BMJ Health & Care Informatics最新文献

{"title":"Missing infrastructure for real-world predictive AI impact.","authors":"Benjamin I Perry, Ben Hammond, Dominic Oliver, Ashley Mushambi, Carina Andrews, Niels Peek, Ben Van Calster, Laure Wynants, Annabel E L Walsh, Joseph E Alderman, Shuqing Si, Alastair K Denniston, Simon J Griffin, Joie Ensor","doi":"10.1136/bmjhci-2025-101954","DOIUrl":"10.1136/bmjhci-2025-101954","url":null,"abstract":"<p><p>Health systems are under increasing pressure as chronic disease becomes the dominant burden of care. Addressing this requires earlier detection and prevention, and predictive artificial intelligence (AI) tools are a demonstrable means of targeting interventions more effectively. Despite extensive research activity and substantial public funding, few predictive AI tools are translated effectively into routine care. We argue that the primary translational barrier is now organisational rather than solely methodological: without defined institutional ownership and long-term stewardship, including to support obtaining and maintaining regulatory certification and navigating governance structures, predictive AI tools cannot progress beyond print. Three interdependent recommendations are proposed to address this problem.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-06-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13288887/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148283085","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Open-source large language model-based on-premises pipeline for automated data extraction from unstructured electronic health records: a pilot study.","authors":"Vasileios Ntinopoulos, Hector Rodriguez Cetina Biefer, Laura Rings, Rodney Alexander Rosalia, Omer Dzemali","doi":"10.1136/bmjhci-2025-101956","DOIUrl":"10.1136/bmjhci-2025-101956","url":null,"abstract":"<p><strong>Objectives: </strong>We evaluated an on-premises, open-source large language model (LLM)-based data extraction pipeline for automated data extraction from unstructured electronic health records (EHRs).</p><p><strong>Methods: </strong>Automated script-based EHR data preprocessing extracted 50 medical texts in German, which were entered into the LLM pipeline. 4-bit and 8-bit quantizations of 14 mid-sized LLMs (30B-90B parameters) were evaluated in 6 information extraction, 11 binary classification and 5 multilevel classification tasks comprising all variables of the European System for Cardiac Operative Risk Evaluation 2 (EuroSCORE 2) model for 1100 predictions each. LLM response consistency was assessed over three same-prompt iterations.</p><p><strong>Results: </strong>In overall accuracy, Qwen3-30b-a3b-q8 presented the highest value (0.954) and 13 LLMs had values over 0.90. In information extraction accuracy, 12 LLMs exhibited a value of 1.0 and all 14 LLMs had values over 0.96. In binary classification accuracy, Llama3.2-vision-90b-q4 exhibited the highest value (0.972), 5 LLMs had values of at least 0.95 and all 14 LLMs showed values over 0.93. In multilevel classification accuracy, Qwen3-30b-a3b-q8 exhibited the highest value (0.940), four LLMs had values over 0.90 and all LLMs presented values over 0.80. Nine LLMs exhibited perfect response consistency and the remaining five LLMs had a Krippendorff's alpha value of 0.999.</p><p><strong>Discussion: </strong>Multiple LLMs exhibited high accuracy in information extraction, binary classification, multilevel classification and response consistency and seem able to reliably automate data extraction from EHRs.</p><p><strong>Conclusion: </strong>This pilot study demonstrates the feasibility of on-premises, privacy-preserving, LLM-based automated EHR data extraction pipelines. Larger-scope studies are warranted to validate their potential in healthcare.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13239439/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148142615","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Routine data are not 'dirty' data: how to approach using routine data for research.","authors":"Grace Duffy, Alice Cunningham, Vishal Sharma, Tom Lawton","doi":"10.1136/bmjhci-2025-102018","DOIUrl":"10.1136/bmjhci-2025-102018","url":null,"abstract":"<p><p>The rise of electronic patient records has led to an abundance of routine healthcare data, and the development of NHS Secure Data Environments (SDEs) is transforming their potential for research. However, routine data differ fundamentally from research data; they are generated to support care processes rather than answer predefined research questions. This creates challenges when they are analysed without understanding the context in which they were produced.Using real examples from practice, this paper illustrates the variety of ways routine data can be shaped and how this leads to misinterpretation if treated as merely 'dirty' research data. These examples underpin practical strategies for researchers, including analysis at or by source and working closely with local data providers to understand how a dataset was created.We argue that SDEs must foster strong links between data users and creators to preserve data provenance and support accurate interpretation. By approaching routine data as meaningful records of healthcare processes-rather than flawed research datasets-researchers can fully realise their potential to drive healthcare improvement.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13223642/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148052394","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Machine learning-based prediction of a high-risk kidney function trajectory class after acute kidney injury.","authors":"Chien-Liang Liu, You-Lin Tain, Chih-Chien Lin, Yueh-Lung Peng, Chien-Ning Hsu","doi":"10.1136/bmjhci-2025-101578","DOIUrl":"10.1136/bmjhci-2025-101578","url":null,"abstract":"<p><strong>Objectives: </strong>Monitoring kidney function after acute kidney injury (AKI) hospitalisation is essential for identifying patients at risk of rapid progression. This study developed a machine learning (ML) model to predict membership in a high-risk longitudinal estimated glomerular filtration rate (eGFR) trajectory associated with subsequent kidney replacement therapy (KRT) initiation.</p><p><strong>Methods: </strong>We conducted a cohort study of 88 632 adults with ≥2 post-discharge eGFR measurements between 2010 and 2017. Joint latent class mixed models were used to identify dynamic eGFR trajectories from 665 499 measurements. Patient-level data splitting (80/20) was strictly enforced. Extreme gradient boosting (XGBoost) and random forest models were trained using clinical features within 3 months after discharge to predict membership in the highest-risk eGFR trajectory class.</p><p><strong>Results: </strong>Three eGFR trajectories were identified: Steady Low (high-risk phenotype), Slow Decline and Recovery. The Steady Low group showed the strongest association with KRT. The 10-variable XGBoost model demonstrated excellent discrimination (AUROC 0.974, sensitivity 0.913, specificity 0.919). The eGFR slope from discharge to 3 months was the most influential predictor of high-risk progression.</p><p><strong>Discussion: </strong>Although the study was limited to a single healthcare system, this ML framework identified patients at risk of rapid kidney function decline using routinely available clinical data. The model supports early risk stratification after AKI and may enable more timely kidney-protective interventions. External validation, calibration assessment and benchmarking against simpler eGFR-based rules remain important next steps.</p><p><strong>Conclusion: </strong>A 10-variable XGBoost model accurately identifies patients with deteriorative eGFR trajectories (Steady Low) early after discharge, providing a practical framework for personalised post-AKI kidney care.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13218091/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148041093","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Towards a feature-based regulatory benchmark for user-centred SaMD design in cancer care.","authors":"Saba Kheirinejad, Christopher L Brett, Robert L Davis, David L Schwartz, Arash Shaban-Nejad","doi":"10.1136/bmjhci-2025-101878","DOIUrl":"10.1136/bmjhci-2025-101878","url":null,"abstract":"<p><p>Software as a medical device (SaMD) is increasingly central to cancer care, yet current regulatory frameworks from the International Medical Device Regulators, the US Food and Drug Administration and the European Medical Device Regulation lack clear premarket standards for evaluating usability, accessibility and emotional/cultural appropriateness for intended users. This gap is particularly problematic in oncology, where patients, caregivers and clinicians navigate high cognitive load, emotional stress and complex care coordination that can make poorly designed tools ineffective or even harmful.This paper introduces a structured, feature-based benchmark grounded in human-centred design and user-centred design principles to address this critical gap and enable integration into early regulatory review. The benchmark defines measurable design criteria for premarket review, with a focus on usability, accessibility, emotional safety and clinical applicability, providing both developers and regulators with a practical checklist. By embedding these standards early in the product lifecycle, the framework supports regulatory consistency, reduces postmarket usability failures and improves patient trust, engagement and clinical outcomes.While developed for oncology care, the benchmark principles are adaptable across diseases, offering a replicable pathway to align SaMD design with both clinical requirements and the real-world needs of patients and healthcare systems.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13201996/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148004585","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Beyond the digital health tsunami: reaffirming biomedical and health informatics as the scientific foundation of digital healthcare transformation.","authors":"Brigitte Séroussi","doi":"10.1136/bmjhci-2025-102025","DOIUrl":"10.1136/bmjhci-2025-102025","url":null,"abstract":"<p><p>Digital health (DH) has emerged as the dominant terminology for technology-enabled healthcare, reshaping institutional discourse, educational programmes and policy agendas. DH is often portrayed as a novel discipline that could supersede Biomedical and Health Informatics (BMHI), creating conceptual ambiguity and the risk of weakening scientific rigour in the design, evaluation and governance of digital interventions. This paper clarifies BMHI's disciplinary identity, characterises DH as a practice-oriented and implementation-oriented domain, and establishes the epistemic relationship between the two in order to inform academic strategy, curricula and professional identity. A comparative conceptual analysis examined scope, theoretical foundations, methodological approaches, educational frameworks, professional structures and stakeholder ecosystems in both domains, informed by recent debates on degree programmes and workforce training. We found that BMHI is a mature scientific discipline with explicit theoretical foundations encompassing data representation, knowledge modelling, clinical reasoning, decision support, interoperability, sociotechnical analysis, evaluation science, safety, ethics and governance. DH, by contrast, represents an implementation-oriented translational domain that operationalises BMHI through products, services and organisational transformation (eg, telehealth, mobile applications, artificial intelligence-enabled clinical services). While DH has amplified visibility and accelerated adoption, it lacks autonomous theoretical foundations and depends on BMHI for scientific validity, safety and sustainability. Three evolutionary scenarios are examined: DH absorbing BMHI (epistemically incoherent), convergence as parallel disciplines (conceptually flawed), and reintegration of DH within BMHI as its applied layer (scientifically coherent and strategically sustainable) but requiring disciplinary communication, curricular alignment and institutional leadership explicitly positioning DH as a translational layer of BMHI.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13202130/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148004505","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Barrier check study: why predictive machine learning struggles to reach the operating room.","authors":"Sara Ben Hmido, Chiara Garita, Frank Bloemers, Sietske Rozie, Kak Khee Yeung, Geert Kazemier, Freek Daams","doi":"10.1136/bmjhci-2025-101910","DOIUrl":"10.1136/bmjhci-2025-101910","url":null,"abstract":"<p><strong>Objectives: </strong>Machine learning (ML) has the potential to enhance surgical decision-making through real-time risk prediction and personalised care yet clinical implementation remains limited. This study aimed to identify and categorise the key barriers to implementing ML tools in surgical practice.</p><p><strong>Methods: </strong>A nationwide qualitative survey was conducted among stakeholders involved in ML development or clinical implementation in healthcare. Participants provided open-ended responses describing perceived barriers, their impact and potential solutions. Responses were analysed using inductive thematic content analysis.</p><p><strong>Results: </strong>95 participants from 53 organisations submitted 178 individual barrier entries. The most frequently reported barriers included limited clinician trust and artificial intelligence literacy, challenges integrating ML into surgical workflows, regulatory misalignment, insufficient data availability and quality, and unclear economic value. Barriers were highly interconnected, with challenges in one domain often reinforcing others. Proposed solutions emphasised interdisciplinary collaboration, early end-user involvement, improved transparency and training, privacy-preserving data-sharing approaches such as federated learning, and better alignment between regulatory, institutional and clinical structures.</p><p><strong>Conclusion: </strong>The findings indicate that ML implementation in surgery is a socio-technical challenge rather than a purely technical one. Barriers often arise from misalignment between responsibility, decision authority and accountability across technical, clinical and organisational domains, particularly in time-critical surgical settings.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13185021/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147970228","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Early sepsis prediction using a hybrid LSTM-GAT model: a study on the PhysioNet 2019 dataset.","authors":"Bahar Khorram, Samaneh Kouchaki","doi":"10.1136/bmjhci-2025-101762","DOIUrl":"10.1136/bmjhci-2025-101762","url":null,"abstract":"<p><strong>Objective: </strong>Sepsis is a potentially fatal systemic response to infection, in which early clinical intervention is critical to reduce mortality. This study presents a hybrid deep learning model that combines temporal and structural information from clinical data to improve early sepsis prediction.</p><p><strong>Methods: </strong>We used data from the PhysioNet/Computing in Cardiology Challenge 2019 to predict sepsis onset up to 12 hours in advance. We developed a hybrid model integrating long short-term memory (LSTM) networks and graph attention networks (GAT) to capture temporal dynamics and intervariable relationships. Performance was compared with three baseline models. To ensure robustness, all models were trained using five repeated train-test splits with different random seeds.</p><p><strong>Results: </strong>The dataset includes 40 336 adult ICU patients. Of all the patients, 2932 developed sepsis during their stay. Each patient's data includes hourly data on 40 clinical variables, including vital signs, laboratory results and demographic information. The LSTM-GAT model achieved an area under the receiver operating characteristic curve (AUROC) of 0.853±0.005, an F1-score of 0.627±0.006 and a specificity of 0.872±0.007, outperforming baseline models. Despite being trained on fixed temporal windows, the model generalised well across multiple prediction horizons without retraining.</p><p><strong>Discussion: </strong>By integrating temporal and structural representations, the proposed approach achieves improved predictive performance compared with the baseline. This capability may support earlier identification of high-risk patients and enhance timely clinical decision-making in critical care environments.</p><p><strong>Conclusions: </strong>The proposed model demonstrates the advantage of combining sequence and graph-based methods. It offers a promising tool for real-time clinical decision support in sepsis detection.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13150876/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147833345","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Self-regulating the use of large language models in clinical practice: a risk-stratified approach.","authors":"Milan Mohammad, Espen Jimenez-Solem, Michael Hejmadi, Andreas Pihl","doi":"10.1136/bmjhci-2025-101921","DOIUrl":"10.1136/bmjhci-2025-101921","url":null,"abstract":"<p><p>The rapid integration of large language models (LLMs) into clinical practice offers promising benefits, including assistance with documentation, decision support and patient communication. However, these advantages are tempered by concerns around model accuracy, data privacy, clinician trust and regulatory responsibility. In this comment, we propose a clinician-led, risk-stratified framework to guide responsible adoption of LLMs in healthcare. The framework categorises applications into four risk tiers: low risk, moderate risk, high risk and critical risk. Each tier demands tailored oversight, validation and governance, with increasing levels of statutory regulatory scrutiny (eg., European Union Artificial Intelligence Act and U.S. Food and Drug Administration) and clinical supervision. We argue that proactive self-regulation combined with ongoing quality management and clinician education is essential to safely integrate LLMs into ongoing care.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-05-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13150866/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147833325","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"AI-generated clinical summaries: errors and susceptibility to speech and speaker variability.","authors":"Thomas C Draper, Jason Leake, Timothy Cox, Kathryn Lamb-Riddell, Benjamin E Johns, John McCormick, Stephen Trowell, Janice Kiely, Richard Luxton","doi":"10.1136/bmjhci-2025-101918","DOIUrl":"10.1136/bmjhci-2025-101918","url":null,"abstract":"<p><strong>Objectives: </strong>To evaluate whether variability in patients' communication style (personality), international English-accents (human and synthetic) and speech impairments affects the accuracy of a Clinical AI Scribe (CAIS) and identify where performance degrades to inform pre-deployment validation and monitoring.</p><p><strong>Methods: </strong>We conducted simulated primary-care consultations using trained actors. For personality types, four scenarios were enacted, each with five patient-personality types. For accents, transcripts of consultations were used to generate combinations of seven accents across five scenarios. The CAIS produced summaries that were compared with transcripts, and errors classified as omissions, factual inaccuracies or hallucinations. For speech impairments, public recordings representing five profiles were transcribed and word-recognition accuracy calculated.</p><p><strong>Results: </strong>Personality types showed no statistically significant differences in errors (all p>0.05). Extraversion had the highest total errors (median 3.5). Across accents, comparisons were non-significant for both patient and doctor voices (patients: p=0.851; doctors: p=0.980). Omissions predominated, with low rates of hallucinations and factual inaccuracies. Omissions were slightly higher for Chinese-accented and Indian-accented doctors (both medians 3.0). Conversely, speech impairments differed: cleft palate and vowel disorders were near-perfect, whereas phonological impairment markedly reduced recognition (p<0.001).</p><p><strong>Discussion: </strong>Operationally, CAIS deployment should include clinician-in-the-loop verification, subgroup performance monitoring (accents, impairments) and predefined 'switch-off' criteria for severe phonological patterns. High-quality synthetic voices are a pragmatic proxy for accent testing when balanced corpora are unavailable.</p><p><strong>Conclusions: </strong>Under controlled conditions CAIS performance was broadly stable across communication styles and most accents, but vulnerable to specific speech characteristics, particularly phonological impairment, in this single-system simulation study.</p>","PeriodicalId":9050,"journal":{"name":"BMJ Health & Care Informatics","volume":"33 1","pages":""},"PeriodicalIF":6.2,"publicationDate":"2026-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13110523/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147761600","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}