To highlight the recent advancements in understanding the influence of psychological factors on the causation and management of obesity, which holds significance for clinical practice. This review explores developments in understanding psychological risk factors, sequelae, and treatments for obesity. Despite good evidence for psychological therapies in weight management, there are no standardized protocols for assessing patients requiring metabolic and bariatric surgery. Psychological therapies are synergistic with obesity medications. Obesity is a complex health issue with psychological dimensions. Stress, emotional dysregulation, and cognitive factors contribute to obesity. Stress's physiological impact on adipose tissue distribution and metabolic function, mediated by cortisol, demonstrates this interaction. Obesity leads to psychological consequences, including depression, low self-esteem, and reduced quality of life. The relationship between depression and obesity is modulated by demographic factors and biological mechanisms. Body composition reflects interactions between habits and cultural ideals, and medical models may increase stigma. Psychological interventions like cognitive-behavioral therapy and motivational interviewing effectively maintain weight loss. Psychological assessments before bariatric surgery are crucial for identifying mental health issues. This review highlights psychological dimensions in obesity prevention and treatment strategies.

The Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons questions are widely used by neurosurgical residents to prepare for written board examinations. Recently, these questions have also served as benchmarks for evaluating large language models' (LLMs) neurosurgical knowledge. LLMs show significant promise for transforming neurosurgical practice; however, they are susceptible to in-text distractions and confounding factors. Given the increasing use of generative artificial intelligence and ambient dictation technologies, clinical text is at a larger risk for the inclusion of extraneous details. The aim of this study was to assess the performance of state-of-the-art LLMs on neurosurgery board-like questions and to evaluate their robustness to the inclusion of distractor statements. A comprehensive evaluation was conducted using 28 state-of-the-art LLMs. These models were tested on 2904 neurosurgery board examination questions derived from the Congress of Neurological Surgeons Self-Assessment for Neurological Surgeons. In addition, the study introduced a distraction framework to assess the fragility of these models. The framework incorporated simple, irrelevant distractor statements containing polysemous words with clinical meanings used in nonclinical contexts to determine the extent to which such distractions degrade model performance on standard medical benchmarks. Six of the 28 tested LLMs achieved board-passing outcomes, with the top-performing models scoring over 15.7% above the passing threshold. When exposed to distractions, accuracy across various model architectures was significantly reduced-by as much as 20.4%-with 1 model failing that had previously passed. Both general-purpose and medical open-source models experienced greater performance declines compared with proprietary variants when subjected to the added distractors. While current LLMs demonstrate an impressive ability to answer neurosurgery board-like examination questions, their performance is markedly vulnerable to extraneous, distracting information. These findings underscore the critical need for developing novel mitigation strategies aimed at bolstering LLM resilience against in-text distractions, particularly for safe and effective clinical deployment.

Medical image classification models often lack validation across diverse datasets, limiting their generalization in clinical settings. This study evaluates and optimizes an enhanced DenseNet-121 model, integrating dilated convolutions and Squeeze-and-Excitation (SE) blocks, for multi-modal medical image classification. We assess its robustness across MRI, CT, and histopathology datasets, focusing on cross-domain and cross-modality performance. Experiments reveal strong in-domain results but significant degradation in cross-modality tasks (e.g., MRI-to-CT accuracy drops to ~0.5). To address this, we propose two strategies: (1) multi-modal joint training, which boosts cross-modality accuracy to 0.87, and (2) CycleGAN-based modality translation, improving performance to 0.7. Grad-CAM visualizations confirm the models focus on clinically relevant regions, enhancing interpretability. Findings highlight the superiority of multi-modal training while demonstrating CycleGANs utility when target-domain data is scarce. Future work should explore larger multi-center datasets and advanced domain adaptation to further improve robustness.

Synthetic data generation plays a crucial role in medical research by mitigating privacy concerns and enabling large-scale patient data analysis. This study presents a Graph Convolutional Neural Network combined with a Beta-Variational Autoencoder (GCN-β-VAE) framework for generating synthetic Abdominal Aortic Aneurysms (AAA). Using a small real-world dataset, our approach extracts key anatomical features and captures complex statistical relationships within a compact disentangled latent space. To address data limitations, low-impact data augmentation based on Procrustes analysis was employed, preserving anatomical integrity. The generation strategies, both deterministic and stochastic, manage to enhance data diversity while ensuring realism. Compared to PCA-based approaches, our model performs more robustly on unseen data by capturing complex, nonlinear anatomical variations. This enables more comprehensive clinical and statistical analyses than the original dataset alone. The resulting synthetic AAA dataset preserves patient privacy while providing a scalable foundation for medical research, device testing, and computational modeling.

Large language models (LLMs) have demonstrated impressive capabilities in medical domains, yet their ability to handle the specialized reasoning patterns required in clinical neurology warrants systematic evaluation. Neurological assessment presents distinctive challenges that combine anatomical localization, temporal pattern recognition, and nuanced symptom interpretation—cognitive processes that are specifically tested in board certification examinations. We developed a comprehensive benchmark comprising 305 questions from Israeli Board Certification Exams in Neurology and classified each along three dimensions of complexity: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs of varying architectures and specializations using this benchmark, testing base models, retrieval-augmented generation (RAG) enhancement, and a novel multi-agent system. Our analysis revealed significant performance variation across models and methodologies. The OpenAI-o1 model achieved the highest base performance (90.9% accuracy), while specialized medical models performed surprisingly poorly (52.9% for Meditron-70B). RAG enhancement provided variable benefits across models; substantial improvements for mid-tier models like GPT-4o (80.5% to 87.3%) and smaller models, but limited effectiveness on the highest complexity questions regardless of model size. In contrast, our multi-agent framework—which decomposes neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation—achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy compared to 69.5% for its base model, with particularly substantial gains on level 3 complexity questions across all dimensions. External validation on MedQA revealed dataset-specific RAG effects: while RAG improved board certification performance, it showed minimal benefit on MedQA questions (LLaMA 3.3-70B: + 1.4% vs + 3.9% on board exams), reflecting alignment between our specialized neurology textbook and board examination content rather than the broader medical knowledge required for MedQA. Most notably, the multi-agent approach transformed inconsistent subspecialty performance into remarkably uniform excellence, effectively addressing the neurological reasoning challenges that persisted even with RAG enhancement. We further validated our approach using an independent dataset comprising 155 neurological cases extracted from MedQA. The results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning offering promising directions for AI assistance in challenging clinical contexts.

Large language models (LLMs) have demonstrated impressive capabilities in medical domains, yet their ability to handle the specialized reasoning patterns required in clinical neurology warrants systematic evaluation. Neurological assessment presents distinctive challenges that combine anatomical localization, temporal pattern recognition, and nuanced symptom interpretation-cognitive processes that are specifically tested in board certification examinations. We developed a comprehensive benchmark comprising 305 questions from Israeli Board Certification Exams in Neurology and classified each along three dimensions of complexity: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs of varying architectures and specializations using this benchmark, testing base models, retrieval-augmented generation (RAG) enhancement, and a novel multi-agent system. Our analysis revealed significant performance variation across models and methodologies. The OpenAI-o1 model achieved the highest base performance (90.9% accuracy), while specialized medical models performed surprisingly poorly (52.9% for Meditron-70B). RAG enhancement provided variable benefits across models; substantial improvements for mid-tier models like GPT-4o (80.5% to 87.3%) and smaller models, but limited effectiveness on the highest complexity questions regardless of model size. In contrast, our multi-agent framework-which decomposes neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation-achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy compared to 69.5% for its base model, with particularly substantial gains on level 3 complexity questions across all dimensions. External validation on MedQA revealed dataset-specific RAG effects: while RAG improved board certification performance, it showed minimal benefit on MedQA questions (LLaMA 3.3-70B: + 1.4% vs + 3.9% on board exams), reflecting alignment between our specialized neurology textbook and board examination content rather than the broader medical knowledge required for MedQA. Most notably, the multi-agent approach transformed inconsistent subspecialty performance into remarkably uniform excellence, effectively addressing the neurological reasoning challenges that persisted even with RAG enhancement. We further validated our approach using an independent dataset comprising 155 neurological cases extracted from MedQA. The results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning offering promising directions for AI assistance in challenging clinical contexts.

Purpose This study explores service providers’ perspectives on factors limiting ART adherence among persons with disabilities living with HIV in Nigeria (PWDLWHIV). It argues that adherence is hindered by stigma, prejudiced views from service providers, and insufficient tailored programming for PWDLWHIV in the HIV response. Methods Thirteen service providers from six Nigerian states participated in in-depth interviews. Participants included medical doctors, nurses, clinical pharmacists, community development workers, and adherence counsellors. Interviews were analysed using Reflexive Thematic Analysis. Results Findings revealed unhelpful sociocultural beliefs about PWDLWHIV persist and are drivers of stigma that negatively impacts ART adherence. The lack of disability-disaggregated data has contributed to the non-recognition of PWDs as a key population. Consequently, there is limited attention to targeted programming and HIV funding for PWDs. Further, service providers held stigmatising attitudes towards PWDs, underpinned by the medical model of disability. These discriminatory views can affect the quality of ART services PWDLWHIV receive. Conclusion Achieving Nigeria’s UN HIV targets of 95-95-95 necessitates a stronger emphasis on supporting PWDLWHIV, who are currently not recognised as a key population in the HIV response. By addressing this gap, Nigeria can position itself as a global leader in accelerating progress in its HIV response.

BackgroundEarly intervention is critical for optimising outcomes in children with neurodevelopmental disorders. The International Classification of Functioning, Disability and Health (ICF) identifies the role of family, interventions through early detection and intervention services within environmental and personal factors. This study explored the perspectives of professionals regarding the implementation and effectiveness of family-centred early intervention within the Sri Lankan context and evaluated the applicability of the ICF model in this setting.MethodsA qualitative, phenomenological study was conducted with 30 professionals from healthcare and education settings. Participants were selected through purposive sampling based on their experience in child development. Data was collected through semistructured interviews, transcribed and analysed thematically to identify key themes and subthemes. The validity of the data was ensured through thick descriptions and member checking.ResultsAnalysis revealed that Sri Lankan early intervention practices show a positive trajectory, more emphasis is placed on the medical model and curative care, with less focus on preventive care and child activity and participation. Participants identified family-centred early intervention as a necessary and potentially effective approach for the Sri Lankan context. Key promoters included professional knowledge, environmental modifications and access to information, while socioeconomic barriers, limited resources and difficulties in implementing the ICF model effectively were challenges.ConclusionThe findings highlighted resources available within the Sri Lankan health and education system to introduce family-centred early intervention and identified families as a key resource as framed within the ICF model. Therefore, it is essential to adopt culturally sensitive methods to implement and sustain such programmes.

Foundation models, pre-trained on extensive datasets, have significantly advanced machine learning by providing robust and transferable embeddings applicable to various domains, including medical imaging diagnostics. This study evaluates the utility of embeddings derived from both general-purpose and medical domain-specific foundation models for training lightweight adapter models in multi-class radiography classification, focusing specifically on tube placement assessment and related findings, with comparison to the end-to-end training of an established convolutional neural network. A dataset comprising 8842 radiographs classified into seven distinct categories was employed to extract embeddings using seven foundation models: DenseNet121, BiomedCLIP, Med-Flamingo, MedImageInsight, MedSigLIP, Rad-DINO, and CXR-Foundation. Adapter models were subsequently trained using classical machine learning algorithms, including K-nearest neighbors (KNN), logistic regression (LR), support vector machines (SVM), random forest (RF), and multi-layer perceptron (MLP). Among these combinations, MedImageInsight embeddings paired with an SVM or MLP adapter yielded the highest mean area under the curve (mAUC) at 93.1%, followed closely by MedSigLIP with MLP (91.0%), Rad-DINO with SVM (90.7%), and CXR-Foundation with LR (88.6%), achieving a higher mAUC score than a fully finetuned convolutional neural network, DenseNet121 (87.2%). In comparison, BiomedCLIP and DenseNet121 exhibited moderate performance with SVM, obtaining mAUC scores of 82.8% and 81.1%, respectively, whereas Med-Flamingo delivered the lowest performance at 78.5% when combined with RF. Significant differences were found between each embedding model and MedImageInsight using the Wilcoxon signed-rank test at the significance level 0.05 (before Bonferroni correction). Notably, most adapter models demonstrated computational efficiency, achieving training within minutes and inference within seconds on CPU, underscoring their practicality for clinical applications. Furthermore, fairness analysis on adapters trained on MedImageInsight-derived embeddings indicated minimal disparities, with gender differences in performance within 1.8% and standard deviations across age groups not exceeding 1.4%. Further analysis indicated there is no significant difference across gender and age at a significance level of 0.05. These findings confirm that foundation model embeddings-especially those from MedImageInsight-facilitate accurate, computationally efficient, and equitable diagnostic classification using lightweight adapters for radiographic image analysis.

This study employs Latent Dirichlet Allocation (LDA) to analyze research trends in deep learning across engineering, natural sciences, and social sciences from 2020 to 2024. Using a corpus of 3,000 research paper titles, latent thematic structures were extracted to identify the major research directions within each field. The analysis uncovered four prominent topics per domain, revealing clear disciplinary differences in thematic emphasis and levels of methodological maturity. Engineering research predominantly addressed automation technologies, intelligent control systems, and real-time optimization. In contrast, natural science studies focused heavily on medical imaging, computational modeling, and data-driven scientific discovery. Social science research demonstrated an increasing integration of deep learning with ensemble modeling, prediction frameworks, and algorithmic decision processes. By offering a comparative view across disciplines, this study highlights both shared and divergent trajectories in deep learning research. The findings also suggest several future research directions, including the advancement of explainable AI techniques, the incorporation of multimodal data sources, and the development of domain-specific methodological adaptations to improve applicability and interpretability.

Cranial imaging diagnosis analyzes lesions in cranial CT images, automatically determining information such as lesion type, providing physicians with diagnostic recommendations, and saving valuable time during urgent medical events like stroke. Most current mainstream diagnostic generation methods rely on single-modal data for analysis, such as imaging reports or a single CT slice. However, relying on a single CT slice leads to loss of lesion information which across different CT slices. In addition, imaging reports often contain only coarse descriptions of abnormal regions without detailed pixel-level features, resulting in insufficient lesion characterization. Moreover, some works directly use 3D CNN to process entire 3D CT scans, but using only 3D CNN makes it challenging to accurately identify very small lesions. Therefore, this paper proposes a multimodal diagnostic model called MM-CD (MultiModal Craniocerebral Diagnose), which integrates imaging reports and cranial 3D CT findings for joint diagnosis. Specifically, this study first uses a 2D image pretrained model in combination with a vertical-dimension weight generation module for the cranial 3D CT images, enabling the model to focus on abnormal CT slices. Next, a multi-scale image fusion module is designed to effectively consolidates lesion descriptions from multiple CT slices into a single slice. Followed by a self-attention mechanism, this paper integrates CT information with the imaging report to construct a more comprehensive diagnostic reference. This clinically oriented design aims to lower missed-diagnosis rates for small, spatially sparse lesions and to shorten door-to-treatment intervals in acute care. Experimental results on a real clinical dataset show that the method improves overall accuracy by 1.65% compared to existing state-of-the-art medical multimodal models.

Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions.

Towards practical federated learning and evaluation for medical prediction models.

A universal medical imaging modality translation model in brain and head-and-neck radiotherapy

Role models are widely regarded as central to medical education, influencing professional identity, specialty choice, and socialization. Given the consistent emphasis on role models within medical education, it is imperative that assumed benefits are substantiated with empirical evidence. This study investigates medical students' role models and their associations with academic performance and sense of belonging, comparing these associations with those observed in non-medical students. We conducted a cross-sectional survey of 37,043 students at Utrecht University in Spring 2022 (final analytic sample: n = 3,474; medical students: n = 478). Students reported whether they had role models and how many. Sense of belonging was measured using the Sense of Belonging Scales (SOBS), assessing perceived faculty understanding, peer support, and classroom comfort. Academic performance was self-reported as GPA. Statistical analyses included chi-square tests, independent-samples t-tests, Mann-Whitney U tests, and Spearman correlations, with Bonferroni corrections for multiple comparisons. Among medical students, 54.4% identified one or more role models, a higher proportion than most other faculties. For medical students who had role models, the mean number of role models per student was 2.82 (SD = 1.94), which was higher than in most other faculties. Having role models and the number of role models were positively associated with all subconstructs of sense of belonging, particularly peer support. Associations with GPA were weak and did not remain significant after correction for multiple testing. Cross-disciplinary comparisons indicated that the prevalence and impact of role models vary by faculty: Veterinary Sciences showed similar role model prevalence to Medicine but no significant associations with outcomes, highlighting context-dependent effects. This study adds empirical weight to the prominent role of role models in medical education, as the majority of medical students reported having role models and these were positively associated with their sense of belonging, particularly in terms of peer support. Comparisons with other faculties show similar or even stronger associations, and other faculties show no significant relationships. Together, these results indicate that role models are likely context-dependent rather than exclusive to medicine. Future research should explore causal mechanisms, mediating factors, and longitudinal effects of role models across diverse academic contexts.

This study compared the performance of classical feature-based machine learning models (CMLs) and large language models (LLMs) in predicting COVID-19 mortality using high-dimensional tabular data from 9,134 patients across four hospitals. Seven CML models, including XGBoost and random forest (RF), were evaluated alongside eight LLMs, such as GPT-4 and Mistral-7b, which performed zero-shot classification on text-converted structured data. Additionally, Mistral-7b was fine-tuned using the QLoRA approach. XGBoost and RF demonstrated superior performance among CMLs, achieving F1 scores of 0.87 and 0.83 for internal and external validation, respectively. GPT-4 led the LLM category with an F1 score of 0.43, while fine-tuning Mistral-7b significantly improved its recall from 1% to 79%, yielding a stable F1 score of 0.74 during external validation. Although LLMs showed moderate performance in zero-shot classification, fine-tuning substantially enhanced their effectiveness, potentially bridging the gap with CML models. However, CMLs still outperformed LLMs in handling high-dimensional tabular data tasks. This study highlights the potential of both CMLs and fine-tuned LLMs in medical predictive modeling, while emphasizing the current superiority of CMLs for structured data analysis.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-26705-7.

The difficulties experienced in accessing cadavers worldwide are increasing the demand for technology-supported solutions. Virtual cadavers can ensure continuity of education by providing students with the opportunity to explore and examine realistic anatomical structures in detail without any geographical or physical restrictions. In this context, the development of alternative methods such as augmented reality (AR) or virtual reality (VR) tools is of great importance for students to continue their education. In the current study, the effectiveness of a medical education model enriched through metaverse-supported virtual cadavers was examined in order to contribute to the seamless learning experience of students. Consistent with the aim of the study, the effects of this model on students' academic achievement, attitudes towards the course, and academic motivation were evaluated. The study adopted a mixed methodology incorporating qualitative and quantitative techniques. An achievement test, anatomy attitude scale, and academic motivation scale were employed as data collection instrument in the study. In addition, interviews were conducted with the students in the experimental group to examine their experiences with virtual cadavers in the metaverse environment in depth. This study was conducted with the participation of 110 first-year medical students studying at a state university in Türkiye. Within the scope of the study, the students were divided into two groups as experimental and control groups. Although the control group followed the traditional anatomy curriculum, the experimental group performed some activities on virtual cadaver models in the metaverse environment in addition to the curriculum. The research results revealed that the academic achievement, motivation and attitude levels of the students in the experimental group increased more than those in the control group. The students emphasized that metaverse-supported virtual cadaver activities have the potential to increase students' course success, attitudes towards the course and motivation. The findings show that more comprehensive and in-depth research is needed on the potential effects of metaverse-supported virtual cadaver applications in education. This is an important step to increase the efficiency of metaverse applications in education.

BackgroundBRAF status is crucial for treating pediatric low-grade gliomas (pLGG) and can be assessed non-invasively from segmented tumor regions on MRI using machine learning (ML). However, there are limitations to manual and automated tumor segmentations. This study assessed the performance of automated segmentation algorithms and a segmentation-free ML classification pipeline.MethodsMolecularly characterized tumors and whole-brain FLAIR MR images were collected from 455 patients with pLGG treated between 1999 and 2023 at a children’s hospital. Three medical segmentation models, TransBTS, MedNeXt, and MedicalNet, were evaluated. Next, we developed a model to identify BRAF status from whole-brain FLAIR MRI, without any reliance on segmentations. We then implemented a novel pretraining regimen that embedded segmentation knowledge into the whole-brain FLAIR MRI classification model. Finally, we trained and evaluated a baseline model that used semiautomatic whole tumor volume segmentations as inputs.ResultsHere we show that the MedNeXt segmentation model (mean Dice score: 0.555) outperformed MedicalNet (0.516) and TransBTS (0.449) (p < 0.05 for all comparisons). The MedNeXt classification model achieved a one-vs-rest area under the ROC curve of 0.741 using the whole brain FLAIR sequence as an input, without any segmentation knowledge. This was improved to 0.772 through pretraining on the segmentation task, which was not significantly different from the baseline semiautomatic whole tumor volume segmentation-based model (0.756, p-value: 0.141).ConclusionsBRAF status can be assessed non-invasively using ML models based on whole-brain FLAIR sequences. Dependence on inconsistent manual or automated segmentations can be reduced by integrating tumor region information into the model through pretraining.

A prompt framework for enhancing LLM-based explainability of medical machine learning models: an intensive care unit application

In recent years, artificial intelligence (AI) has been actively implemented into various fields of medicine, trans-forming the diagnosis, treatment and prognosis of diseases. This article provides an overview of current re-search by foreign authors on the role of AI in healthcare. Key areas are considered, including medical image analysis, predictive modeling, personalized medicine, and automation of clinical processes. Among the re-viewed works is a study by F. Amisha, P. Malik, M. Patania and V.K. Rathaur, dedicated to the use of AI in the diagnosis of cancer, as well as the works of K. Kulikovski and Z. Ahmad and colleagues, analyzing the ethical and legal aspects of the introduction of AI. The contributions of D. Holzner et al. cover the use of machine learn-ing in epidemiological research, while F. Gama, D. Tiskbo et al. consider AI in the management of hospital sys-tems. Special attention is paid to the challenges associated with the integration of AI into clinical practice, in-cluding issues of trust, interpretability of algorithms and data protection. The article summarizes the current achievements and prospects for the development of AI in medicine, based on the opinions of leading experts in the field.