James Th Teo

Liver transplantation (LT) is a complex surgical intervention, with intricate decision-making and risks at all stages before, during and after the procedure itself. Artificial intelligence (AI) may assist the process but has seen limited use. This scoping review assessed primary articles reporting on clinical AI in LT in MEDLINE and Embase from 2000 to September 2026. 196 articles were identified: focusing on the pre-operative period before transplant (21.9%), organ retrieval (9.7%), intra-operative issues (3.6%), short-term post-operative complications (16.8%), long-term post-operative complications (45.9%) and natural language processing (2.0%). 95.9% had retrospective designs and only 12.8% had direct multi-centre data, with the rest from national registries or single centres. 13.8% of studies externally validated their AI tools. Only one study performed a prospective evaluation of their tool with a randomized controlled trial. Progress towards clinically deployable tools will require harmonised, higher-granularity datasets, consistent reporting, and prospective evaluation within real-world transplant workflows.

UNSTRUCTURED Large language models (LLMs) are increasingly explored for clinical documentation, information synthesis, health communication, and decision-support tasks. Evaluation must address the quality of their outputs in clinically meaningful terms, not only whether a response reaches an expected answer. Endpoint correctness can conceal important failures in problem representation, differential prioritisation, response to new evidence, temporal synthesis, uncertainty, and safety. Existing clinical benchmarks address overlapping aspects of these qualities through physician-authored criteria and task-specific scoring, but organise them in different ways. Medical education offers established approaches to making selected aspects of clinical reasoning observable for assessment. Its instruments were developed for human learners: their validity does not transfer automatically to LLMs, and a generated explanation does not establish how a model produced its answer. We therefore use medical-education assessment traditions as sources of constructs and task-design logic, not as prevalidated scoring tools for model outputs. We propose a provisional seven-domain framework for evaluating expressed clinical reasoning: the clinically defensible organisation, interpretation, and updating of available patient information in a generated response. The domains cover accuracy and grounding, validity and completeness, diagnostic and management reasoning, temporal reasoning, uncertainty and updating, clinical safety, and communication. They are organised into an evidential foundation, an expressed-reasoning core, and clinical guardrails. Proposed five-point behavioural anchors are applied only to relevant domains and components. Case-specific reference criteria permit defensible alternatives, and judgements are made using evidence available at the specified decision point. Prespecified safety-critical breaches are flagged separately and cannot be compensated for by strong scores elsewhere. The framework is intended for offline evaluation, error analysis, and benchmark design. It is not a validated measurement instrument, a measure of internal model reasoning, or evidence of readiness for clinical deployment. We outline priorities for independent content review, feasibility testing, domain-level agreement assessment, and evaluation of automated judging before stronger measurement claims are made.

Background: Laboratory-based frailty indices (FI-Lab) have shown promise in geriatric medicine research. We aimed, in diverse samples, to determine the optimal construction of an FI-Lab for acute care and evaluate its validity as a measure of latent health status across the adult life span. Methods and findings: Our retrospective multicohort study used emergency department encounters in Boston, USA (MIMIC-IV-ED; 2011–2019) and London, UK (King’s College Hospital (KCH); 2017–2020); and a community-based cohort in the UK (UK Biobank; 2006–2010). We evaluated FI-Lab configurations by varying test selection strategies, number of tests, and minimum test thresholds. Sensitivity analyses examined performance across age, ethnicities, sexes, and model types to identify potential blind spots. The primary outcome was 1-year all-cause mortality. We analysed 227,736 visits (113,032 patients) from MIMIC-IV-ED; 152,305 visits (96,843 patients) from KCH; and 492,703 UK Biobank. An FI-Lab calculated from 25 commonly ordered tests (minimum 15 required) demonstrated hazard ratios for 1-year mortality approaching those of chronological age and exceeding the National Early Warning Score 2 (NEWS2). In combined models, hazard ratios for FI-Lab per standard deviation increase were 2.04 (95 %CI [1.99, 2.09]; p < 0.001) in MIMIC-IV-ED, 2.55 (95 %CI [2.42, 2.69]; p < 0.001) in KCH, and 1.87 (95% CI [1.80, 1.94]; p < 0.001) in UK Biobank. Performance was consistent across subject groupings. Discrimination plateaued at 20–40 tests. This was a retrospective study; future work exploring the impact on everyday use requires prospective evaluation, potentially in a randomised controlled trial. Conclusions: The FI-Lab provides an automated, scalable measure of patient vulnerability that is robust across healthcare settings and populations. A core set of 20–40 commonly ordered tests is sufficient for signal capture. It offers a pragmatic complement to clinical judgement without requiring manual data entry or additional tests.

Background: Artificial intelligence (AI) is rapidly transforming clinical neurology, offering significant potential to enhance diagnosis, treatment, and disease management. Despite this transformative promise, AI adoption in neurology remains limited due to a lack of a reliable evidence base to support its use, as well as educational, ethical, methodological, and regulatory barriers. Furthermore, currently, there is no standardized educational framework for the application of AI in clinical neurology across Europe. Methods: To address this gap, the European Academy of Neurology (EAN) has established a dedicated multidisciplinary Task Force (TF) on AI in Clinical Neurology. This TF involves neurologists, neurology trainees, medical students, computer and data scientists, ethicists, patient representatives, and regulatory experts. Results: The AI TF has designed a modular curriculum covering technical AI foundations and models, clinical applications, ethical implications, and regulatory compliance. Planned educational resources include e‐learning modules, podcasts, masterclasses, and interactive sessions at EAN congresses. Also, a recently conducted Europe‐wide survey supported the TF in identifying current knowledge levels and educational and systemic barriers, informing the design of targeted interventions. In parallel, the TF will formulate recommendations addressing the ethical and regulatory implications of AI use in neurology, tailored to the specific needs of clinicians. Conclusions: The EAN TF on AI in Clinical Neurology represents a strategic initiative to enable responsible AI integration through multidisciplinary collaboration. By closing educational gaps, establishing clear ethical standards, and facilitating stakeholder engagement, the TF aims to empower neurologists to confidently and ethically adopt AI technologies, ultimately improving patient care across Europe.

Background and aims: Heart failure with preserved ejection fraction (HFpEF) phenotypes have been characterised as static entities in trial populations yet their temporal dynamics remain unexplored in routine clinical practice. We aimed to identify HFpEF phenogroups from real-world electronic health records (EHRs), characterise their longitudinal stability and determine whether phenogroup trajectories were associated with differences in mortality. Methods: A natural language processing (NLP) pipeline identified HFpEF cases from multisite EHRs (2010–2022). Latent class analysis (LCA) defined baseline phenogroups and annual supervised random forest (RF) classification using time-updated values of the same baseline LCA variables tracked individual transitions over five years. Competing risks regression quantified transition versus death probabilities. Time-varying Cox models assessed mortality risk associated with current (vs baseline) phenogroup status. Phenogroups were independently re-derived by LCA in an external cohort and compared with RF predictions. Results: Among 2223 patients (60% female, median age 75 years (IQR 63–83); left ventricular ejection fraction (LVEF) 60.1%; follow-up 4.0 years (IQR 2.2–6.1), we identified four phenogroups: young-low comorbidity (22%), obesity-predominant (24%), elderly atrial dysfunction (32%) and cardiovascular-kidney-metabolic (23%). Young-low comorbidity had the highest transition probability (47.7%), while cardiovascular-kidney-metabolic and elderly-atrial dysfunction were highly stable (3.2% and 12.1% transitioned, respectively). Within young-low comorbidity, those who transitioned had greater baseline cardiac abnormalities, including higher E/e′, left ventricle mass and wall thickness. Adjusted mortality risk was highest in cardiovascular-kidney-metabolic (HR 1.53, 95% CI 1.22 to 1.92, p<0.001); current phenogroup status discriminated mortality risk modestly better than baseline classification alone (C-index 0.659 vs 0.648). In an independent external cohort (n=3349), RF predictions agreed strongly with independently derived LCA phenogroup labels (C-statistics 0.891–0.953). Conclusion: Longitudinal NLP-based phenotyping of real-world EHRs identified four reproducible HFpEF phenogroups with distinct trajectories, distinguishing an early progressive disease state from stable advanced phenotypes. Earlier HFpEF detection using artificial intelligence-driven EHR tools could facilitate phenotype-targeted treatment of patients at potentially modifiable stages.

Artificial intelligence (AI) applications in neurology have reached an inflection point. Despite US Food and Drug Administration approval of numerous algorithms in neuroimaging, neurophysiology, genetics and chatbots, their real-world impact remains limited. This disconnect between research promise and clinical reality represents a gap in understanding how to translate AI algorithms into clinical benefit for patients globally. In this Perspective, we examine the challenges that prevent clinical AI use in neurology moving beyond pilot studies towards meaningful clinical impact. We consider the steps required in the process of translation, including research, validation of AI models, regulatory approval pathways and clinical implementation. We discuss implementation of AI models as stand-alone products versus embedded platforms, and the requirements for sustainable deployment. Beyond traditional clinical decision support tools, we examine paraclinical applications of AI, including chatbots and ambient voice documentation. We recommend expanding capacity for prospective validation and scaling by implementing and validating technologies across multiple sites and countries, which requires infrastructure from long-term partnerships. Neurology must shift from asking whether AI can work to understanding how to use it safely at scale.

Objective: Digital health apps and patient portals are proposed as part of the drive from ‘analogue to digital’ care for the National Health Service (NHS) 10-Year Plan. Without mitigation strategies, digital inequalities could arise as a result, and more evidence is needed to understand how to mitigate this. Methods: As part of an equality impact assessment, a retrospective cross-sectional analysis was conducted examining patient portal activation among patients invited to outpatient appointments at two large south-east London Hospital Trusts between 1 May and 1 November 2024. Results: Of the 503 688 patients invited to attend outpatient clinics during the study period, 52.7% had activated the patient portal. Availability of email contact details was the strongest determinant of onboarding likelihood (OR 10.86). Multivariate logistic regression models showed that the following groups were less likely to activate the patient portal: men (OR 0.84), individuals at the extremes of age (71–80 or 11–20 years), those of mixed or undefined ethnicity (OR 0.58), those of black ethnicity (OR 0.62) and those with the highest degree of socioeconomic deprivation (Index of Multiple Deprivation group 1; OR 0.68). Conclusion: This large-scale roll-out of a digital health portal provides empirical evidence of factors that drive digital inequalities for patients of two major London NHS Trusts. The observed disparities across demographic and socioeconomic dimensions and simple reliable digital contact mechanisms highlight the risk that digital healthcare initiatives may inadvertently produce new types of inequalities.

Aims: Heart failure with preserved ejection fraction (HFpEF) accounts for approximately half of all heart failure cases, with high levels of morbidity and mortality. However, many patients who meet diagnostic criteria for HFpEF do not have a documented diagnosis, particularly in non-White populations where conventional risk scores may underestimate risk. Our aim was to develop and validate a diagnostic prediction model to detect HFpEF based on ESC criteria, AIM-HFpEF. Methods and results: We applied natural language processing (NLP) and machine learning methods to routinely collected electronic health record (EHR) data from a tertiary centre hospital trust in London, UK, to derive the AIM-HFpEF model. We then externally validated the model and performed benchmarking against existing HFpEF prediction models (H2FPEF and HFpEF-ABA) for diagnostic power on the entire external cohort and in patients of non-White ethnicity and patients from areas of increased socioeconomic deprivation. An XGBoost model combining demographic, clinical, and echocardiogram data showed strong diagnostic performance in the derivation dataset [n = 3173, AUC = 0.88, (95% CI, 0.85–0.91)] and validation cohort [n = 5383, AUC: 0.88 (95% CI, 0.86–0.90)]. Diagnostic performance was maintained in patients of non-White ethnicity [AUC = 0.89 (95% CI, 0.85–0.93)] and patients from areas of high socioeconomic deprivation [AUC = 0.90 (95% CI, 0.85–0.95)]. In contrast, AIM-HFpEF demonstrated favourable performance relative to the H2FPEF and HFpEF-ABA models. AIM-HFpEF model probabilities were associated with an increased risk of death, hospitalization, and stroke in the external validation cohort (P < 0.001, P = 0.01, P < 0.001, respectively, for highest vs. middle tertile). Conclusion: AIM-HFpEF represents a validated equitable diagnostic model for HFpEF, which can be embedded within an EHR to allow for fully automated HFpEF detection.