Zhecheng Sheng, Xiruo Ding, Brian Hur, Changye Li, Trevor Cohen, Serguei V. S. Pakhomov. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

OBJECTIVE: Semantic coherence in speech is characterized by a logical, connected flow of ideas. A lack of coherence in speech may reflect disorganized thinking, a core feature of psychosis in schizophrenia spectrum disorders (SSDs). Developing tools that could help with automated assessment of semantic coherence in language could facilitate early detection of SSDs and improved monitoring of symptoms, enabling more timely intervention. Large language models (LLMs) have demonstrated strong capabilities on numerous language-centric tasks and have shown promise for analyzing semantic coherence due to the natural fit between their innate measures of language perplexity and the surprising turns that incoherent narrative often takes. This study aims to develop a novel representation and associated measure of semantic coherence using LLM-based perplexity metrics and to compare this measure with traditional vector distance-based coherence metrics. METHOD: We evaluated "bag" and "chain" models based on LLM perplexities as measures of semantic coherence. Regression models were trained using both single and paired combinations of perplexity- and proximity-based features to predict human ratings of semantic coherence using standardized instruments. Performance was evaluated on held-out examples from a training set of speeches from individuals experiencing psychotic symptoms and a test set of clinical interviews with patients diagnosed with SSDs, both with labels from human assessments of disorganized thinking severity. RESULTS: The best performance was achieved using a combination of perplexity and proximity features, yielding a Spearman correlation with human ratings of 0.61 (vs. 0.56 with proximity features alone) on leave-one-out cross-validation in the training set, and 0.54 (vs. 0.52 with proximity features alone) on the test set. CONCLUSION: We developed novel methods for assessing semantic coherence using LLM perplexities and found them complementary to proximity-based methods. Combined, these methods showed improved performance across two datasets, highlighting LLM's potential in enhancing automated diagnosis and monitoring of SSDs.

Deep learning (DL) techniques involving finetuning large numbers of model parameters have delivered impressive performance on the task of discriminating between language produced by cognitively healthy individuals, and those with Alzheimer's disease (AD).However, questions remain about their ability to generalize beyond the small reference sets that are publicly available for research.As an alternative to fitting model parameters directly, we propose a novel method by which a Transformer DL model (GPT-2) pre-trained on general English text is paired with an artificially degraded version of itself (GPT-D), to compute the ratio between these two models' perplexities on language from cognitively healthy and impaired individuals.This technique approaches state-ofthe-art performance on text data from a widely used "Cookie Theft" picture description task, and unlike established alternatives also generalizes well to spontaneous conversations.Furthermore, GPT-D generates text with characteristics known to be associated with AD, demonstrating the induction of dementia-related linguistic anomalies.Our study is a step toward better understanding of the relationships between the inner workings of generative neural language models, the language that they produce, and the deleterious effects of dementia on human speech and language characteristics.

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: : Developing clinical natural language processing systems often requires access to many clinical documents, which are not widely available to the public due to privacy and security concerns. To address this challenge, we propose to develop methods to generate synthetic clinical notes and evaluate their utility in real clinical natural language processing tasks. MATERIALS AND METHODS: : We implemented 4 state-of-the-art text generation models, namely CharRNN, SegGAN, GPT-2, and CTRL, to generate clinical text for the History and Present Illness section. We then manually annotated clinical entities for randomly selected 500 History and Present Illness notes generated from the best-performing algorithm. To compare the utility of natural and synthetic corpora, we trained named entity recognition (NER) models from all 3 corpora and evaluated their performance on 2 independent natural corpora. RESULTS: : Our evaluation shows GPT-2 achieved the best BLEU (bilingual evaluation understudy) score (with a BLEU-2 of 0.92). NER models trained on synthetic corpus generated by GPT-2 showed slightly better performance on 2 independent corpora: strict F1 scores of 0.709 and 0.748, respectively, when compared with the NER models trained on natural corpus (F1 scores of 0.706 and 0.737, respectively), indicating the good utility of synthetic corpora in clinical NER model development. In addition, we also demonstrated that an augmented method that combines both natural and synthetic corpora achieved better performance than that uses the natural corpus only. CONCLUSIONS: : Recent advances in text generation have made it possible to generate synthetic clinical notes that could be useful for training NER models for information extraction from natural clinical notes, thus lowering the privacy concern and increasing data availability. Further investigation is needed to apply this technology to practice.

In recent years there has been a burgeoning interest in the use of computational methods to distinguish between elicited speech samples produced by patients with dementia, and those from healthy controls.The difference between perplexity estimates from two neural language models (LMs) -one trained on transcripts of speech produced by healthy participants and the other trained on transcripts from patients with dementia -as a single feature for diagnostic classification of unseen transcripts has been shown to produce state-of-the-art performance.However, little is known about why this approach is effective, and on account of the lack of case/control matching in the most widely-used evaluation set of transcripts (De-mentiaBank), it is unclear if these approaches are truly diagnostic, or are sensitive to other variables.In this paper, we interrogate neural LMs trained on participants with and without dementia using synthetic narratives previously developed to simulate progressive semantic dementia by manipulating lexical frequency.We find that perplexity of neural LMs is strongly and differentially associated with lexical frequency, and that a mixture model resulting from interpolating control and dementia LMs improves upon the current state-of-the-art for models trained on transcript text exclusively.

Large observational data networks that leverage routine clinical practice data in electronic health records (EHRs) are critical resources for research on coronavirus disease 2019 (COVID-19). Data normalization is a key challenge for the secondary use of EHRs for COVID-19 research across institutions. In this study, we addressed the challenge of automating the normalization of COVID-19 diagnostic tests, which are critical data elements, but for which controlled terminology terms were published after clinical implementation. We developed a simple but effective rule-based tool called COVID-19 TestNorm to automatically normalize local COVID-19 testing names to standard LOINC (Logical Observation Identifiers Names and Codes) codes. COVID-19 TestNorm was developed and evaluated using 568 test names collected from 8 healthcare systems. Our results show that it could achieve an accuracy of 97.4% on an independent test set. COVID-19 TestNorm is available as an open-source package for developers and as an online Web application for end users (https://clamp.uth.edu/covid/loinc.php). We believe that it will be a useful tool to support secondary use of EHRs for research on COVID-19.

Existing general clinical natural language processing (NLP) systems such as MetaMap and Clinical Text Analysis and Knowledge Extraction System have been successfully applied to information extraction from clinical text. However, end users often have to customize existing systems for their individual tasks, which can require substantial NLP skills. Here we present CLAMP (Clinical Language Annotation, Modeling, and Processing), a newly developed clinical NLP toolkit that provides not only state-of-the-art NLP components, but also a user-friendly graphic user interface that can help users quickly build customized NLP pipelines for their individual applications. Our evaluation shows that the CLAMP default pipeline achieved good performance on named entity recognition and concept encoding. We also demonstrate the efficiency of the CLAMP graphic user interface in building customized, high-performance NLP pipelines with 2 use cases, extracting smoking status and lab test values. CLAMP is publicly available for research use, and we believe it is a unique asset for the clinical NLP community.

MOTIVATION: Automatically quantifying semantic similarity and relatedness between clinical terms is an important aspect of text mining from electronic health records, which are increasingly recognized as valuable sources of phenotypic information for clinical genomics and bioinformatics research. A key obstacle to development of semantic relatedness measures is the limited availability of large quantities of clinical text to researchers and developers outside of major medical centers. Text from general English and biomedical literature are freely available; however, their validity as a substitute for clinical domain to represent semantics of clinical terms remains to be demonstrated. RESULTS: We constructed neural network representations of clinical terms found in a publicly available benchmark dataset manually labeled for semantic similarity and relatedness. Similarity and relatedness measures computed from text corpora in three domains (Clinical Notes, PubMed Central articles and Wikipedia) were compared using the benchmark as reference. We found that measures computed from full text of biomedical articles in PubMed Central repository (rho = 0.62 for similarity and 0.58 for relatedness) are on par with measures computed from clinical reports (rho = 0.60 for similarity and 0.57 for relatedness). We also evaluated the use of neural network based relatedness measures for query expansion in a clinical document retrieval task and a biomedical term word sense disambiguation task. We found that, with some limitations, biomedical articles may be used in lieu of clinical reports to represent the semantics of clinical terms and that distributional semantic methods are useful for clinical and biomedical natural language processing applications. AVAILABILITY AND IMPLEMENTATION: The software and reference standards used in this study to evaluate semantic similarity and relatedness measures are publicly available as detailed in the article. CONTACT: [email protected] information: Supplementary data are available at Bioinformatics online.

OBJECTIVE: In recognition of potential barriers that may inhibit the widespread adoption of biomedical software, the 2014 i2b2 Challenge introduced a special track, Track 3 - Software Usability Assessment, in order to develop a better understanding of the adoption issues that might be associated with the state-of-the-art clinical NLP systems. This paper reports the ease of adoption assessment methods we developed for this track, and the results of evaluating five clinical NLP system submissions. MATERIALS AND METHODS: A team of human evaluators performed a series of scripted adoptability test tasks with each of the participating systems. The evaluation team consisted of four "expert evaluators" with training in computer science, and eight "end user evaluators" with mixed backgrounds in medicine, nursing, pharmacy, and health informatics. We assessed how easy it is to adopt the submitted systems along the following three dimensions: communication effectiveness (i.e., how effective a system is in communicating its designed objectives to intended audience), effort required to install, and effort required to use. We used a formal software usability testing tool, TURF, to record the evaluators' interactions with the systems and 'think-aloud' data revealing their thought processes when installing and using the systems and when resolving unexpected issues. RESULTS: Overall, the ease of adoption ratings that the five systems received are unsatisfactory. Installation of some of the systems proved to be rather difficult, and some systems failed to adequately communicate their designed objectives to intended adopters. Further, the average ratings provided by the end user evaluators on ease of use and ease of interpreting output are -0.35 and -0.53, respectively, indicating that this group of users generally deemed the systems extremely difficult to work with. While the ratings provided by the expert evaluators are higher, 0.6 and 0.45, respectively, these ratings are still low indicating that they also experienced considerable struggles. DISCUSSION: The results of the Track 3 evaluation show that the adoptability of the five participating clinical NLP systems has a great margin for improvement. Remedy strategies suggested by the evaluators included (1) more detailed and operation system specific use instructions; (2) provision of more pertinent onscreen feedback for easier diagnosis of problems; (3) including screen walk-throughs in use instructions so users know what to expect and what might have gone wrong; (4) avoiding jargon and acronyms in materials intended for end users; and (5) packaging prerequisites required within software distributions so that prospective adopters of the software do not have to obtain each of the third-party components on their own.

OBJECTIVE: To create a sense inventory of abbreviations and acronyms from clinical texts. METHODS: The most frequently occurring abbreviations and acronyms from 352,267 dictated clinical notes were used to create a clinical sense inventory. Senses of each abbreviation and acronym were manually annotated from 500 random instances and lexically matched with long forms within the Unified Medical Language System (UMLS V.2011AB), Another Database of Abbreviations in Medline (ADAM), and Stedman's Dictionary, Medical Abbreviations, Acronyms & Symbols, 4th edition (Stedman's). Redundant long forms were merged after they were lexically normalized using Lexical Variant Generation (LVG). RESULTS: The clinical sense inventory was found to have skewed sense distributions, practice-specific senses, and incorrect uses. Of 440 abbreviations and acronyms analyzed in this study, 949 long forms were identified in clinical notes. This set was mapped to 17,359, 5233, and 4879 long forms in UMLS, ADAM, and Stedman's, respectively. After merging long forms, only 2.3% matched across all medical resources. The UMLS, ADAM, and Stedman's covered 5.7%, 8.4%, and 11% of the merged clinical long forms, respectively. The sense inventory of clinical abbreviations and acronyms and anonymized datasets generated from this study are available for public use at http://www.bmhi.umn.edu/ihi/research/nlpie/resources/index.htm ('Sense Inventories', website). CONCLUSIONS: Clinical sense inventories of abbreviations and acronyms created using clinical notes and medical dictionary resources demonstrate challenges with term coverage and resource integration. Further work is needed to help with standardizing abbreviations and acronyms in clinical care and biomedicine to facilitate automated processes such as text-mining and information extraction.

Family history information has emerged as an increasingly important tool for clinical care and research. While recent standards provide for structured entry of family history, many clinicians record family history data in text. The authors sought to characterize family history information within clinical documents to assess the adequacy of existing models and create a more comprehensive model for its representation. Models were evaluated on 100 documents containing 238 sentences and 410 statements relevant to family history. Most statements were of family member plus disease or of disease only. Statement coverage was 91%, 77%, and 95% for HL7 Clinical Genomics Family History Model, HL7 Clinical Statement Model, and the newly created Merged Family History Model, respectively. Negation (18%) and inexact family member specification (9.5%) occurred commonly. Overall, both HL7 models could represent most family history statements in clinical reports; however, refinements are needed to represent the full breadth of family history data.

We examine the feasibility of a machine learning approach to identification of foot examination (FE) findings from the unstructured text of clinical reports. A Support Vector Machine (SVM) based system was constructed to process the text of physical examination sections of in- and out-patient clinical notes to identify if the findings of structural, neurological, and vascular components of a FE revealed normal or abnormal findings or were not assessed. The system was tested on 145 randomly selected patients for each FE component using 10-fold cross validation. The accuracy was 80%, 87% and 88% for structural, neurological, and vascular component classifiers, respectively. Our results indicate that using machine learning to identify FE findings from clinical reports is a viable alternative to manual review and warrants further investigation. This application may improve quality and safety by providing inexpensive and scalable methodology for quality and risk factor assessments at the point of care.

OBJECTIVE: Human classification of diagnoses is a labor intensive process that consumes significant resources. Most medical practices use specially trained medical coders to categorize diagnoses for billing and research purposes. METHODS: We have developed an automated coding system designed to assign codes to clinical diagnoses. The system uses the notion of certainty to recommend subsequent processing. Codes with the highest certainty are generated by matching the diagnostic text to frequent examples in a database of 22 million manually coded entries. These code assignments are not subject to subsequent manual review. Codes at a lower certainty level are assigned by matching to previously infrequently coded examples. The least certain codes are generated by a naïve Bayes classifier. The latter two types of codes are subsequently manually reviewed. MEASUREMENTS: Standard information retrieval accuracy measurements of precision, recall and f-measure were used. Micro- and macro-averaged results were computed. RESULTS At least 48% of all EMR problem list entries at the Mayo Clinic can be automatically classified with macro-averaged 98.0% precision, 98.3% recall and an f-score of 98.2%. An additional 34% of the entries are classified with macro-averaged 90.1% precision, 95.6% recall and 93.1% f-score. The remaining 18% of the entries are classified with macro-averaged 58.5%. CONCLUSION: Over two thirds of all diagnoses are coded automatically with high accuracy. The system has been successfully implemented at the Mayo Clinic, which resulted in a reduction of staff engaged in manual coding from thirty-four coders to seven verifiers.

Text normalization is an important aspect of successful information retrieval from medical documents such as clinical notes, radiology reports and discharge summaries. In the medical domain, a significant part of the general problem of text normalization is abbreviation and acronym disambiguation. Numerous abbreviations are used routinely throughout such texts and knowing their meaning is critical to data retrieval from the document. In this paper I will demonstrate a method of automatically generating training data for Maximum Entropy (ME) modeling of abbreviations and acronyms and will show that using ME modeling is a promising technique for abbreviation and acronym normalization. I report on the results of an experiment involving training a number of ME models used to normalize abbreviations and acronyms on a sample of 10,000 rheumatology notes with ~89% accuracy.

We propose a technique for improving language modeling for automated speech recognition of medical dictations by interpolating finished text (25M words) with small human-generated literal or/and machine-generated semiliteral corpora. By building and testing interpolated (ILM) with literal (LILM), semiliteral (SILM) and partial (PILM) corpora, we show that both perplexity and recognition results improve significantly with LILM and SILM; the two yielding very close results. 1.

Filled pauses are characteristic of spontaneous speech and can present considerable problems for speech recognition by being often recognized as short words. An um can be recognized as thumb or arm if the recognizer's language model does not adequately represent FP's. Recognition of quasi-spontaneous speech (medical dictation) is subject to this problem as well. Results from medical dictations by 21 family practice physicians show that using an FP model trained on the corpus populated with FP's produces overall better result than a model trained on a corpus that excluded FP's or a corpus that had random FP's.