Free-text Descriptions Research Articles

Trans and gender diverse identities in adolescent health research: making the most of imperfect data

Advancing adolescent health research necessitates deliberate design and analysis that accurately captures the rapidly evolving world in which adolescents live and the ways in which they understand and express themselves...

Vision language models in ophthalmology.

Vision Language Models are an emerging paradigm in artificial intelligence that offers the potential to natively analyze both image and textual data simultaneously, within a single model. The fusion of these two modalities is of particular relevance to ophthalmology, which has historically involved specialized imaging techniques such as angiography, optical coherence tomography, and fundus photography, while also interfacing with electronic health records that include free text descriptions. This review then surveys the fast-evolving field of Vision Language Models as they apply to current ophthalmologic research and practice. Although models incorporating both image and text data have a long provenance in ophthalmology, effective multimodal Vision Language Models are a recent development exploiting advances in technologies such as transformer and autoencoder models. Vision Language Models offer the potential to assist and streamline the existing clinical workflow in ophthalmology, whether previsit, during, or post-visit. There are, however, also important challenges to be overcome, particularly regarding patient privacy and explainability of model recommendations.

Reversing accidents involving pedestrians: an epidemiological investigation and cluster analysis using a novel database derived from Swedish National Register data

Objectives In contemporary urban environments, the intersection between Vulnerable Road Users, specifically unprotected pedestrians, and motor vehicles poses a persistent challenge to road safety. Reversing accidents, characterized by vehicles moving backward, often result in severe consequences for pedestrians due to limited visibility and inherent blind spots for the driver. This paper aims to provide a thorough examination of reversing accidents involving pedestrians in Sweden between 2000–2021; shedding light on the magnitude of the problem, contributing factors, consequences, and potential mitigation strategies. Method Using a national database, reversing accidents were compiled by classifying free text descriptions of traffic accidents reported to the Swedish Traffic Accident Data Acquisition. The database was then used to describe reversing accidents and their consequences for struck pedestrians using descriptive epidemiology and cluster analysis. Results The results show that reversing accidents accounted for 12% of all pedestrians injured in a collision with a motor vehicle in Sweden during the studied period. In terms of personal characteristics, the struck pedestrians were more often female and of old or young age, whilst the drivers of the reversing vehicles were more often between 18–54 years and men. Most accidents were non-fatal, with only roughly 2% resulting in fatalities. Through a cluster analysis, four distinct accident types were identified. Two of these were identified as particularly important to prioritize in future preventative work: accidents that are characterized by occur during daylight hours in the urban environment (often low speeds) and often result in fatality or serious injury. Conclusions Due to the low speeds, reversing accidents are rarely fatal. However, they account for a large proportion of accidents with pedestrians and can be grouped into clearly distinguishable accident types that can function as templates in road safety development in Sweden.

Insights into decision‐maker's perceptions of good versus bad decisions in emergency services—A modified Delphi study

AbstractDecision‐making in emergency situations, such as those faced by fire, police, and health service personnel, presents unique challenges due to the high‐stakes and time‐pressured environment. Here we aim to better understand what emergency responders regard as constituting ‘good’ and ‘bad’ decisions in emergency situations. We administered a modified Delphi study, eliciting opinions from decision‐makers across all these sectors towards consensus around the key elements of good and bad decision‐making. Participants were first asked to define what makes a ‘good’ and a ‘bad’ decision, and subsequently to identify the top five most important elements of each. While consensus was not found, important insights were identified that can assist improve the standard of decision making at all levels of emergency response and management. We observed (i) a lack of a common understanding between participants of what a decision is, and how a decision differs from pre‐decision and postdecision components; (ii) responses varied according to whether a free text description or the identification of separate elements was requested; (iii) respondents valued ‘goodness’ across different and at times unrelated components a scaled measure of decision quality as opposed to a binary evaluation of ‘rightness’; and (iv) pre‐ and postdecision elements are considered more important than the decision itself when determining the quality of a decision. To address the issues highlighted by the study we recommend improvements in training, and improvements to organization doctrine related to decision making, risk tolerance, assessment of decision quality and development of intent‐based or principle‐based operational guidance. Ultimately, incident controllers from all services must remain cognizant that they will face considerable scrutiny if they cannot explain how they arrived at the decision they made.

SemOntoMap: A Hybrid Approach for Semantic Annotation of Clinical Texts.

This study addresses the challenge of leveraging free-text descriptions in Electronic Health Records (EHR) for clinical research and healthcare improvement. Despite the potential of this data, its direct interpretation by computers is limited. Semantic annotation emerges as a method to make EHR free text machine-interpretable but struggles with specific domain ontologies and faces heightened difficulties in psychiatry. To tackle these challenges, this study proposes a system based on unsupervised learning techniques to extract entities and their relationships, aligning them with a domain ontology. The effectiveness of this system has been validated within PsyCARE project by analyzing 60 patient discharge summaries.

The Safety Implications of Information Technology in Nursing: Japanese Incident Data Analysis.

This study delves into the impact of Information Technology (IT) on nursing practice in Japan, focusing on patient safety within the 2021-2022 Japanese Medical Accident Report Data. The research aims to understand how IT factors contribute to nursing-related medical incidents in a healthcare landscape rapidly integrating IT. The study identifies IT-related incidents through a retrospective analysis of medical incident reports, primarily in nursing, by analyzing categorized data and free-text descriptions for IT-related keywords. The findings indicate significant IT-related issues, with 'Other EHR Related' problems (36%) and 'EHR Reporting' errors (25%) being the most prevalent. These incidents often involve challenges in patient identification and medication management. The study suggests improvements like enhanced verification processes and automated systems to mitigate these risks. Conclusively, it underscores the dual nature of IT in nursing: while it holds the potential to enhance patient care, it also introduces challenges that necessitate specialized informatics expertise to ensure its beneficial integration into nursing practices.

Decoding the language of first impressions: Comparing models of first impressions of faces derived from free-text descriptions and trait ratings.

First impressions formed from facial appearance predict important social outcomes. Existing models of these impressions indicate they are underpinned by dimensions of Valence and Dominance, and are typically derived by applying data reduction methods to explicit ratings of faces for a range of traits. However, this approach is potentially problematic because the trait ratings may not fully capture the dimensions on which people spontaneously assess faces. Here, we used natural language processing to extract 'topics' directly from participants' free-text descriptions (i.e., their first impressions) of 2222 face images. Two topics emerged, reflecting first impressions related to positive emotional valence and warmth (Topic 1) and negative emotional valence and potential threat (Topic 2). Next, we investigated how these topics were related to Valence and Dominance components derived from explicit trait ratings. Collectively, these components explained only ~44% of the variance in the topics extracted from free-text descriptions and suggested that first impressions are underpinned by correlated valence dimensions that subsume the content of existing trait-rating-based models. Natural language offers a promising new avenue for understanding social cognition, and future work can examine the predictive utility of natural language and traditional data-driven models for impressions in varying social contexts.

#1657 Impact of cumulative fluid and sodium imbalance exposure on cause-specific mortality in hemodialysis

Abstract Background and Aims Even mild fluid overload and prolonged exposure to salt excess showed to be associated with all-cause mortality in hemodialysis patients. This analysis aims to estimate the effect of fluid and sodium imbalance on cause-specific mortality. Method The study followed 68,196 incident hemodialysis patients from 875 clinics in 25 countries for a decade (2010-2020) using the European Clinical Database 5. Fatal events (21,644) were classified through ICD10- codes and free text descriptions into different endpoints: cardiac death (7,926 events)—with subcomponents sudden cardiac death (2,815 events), coronary death (1,705 events) and heart failure death (3,248 events); cerebrovascular death (1,516 events), non-cardiovascular death (10,898 events)—with subcomponents death by sepsis (2,054) or other infection (1,217 events), cancer death (2,095), other non-cardiovascular death (5,532 events) and death of unknown origin (1,304 events). Time-varying cumulative exposure times (in months) of relative fluid overload/-depletion, hypo-/hypernatremia and low dialysate sodium (≤ 138) were assessed and hazard ratios (HR) were calculated using a multivariate Cox model. Results Fluid overload showed the most consistent exposure risk associations for cardiovascular endpoints with strongest associations in cardiac causes (arrythmia death: HR peak at 7.5 (95% CI: 4.2-3.6), heart failure death: HR peak at 4.2 (95% CI: 3.3-5.4) and sudden cardiac death: HR peak at 3.8 (95% CI: 2.6-4.5)). Fluid depletion was less important but consistently associated with all causes of death. By contrast, Hyponatremia showed heterogeneity of cause specific effects. We found slightly stronger associations for non- cardiovascular causes: Death by sepsis (HR peak at 2.3 (95% CI: 1.8-2.8)) or death by other infections (HR peak at 2.1 (95% CI: 1.6-2.7)) showed the strongest cumulative exposure risk association pattern. Hypernatremia was less important with substantially smaller effects and slightly higher associations for non-cardiovascular endpoints. For low dialysate sodium, we observed evidence of heterogeneity in cause specific effects. The risk of death due to heart failure was notably pronounced with low dialysate sodium (HR peak at 5.8 (95% CI: 1.9-17.9)), while other endpoints did not show this effect. Sensitivity analysis in patients with 48 or more months of follow-up indicated that HR in higher exposure categories might be biased downwards due to non-survival of high-risk individuals. Conclusion Cause-specific hazard models confirm the role of fluid and sodium imbalance as independently acting risk factors for mortality. Fluid overload has the most pronounced effect on cardiac death, especially death by heart failure, arrhythmia, or sudden death. In contrast, hyponatremia is a more critical risk factor for non-cardiovascular death, particularly death caused by sepsis or other infections. The significantly increased risk in patients on low dialysate sodium prescriptions is primarily driven by a dramatic increase in risk of death by heart failure.

Oil and Gas Extraction Industry Workers' Compensation Claims and Proposed Safety Interventions.

This study analyzed Ohio workers' compensation data to identify potential prevention strategies for common oil and gas extraction industry claims. Claim rates for 2001-2018 were calculated per full-time equivalent employee. Descriptive analyses on free-text descriptions of lost-time (LT) claims (>7 days away from work) identified common characteristics among claims and injured workers. Among 3134 claims, 860 (27%) were LT. The industry group, drilling contractors, experienced the highest LT claims rate, whereas the cost from servicing contractors was the highest. Contact with objects and equipment caused the highest LT claims rate. The most frequent LT occupation was roustabout, and the most frequent LT work activity was material handling. Transportation incidents caused most fatalities and hospitalizations. Over half of LT claims were from short-tenured workers. Both proven and innovative approaches are needed to reduce severe workers' compensation claims in this industry.

Sexist, Racist, and Homophobic Violence against Paramedics in a Single Canadian Site.

Violence against paramedics is widely recognized as a serious, but underreported, problem. While injurious physical attacks on paramedics are generally reported, non-physical violence is less likely to be documented. Verbal abuse can be very distressing, particularly if the harassment targets personal or cultural identities, such as race, ethnicity, gender, or sexual orientation. Leveraging a novel, point-of-event reporting process, our objective was to estimate the prevalence of harassment on identity grounds against paramedics in a single paramedic service in Ontario, Canada, and assess its potentially differential impact on emotional distress. In an analysis of 502 reports filed between 1 February 2021 and 28 February 2022, two paramedic supervisors independently coded the free-text narrative descriptions of violent encounters for themes suggestive of sexism, racism, and homophobia. We achieved high inter-rater agreement across the dimensions (k = 0.73-0.83), and after resolving discrepant cases, we found that one in four violent reports documented abuse on at least one of the identity grounds. In these cases, paramedics were 60% more likely to indicate being emotionally distressed than for other forms of violence. Our findings offer unique insight into the type of vitriol paramedics experience over the course of their work and its potential for psychological harm.

Development of a scoring system to quantify errors from semantic characteristics in incident reports

ObjectivesIncident reporting systems are widely used to identify risks and enable organisational learning. Free-text descriptions contain important information about factors associated with incidents. This study aimed to develop error scores...

Identification of mycoplasma pneumonia in children based on fusion of multi-modal clinical free-text description and structured test data.

Mycoplasma pneumonia may lead to hospitalizations and pose life-threatening risks in children. The automated identification of mycoplasma pneumonia from electronic medical records holds significant potential for improving the efficiency of hospital resource allocation. In this study, we proposed a novel method for identifying mycoplasma pneumonia by integrating multi-modal features derived from both free-text descriptions and structured test data in electronic medical records. Our approach begins with the extraction of free-text and structured data from clinical records through a systematic preprocessing pipeline. Subsequently, we employ a pre-trained transformer language model to extract features from the free-text, while multiple additive regression trees are used to transform features from the structured data. An attention-based fusion mechanism is then applied to integrate these multi-modal features for effective classification. We validated our method using clinic records of 7157 patients, retrospectively collected for training and testing purposes. The experimental results demonstrate that our proposed multi-modal fusion approach achieves significant improvements over other methods across four key performance metrics.

Accurate, Robust, and Scalable Machine Abstraction of Mayo Endoscopic Subscores From Colonoscopy Reports.

The Mayo endoscopic subscore (MES) is an important quantitative measure of disease activity in ulcerative colitis. Colonoscopy reports in routine clinical care usually characterize ulcerative colitis disease activity using free text description, limiting their utility for clinical research and quality improvement. We sought to develop algorithms to classify colonoscopy reports according to their MES. We annotated 500 colonoscopy reports from 2 health systems. We trained and evaluated 4 classes of algorithms. Our primary outcome was accuracy in identifying scorable reports (binary) and assigning an MES (ordinal). Secondary outcomes included learning efficiency, generalizability, and fairness. Automated machine learning models achieved 98% and 97% accuracy on the binary and ordinal prediction tasks, outperforming other models. Binary models trained on the University of California, San Francisco data alone maintained accuracy (96%) on validation data from Zuckerberg San Francisco General. When using 80% of the training data, models remained accurate for the binary task (97% [n = 320]) but lost accuracy on the ordinal task (67% [n = 194]). We found no evidence of bias by gender (P = .65) or area deprivation index (P = .80). We derived a highly accurate pair of models capable of classifying reports by their MES and recognizing when to abstain from prediction. Our models were generalizable on outside institution validation. There was no evidence of algorithmic bias. Our methods have the potential to enable retrospective studies of treatment effectiveness, prospective identification of patients meeting study criteria, and quality improvement efforts in inflammatory bowel diseases.

Bootstrapping Large Language Models for Radiology Report Generation

Radiology report generation (RRG) aims to automatically generate a free-text description from a specific clinical radiograph, e.g., chest X-Ray images. Existing approaches tend to perform RRG with specific models trained on the public yet limited data from scratch, where they often lead to inferior performance owing to the problem of inefficient capabilities in both aligning visual and textual features and generating informative reports accordingly. Currently, large language models (LLMs) offered a promising solution to text generation with their power in learning from big data, especially for cross-modal scenarios such as RRG. However, most existing LLMs are pre-trained on general data, and suffer from the same problem of conventional approaches caused by knowledge gap between general and medical domain if they are applied to RRG. Therefore in this paper, we propose an approach to bootstrapping LLMs for RRG with a in-domain instance induction and a coarse-to-fine decoding process. Specifically, the in-domain instance induction process learns to align the LLM to radiology reports from general texts through contrastive learning. The coarse-to-fine decoding performs a text elevating process for those reports from the ranker, further enhanced with visual features and refinement prompts. Experimental results on two prevailing RRG datasets, namely, IU X-Ray and MIMIC-CXR, demonstrate the superiority of our approach to previous state-of-the-art solutions. Further analyses illustrate that, for the LLM, the induction process enables it to better align with the medical domain and the coarse-to-fine generation allows it to conduct more precise text generation.

GO2Sum: generating human-readable functional summary of proteins from GO terms

Understanding the biological functions of proteins is of fundamental importance in modern biology. To represent a function of proteins, Gene Ontology (GO), a controlled vocabulary, is frequently used, because it is easy to handle by computer programs avoiding open-ended text interpretation. Particularly, the majority of current protein function prediction methods rely on GO terms. However, the extensive list of GO terms that describe a protein function can pose challenges for biologists when it comes to interpretation. In response to this issue, we developed GO2Sum (Gene Ontology terms Summarizer), a model that takes a set of GO terms as input and generates a human-readable summary using the T5 large language model. GO2Sum was developed by fine-tuning T5 on GO term assignments and free-text function descriptions for UniProt entries, enabling it to recreate function descriptions by concatenating GO term descriptions. Our results demonstrated that GO2Sum significantly outperforms the original T5 model that was trained on the entire web corpus in generating Function, Subunit Structure, and Pathway paragraphs for UniProt entries.

Senses, emotions and wellbeing in forest recreation and tourism

ABSTRACT Outdoor tourism can provide nature therapies. Implementation research for mental healthcare aims to establish psychological mechanisms and specify designs, doses and durations for medical trials. As one component, we tested simultaneous links from senses to emotions, and emotions to self-perceived wellbeing, for 524 tourists at 27 forest reserves in China. We used quantitative surveys, in person, with: 5-point Likert-type scales to measure links; free-text descriptions of their most memorable sensory and emotional experiences during these forest visits; and 7-item Personal Wellbeing Index (PWI) and 10-item Kessler stress index (K10) to measure mental health and wellbeing. We found that Chinese forest tourists perceive strongly (p < 0.00001) that sensory experiences drive emotions, and emotions determine wellbeing. Senses reported were: sight by 86%, sound by 36%, smell by 41%, touch by 18%, temperature by 42%, and humidity by 20%. The number of senses engaged during forest experiences ranged from 1, for 39%, to 6, for 9%. The K10 stress-reduction measure of wellbeing was significantly (p = 0.0272) but weakly (R 2 = 0.009) correlated with the number of senses engaged. Future research on mechanisms for mental health benefits to outdoor tourists, and tourist choreography by tour guides and trail signs, should therefore focus on intensity of senses and emotions.

Application of natural language processing for spill reduction in an exploration and production company

Spill reduction is a key focus for exploration and production (E&P) companies to protect people, the environment, and to maintain oil production. Learning from previous incidents enable companies to identify ways to lower the severity and frequency of spills. Root cause failure analysis (RCFA) is a key element in process safety management, and this is often focused on high severity incidents because of the time required to properly conduct the analysis. Incidents that have lower severity occur more frequently, but typically less time and resources are spent analyzing these events. Learning from lower severity, high frequency events can be useful in improving process safety programs and reduce spill frequency and impacts because the data set is more representative of the big picture. Natural language processing (NLP) is applied to identify causes and contributing factors of spills and leaks associated with E&P activity to reduce the frequency and severity of these events. Reviewing incident reports allow us to identify trends and areas of improvement, and NLP can be used to process the data quickly and improve data quality. Incident reports are typically free text that includes a description and cause of the event. One challenge with free-text descriptions is that further analysis and interpretation is needed in order to effectively analyze, trend, and learn from the available data. In addition, identifying the causes and contributing factors of a spill accurately can be time intensive. Two natural language processing methods, rule-based entity extraction and a machine learning model approach have been utilized to analyze a greater number of incident reports, which enables us to pinpoint causes of leaks, reduce the frequency of spills, and mitigate the impact of these events. This paper presents the challenges associated with learning from incident reports, particularly spills, and offers solutions that natural language processing can provide to improve data quality and contribute to spill reduction efforts in the context of an exploration and production company.

Vision-Language Model for Generating Textual Descriptions From Clinical Images: Model Development and Validation Study.

The automatic generation of radiology reports, which seeks to create a free-text description from a clinical radiograph, is emerging as a pivotal intersection between clinical medicine and artificial intelligence. Leveraging natural language processing technologies can accelerate report creation, enhancing health care quality and standardization. However, most existing studies have not yet fully tapped into the combined potential of advanced language and vision models. The purpose of this study was to explore the integration of pretrained vision-language models into radiology report generation. This would enable the vision-language model to automatically convert clinical images into high-quality textual reports. In our research, we introduced a radiology report generation model named ClinicalBLIP, building upon the foundational InstructBLIP model and refining it using clinical image-to-text data sets. A multistage fine-tuning approach via low-rank adaptation was proposed to deepen the semantic comprehension of the visual encoder and the large language model for clinical imagery. Furthermore, prior knowledge was integrated through prompt learning to enhance the precision of the reports generated. Experiments were conducted on both the IU X-RAY and MIMIC-CXR data sets, with ClinicalBLIP compared to several leading methods. Experimental results revealed that ClinicalBLIP obtained superior scores of 0.570/0.365 and 0.534/0.313 on the IU X-RAY/MIMIC-CXR test sets for the Metric for Evaluation of Translation with Explicit Ordering (METEOR) and the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) evaluations, respectively. This performance notably surpasses that of existing state-of-the-art methods. Further evaluations confirmed the effectiveness of the multistage fine-tuning and the integration of prior information, leading to substantial improvements. The proposed ClinicalBLIP model demonstrated robustness and effectiveness in enhancing clinical radiology report generation, suggesting significant promise for real-world clinical applications.

Exploratory evaluation of spinal cord stimulation with dynamic pulse patterns: a promising approach to improve stimulation sensation, coverage of pain areas, and expected pain relief.

The societal burden of chronic pain and the contribution-in-part to the opioid crisis, is a strong motivation to improve and expand non-addictive treatments, including spinal cord stimulation (SCS). For several decades standard SCS has consisted in delivery of tonic pulses with static parameter settings in frequency, pulse width, and amplitude. These static parameters have limited ability to personalize the quality of paresthesia, the dermatomal coverage, and thus may affect SCS efficacy. Further, static settings may contribute to the build-up of tolerance or loss of efficacy of the therapy over time in some patients. We conducted an acute exploratory study to evaluate the effects of SCS using time-dynamic pulses as compared to time-static (conventional tonic) stimulation pulses, with the hypotheses that dynamic pulse SCS may enable beneficial tailoring of the sensation and the patient's expectation for better pain relief with SCS. During a single clinic visit, consented subjects undergoing a standard SCS trial had their implanted leads temporarily connected to an investigational external stimulator capable of delivering time-static and six categories of time-dynamic pulse sequences, each characterized by continuously varying a stimulation parameter. Study subjects provided several assessments while blinded to the stimulation pattern, including: drawing of paresthesia maps, descriptions of sensation, and ratings for comfort and helpfulness to pain relief. Even without optimization of the field location, a majority of subjects rated sensations from dynamic stimulation as better or equal to that of static stimulation for comfortableness and for helpfulness to pain relief. The initial data showed a gender and/or pain dermatomal location related preference to a stimulation pattern. In particular, female subjects and subjects with pain at higher dermatomes tended to rank the sensation from dynamic stimulation better. Dynamic stimulation produced greater pain coverage without optimization; in 70% (9/13) of subjects, maximal pain coverage was achieved with a dynamic stimulation pattern. There was also greater variety in the words used by patients to describe stimulation sensation in the free text and free form verbal descriptions associated with dynamic stimulation. With the same electrode configuration and comparable parameter settings, acute SCS using dynamic pulses produced more positive ratings, expanded paresthesia coverage, and greater variation in sensation as compared to SCS using static pulses, suggesting that dynamic stimulation has the potential to improve capabilities of SCS for the treatment of chronic pain. Further study is warranted. This study was registered at ClinicalTrials.gov under ID NCT02988713, November 2016 (URL: https://clinicaltrials.gov/ct2/show/NCT02988713).

Anomaly detection in trade declarations using deep learning techniques: A risk-assessment approach to identify misclassification and incorrect valuation

In Hong Kong, merchandise trade statistics are compiled based on the commodity information given on the trade declarations submitted by traders. Due to the complexity of the standardised commodity classification system (i.e. Hong Kong Harmonized System, or HKHS in short), there are often reporting errors, especially in the commodity codes and quantities. With around 20 million declarations received annually, the availability of this big data source motivates us to adopt deep learning techniques to detect the reporting errors. This paper proposes a mechanism consisting of three deep learning models for checking the commodity code, quantity and value, which offers an end-to-end solution to data quality assurance for declarations. The results show that the proposed mechanism could enhance the accuracy of error detection, which is conducive to improving the quality of trade statistics. With the use of text analytics techniques, the mechanism could fully utilise free-text commodity descriptions declared by traders to check the accuracy of the declared information comprehensively. It also overcomes some limitations of the traditional rule-based models. The whole study demonstrates the potential of using deep learning approach in quality assurance of existing statistical systems for official statistics.