Data Curation Research Articles

Large language models outperform traditional natural language processing methods in extracting patient-reported outcomes in IBD.

Patient-reported outcomes (PROs) are vital in assessing disease activity and treatment outcomes in inflammatory bowel disease (IBD). However, manual extraction of these PROs from the free-text of clinical notes is burdensome. We aimed to improve data curation from free-text information in the electronic health record, making it more available for research and quality improvement. This study aimed to compare traditional natural language processing (tNLP) and large language models (LLMs) in extracting three IBD PROs (abdominal pain, diarrhea, fecal blood) from clinical notes across two institutions. Clinic notes were annotated for each PRO using preset protocols. Models were developed and internally tested at the University of California San Francisco (UCSF), and then externally validated at Stanford University. We compared tNLP and LLM-based models on accuracy, sensitivity, specificity, positive and negative predictive value. Additionally, we conducted fairness and error assessments. Inter-rater reliability between annotators was >90%. On the UCSF test set (n=50), the top-performing tNLP models showcased accuracies of 92% (abdominal pain), 82% (diarrhea) and 80% (fecal blood), comparable to GPT-4, which was 96%, 88%, and 90% accurate, respectively. On external validation at Stanford (n=250), tNLP models failed to generalize (61-62% accuracy) while GPT-4 maintained accuracies >90%. PaLM-2 and GPT-4 showed similar performance. No biases were detected based on demographics or diagnosis. LLMs are accurate and generalizable methods for extracting PROs. They maintain excellent accuracy across institutions, despite heterogeneity in note templates and authors. Widespread adoption of such tools has the potential to enhance IBD research and patient care.

Cetacean passive acoustic network in the Belgian part of the North sea

In 2016, a design for detecting harbor porpoises (Phocoena phocoena) in the Belgian part of the North Sea (BPNS) was developed in the framework of the LifeWatch project using Passive Acoustic Monitoring (PAM) loggers. In 2018, the network of deployments fully matured, and eight locations across the BPNS are presently housed with a C-POD (Chelonia Ltd., UK), a PAM logger moored on the seafloor using a multi-use platform. The PAM logger continuously listens for harbor porpoise click trains, but only stores the click parameters. This paper presents the generated data series of harbor porpoise click trains at a minute-resolution and the details on the individual click trains. The field recordings, methodology and processing are described, along with its data curation, integration and quality control. The data are annually published online in a standardized format, accompanied with the metadata description, and labelled with a digital object identifier for traceability. Data are published under a CC-BY license, allowing the use of data under the condition of providing reference to the original source.

A multimodal neural network with gradient blending improves predictions of survival and metastasis in sarcoma

The objective of this study is to develop a multimodal neural network (MMNN) model that analyzes clinical variables and MRI images of a soft tissue sarcoma (STS) patient, to predict overall survival and risk of distant metastases. We compare the performance of this MMNN to models based on clinical variables alone, radiomics models, and an unimodal neural network. We include patients aged 18 or older with biopsy-proven STS who underwent primary resection between January 1st, 2005, and December 31st, 2020 with complete outcome data and a pre-treatment MRI with both a T1 post-contrast sequence and a T2 fat-sat sequence available. A total of 9380 MRI slices containing sarcomas from 287 patients are available. Our MMNN accepts the entire 3D sarcoma volume from T1 and T2 MRIs and clinical variables. Gradient blending allows the clinical and image sub-networks to optimally converge without overfitting. Heat maps were generated to visualize the salient image features. Our MMNN outperformed all other models in predicting overall survival and the risk of distant metastases. The C-Index of our MMNN for overall survival is 0.77 and the C-Index for risk of distant metastases is 0.70. The provided heat maps demonstrate areas of sarcomas deemed most salient for predictions. Our multimodal neural network with gradient blending improves predictions of overall survival and risk of distant metastases in patients with soft tissue sarcoma. Future work enabling accurate subtype-specific predictions will likely utilize similar end-to-end multimodal neural network architecture and require prospective curation of high-quality data, the inclusion of genomic data, and the involvement of multiple centers through federated learning.

CultIS : Web-based Platform for Intelligent Cultural Content Management

The paper presents the research, development, and some of the major implementations to date of the web-based platform for intelligent cultural content management CultIS. It is an innovative environment providing descriptive cataloging, subject indexing, intelligent data retrieval, curation and collection development aimed at meeting the functional needs associated with the growing expectations of users of cultural institutions such as libraries, museums and other providers of cultural and historical heritage. Based on the implementations of CultIS for different purposes and contexts, not only its effectiveness, stability and reliability are tested, but the results also serve for its constant development and improvement.

Towards artificial intelligence-based disease prediction algorithms that comprehensively leverage and continuously learn from real-world clinical tabular data systems.

This manuscript presents a proof-of-concept for a generalizable strategy, the full algorithm, designed to estimate disease risk using real-world clinical tabular data systems, such as electronic health records (EHR) or claims databases. By integrating classic statistical methods and modern artificial intelligence techniques, this strategy automates the production of a disease prediction model that comprehensively reflects the dynamics contained within the underlying data system. Specifically, the full algorithm parses through every facet of the data (e.g., encounters, diagnoses, procedures, medications, labs, chief complaints, flowsheets, vital signs, demographics, etc.), selects which factors to retain as predictor variables by evaluating the data empirically against statistical criteria, structures and formats the retained data into time-series, trains a neural network-based prediction model, then subsequently applies this model to current patients to generate risk estimates. A distinguishing feature of the proposed strategy is that it produces a self-adaptive prediction system, capable of evolving the prediction mechanism in response to changes within the data: as newly collected data expand/modify the dataset organically, the prediction mechanism automatically evolves to reflect these changes. Moreover, the full algorithm operates without the need for a-priori data curation and aims to harness all informative risk and protective factors within the real-world data. This stands in contrast to traditional approaches, which often rely on highly curated datasets and domain expertise to build static prediction models based solely on well-known risk factors. As a proof-of-concept, we codified the full algorithm and tasked it with estimating 12-month risk of initial stroke or myocardial infarction using our hospital's real-world EHR. A 66-month pseudo-prospective validation was conducted using records from 558,105 patients spanning April 2015 to September 2023, totalling 3,424,060 patient-months. Area under the receiver operating characteristic curve (AUROC) values ranged from .830 to .909, with an improving trend over time. Odds ratios describing model precision for patients 1-100 and 101-200 (when ranked by estimated risk) ranged from 15.3 to 48.1 and 7.2 to 45.0, respectively, with both groups showing improving trends over time. Findings suggest the feasibility of developing high-performing disease risk calculators in the proposed manner.

Data Management in Biobanking: Strategies, Challenges, and Future Directions.

Biobanking plays a pivotal role in biomedical research by providing standardized processing, precise storing, and management of biological sample collections along with the associated data. Effective data management is a prerequisite to ensure the integrity, quality, and accessibility of these resources. This review provides a current landscape of data management in biobanking, discussing key challenges, existing strategies, and potential future directions. We explore multiple aspects of data management, including data collection, storage, curation, sharing, and ethical considerations. By examining the evolving technologies and methodologies in biobanking, we aim to provide insights into addressing the complexities and maximizing the utility of biobank data for research and clinical applications.

FlaHMM: unistrand flamenco-like piRNA cluster prediction in Drosophila species using hidden Markov models.

PIWI-interacting RNAs (piRNAs) are a class of small non-coding RNAs that are essential for transposon control in animal gonads. In Drosophila ovarian somatic cells, piRNAs are transcribed from large genomic regions called piRNA clusters, which are enriched for transposon fragments and act as a memory of past invasions. Despite being widely present across Drosophila species, somatic piRNA clusters are difficult to identify and study due to their lack of sequence conservation and limited synteny. Current identification methods rely on either extensive manual curation or availability of high-throughput small RNA sequencing data, limiting large-scale comparative studies. We now present FlaHMM, a hidden Markov model developed to automate genomic annotation of flamenco-like unistrand piRNA clusters in Drosophila species, requiring only a genome assembly and transposon annotations. FlaHMM uses transposable element content across 5- or 10-kb bins, which can be calculated from genome sequence alone, and is thus able to detect candidate piRNA clusters without the need to obtain flies and experimentally perform small RNA sequencing. We show that FlaHMM performs on par with piRNA-guided or manual methods, and thus provides a scalable and efficient approach to piRNA cluster annotation in new genome assemblies. FlaHMM is freely available at https://github.com/Hannon-lab/FlaHMM under an MIT licence.

Empowering Research With the American Heart Association Get With The Guidelines Registries Through Integration of a Database and Research Tools.

The American Heart Association's Get With The Guidelines (GWTG) has emerged as a vital resource in advancing the standards and practices of inpatient care across stroke, heart failure, coronary artery disease, atrial fibrillation, and resuscitation focus areas. The GWTG registry data have also created new opportunities for secondary use of real-world clinical data in biomedical research. Our goal was to implement a scalable database with an integrated user interface (UI) to improve GWTG data management and accessibility. The curation of registry data begins by going through a data processing and quality control pipeline programmed in Python. This pipeline includes data cleaning and record exclusion, variable derivation and unit harmonization, limited data set preparation, and documentation generation of the registry data. The database was built using PostgreSQL, and integrations between the database and the UI were built using the Django Web Framework in Python. Smaller subsets of data were created using SQLite database files for distribution purposes. Use cases of these tools are provided in the article. We implemented an automated data curation pipeline, centralized database, and UI application for the American Heart Association GWTG registry data. The database and the UI are accessible through a Precision Medicine Platform workspace. As of March 2024, the database contains over 13.2 million cleaned GWTG patient records. The SQLite subsets benefit researchers by optimizing data extraction and manipulation using Structured Query Language. The UI improves accessibility for nontechnical researchers by presenting data in a user-friendly tabular format with intuitive filtering options. With the implementation of the GWTG database and UI application, we addressed data management and accessibility concerns despite its growing scale. We have launched tools to provide streamlined access and accessibility of GWTG registry data to all researchers, regardless of familiarity or experience in coding.

Patterns of pharmacogenetic variation in nine biogeographic groups.

Frequencies of pharmacogenetic (PGx) variants are known to differ substantially across populations but much of the available PGx literature focuses on one or a few population groups, often defined in nonstandardized ways, or on a specific gene or variant. Guidelines produced by the Clinical Pharmacogenetic Implementation Consortium (CPIC) provide consistent methods of literature extraction, curation, and reporting, including comprehensive curation of allele frequency data across nine defined "biogeographic groups" from the PGx literature. We extracted data from 23 CPIC guidelines encompassing 19 genes to compare the sizes of the populations from each group and allele frequencies of altered function alleles across groups. The European group was the largest in the curated literature for 16 of the 19 genes, while the American and Oceanian groups were the smallest. Nearly 200 alleles were detected in nonreference groups that were not reported in the largest (reference) group. The genes CYP2B6 and CYP2C9 were more likely to have higher frequencies of altered function alleles in nonreference groups compared to the reference group, while the genes CYP4F2, DPYD, SLCO1B1, and UGT1A1 were less likely to have higher frequencies in nonreference groups. PGx allele frequencies and function differ substantially across nine biogeographic groups, all but two of which are underrepresented in available PGx data. Awareness of these differences and increased efforts to characterize the breadth of global PGx variation are needed to ensure that implementation of PGx-guided drug selection does not further widen existing health disparities among populations currently underrepresented in PGx data.

GEPREP: A comprehensive data atlas of RNA-seq-based gene expression profiles of exercise responses

BackgroundPhysical activity can regulate and affect gene expression in multiple tissues and cells. Recently, with the development of next-generation sequencing, a large number of RNA-sequencing (RNA-seq)-based gene expression profiles about physical activity have been shared in public resources; however, they are poorly curated and underutilized. To tackle this problem, we developed a data atlas of such data through comprehensive data collection, curation, and organization. MethodsThe data atlas, termed gene expression profiles of RNA-seq-based exercise responses (GEPREP), was built on a comprehensive collection of high-quality RNA-seq data on exercise responses. The metadata of each sample were manually curated. Data were uniformly processed and batch effects corrected. All the information was well organized in an easy-to-use website for free search, visualization, and download. ResultsGEPREP now includes 69 RNA-seq datasets of pre- and post-exercise, comprising 26 human datasets (1120 samples) and 43 mouse datasets (1006 samples). Specifically, there were 977 (87.2 %) human samples of skeletal muscle and 143 (12.8 %) human samples of blood. There were also samples across 9 mice tissues with skeletal muscle (359, 35.7 %) and brain (280, 27.8 %) accounting for the main fractions. Metadata—including subject, exercise interventions, sampling sites, and post-processing methods—are also included. The metadata and gene expression profiles are freely accessible at http://www.geprep.org.cn/. ConclusionGEPREP is a comprehensive data atlas of RNA-seq-based gene expression profiles responding to exercise. With its reliable annotations and user-friendly interfaces, it has the potential to deepen our understanding of exercise physiology.

Impact of reduced sampling frequency of illicit drug wastewater monitoring in the Netherlands

Drug consumption estimates are traditionally based on surveys or information from police seizures. Alternatively, residues of illicit drugs in untreated wastewater (influent) can be used to calculate mass loads and subsequently estimate drug consumption in the community throughout the week. For this purpose, wastewater is commonly sampled for seven consecutive days within the Sewage analysis CORe group Europe (SCORE), while other sampling schemes may be implemented in long-term studies outside this consortium. The current study demonstrates how sampling frequency of illicit drug residues in the influent of wastewater treatment plants (WWTPs) affects the derived weekly average. Thirty WWTPs were sampled over the course of 12 years and influents were analyzed for five drugs (metabolites): 3,4-methylenedioxymethamphetamine (MDMA), methamphetamine, amphetamine, benzoylecgonine (a metabolite of cocaine), and 11-nor-9-Carboxy tetrahydrocannabinol (THC-COOH). Subsequently, small and large WWTPs were grouped with a threshold of 100,000 inhabitants. After data curation, standardized loads were calculated (mg/d per 1000 inhabitants). Weekly averages of loads of the drug residues were calculated based on six scenarios (sampling one to six weekdays) and compared to the weekly average in the control situation (sampling seven weekdays) in a Monte Carlo simulation. Results indicate that drug residues with more dynamic loads over a week require more frequent sampling. The analysis illustrates that a decreased sampling frequency (4 or 5 days per week) still leads to a representative weekly average for all drugs tested when a deviation up to a factor of 1.25 is deemed acceptable. However, knowledge on typical levels is necessary to define outliers. We therefore recommend to study dynamics in drug residue loads for WWTPs before reducing sampling frequency in long term monitoring programs.

Going global by going local: Impacts and opportunities of geographically focused data integration

Abstract Biodiversity conservation is a global challenge that requires the integration of global and local data. Expanding global data infrastructures have opened unprecedented opportunities for biodiversity data storage, curation, and dissemination. Within one such infrastructure—the Global Biodiversity Information Facility (GBIF)—these benefits are achieved by aggregating data from over 100 regional infrastructure nodes. Such, regional biodiversity infrastructures benefit scientific communities in ways that exceed their core function of contributing to global data aggregation, but these additional scientific impacts are rarely quantified. To fill this gap, we characterize the scientific impact of the Atlas of Living Australia, one of the oldest and largest GBIF nodes, as a case study of a regional biodiversity information facility. Our discussion reveals the multifaceted impact of the regional biodiversity data infrastructure. We showcase the global importance of such infrastructures, data sets, and collaborations.

Physical Health Checks and Follow-Up Care in Deprived and Ethnically Diverse People With Severe Mental Illness: Co-Designed Recommendations for Better Care.

There is wide variation in premature mortality rates in adults with severe mental illness (SMI) across London, with Tower Hamlets (a highly deprived and ethnically diverse area) scoring the highest. To identify examples of best practice and co-design recommendations for improving physical health checks and follow-up care amongst people with SMI in Tower Hamlets. Data were collected through online questionnaires (using SMI physical health best practice checklists), one-on-one interviews (n = 7) and focus groups (n = 3) with general practices, secondary mental health services, commissioners and leads of community services and public health programmes, experts by experience and community, voluntary and social enterprise organisations in Tower Hamlets. Data were analysed using deductive and inductive thematic analysis. Twenty-two participants representing 15 general practices (out of 32), secondary mental health services, commissioners and public health leads completed the online questionnaires. Twenty-one participants took part in interviews and focus groups. Examples of best practice included cleaning and validating the SMI register regularly by general practices, knowing the number of patients who had been offered and/or received physical health checks, having clear pathways to community and specialist care services, using various communication methodsand having a key performance indicator (KPI) for tailored smoking cessation services for people with SMI. Recommendations included adopting evidence-informed frameworks for risk stratification and utilising the wider primary care workforce with specific training to follow up on results, offer interventionsand support navigating pathways and taking up follow-up care. Incentivising schemes were needed to deliver additional physical health check components such as oral health, cancer screening, Covid-19 vaccination and sexual health checks. Including KPIs in other community services' specifications with reference to SMI people was warranted. Further engagement with experts by experience and staff training were needed. The present initiative identified best practice examples and co-designed recommendations for improving physical health checks and follow-up care in deprived and ethnically diverse people with SMI. This initiative was supported by three experts with experience, and two community organisations, who were involved in data curation and interpretation, development of recommendationsand/or dissemination activities including writing this manuscript.

Perspective: Better Privacy Theorists for Better Data Stewards

The U.S. Census Bureau's use of differential privacy (DP) fundamentally changed how academic DP researchers perform outreach with official statistics stakeholders. In this perspectives piece, I propose ways for us in this community to improve those processes by being more receptive to the practical concerns raised by building DP systems. First, I discuss how academic DP work fundamentally differs from the policy decisions needed to implement DP systems and why this distinction has political consequences. Through examples and discussions from workshops, I show how the DP community largely asked applied stakeholders to communicate on DP's theoretical terms, when such an ask foreclosed important considerations relevant for the Census Bureau's policy problems. Second, I discuss how existing polarization between theoretical and empirical privacy researchers unintentionally seeped into the ways we communicated about DP, pointing to why both perspectives are necessary in different ways for policy conversations. Finally, I conclude by discussing how these issues are not unique to data privacy work, but instead reflect structural problems in translating theoretical science into practice. These ideas are presented in service of a single goal: to ensure DP theory supports substantive, privacy-aware data processing and dissemination in practice for essential data curators.

Clinical and biobehavioral phenotypic assessments and data harmonization for the RE-JOIN research consortium: Recommendations for common data element selection

BackgroundThe Restoring Joint Health and Function to Reduce Pain (RE-JOIN) Consortium is part of the Helping to End Addiction Long-term® (HEAL) Initiative. HEAL is an ambitious, NIH-wide initiative to speed scientific solutions to stem the national opioid public health crisis. The RE-JOIN consortium’s over-arching goal is to define how chronic joint pain-mediating neurons innervate different articular and peri-articular tissues, with a focus on the knee and temporomandibular joints (TMJ) across species employing the latest neuroscience approaches. The aim of this manuscript is to elucidate the human data gathered by the RE-JOIN consortium, as well as to expound upon its underlying rationale and the methodologies and protocols for harmonization and standardization that have been instituted by the RE-JOIN Consortium. MethodsThe consortium-wide human models working subgroup established the RE-JOIN minimal harmonized data elements that will be collected across all human studies and set the stage to develop parallel pre-clinical data collection standards. Data harmonization considerations included requirements from the HEAL program and recommendations from the consortium’s researchers and experts on informatics, knowledge management, and data curation. ResultsMultidisciplinary experts − including preclinical and clinical researchers, with both clinician-scientists- developed the RE-JOIN’s Minimal Human Data Standard with required domains and outcome measures to be collected across projects and institutions. The RE-JOIN minimal data standard will include HEAL Common Data Elements (CDEs) (e.g., standardized demographics, general pain, psychosocial and functional measures), and RE-JOIN common data elements (R-CDE) (i.e., both general and joint-specific standardized and clinically important self-reported pain and function measures, as well as pressure pain thresholds part of quantitative sensory testing). In addition, discretionary, site-specific measures will be collected by individual institutions (e.g., expanded quantitative sensory testing and gait biomechanical assessments), specific to the knee or TMJ. Research teams will submit datasets of standardized metadata to the RE-JOIN Data Coordinating Center (DCG) via a secure cloud-based central data repository and computing infrastructure for researchers to share and conduct analyses on data collected by or acquired for RE-JOIN. RE-JOIN datasets will have protected health information (PHI) removed and be publicly available on the SPARC portal and accessible through the HEAL Data Ecosystem. ConclusionData Harmonization efforts provide the multidisciplinary consortium with an opportunity to effectively collaborate across decentralized research teams, and data standardization sets the framework for efficient future analyses of RE-JOIN data collected by the consortium. The harmonized phenotypic information obtained will significantly enhance our understanding of the neurobiology of the pain-pathology relationships in humans, providing valuable insights for comparison with pre-clinical models.

Reflexive Data Curation: Opportunities and Challenges for Embracing Uncertainty in Human-AI Collaboration

This article presents findings from a Research through Design investigation focusing on a reflexive approach to data curation and the use of generative AI in design and creative practices. Using binary gender categories manifested in children’s toys as a context, we examine three design experiments aimed at probing how designers can cultivate a reflexive human-AI practice to confront and challenge their internalized biases. Our goal is to underscore the intricate interplay between the designer, AI technology, and publicly held imaginaries and to offer an initial set of tactics for how personal biases and societal norms can be illuminated through interactions with AI. We conclude by proposing that designers not only bear the responsibility of grappling critically with the complexities of AI but also possess the opportunity to creatively harness the limitations of technology to craft a reflexive data curation that encourages profound reflections and awareness within design processes.

Cost-efficient prompt engineering for unsupervised entity resolution in the product matching domain

Entity Resolution (ER) is the problem of semi-automatically determining when two entities refer to the same underlying entity, with applications ranging from healthcare to e-commerce. Traditional ER solutions required considerable manual expertise, including domain-specific feature engineering, as well as identification and curation of training data. Recently released large language models (LLMs) provide an opportunity to make ER more seamless and domain-independent. Because of LLMs’ pre-trained knowledge, the matching step in ER can be made easier by just prompting. However, it is also well known that LLMs can pose risks, that the quality of their outputs can depend on how prompts are engineered, and that the cost of using LLMs can be significant. Unfortunately, a systematic experimental study on the effects of different prompting methods and their respective cost for solving domain-specific entity matching using LLMs, like ChatGPT, has been lacking thus far. This paper aims to address this gap by conducting such a study. We consider some relatively simple and cost-efficient ER prompt engineering methods and apply them to perform product matching on two real-world datasets widely used in the community. We select two well-known e-commerce datasets and provide extensive experimental results to show that an LLM like GPT3.5 is viable for high-performing product matching and, interestingly, that more complicated and detailed (and hence, expensive) prompting methods do not necessarily outperform simpler approaches. We provide brief discussions on qualitative and error analysis, including a study of the inter-consistency of different prompting methods to determine whether they yield stable outputs. Finally, we consider some limitations of LLMs when used as a product matcher in potential real-world e-commerce applications.

ENTRANT: A Large Financial Dataset for Table Understanding

Tabular data is a way to structure, organize, and present information conveniently and effectively. Real-world tables present data in two dimensions by arranging cells in matrices that summarize information and facilitate side-by-side comparisons. Recent research efforts aim to train large models to understand structured tables, a process that enables knowledge transfer in various downstream tasks. Model pre-training, though, requires large datasets, conveniently formatted to reflect cell and table characteristics. This paper presents ENTRANT, a financial dataset that comprises millions of tables, which are transformed to reflect cell attributes, as well as positional and hierarchical information. Hence, they facilitate, among other things, pre-training tasks for table understanding with deep learning methods. The dataset provides table and cell information along with the corresponding metadata in a machine-readable format. We have automated all data processing and curation and technically validated the dataset through unit testing of high code coverage. Finally, we demonstrate the use of the dataset in a pre-training task of a state-of-the-art model, which we use for downstream cell classification.

Artificial Intelligence for Radiation Treatment Planning: Bridging Gaps From Retrospective Promise to Clinical Reality

Artificial intelligence (AI) radiation therapy (RT) planning holds promise for enhancing the consistency and efficiency of the RT planning process. Despite technical advancements, the widespread integration of AI into RT treatment planning faces challenges. The transition from controlled retrospective environments to real-world clinical settings introduces heightened scrutiny from clinical end users, potentially leading to decreased clinical acceptance.Key considerations for implementing AI RT planning include ensuring the AI model performance aligns with clinical standards, using high-quality training data, and incorporating sufficient data variation through meticulous curation by clinical experts. Beyond technical aspects, factors such as potential biases and the level of trust clinical end users place in AI may present unforeseen obstacles for real-world clinical use.Addressing these challenges requires bridging education and expertise gaps among clinical end users, enabling them to confidently embrace and utilize AI for routine RT planning. By fostering a better understanding of AI capabilities, building trust, and providing comprehensive training, the promises of AI RT planning can be a reality in the clinical setting.This article assesses the current clinical use of AI RT planning and explores challenges and considerations for bridging gaps in knowledge and expertise for AI operationalization, with focus on training data curation, workflow integration, explainability, bias, and domain knowledge. Remaining challenges in clinical implementation of AI RT treatment planning are examined in the context of trust building approaches.

Curation is Communal: Transparency, Trust, and (In)visible Labour

Research about trust and transparency within the realm of research data management and sharing typically centres on accreditation and compliance. Missing from many of these conversations are the social systems and enabling structures that are built on interpersonal connections. As members of the Data Curation Network (DCN), a consortium of United States-based institutional and non-profit data repositories, we have experienced first-hand the effort required to develop and sustain interpersonal trust and the benefits it provides to curation. In this paper, we reflect on the well-documented realities of curator and labour invisibility; the importance of fostering active communities (such as the DCN); and how trust, vulnerability and connectivity among colleagues leads to better curation practices. Through an investigation into data curators in the DCN, we found that, while curation can be isolating and invisible work, having a network of trusted peers helps alleviate these burdens and makes us better curators. We conclude with practical suggestions for implementing trust and transparency in relationships with colleagues and researchers.