OBJECTIVES: This study explores the use of advanced natural language processing (NLP) techniques to enhance food classification and dietary analysis using raw text input from a diet tracking app. MATERIALS AND METHODS: The study was conducted in 3 stages: data collection, framework development, and application. Data were collected from a 12-week randomized controlled trial (RCT: NCT04259632), in which participants recorded their meals in free-text format using the myCircadianClock app. Only de-identified data were used. We developed nutrition-focused retrieval-augmented generation (NutriRAG), an NLP framework that uses a retrieval-augmented generation approach to enhance food classification from free-text inputs. The framework retrieves relevant examples from a curated database and then leverages large language models, such as GPT-4, to classify user-recorded food items into predefined categories without fine-tuning. NutriRAG was then applied to data from the RCT, which included 77 adults with obesity recruited from the Twin Cities metro area and randomized into 3 intervention groups: time-restricted eating (TRE, 8-hs eating window), caloric restriction (CR, 15% reduction), and unrestricted eating. RESULTS: NutriRAG significantly enhanced classification accuracy and helped to analyze dietary habits, as noted by the retrieval-augmented GPT-4 model achieving a micro-F1 score of 82.24. Both interventions showed dietary alterations: CR participants ate fewer snacks and sugary foods, while TRE participants reduced nighttime eating. CONCLUSION: By using artificial intelligence, NutriRAG marks a substantial advancement in food classification and dietary analysis of nutritional assessments. The findings highlight NLP's potential to personalize nutrition and manage diet-related health issues, suggesting further research to expand these models for wider use.

Huixue Zhou, Hengrui Gu, Zaifu Zhan, Xi Liu, Kaixiong Zhou, Yongkang Xiao, Mingfu Liang, Srinivas Prasad Govindan, Piyush Chawla, Jiyan Yang, Xiangfei Meng, Huayu Li, Buyun Zhang, Liang Luo, Wen-Yen Chen, Yiping Han, Bo Long, Rui Zhang, Tianlong Chen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Retrieval-Augmented Generation (RAG) has been empirically shown to enhance the performance of large language models (LLMs) in knowledge-intensive domains such as healthcare, finance, and legal contexts. Given a query, RAG retrieves relevant documents from a corpus and integrates them into the LLMs’ generation process. In this study, we investigate the adversarial robustness of RAG, focusing specifically on examining the retrieval system. First, across 225 different setup combinations of corpus, retriever, query, and targeted information, we show that retrieval systems are vulnerable to universal poisoning attacks in medical Q&A. In such attacks, adversaries generate poisoned documents containing a broad spectrum of targeted information, such as personally identifiable information. When these poisoned documents are inserted into a corpus, they can be accurately retrieved by any users, as long as attacker-specified queries are used. To understand this vulnerability, we discovered that the deviation from the query’s embedding to that of the poisoned document tends to follow a pattern in which the high similarity between the poisoned document and the query is retained, thereby enabling precise retrieval. Based on these findings, we develop a new detection-based defense to ensure the safe use of RAG. Through extensive experiments spanning various Q&A domains, we observed that our proposed method consistently achieves excellent detection rates in nearly all cases.

The integration of large language models (LLMs) into robotic task planning has unlocked better reasoning capabilities for complex, long-horizon workflows. However, ensuring safety in LLM-driven plans remains a critical challenge, as these models often prioritize task completion over risk mitigation. This paper introduces SAFER (Safety-Aware Framework for Execution in Robotics), a multi-LLM framework designed to embed safety awareness into robotic task planning. SAFER employs a Safety Agent that operates alongside the primary task planner, providing safety feedback. Additionally, we introduce LLM-as-a-Judge, a novel metric leveraging LLMs as evaluators to quantify safety violations within generated task plans. Our framework integrates safety feedback at multiple stages of execution, enabling real-time risk assessment, proactive error correction, and transparent safety evaluation. We also integrate a control framework using Control Barrier Functions (CBFs) to ensure safety guarantees within SAFER’s task planning. We evaluated SAFER against state-of-the-art LLM planners on complex long-horizon tasks involving heterogeneous robotic agents, demonstrating its effectiveness in reducing safety violations while maintaining task efficiency. We also verify the task planner and safety planner through actual hardware experiments involving multiple robots and a human.

OBJECTIVE: Access to firearms is associated with increased suicide risk. Our aim was to develop a natural language processing approach to characterizing firearm access in clinical records. MATERIALS AND METHODS: We used clinical notes from 36 685 Veterans Health Administration (VHA) patients between April 10, 2023 and April 10, 2024. We expanded preexisting firearm term sets using subject matter experts and generated 250-character snippets around each firearm term appearing in notes. Annotators labeled 3000 snippets into three classes. Using these annotated snippets, we compared four nonneural machine learning models (random forest, bagging, gradient boosting, logistic regression with ridge penalization) and two versions of Bidirectional Encoder Representations from Transformers, or BERT (specifically, BioBERT and Bio-ClinicalBERT) for classifying firearm access as "definite access", "definitely no access", or "other". RESULTS: Firearm terms were identified in 36 685 patient records (41.3%), 33.7% of snippets were categorized as definite access, 9.0% as definitely no access, and 57.2% as "other". Among models classifying firearm access, five of six had acceptable performance, with BioBERT and Bio-ClinicalBERT performing best, with F1s of 0.876 (95% confidence interval, 0.874-0.879) and 0.896 (95% confidence interval, 0.894-0.899), respectively. DISCUSSION AND CONCLUSION: Firearm-related terminology is common in the clinical records of VHA patients. The ability to use text to identify and characterize patients' firearm access could enhance suicide prevention efforts, and five of our six models could be used to identify patients for clinical interventions.

OBJECTIVES: To optimize in-context learning in biomedical natural language processing by improving example selection. MATERIALS AND METHODS: We introduce a novel multi-mode retrieval-augmented generation (MMRAG) framework, which integrates 4 retrieval strategies: (1) Random Mode, selecting examples arbitrarily; (2) Top Mode, retrieving the most relevant examples based on similarity; (3) Diversity Mode, ensuring variation in selected examples; and (4) Class Mode, selecting category-representative examples. This study evaluates MMRAG on 3 core biomedical NLP tasks: Named Entity Recognition (NER), Relation Extraction (RE), and Text Classification (TC). The datasets used include BC2GM for gene and protein mention recognition (NER), DDI for drug-drug interaction extraction (RE), GIT for general biomedical information extraction (RE), and HealthAdvice for health-related text classification (TC). The framework is tested with 2 large language models (Llama-2-7B and Llama-3-8B) and 3 retrievers (Contriever, MedCPT, and BGE-Large) to assess performance across different retrieval strategies. RESULTS: The results from the Random Mode indicate that providing more examples in the prompt improves the model's generation performance. Meanwhile, Top Mode and Diversity Mode significantly outperform Random Mode on the RE (DDI) task, achieving an F1 score of 0.9669-a 26.4% improvement. Among the 3 retrievers tested, Contriever outperformed the other 2 in a greater number of experiments. Additionally, Llama 2 and Llama 3 demonstrated varying capabilities across different tasks, with Llama 3 showing a clear advantage in handling NER tasks. CONCLUSION: MMRAG effectively enhances biomedical in-context learning by refining example selection, mitigating data scarcity issues, and demonstrating superior adaptability for NLP-driven healthcare applications.

OBJECTIVE: To develop an advanced multi-task large language model (LLM) framework for extracting diverse types of information about dietary supplements (DSs) from clinical records. METHODS: We focused on 4 core DS information extraction tasks: named entity recognition (2 949 clinical sentences), relation extraction (4 892 sentences), triple extraction (2 949 sentences), and usage classification (2 460 sentences). To address these tasks, we introduced the retrieval-augmented multi-task information extraction (RAMIE) framework, which incorporates: (1) instruction fine-tuning with task-specific prompts; (2) multi-task training of LLMs to enhance storage efficiency and reduce training costs; and (3) retrieval-augmented generation, which retrieves similar examples from the training set to improve task performance. We compared the performance of RAMIE to LLMs with instruction fine-tuning alone and conducted an ablation study to evaluate the individual contributions of multi-task learning and retrieval-augmented generation to overall performance improvements. RESULTS: Using the RAMIE framework, Llama2-13B achieved an F1 score of 87.39 on the named entity recognition task, reflecting a 3.51% improvement. It also excelled in the relation extraction task with an F1 score of 93.74, a 1.15% improvement. For the triple extraction task, Llama2-7B achieved an F1 score of 79.45, representing a significant 14.26% improvement. MedAlpaca-7B delivered the highest F1 score of 93.45 on the usage classification task, with a 0.94% improvement. The ablation study highlighted that while multi-task learning improved efficiency with a minor trade-off in performance, the inclusion of retrieval-augmented generation significantly enhanced overall accuracy across tasks. CONCLUSION: The RAMIE framework demonstrates substantial improvements in multi-task information extraction for DS-related data from clinical records.

Retrieval-augmented generation (RAG) involves a solution by retrieving knowledge from an established database to enhance the performance of large language models (LLM). , these models retrieve information at the sentence or paragraph level, potentially introducing noise and affecting the generation quality. To address these issues, we propose a novel BiomedRAG framework that directly feeds automatically retrieved chunk-based documents into the LLM. Our evaluation of BiomedRAG across four biomedical natural language processing tasks using eight datasets demonstrates that our proposed framework not only improves the performance by 9.95% on average, but also achieves state-of-the-art results, surpassing various baselines by 4.97%. BiomedRAG paves the way for more accurate and adaptable LLM applications in the biomedical domain.

The increasing reliance on Large Language Models (LLMs) for health information seeking can pose severe risks due to the potential for misinformation and the complexity of these topics. This paper introduces KnowNet a visualization system that integrates LLMs with Knowledge Graphs (KG) to provide enhanced accuracy and structured exploration. Specifically, for enhanced accuracy, KnowNet extracts triples (e.g., entities and their relations) from LLM outputs and maps them into the validated information and supported evidence in external KGs. For structured exploration, KnowNet provides next-step recommendations based on the neighborhood of the currently explored entities in KGs, aiming to guide a comprehensive understanding without overlooking critical aspects. To enable reasoning with both the structured data in KGs and the unstructured outputs from LLMs, KnowNet conceptualizes the understanding of a subject as the gradual construction of graph visualization. A progressive graph visualization is introduced to monitor past inquiries, and bridge the current query with the exploration history and next-step recommendations. We demonstrate the effectiveness of our system via use cases and expert interviews.

Zero-shot link prediction (ZSLP) on knowledge graphs aims at automatically identifying relations between given entities. Existing methods primarily employ auxiliary information to predict tail entity given head entity and its relation, yet face challenges due to the occasional unavailability of such detailed information and the inherent simplicity of predicting tail entities based on semantic similarities. Even though Large Language Models (LLMs) offer a promising solution to predict unobserved relations between the head and tail entity in a zero-shot manner, their performance is still restricted due to the inability to leverage all the (exponentially many) paths' information between two entities, which are critical in collectively indicating their relation types. To address this, in this work, we introduce a Condensed Transition Graph Framework for Zero-Shot Link Prediction (CTLP), which encodes all the paths' information in linear time complexity to predict unseen relations between entities, attaining both efficiency and information preservation. Specifically, we design a condensed transition graph encoder with theoretical guarantees on its coverage, expressiveness, and efficiency. It is learned by a transition graph contrastive learning strategy. Subsequently, we design a soft instruction tuning to learn and map the all-path embedding to the input of LLMs. Experimental results show that our proposed CTLP method achieves state-of-the-art performance on three standard ZSLP datasets.11The code is available here: https://github.com/ToneLi/Graph_LLM_link_predcition

BACKGROUND: Error analysis plays a crucial role in clinical concept extraction, a fundamental subtask within clinical natural language processing (NLP). The process typically involves a manual review of error types, such as contextual and linguistic factors contributing to their occurrence, and the identification of underlying causes to refine the NLP model and improve its performance. Conducting error analysis can be complex, requiring a combination of NLP expertise and domain-specific knowledge. Due to the high heterogeneity of electronic health record (EHR) settings across different institutions, challenges may arise when attempting to standardize and reproduce the error analysis process. OBJECTIVES: This study aims to facilitate a collaborative effort to establish common definitions and taxonomies for capturing diverse error types, fostering community consensus on error analysis for clinical concept extraction tasks. MATERIALS AND METHODS: We iteratively developed and evaluated an error taxonomy based on existing literature, standards, real-world data, multisite case evaluations, and community feedback. The finalized taxonomy was released in both .dtd and .owl formats at the Open Health Natural Language Processing Consortium. The taxonomy is compatible with several different open-source annotation tools, including MAE, Brat, and MedTator. RESULTS: The resulting error taxonomy comprises 43 distinct error classes, organized into 6 error dimensions and 4 properties, including model type (symbolic and statistical machine learning), evaluation subject (model and human), evaluation level (patient, document, sentence, and concept), and annotation examples. Internal and external evaluations revealed strong variations in error types across methodological approaches, tasks, and EHR settings. Key points emerged from community feedback, including the need to enhancing clarity, generalizability, and usability of the taxonomy, along with dissemination strategies. CONCLUSION: The proposed taxonomy can facilitate the acceleration and standardization of the error analysis process in multi-site settings, thus improving the provenance, interpretability, and portability of NLP models. Future researchers could explore the potential direction of developing automated or semi-automated methods to assist in the classification and standardization of error analysis.

OBJECTIVES: This article aims to enhance the performance of larger language models (LLMs) on the few-shot biomedical named entity recognition (NER) task by developing a simple and effective method called Retrieving and Chain-of-Thought (RT) framework and to evaluate the improvement after applying RT framework. MATERIALS AND METHODS: Given the remarkable advancements in retrieval-based language model and Chain-of-Thought across various natural language processing tasks, we propose a pioneering RT framework designed to amalgamate both approaches. The RT approach encompasses dedicated modules for information retrieval and Chain-of-Thought processes. In the retrieval module, RT discerns pertinent examples from demonstrations during instructional tuning for each input sentence. Subsequently, the Chain-of-Thought module employs a systematic reasoning process to identify entities. We conducted a comprehensive comparative analysis of our RT framework against 16 other models for few-shot NER tasks on BC5CDR and NCBI corpora. Additionally, we explored the impacts of negative samples, output formats, and missing data on performance. RESULTS: Our proposed RT framework outperforms other LMs for few-shot NER tasks with micro-F1 scores of 93.50 and 91.76 on BC5CDR and NCBI corpora, respectively. We found that using both positive and negative samples, Chain-of-Thought (vs Tree-of-Thought) performed better. Additionally, utilization of a partially annotated dataset has a marginal effect of the model performance. DISCUSSION: This is the first investigation to combine a retrieval-based LLM and Chain-of-Thought methodology to enhance the performance in biomedical few-shot NER. The retrieval-based LLM aids in retrieving the most relevant examples of the input sentence, offering crucial knowledge to predict the entity in the sentence. We also conducted a meticulous examination of our methodology, incorporating an ablation study. CONCLUSION: The RT framework with LLM has demonstrated state-of-the-art performance on few-shot NER tasks.

IMPORTANCE: Reinforcement learning (RL) represents a pivotal avenue within natural language processing (NLP), offering a potent mechanism for acquiring optimal strategies in task completion. This literature review studies various NLP applications where RL has demonstrated efficacy, with notable applications in healthcare settings. OBJECTIVES: To systematically explore the applications of RL in NLP, focusing on its effectiveness in acquiring optimal strategies, particularly in healthcare settings, and provide a comprehensive understanding of RL's potential in NLP tasks. MATERIALS AND METHODS: Adhering to the PRISMA guidelines, an exhaustive literature review was conducted to identify instances where RL has exhibited success in NLP applications, encompassing dialogue systems, machine translation, question-answering, text summarization, and information extraction. Our methodological approach involves closely examining the technical aspects of RL methodologies employed in these applications, analyzing algorithms, states, rewards, actions, datasets, and encoder-decoder architectures. RESULTS: The review of 93 papers yields insights into RL algorithms, prevalent techniques, emergent trends, and the fusion of RL methods in NLP healthcare applications. It clarifies the strategic approaches employed, datasets utilized, and the dynamic terrain of RL-NLP systems, thereby offering a roadmap for research and development in RL and machine learning techniques in healthcare. The review also addresses ethical concerns to ensure equity, transparency, and accountability in the evolution and application of RL-based NLP technologies, particularly within sensitive domains such as healthcare. DISCUSSION: The findings underscore the promising role of RL in advancing NLP applications, particularly in healthcare, where its potential to optimize decision-making and enhance patient outcomes is significant. However, the ethical challenges and technical complexities associated with RL demand careful consideration and ongoing research to ensure responsible and effective implementation. CONCLUSIONS: By systematically exploring RL's applications in NLP and providing insights into technical analysis, ethical implications, and potential advancements, this review contributes to a deeper understanding of RL's role for language processing.

OBJECTIVE: This study leverages the rich diversity of the All of Us Research Program (All of Us)'s dataset to devise a predictive model for cardiovascular disease (CVD) in breast cancer (BC) survivors. Central to this endeavor is the creation of a robust data integration pipeline that synthesizes electronic health records (EHRs), patient surveys, and genomic data, while upholding fairness across demographic variables. MATERIALS AND METHODS: We have developed a universal data wrangling pipeline to process and merge heterogeneous data sources of the All of Us dataset, address missingness and variance in data, and align disparate data modalities into a coherent framework for analysis. Utilizing a composite feature set including EHR, lifestyle, and social determinants of health (SDoH) data, we then employed Adaptive Lasso and Random Forest regression models to predict 6 CVD outcomes. The models were evaluated using the c-index and time-dependent Area Under the Receiver Operating Characteristic Curve over a 10-year period. RESULTS: The Adaptive Lasso model showed consistent performance across most CVD outcomes, while the Random Forest model excelled particularly in predicting outcomes like transient ischemic attack when incorporating the full multi-model feature set. Feature importance analysis revealed age and previous coronary events as dominant predictors across CVD outcomes, with SDoH clustering labels highlighting the nuanced impact of social factors. DISCUSSION: The development of both Cox-based predictive model and Random Forest Regression model represents the extensive application of the All of Us, in integrating EHR and patient surveys to enhance precision medicine. And the inclusion of SDoH clustering labels revealed the significant impact of sociobehavioral factors on patient outcomes, emphasizing the importance of comprehensive health determinants in predictive models. Despite these advancements, limitations include the exclusion of genetic data, broad categorization of CVD conditions, and the need for fairness analyses to ensure equitable model performance across diverse populations. Future work should refine clinical and social variable measurements, incorporate advanced imputation techniques, and explore additional predictive algorithms to enhance model precision and fairness. CONCLUSION: This study demonstrates the liability of the All of Us's diverse dataset in developing a multi-modality predictive model for CVD in BC survivors risk stratification in oncological survivorship. The data integration pipeline and subsequent predictive models establish a methodological foundation for future research into personalized healthcare.

OBJECTIVE: To construct an exhaustive Complementary and Integrative Health (CIH) Lexicon (CIHLex) to help better represent the often underrepresented physical and psychological CIH approaches in standard terminologies, and to also apply state-of-the-art natural language processing (NLP) techniques to help recognize them in the biomedical literature. MATERIALS AND METHODS: We constructed the CIHLex by integrating various resources, compiling and integrating data from biomedical literature and relevant sources of knowledge. The Lexicon encompasses 724 unique concepts with 885 corresponding unique terms. We matched these concepts to the Unified Medical Language System (UMLS), and we developed and utilized BERT models comparing their efficiency in CIH named entity recognition to well-established models including MetaMap and CLAMP, as well as the large language model GPT3.5-turbo. RESULTS: Of the 724 unique concepts in CIHLex, 27.2% could be matched to at least one term in the UMLS. About 74.9% of the mapped UMLS Concept Unique Identifiers were categorized as "Therapeutic or Preventive Procedure." Among the models applied to CIH named entity recognition, BLUEBERT delivered the highest macro-average F1-score of 0.91, surpassing other models. CONCLUSION: Our CIHLex significantly augments representation of CIH approaches in biomedical literature. Demonstrating the utility of advanced NLP models, BERT notably excelled in CIH entity recognition. These results highlight promising strategies for enhancing standardization and recognition of CIH terminology in biomedical contexts.

OBJECTIVE: To investigate the demonstration in large language models (LLMs) for biomedical relation extraction. This study introduces a framework comprising three types of adaptive tuning methods to assess their impacts and effectiveness. MATERIALS AND METHODS: Our study was conducted in two phases. Initially, we analyzed a range of demonstration components vital for LLMs' biomedical data capabilities, including task descriptions and examples, experimenting with various combinations. Subsequently, we introduced the LLM instruction-example adaptive prompting (LEAP) framework, including instruction adaptive tuning, example adaptive tuning, and instruction-example adaptive tuning methods. This framework aims to systematically investigate both adaptive task descriptions and adaptive examples within the demonstration. We assessed the performance of the LEAP framework on the DDI, ChemProt, and BioRED datasets, employing LLMs such as Llama2-7b, Llama2-13b, and MedLLaMA_13B. RESULTS: Our findings indicated that Instruction + Options + Example and its expanded form substantially improved F1 scores over the standard Instruction + Options mode for zero-shot LLMs. The LEAP framework, particularly through its example adaptive prompting, demonstrated superior performance over conventional instruction tuning across all models. Notably, the MedLLAMA_13B model achieved an exceptional F1 score of 95.13 on the ChemProt dataset using this method. Significant improvements were also observed in the DDI 2013 and BioRED datasets, confirming the method's robustness in sophisticated data extraction scenarios. CONCLUSION: The LEAP framework offers a compelling strategy for enhancing LLM training strategies, steering away from extensive fine-tuning towards more dynamic and contextually enriched prompting methodologies, showcasing in biomedical relation extraction.

OBJECTIVE: Relation extraction is an essential task in the field of biomedical literature mining and offers significant benefits for various downstream applications, including database curation, drug repurposing, and literature-based discovery. The broad-coverage natural language processing (NLP) tool SemRep has established a solid baseline for extracting subject-predicate-object triples from biomedical text and has served as the backbone of the Semantic MEDLINE Database (SemMedDB), a PubMed-scale repository of semantic triples. While SemRep achieves reasonable precision (0.69), its recall is relatively low (0.42). In this study, we aimed to enhance SemRep using a relation classification approach, in order to eventually increase the size and the utility of SemMedDB. METHODS: We combined and extended existing SemRep evaluation datasets to generate training data. We leveraged the pre-trained PubMedBERT model, enhancing it through additional contrastive pre-training and fine-tuning. We experimented with three entity representations: mentions, semantic types, and semantic groups. We evaluated the model performance on a portion of the SemRep Gold Standard dataset and compared it to SemRep performance. We also assessed the effect of the model on a larger set of 12K randomly selected PubMed abstracts. RESULTS: score of 0.70. Assessment on 12K abstracts shows that the model could double the size of SemMedDB, when applied to entire PubMed. We also manually assessed the quality of 506 triples predicted by the model that SemRep had not previously identified, and found that 67% of these triples were correct. CONCLUSION: These findings underscore the promise of our model in achieving a more comprehensive coverage of relationships mentioned in biomedical literature, thereby showing its potential in enhancing various downstream applications of biomedical literature mining. Data and code related to this study are available at https://github.com/Michelle-Mings/SemRep_RelationClassification.

Despite recent methodology advancements in clinical natural language processing (NLP), the adoption of clinical NLP models within the translational research community remains hindered by process heterogeneity and human factor variations. Concurrently, these factors also dramatically increase the difficulty in developing NLP models in multi-site settings, which is necessary for algorithm robustness and generalizability. Here, we reported on our experience developing an NLP solution for Coronavirus Disease 2019 (COVID-19) signs and symptom extraction in an open NLP framework from a subset of sites participating in the National COVID Cohort (N3C). We then empirically highlight the benefits of multi-site data for both symbolic and statistical methods, as well as highlight the need for federated annotation and evaluation to resolve several pitfalls encountered in the course of these efforts.

BACKGROUND: To date, there are no effective treatments for most neurodegenerative diseases. Knowledge graphs can provide comprehensive and semantic representation for heterogeneous data, and have been successfully leveraged in many biomedical applications including drug repurposing. Our objective is to construct a knowledge graph from literature to study the relations between Alzheimer's disease (AD) and chemicals, drugs and dietary supplements in order to identify opportunities to prevent or delay neurodegenerative progression. We collected biomedical annotations and extracted their relations using SemRep via SemMedDB. We used both a BERT-based classifier and rule-based methods during data preprocessing to exclude noise while preserving most AD-related semantic triples. The 1,672,110 filtered triples were used to train with knowledge graph completion algorithms (i.e., TransE, DistMult, and ComplEx) to predict candidates that might be helpful for AD treatment or prevention. RESULTS: Among three knowledge graph completion models, TransE outperformed the other two (MR = 10.53, Hits@1 = 0.28). We leveraged the time-slicing technique to further evaluate the prediction results. We found supporting evidence for most highly ranked candidates predicted by our model which indicates that our approach can inform reliable new knowledge. CONCLUSION: This paper shows that our graph mining model can predict reliable new relationships between AD and other entities (i.e., dietary supplements, chemicals, and drugs). The knowledge graph constructed can facilitate data-driven knowledge discoveries and the generation of novel hypotheses.

OBJECTIVE: Accurate extraction of breast cancer patients' phenotypes is important for clinical decision support and clinical research. This study developed and evaluated cancer domain pretrained CancerBERT models for extracting breast cancer phenotypes from clinical texts. We also investigated the effect of customized cancer-related vocabulary on the performance of CancerBERT models. MATERIALS AND METHODS: A cancer-related corpus of breast cancer patients was extracted from the electronic health records of a local hospital. We annotated named entities in 200 pathology reports and 50 clinical notes for 8 cancer phenotypes for fine-tuning and evaluation. We kept pretraining the BlueBERT model on the cancer corpus with expanded vocabularies (using both term frequency-based and manually reviewed methods) to obtain CancerBERT models. The CancerBERT models were evaluated and compared with other baseline models on the cancer phenotype extraction task. RESULTS: All CancerBERT models outperformed all other models on the cancer phenotyping NER task. Both CancerBERT models with customized vocabularies outperformed the CancerBERT with the original BERT vocabulary. The CancerBERT model with manually reviewed customized vocabulary achieved the best performance with macro F1 scores equal to 0.876 (95% CI, 0.873-0.879) and 0.904 (95% CI, 0.902-0.906) for exact match and lenient match, respectively. CONCLUSIONS: The CancerBERT models were developed to extract the cancer phenotypes in clinical notes and pathology reports. The results validated that using customized vocabulary may further improve the performances of domain specific BERT models in clinical NLP tasks. The CancerBERT models developed in the study would further help clinical decision support.

OBJECTIVE: Develop a novel methodology to create a comprehensive knowledge graph (SuppKG) to represent a domain with limited coverage in the Unified Medical Language System (UMLS), specifically dietary supplement (DS) information for discovering drug-supplement interactions (DSI), by leveraging biomedical natural language processing (NLP) technologies and a DS domain terminology. MATERIALS AND METHODS: We created SemRepDS (an extension of an NLP tool, SemRep), capable of extracting semantic relations from abstracts by leveraging a DS-specific terminology (iDISK) containing 28,884 DS terms not found in the UMLS. PubMed abstracts were processed using SemRepDS to generate semantic relations, which were then filtered using a PubMedBERT model to remove incorrect relations before generating SuppKG. Two discovery pathways were applied to SuppKG to identify potential DSIs, which are then compared with an existing DSI database and also evaluated by medical professionals for mechanistic plausibility. RESULTS: SemRepDS returned 158.5% more DS entities and 206.9% more DS relations than SemRep. The fine-tuned PubMedBERT model (significantly outperformed other machine learning and BERT models) obtained an F1 score of 0.8605 and removed 43.86% of semantic relations, improving the precision of the relations by 26.4% over pre-filtering. SuppKG consists of 56,635 nodes and 595,222 directed edges with 2,928 DS-specific nodes and 164,738 edges. Manual review of findings identified 182 of 250 (72.8%) proposed DS-Gene-Drug and 77 of 100 (77%) proposed DS-Gene1-Function-Gene2-Drug pathways to be mechanistically plausible. DISCUSSION: With added DS terminology to the UMLS, SemRepDS has the capability to find more DS-specific semantic relationships from PubMed than SemRep. The utility of the resulting SuppKG was demonstrated using discovery patterns to find novel DSIs. CONCLUSION: For the domain with limited coverage in the traditional terminology (e.g., UMLS), we demonstrated an approach to leverage domain terminology and improve existing NLP tools to generate a more comprehensive knowledge graph for the downstream task. Even this study focuses on DSI, the method may be adapted to other domains.

Statistical modeling of outcomes based on a patient's presenting symptoms (symptomatology) can help deliver high quality care and allocate essential resources, which is especially important during the COVID-19 pandemic. Patient symptoms are typically found in unstructured notes, and thus not readily available for clinical decision making. In an attempt to fill this gap, this study compared two methods for symptom extraction from Emergency Department (ED) admission notes. Both methods utilized a lexicon derived by expanding The Center for Disease Control and Prevention's (CDC) Symptoms of Coronavirus list. The first method utilized a word2vec model to expand the lexicon using a dictionary mapping to the Uni ed Medical Language System (UMLS). The second method utilized the expanded lexicon as a rule-based gazetteer and the UMLS. These methods were evaluated against a manually annotated reference (f1-score of 0.87 for UMLS-based ensemble; and 0.85 for rule-based gazetteer with UMLS). Through analyses of associations of extracted symptoms used as features against various outcomes, salient risks among the population of COVID-19 patients, including increased risk of in-hospital mortality (OR 1.85, p-value < 0.001), were identified for patients presenting with dyspnea. Disparities between English and non-English speaking patients were also identified, the most salient being a concerning finding of opposing risk signals between fatigue and in-hospital mortality (non-English: OR 1.95, p-value = 0.02; English: OR 0.63, p-value = 0.01). While use of symptomatology for modeling of outcomes is not unique, unlike previous studies this study showed that models built using symptoms with the outcome of in-hospital mortality were not significantly different from models using data collected during an in-patient encounter (AUC of 0.9 with 95% CI of [0.88, 0.91] using only vital signs; AUC of 0.87 with 95% CI of [0.85, 0.88] using only symptoms). These findings indicate that prognostic models based on symptomatology could aid in extending COVID-19 patient care through telemedicine, replacing the need for in-person options. The methods presented in this study have potential for use in development of symptomatology-based models for other diseases, including for the study of Post-Acute Sequelae of COVID-19 (PASC).

OBJECTIVE: We sought to demonstrate the feasibility of utilizing deep learning models to extract safety signals related to the use of dietary supplements (DSs) in clinical text. MATERIALS AND METHODS: Two tasks were performed in this study. For the named entity recognition (NER) task, Bi-LSTM-CRF (bidirectional long short-term memory conditional random field) and BERT (bidirectional encoder representations from transformers) models were trained and compared with CRF model as a baseline to recognize the named entities of DSs and events from clinical notes. In the relation extraction (RE) task, 2 deep learning models, including attention-based Bi-LSTM and convolutional neural network as well as a random forest model were trained to extract the relations between DSs and events, which were categorized into 3 classes: positive (ie, indication), negative (ie, adverse events), and not related. The best performed NER and RE models were further applied on clinical notes mentioning 88 DSs for discovering DSs adverse events and indications, which were compared with a DS knowledge base. RESULTS: For the NER task, deep learning models achieved a better performance than CRF, with F1 scores above 0.860. The attention-based Bi-LSTM model performed the best in the RE task, with an F1 score of 0.893. When comparing DS event pairs generated by the deep learning models with the knowledge base for DSs and event, we found both known and unknown pairs. CONCLUSIONS: Deep learning models can detect adverse events and indication of DSs in clinical notes, which hold great potential for monitoring the safety of DS use.

OBJECTIVE: To build a knowledge base of dietary supplement (DS) information, called the integrated DIetary Supplement Knowledge base (iDISK), which integrates and standardizes DS-related information from 4 existing resources. MATERIALS AND METHODS: iDISK was built through an iterative process comprising 3 phases: 1) establishment of the content scope, 2) development of the data model, and 3) integration of existing resources. Four well-regarded DS resources were integrated into iDISK: The Natural Medicines Comprehensive Database, the "About Herbs" page on the Memorial Sloan Kettering Cancer Center website, the Dietary Supplement Label Database, and the Natural Health Products Database. We evaluated the iDISK build process by manually checking that the data elements associated with 50 randomly selected ingredients were correctly extracted and integrated from their respective sources. RESULTS: iDISK encompasses a terminology of 4208 DS ingredient concepts, which are linked via 6 relationship types to 495 drugs, 776 diseases, 985 symptoms, 605 therapeutic classes, 17 system organ classes, and 137 568 DS products. iDISK also contains 7 concept attribute types and 3 relationship attribute types. Evaluation of the data extraction and integration process showed average errors of 0.3%, 2.6%, and 0.4% for concepts, relationships and attributes, respectively. CONCLUSION: We developed iDISK, a publicly available standardized DS knowledge base that can facilitate more efficient and meaningful dissemination of DS knowledge.

OBJECTIVE: We sought to assess the need for additional coverage of dietary supplements (DS) in the Unified Medical Language System (UMLS) by investigating (1) the overlap between the integrated DIetary Supplements Knowledge base (iDISK) DS ingredient terminology and the UMLS and (2) the coverage of iDISK and the UMLS over DS mentions in the biomedical literature. MATERIALS AND METHODS: We estimated the overlap between iDISK and the UMLS by mapping iDISK to the UMLS using exact and normalized strings. The coverage of iDISK and the UMLS over DS mentions in the biomedical literature was evaluated via a DS named-entity recognition (NER) task within PubMed abstracts. RESULTS: The coverage analysis revealed that only 30% of iDISK terms can be matched to the UMLS, although these cover over 99% of iDISK concepts. A manual review revealed that a majority of the unmatched terms represented new synonyms, rather than lexical variants. For NER, iDISK nearly doubles the precision and achieves a higher F1 score than the UMLS, while maintaining a competitive recall. DISCUSSION: While iDISK has significant concept overlap with the UMLS, it contains many novel synonyms. Furthermore, almost 3000 of these overlapping UMLS concepts are missing a DS designation, which could be provided by iDISK. The NER experiments show that the specialization of iDISK is useful for identifying DS mentions. CONCLUSIONS: Our results show that the DS representation in the UMLS could be enriched by adding DS designations to many concepts and by adding new synonyms.

OBJECTIVES: The study sought to test the feasibility of using Twitter data to assess determinants of consumers' health behavior toward human papillomavirus (HPV) vaccination informed by the Integrated Behavior Model (IBM). MATERIALS AND METHODS: We used 3 Twitter datasets spanning from 2014 to 2018. We preprocessed and geocoded the tweets, and then built a rule-based model that classified each tweet into either promotional information or consumers' discussions. We applied topic modeling to discover major themes and subsequently explored the associations between the topics learned from consumers' discussions and the responses of HPV-related questions in the Health Information National Trends Survey (HINTS). RESULTS: We collected 2 846 495 tweets and analyzed 335 681 geocoded tweets. Through topic modeling, we identified 122 high-quality topics. The most discussed consumer topic is "cervical cancer screening"; while in promotional tweets, the most popular topic is to increase awareness of "HPV causes cancer." A total of 87 of the 122 topics are correlated between promotional information and consumers' discussions. Guided by IBM, we examined the alignment between our Twitter findings and the results obtained from HINTS. Thirty-five topics can be mapped to HINTS questions by keywords, 112 topics can be mapped to IBM constructs, and 45 topics have statistically significant correlations with HINTS responses in terms of geographic distributions. CONCLUSIONS: Mining Twitter to assess consumers' health behaviors can not only obtain results comparable to surveys, but also yield additional insights via a theory-driven approach. Limitations exist; nevertheless, these encouraging results impel us to develop innovative ways of leveraging social media in the changing health communication landscape.

The widespread prevalence of dietary supplements has drawn extensive attention due to the safety and efficacy issue. Clinical notes document a great amount of detailed information on dietary supplement usage, thus providing a rich source for clinical research on supplement safety surveillance. Identification the use status of dietary supplements is one of the initial steps for the ultimate goal of the supplement safety surveillance. In this study, we built rule-based and machine learning-based classifiers to automatically classify the use status of supplements into four categories: Continuing (C), Discontinued (D), Started (S), and Unclassified (U). In comparison to the machine learning classifier trained on the same datasets, the rule-based classifier showed a better performance with F-measure in the C, D, S, U status of 0.93, 0.98, 0.95, and 0.83, respectively. We further analyzed the errors generated by the rule-based classifier. The classifier can be potentially applied to extract supplement information from clinical notes for supporting research and clinical practice related to patient safety on supplement usage.

Cardiac Resynchronization Therapy (CRT) is an established pacing therapy for heart failure patients. The New York Heart Association (NYHA) classification is often used as a measure of a patient's response to CRT. Identifying NYHA class for heart failure patients in an electronic health record (EHR) consistently, over time, can provide better understanding of the progression of heart failure and assessment of CRT response and effectiveness. However, NYHA is rarely stored in EHR structured data such information is often documented in unstructured clinical notes. In this study, we thus investigated the use of natural language processing (NLP) methods to identify NYHA classification from clinical notes. We collected 6,174 clinical notes that were matched with hospital-specific custom NYHA class diagnosis codes. Machine-learning based methods performed similar with a rule-based method. The best machine-learning method, support vector machine with n-gram features, performed the best (93% F-measure). Further validation of the findings is required.

Clinical notes contain rich information about dietary supplements, which are critical for detecting signals of dietary supplement side effects and interactions between drugs and supplements. One of the important factors of supplement documentation is usage status, such as started and discontinuation. Such information is usually stored in the unstructured clinical notes. We developed a rule-based classifier to identify supplement usage status in clinical notes. The categories referring to the patient's status of supplement use were classified into four classes: Continuing (C), Discontinued (D), Started (S), and Unclassified (U). Clinical notes containing 10 of the most commonly consumed supplements (i.e., alfalfa, echinacea, fish oil, garlic, ginger, ginkgo, ginseng, melatonin, St. John's Wort, and Vitamin E) were retrieved from the University of Minnesota Clinical Data Repository. The gold standard was defined by manually annotating 1000 randomly selected sentences or statements mentioning at least one of these 10 supplements. The rules in the classifier was initially developed on two-thirds of the set of 7 supplements (i.e., alfalfa, garlic, ginger, ginkgo, ginseng, St. John's Wort, and Vitamin E); the performance was evaluated on the remaining one-third of this set. To evaluate the generalizability of rules, we further validated the second testing set on other 3 supplements (i.e., echinacea, fish oil, and melatonin). The performance of the classifier achieved F-measures of 0.95, 0.97, 0.96, and 0.96 for status C, D, S, and U on 7 supplements, respectively. The classifier also showed good generalizability when it was applied to the other 3 supplements with F-measures of 0.96 for C, 0.96 for D, 0.95 for S, and 0.89 for U. This study demonstrated that the classifier can accurately classify supplement usage status, which can be further integrated as a module into the existing natural language processing pipeline for supporting dietary supplement knowledge discovery.