Internal medicine involves high-stakes, time-sensitive decisions (such as triaging acute illnesses, escalating care, providing thromboprophylaxis, planning discharges, and managing chronic diseases) often under uncertainty. Risk stratification tools convert limited bedside data into actionable categories. Predictive analytics, by contrast, draws on richer electronic health record data streams to estimate short- or long-term risks in real time. Model performance varies across sites due to differences in data quality, outcome definitions, and workflows; external validation may reveal miscalibration, while alerts can cause alarm fatigue. This review provides a practical framework for clinicians and clinical leaders to appraise, select, and implement predictive models in internal medicine. We map common use cases (like inpatient deterioration, sepsis, acute kidney injury, venous thromboembolism, and readmission) to the clinical decisions they should trigger, emphasising that prediction without an intervention pathway rarely improves outcomes. We outline what to demand before deployment: a clearly specified target population and time horizon; predictor availability and measurement consistency; evaluation beyond discrimination to include calibration and decision-curve analysis; equity checks across key subgroups; and a life-cycle plan for monitoring dataset shift and calibration drift. We also discuss how we can choose thresholds, interpret probabilities at patient bedside and evaluate clinical impact with studies and continued monitoring. Recommendations are aligned with contemporary reporting and appraisal standards for prediction and AI studies (TRIPOD+AI, PROBAST+AI, CONSORT-AI, DECIDE-AI) and with governance and regulatory principles for AI-enabled medical software. The goal is to help internists translate risk scores and EHR-based predictions into safer, more equitable, and measurably effective care pathways. [1].

Artificial intelligence (AI) comprises computational methods capable of tasks associated with human cognition, and includes specialized subfields such as machine learning, deep learning, convolutional neural networks for image analysis, and large language models for text-based workflows. In dermatology, these methods are increasingly used across research and clinical practice, supporting drug discovery, diagnosis, decision-making, and personalized treatment. In drug discovery, AI can accelerate target identification, support in silico compound design, and predict the effectiveness and safety profiles of compounds. It also offers opportunities for drug repurposing, helping identify candidates for conditions with limited treatment options. In clinical practice, AI enhances teledermatology and imaging-based assessment, providing consistent lesion analysis, severity evaluation, and triage support across malignant, inflammatory, and infectious skin diseases. Beyond diagnosis, AI is beginning to play a significant role in supporting therapeutic decision-making. Clinical decision-support systems and multimodal AI tools assist clinicians by organizing information, suggesting management strategies, monitoring disease progression, and helping align treatment choices with current guidelines. AI models can further estimate treatment response, durability, and the likelihood of therapy adjustment, creating a basis for more individualized care. Despite these advances, significant challenges remain. Bias in training data, limited real-world validation, unclear regulatory frameworks, concerns over interpretability, and the need for patient and clinician trust remain ongoing barriers. This review aims to provide an up-to-date overview of the emerging application domains of AI in dermatology.

Artificial intelligence (AI) has moved from proof-of-concept studies in dermatology to selective, real-world clinical use, particularly in image-based triage, lesion assessment, and workflow augmentation. Dermatology is uniquely suited to AI because much of diagnostic reasoning depends on visual information (clinical photos, dermoscopy, reflectance confocal microscopy, optical coherence tomography) plus context (history, distribution, symptoms, treatments, and risk factors). Yet translation into routine care depends less on eye-catching benchmark accuracy and more on careful validation, robust reference standards, fairness across skin tones and devices, usability within clinical workflows, and continuous post-deployment monitoring for performance and safety. This review summarizes leading diagnostic applications of AI in dermatology, provides a practical framework for clinical validation, and highlights real-world deployment patterns including teledermatology triage, melanoma risk workflows, inflammatory dermatoses severity scoring, and AI-assisted dermatopathology. We emphasize that "AI performance" is not a single number: the intended use (screening vs. diagnosis vs. referral prioritization), operating thresholds, prevalence, human-AI interaction, and downstream clinical actions determine real impact. Finally, we outline implementation considerations (governance, privacy, regulatory expectations, and monitoring) and propose practical steps for clinics and health systems seeking safe adoption.

Artificial intelligence (AI) is rapidly transforming medical research and scholarly publishing, reshaping how scientific knowledge is produced, evaluated, and disseminated. Initially developed as a decision-support tool, AI has evolved into a complex ecosystem encompassing machine learning, deep learning, and large language models, with applications spanning data analysis, diagnostic support, evidence synthesis, manuscript preparation, peer review, and post-publication analytics. These technologies offer substantial benefits, including accelerated research workflows, improved analytical precision, enhanced reproducibility, and expanded access to scientific communication, particularly for early-career investigators and non-native English authors. However, the integration of generative AI introduces significant challenges. Persistent risks include algorithmic bias, hallucinated or misattributed citations, erosion of authorship accountability, confidentiality concerns, and the potential degradation of peer review integrity. As AI-generated outputs increasingly resemble human scholarly work, longstanding norms surrounding authorship, transparency, and responsibility are being reexamined. In response, editorial organizations, journals, and global health authorities have begun to establish governance frameworks emphasizing disclosure, human verification, and ethical boundaries for AI use. This narrative review synthesizes current evidence on the evolution and applications of AI in medical research and publishing, critically examines associated risks and ethical dilemmas, and reviews emerging regulatory and editorial guidance. Finally, it outlines future directions centered on explainable and auditable AI, standardized AI literacy, and hybrid human-AI workflows. Ensuring that AI remains a tool for augmentation rather than replacement will be essential to preserving trust, rigor, and integrity in medical scholarship as these technologies become increasingly embedded in the scientific enterprise.

Artificial intelligence (AI) is increasingly integrated into healthcare systems, offering transformative opportunities in diagnostics, treatment personalization, predictive analytics, and workflow optimization. However, alongside these advancements, AI introduces complex ethical, legal, and regulatory challenges that must be addressed to ensure safe, equitable, and trustworthy implementation. This perspective examines the evolving landscape of AI-based healthcare tools, focusing on critical issues including algorithmic bias, accountability, data ownership, privacy, regulatory adaptation, human oversight, and the emergence of generative AI. Bias within training datasets may reinforce healthcare inequities by producing models that fail to generalize across diverse populations, potentially contributing to misdiagnosis and unequal care delivery. At the same time, the opacity of "black box" algorithms raises significant medico-legal concerns regarding liability and clinician responsibility in AI-assisted decision-making. Expanding use of wearable devices, digital biomarkers, and continuously evolving datasets further complicates questions surrounding informed consent, patient autonomy, and data governance. Regulatory systems also face challenges adapting to AI technologies that are dynamic and continuously updated rather than static medical devices. The article additionally highlights the importance of maintaining meaningful human oversight to prevent overreliance on automated recommendations and preserve clinical judgment. Emerging generative AI tools, including large language models and synthetic image generators, introduce further ethical concerns related to misinformation, transparency, and autonomous guidance. Ultimately, responsible AI integration in medicine requires interdisciplinary collaboration, evolving governance frameworks, and continuous ethical vigilance. AI should remain a supervised assistive technology designed to support, rather than replace, human expertise, transparency, accountability, and patient-centered care.

Health communication is central to prevention and care, yet generic messages frequently fail to achieve real behaviour change. Personalisation offers a way forward by aligning health information with individual characteristics. This chapter examines why one-size-fits-all approaches are limited, how tailored strategies can improve engagement and adherence, and what challenges must be addressed to ensure equity and trust. We also discuss how emerging technologies, including large language models, create new opportunities for scalable, evidence-based personalisation. Ultimately, we argue that personalised health communication is not a refinement but a strategic necessity for improving outcomes across healthcare systems.

Decision-making does not culminate when a choice is made. Instead, humans continue to evaluate their decisions through post-decisional reflection, and when needed, use new evidence to update the original decision. Post-decisional evaluation enables the formation of confidence in a decision, and detection and of revision errors. These processes belong to a class cognitive processes labelled metacognition-the capacity to monitor, evaluate, and regulate one's own cognitive processing. Understanding post-decisional dynamics is especially consequential in clinical decision making, where initial decisions are often made under uncertainty and time pressure, and may need to be revised as new information emerges. In this review, we synthesise theoretical, behavioural, and neuroscientific work on metacognition and post-decisional processing, focusing on evidence accumulation frameworks. We describe how classic decision models, such as drift-diffusion models, can be extended beyond the point of choice to account for confidence, error detection, and changes of mind. We discuss how confidence is dynamically updated by post-decisional information, underpinned by neural signatures of continued evidence accumulation and performance monitoring. We also examine how these processes are shaped by systematic biases, including a confidence-dependent confirmation bias. Finally, we highlight recent computational approaches that reveal different sources of metacognitive bias, including distortions in post-decisional accumulation and differences in how evidence is interpreted, offering new explanations of individual differences in belief updating. In clinical decision-making, such mechanisms may determine whether practitioners appropriately revise or persist with initial judgments, with important implications for diagnostic accuracy and patient outcomes.

Standard decision models are often resisted by clinical practitioners. This resistance can be well justified: Standard decision models can be opaque and complex, featuring overwhelming calculations, yet at the same time being simplistic, unable to handle the ill-defined structures or lack of information that mark medical settings. Fast-and-frugal heuristics are intuitive models of decision making that use few, obtainable pieces of information and combine them in simple ways. More specifically, fast-and-frugal heuristics rely on simple arithmetic and logic, such as summing some-say, no more than five-variables (e.g., the most important possible side effects of a medical treatment) and ordering these variables (e.g., judging which side effects are more important for most patients). One family of these heuristics that has been applied widely and with success to clinical practice is fast-and-frugal trees. This review uses examples to define, discuss, and show how to build fast-and-frugal trees, while providing literature pointers. The review also features two applications of fast-and-frugal trees for supporting decisions in fetal monitoring and assignment to intensive care. Finally, future theory and applications are discussed, with an emphasis on connections and challenges to building accurate and transparent AI for clinical decision making.

Pharmacists, while central to medication safety, face underestimated risks due to their daily exposure to toxic and explosive substances. In compounding pharmacies, handling carcinogenic, mutagenic, and reprotoxic (CMR) substances such as chemotherapy drugs, anesthetic gases, antibiotics, and hormones poses significant health hazards. These substances can cause environmental and secondary contamination through fine powder dispersion or volatile vapors, leading to indirect exposure even with standard protective gear. Explosive and fire risks also abound, especially with volatile solvents like ethanol, acetone, and ethyl ether. These substances are flammable and require careful storage and handling. Accidental chemical reactions, such as the mixing of acids and bases or contact between unstable compounds like picric acid and friction, can result in toxic gas release or explosions. To mitigate these dangers, pharmacies must implement strict preventive measures. This includes the use of fume hoods, appropriate personal protective equipment (PPE), and standardized protocols for hazardous drug preparation. Flammable substances must be stored in ventilated, ATEX-compliant cabinets, and chemicals clearly labeled with updated safety data sheets (SDS). Safe waste management and continuous staff training are also essential. Current challenges, such as drug shortages, force pharmacies to work with unfamiliar raw materials, increasing the risk of accidents. A notable example is potassium clavulanate, a fine powder prone to dust explosions if not carefully managed. Ultimately, ensuring pharmacist safety requires a culture of vigilance, continuous education, and adherence to safety protocols to manage both toxicological and explosive risks effectively.

As clinical information multiplies and patient cases become more intricate, even the most experienced doctors face the limits of solitary expertise. Collective intelligence-the idea that groups of people, and increasingly people together with AI, can make better decisions than individuals-offers a promising way to meet this challenge. We define collective intelligence in medicine and explain why healthcare is a uniquely important domain for its study, given the combination of high complexity and high stakes. We then distinguish different dimensions of collective medical decision-making, and review evidence for the benefits and limits of simple aggregation of judgments ("the wisdom of crowds") and interactive decision-making through deliberation. Finally, we highlight open questions and emerging directions, with particular focus on the expanding role of human-AI collaboration. We argue that the future of collective intelligence in medicine lies not only in classic hybrid intelligence where individual human judgement can be augmented by algorithmic decisions but in truly collaborative intelligence, where deliberation across multiple hybrid teams remain central to managing uncertainty.

Clinicians often prefer to use their clinical judgement instead of relying on statistical algorithms - a phenomenon also observed in other domains of decision making. This review explores this phenomenon using insights from the advice taking literature. One of the most consistent findings of this literature is egocentric advice discounting - the idea that people place more weight on their own judgement than on advice. Advice taking studies typically use the Judge-Advisor System (JAS) to measure how people weigh advice. In JAS studies, participants' estimates are elicited before and after advice is provided. The shift in estimates is often expressed numerically relative to the advice distance, i.e., the difference between initial estimates and advice, using the Weight-of-Advice index. Although advice taking studies have traditionally used lay participants making everyday judgements of low stakes, more recent research has used experts delivering judgements on domain-relevant problems. The review also discusses the popular concept of algorithm aversion, that is, preference for human over algorithmic advice due to intolerance of algorithmic error and comparatively greater acceptance of human error. This concept is not always borne out in research and studies have also observed preference for algorithmic advice. Insights for medical decision making are drawn, although the direct comparison of preference for human vs. algorithmic advice has not yet been investigated in clinicians. Ways to increase algorithmic advice uptake are discussed. Finally, the neglected idea that advice can be used to support learning is discussed with evidence from a range of studies. In an era rife with calls for machines to replace doctors, shifting emphasis from AI-bots to AI-tutors could promote uptake and reduce fears of professional deskilling and automation bias.

Human judgment is often prone to biases, and healthcare professionals are no exception. In clinical environments - characterized by high pressure, time constraints, and information overload - intuitive impressions can sometimes override statistical reasoning, leading to severe consequences such as diagnostic errors. There is an urgent need to identify effective strategies for reducing clinical decision-making errors and improving patient safety. In this paper, we examine the roots of logical fallacies and present promising debiasing interventions designed to mitigate such biases. Especially, we review debiasing training procedures that have been shown to significantly boost logico-mathematical reasoning, producing durable improvements across reasoning tasks and populations. Critically, these improvements occur as early as the initial intuitive stage, allowing faster and more accurate responses - which is particularly relevant to clinical contexts. We also discuss recent alternative procedures that help identify the conditions under which training is most effective. Overall, these findings suggest that short debiasing training can cultivate reliable intuitive judgments, offering a scalable, ecologically valid path to reducing medical decision-making errors.