Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 7 appraisals
Neural networks : the official journal of the International Neural Network Society
MFBD: Model-free backdoor defense based on vision-language pre-trained models
Backdoor attacks pose serious security threats to deep neural networks (DNNs), causing networks to output predetermined target labels when specific trigger patterns are embedded in input images. While existing backdoor defense methods utilize additional clean data or victim network models, their effectiveness in defending against various trigger injection methods is limited. We find that vision-language models, such as BLIP-2 and CLIP, are largely unaffected by backdoor triggers in poisoned images. Based on this observation, we propose MFBD, an effective model-free backdoor defense method leveraging the semantic consistency of dual vision-language descriptors. Specifically, we employ BLIP-2 and CLIP to generate descriptions of an image and its saliency masked image, respectively. We then embed both the descriptions and the image labels into a semantic space using SBERT. To distinguish clean and poisoned images, we compute the cosine similarity between their semantic representations. The image exhibiting a low cosine similarity value is classified as a poisoned image. We evaluate a suiute of backdoor defense methods against various attack methods on three datasets. Experimental results demonstrate that our method can effectively mitigate various complex backdoor threats while maintaining high performance across different datasets.
2 Aug 2026
Read appraisal →Balkan medical journal
Bias and Fairness Across the Healthcare AI Lifecycle: A Clinician-Oriented Review
Artificial intelligence (AI) is increasingly being investigated and, in selected clinical settings, implemented to support diagnosis, triage, and workflow optimization. Although these systems have the potential to improve access, consistency, and efficiency, they may also reproduce or amplify health inequities when bias is introduced during development, evaluation, implementation, or postdeployment use. This clinician-oriented narrative review adopts a practical lifecycle approach to explain how algorithmic unfairness becomes clinically relevant, how clinicians can recognize it, and how institutions can mitigate its impact. We first outline the ethical, clinical, and mathematical dimensions of fairness. We then examine fairness risks and sources of bias across six stages of the healthcare AI lifecycle: problem formulation, data generation, model development, evaluation, implementation, and postdeployment monitoring and governance. Key mechanisms include biased proxy outcomes, unrepresentative or error-prone data and labels, model shortcut learning, hidden stratification, distribution shift, and human-AI interaction effects (e.g., automation bias and alert fatigue), all of which can create feedback loops and contribute to fairness drift over time. For each stage, we identify clinician-facing red flags and practical mitigation strategies, including defining clinically meaningful outcomes, using representative and well-documented datasets, conducting subgroup-stratified evaluations, performing external and prospective validation, justifying decision thresholds, implementing safeguards for human-AI interactions, and maintaining continuous postdeployment monitoring, including postmarket surveillance for regulated medical devices. Fairness cannot be ensured through a single metric, publication, regulatory clearance, or one-time validation. Instead, equitable healthcare AI requires transparent design, rigorous evaluation, local governance, and ongoing monitoring across diverse populations, clinical sites, devices, workflows, and time. Fairness should therefore be regarded as a continuous clinical and institutional responsibility rather than a downstream technical consideration.
1 Aug 2026
Read appraisal →Proceedings of the National Academy of Sciences of the United States of America
The backfiring effect of weak AI safety regulation
Recent policy proposals aim to improve the safety of general-purpose AI, but there is little understanding of the efficacy of different regulatory approaches. We present a strategic model that explores interactions between safety regulation, general-purpose AI technology creators, and domain specialists-those who adapt the technology for specific applications. Our analysis examines how regulatory measures targeting different parts of the AI development chain affect the outcome of this game. Our model assumes AI technology is characterized by two key attributes: safety and performance. The regulator first sets a minimum safety requirement that applies to one or both players. The general-purpose creator then invests in the technology, establishing its initial safety and performance levels. Next, domain specialists refine the AI for their use cases, updating the safety and performance levels and taking the product to market. Resulting revenue is shared between the specialist and generalist. Our analysis reveals two insights: first, weak safety regulation imposed predominantly on domain specialists can backfire. While it might seem logical to regulate AI use cases, our analysis shows that weak regulations targeting domain specialists alone can reduce safety in a large class of parameterizations. Second, in contrast to the previous finding, we observe that stronger, well-placed regulation can mutually benefit all players. When regulators impose appropriate safety standards on both general-purpose AI creators and domain specialists, the regulation can function as a commitment device, leading to safety and performance gains, surpassing what is achievable under no regulation or regulating only one player.
28 July 2026
Read appraisal →International journal for equity in health
A conceptual framework for measuring AI health equity
BACKGROUND: Artificial intelligence (AI) is increasingly embedded in health systems globally and has the potential to improve efficiency, diagnostic accuracy, and decision support. However, its benefits remain unevenly distributed, particularly in low- and middle-income countries (LMICs). Models developed using datasets from specific populations may perform poorly in other settings, reinforcing structural inequities rather than correcting them. OBJECTIVE: This viewpoint proposes a composite framework, the AI in Healthcare Equity Index (AIHEI), to support measurable assessment of equity in health AI systems. FRAMEWORK: The AIHEI is designed to assess equity across five domains: data representation, algorithmic fairness, transparency and explainability, governance and oversight, and community impact and benefit sharing. By generating a standardised score, the index could enable comparisons across technologies, incentivise improvement, and support regulation, procurement, publication, and funding decisions. IMPLICATIONS: Pilots across diverse health domains and geographic settings are needed to assess feasibility, refine domain weighting, and evaluate reliability, reproducibility, and validity. Important challenges include contextual definitions of fairness, data sovereignty, post-deployment monitoring, and the risk of metric gaming. CONCLUSIONS: Quantifying equity in health AI is essential to ensure that AI does not create, widen, or exacerbate existing disparities by neglecting underserved populations. A common, objective measure of AI-related health equity can help move the field from ethical aspiration toward measurable accountability, monitoring, and enforcement.
25 July 2026
Read appraisal →Journal of the American Medical Informatics Association : JAMIA
Clinicians' rationale for editing ambient AI-drafted clinical notes: persistent challenges and implications for improvement
OBJECTIVE: The use of ambient AI documentation tools is rapidly growing in US hospitals and clinics. Such tools generate the first draft of clinical notes from scribed patient-provider conversations, which clinicians can then review and edit before signing into electronic health records (EHR). Understanding how and why clinicians make modifications to AI-generated drafts is critical to improving AI design and clinical efficiency, yet it has been under-studied. This study aims to address this gap. MATERIALS AND METHODS: We conducted semistructured interviews with 30 clinicians from the University of California, Irvine Health who used a commercial ambient AI tool in routine outpatient care. We invited them to describe how and why they edited AI drafts based on both their personal experience and review of some real-world examples identified from our previous studies. RESULTS: Modifications to AI drafts were primarily made to improve clinical accuracy and specialty-specific precision, reduce medico-legal and liability risk, and meet billing, coding, and documentation standards. Such editing was necessary due to reasons such as transcription errors, speaker attribution mistakes, overconfident statements without evidence, missing key clinical details, and AI's lack of information about the patient context. CONCLUSION AND DISCUSSION: Improving ambient AI documentation will require coordinated effort from vendors, institutions, and clinicians. Key targets include core model reliability (eg, transcription accuracy), specialty- and encounter-level customization, clinician-level personalization, more effective EHR integration, and institutional support (eg, training, governance, and standardized review guidance), complemented by clinicians' adaptive communication strategies that strengthen human-AI collaboration.
2 July 2026
Read appraisal →Journal of medical Internet research
Personal Health Large Language Models and the Negotiation of Medical Authority in Clinical Care: Opportunities, Risks, and Governance
Personal health large language models (PH-LLMs) are patient-facing conversational systems that synthesize user-entered information, patient-generated health data, wearable data, and selected personal health records-where users choose to connect them-into personalized, longitudinal, action-oriented health narratives. Unlike generic health chatbots that mainly provide one-off responses to isolated questions, PH-LLMs may generate continuing interpretations, priorities, and candidate next steps that patients bring into clinical encounters. In contrast to electronic health record-tethered clinical artificial intelligence (AI), they often originate outside institutional oversight and may be selected, used, or trusted by patients before professional review. This viewpoint examines how PH-LLMs may reshape the negotiation of medical authority by contributing to a shift from the traditional dyadic clinician-patient relationship toward a triadic model of negotiated authority, in which clinicians may increasingly need to mediate among clinical evidence, patient values, and algorithmic narratives. PH-LLMs may support patient participation by organizing symptoms, contextualizing home-monitoring and wearable data, improving health literacy, assisting chronic disease self-management, and preparing patients for more collaborative visits. Patient-facing AI narratives may also introduce distinct risks. At the individual level, these include inaccurate or incomplete responses arising from imprecise queries or missing context, misinterpretation of otherwise accurate information in the absence of clinical context, and contextually biased or poorly matched advice across demographic, cultural, linguistic, disability-related, or socioeconomic contexts. At the system level, they include authority conflict when AI recommendations diverge from clinical judgment, fragmentation of clinical truth, privacy and data-governance concerns, diffusion of accountability when harm results from advice produced outside clinical governance, and inequitable access to premium tools and continuous monitoring devices. To address these challenges, we propose a 3-layer clinical governance framework for patient-brought PH-LLM narratives. The first layer, evidence and provenance, makes AI-generated narratives epistemically legible by clarifying platform identity, data sources, temporal anchoring, uncertainty, and privacy-relevant data-use and retention conditions. The second layer, clinical arbitration and workflow integration, uses risk-stratified intake, proportionate documentation, escalation triggers, and equity-preserving workflows to embed PH-LLM outputs into routine care. The third layer, competence and accountability, defines the communication competencies, AI literacy supports, institutional responsibilities, vendor accountability, and risk-proportionate verification duties needed for triadic care. This framework is a conceptual and governance-oriented proposal rather than a validated clinical protocol. Future empirical work should evaluate its feasibility, documentation burden, equity effects, clinical safety impact, and acceptability among patients, clinicians, and health systems. Governed through these interdependent layers, PH-LLMs may serve as supporting infrastructure for safer, person-centered longitudinal care.
26 June 2026
Read appraisal →Journal of medical systems
Beyond Validation: Operationalising Post-Deployment Surveillance of AI Medical Devices in Clinical Practice
Artificial intelligence medical devices are increasingly deployed in clinical practice, yet practical approaches to post-deployment monitoring remain poorly defined. We present a structured, decision-oriented approach to monitoring within healthcare institutions, grounded in our own deployment experience. By framing surveillance as a set of interdependent decisions, this model supports effective performance assessment and governance-linked corrective action, enabling safer and more accountable integration of AI into routine clinical care.
21 June 2026
Read appraisal →