electronic health record progress notes in the Medical Information Mart for Intensive Care (MIMIC)-III. T5 and BART are trained on general domain text, and we experiment with a data augmentation method and a domain adaptation pre-training method to increase exposure to medical vocabulary and knowledge. Evaluation methods include ROUGE, BERTScore, cosine similarity on sentence embedding, and F-score on medical concepts. Results show that T5 with domain adaptive pre-training achieves significant performance gains compared to a rule-based system and general domain pre-trained language models, indicating a promising direction for tackling the problem summarization task.

Applying methods in natural language processing on electronic health records (EHR) data has attracted rising interests. Existing corpus and annotation focus on modeling textual features and relation prediction. However, there are a paucity of annotated corpus built to model clinical diagnostic thinking, a processing involving text understanding, domain knowledge abstraction and reasoning. In this work, we introduce a hierarchical annotation schema with three stages to address clinical text understanding, clinical reasoning and summarization. We create an annotated corpus based on a large collection of publicly available daily progress notes, a type of EHR that is time-sensitive, problem-oriented, and well-documented by the format of Subjective, Objective, Assessment and Plan (SOAP). We also define a new suite of tasks, Progress Note Understanding, with three tasks utilizing the three annotation stages. This new suite aims at training and evaluating future NLP models for clinical text understanding, clinical knowledge representation, inference and summarization.

OBJECTIVE: To provide a scoping review of papers on clinical natural language processing (NLP) shared tasks that use publicly available electronic health record data from a cohort of patients. MATERIALS AND METHODS: We searched 6 databases, including biomedical research and computer science literature databases. A round of title/abstract screening and full-text screening were conducted by 2 reviewers. Our method followed the PRISMA-ScR guidelines. RESULTS: A total of 35 papers with 48 clinical NLP tasks met inclusion criteria between 2007 and 2021. We categorized the tasks by the type of NLP problems, including named entity recognition, summarization, and other NLP tasks. Some tasks were introduced as potential clinical decision support applications, such as substance abuse detection, and phenotyping. We summarized the tasks by publication venue and dataset type. DISCUSSION: The breadth of clinical NLP tasks continues to grow as the field of NLP evolves with advancements in language systems. However, gaps exist with divergent interests between the general domain NLP community and the clinical informatics community for task motivation and design, and in generalizability of the data sources. We also identified issues in data preparation. CONCLUSION: The existing clinical NLP tasks cover a wide range of topics and the field is expected to grow and attract more attention from both general domain NLP and clinical informatics community. We encourage future work to incorporate multidisciplinary collaboration, reporting transparency, and standardization in data preparation. We provide a listing of all the shared task papers and datasets from this review in a GitLab repository.

Yanjun Gao, Ting-Hao Huang, Rebecca J. Passonneau. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Pyramid evaluation was developed to assess the content of paragraph length summaries of source texts.A pyramid lists the distinct units of content found in several reference summaries, weights content units by how many reference summaries they occur in, and produces three scores based on the weighted content of new summaries.We present an automated method that is more efficient, more transparent, and more complete than previous automated pyramid methods.It is tested on a new dataset of student summaries, and historical NIST data from extractive summarizers.