Stroke remains a major global public health challenge, representing the second leading cause of death worldwide and a primary contributor to long-term disability. The paradigm “time is brain” underscores the importance of treating stroke patients within the critical window period, ideally within 60 min from symptom onset, to minimize damage and improve outcomes. The integration of artificial intelligence (AI) into stroke imaging has transformed diagnosis and management by increasing speed, accuracy, and efficiency. AI algorithms have been trained to detect acute stroke, assess hemorrhage, detect and quantify midline shifts, calculate automated Alberta Stroke Program Early Computed Tomography Scores, and identify dense middle cerebral artery on non-contrast computed tomography (CT) as well as large vessel occlusions on CT angiograms, with high sensitivity and specificity. AI also aids in treatment guidance and outcome monitoring. This review provides insights into AI applications in acute stroke imaging, including its role in early detection, screening, triage and prioritization, automated image analysis, workflow optimization, and system integration. Despite its benefits, AI adoption faces challenges such as clinical validation, ethical considerations, and integration into existing workflows. Future developments depend on large, diverse, and well-annotated datasets to train more robust AI systems capable of guiding treatment strategies and improving patient outcomes. The seamless integration of cloud-based AI solutions with telereporting platforms has the potential to revolutionize stroke care by enabling rapid, high-quality radiologic interpretation, even in remote locations.

Artificial intelligence and machine learning are advancing rapidly in medical and mental health research, yet clinical publishing remains structurally unprepared to evaluate these technologies with the rigor they demand. Despite the rise of AI-driven models for suicide risk prediction and diagnostic assessment, editorial and peer review processes often lack the technical expertise required to assess methodological validity. Drawing on dual fluency in AI and clinical publishing, this perspective identifies a critical gap at the intersection of innovation and editorial oversight. This article reveals how editorial decisions in high-impact psychiatry journals have dismissed valid methodological concerns as “overly technical” and undermined independent scientific critique, drawing on two case studies: one involving a model that differentiates suicidal from non-suicidal self-harm, and another analyzing speech-based suicide risk assessment. These case studies serve as the foundation for a broader critique of editorial decision-making in clinical publishing, revealing persistent structural blind spots in evaluating AI-integrated research. To prevent the pre-mature adoption of flawed models in clinical care, this perspective proposes targeted reforms: recruiting technically proficient reviewers, mandating transparent methodological reporting, and protecting space for independent post-publication evaluation. Without such changes, the integrity of the field and the safety of patients remain at risk.

Early and accurate detection of Parkinson’s disease (PD) remains a crucial diagnostic challenge with substantial clinical implications, particularly for ensuring effective treatment and patient management. For instance, a group of subjects with scans without evidence of dopaminergic deficit (SWEDD) who are initially diagnosed as PD but exhibit normal single photon emission computed tomography (SPECT) scans. Over time, follow-up assessments often lead to a revised diagnosis of non-PD. In the meantime, these subjects may receive PD-specific medications that can cause more harm than benefit. In this paper, a case study is presented in which machine learning models are developed and trained on SPECT images to distinguish early PD from healthy controls, as well as to differentiate SWEDD cases from early PD. The case study utilizes a well-known, publicly available dataset and explores several machine learning classifiers, including support vector machines, logistic regression, feed forward neural networks, and convolutional neural networks (CNNs). The CNN model gave the best performance in differentiating PD from healthy subjects. All these models demonstrated strong potential for early differentiation of SWEDD cases from PD. These results suggest that the proposed approach could support clinicians in making more accurate and timely diagnostic decisions.

This research explores adapting vision transformers (ViTs) to classify neurodegenerative diseases while ensuring their decision-making process is interpretable. We developed a model to classify ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) positron emission tomography (PET) brain scans into three categories: cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer’s disease (AD). The dataset utilized in this research contains 580 samples of ¹⁸F-FDG PET scans obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The proposed model obtained an F1 score of 81% (macro-average of all classes) on the test dataset, a significant performance improvement compared to the literature. Furthermore, we combined the model’s attention maps with the Automated Anatomical Atlas 3 (AAL3), which represents a digital brain map, to identify the most influential areas on the model’s predictions and to conduct a regions’ importance study as a step toward explainability. We demonstrated that ViTs can achieve competitive performance compared to convolutional neural networks while enabling the development of explainable models without extra computations due to the attention mechanism.

In the contemporary context of the obesity epidemic and its associated comorbidities, early detection of individuals at risk is critical. Artificial intelligence and machine learning techniques offer substantial potential for automating obesity risk assessment, enabling early diagnosis and intervention. However, the development of robust predictive models is often hampered by limited or imbalanced datasets. Synthetic data generation has emerged as a key solution, allowing the expansion and balancing of data while preserving privacy. Recent surveys highlight that the synthetic minority oversampling technique (SMOTE) is a leading method for data generation in obesity detection. In line with this, our study analyzed the Estimation of Obesity Levels dataset, a dataset from the University of California, Irvine repository, focused on dietary habits and physical condition, which suffers from class imbalance. We compared three synthetic data generation approaches: SMOTE—nominal and continuous, variational autoencoders, and conditional tabular generative adversarial network. We trained multiple classifiers on the generated datasets and evaluated their performance. Classifiers trained on data including height and weight (i.e., body mass index [BMI]-related features) achieved F1-scores of up to 98.16%, as expected due to the direct role of BMI in obesity classification. Crucially, models trained without height and weight still achieved an F1-score of 74.48% when synthetic augmentation was used, demonstrating that useful obesity prediction models can be developed even in the absence of explicit anthropometric measures. These results indicate that synthetic data can enable accurate classification when key features are missing or when data are scarce.

The COVID-19 pandemic, one of the most disruptive global health crises in recent history, exposed critical vulnerabilities in existing healthcare infrastructure. Given the likelihood of future pandemics, it is essential to build a resilient, collaborative, synergistic, data-driven, and intelligent digital healthcare software. It should be meticulously designed and selectively curated to enhance early detection, rapid response, and efficient containment of outbreaks. In this article, we propose a federated learning (FL)-based health stack that prioritizes privacy while fostering collaborative intelligence among hospitals or client nodes. Our framework incorporates hierarchical FL, Byzantine-resilient information-theoretic FL (ByITFL), homomorphic encryption, and blockchain-based smart contracts to ensure secure collaboration among healthcare institutions without sharing raw data. Hierarchical FL leverages multilevel model aggregation to enhance model convergence, scalability, and resilience. ByITFL strengthens security by incorporating trust mechanisms and information-theoretic privacy scoring, while blockchain-based smart contracts ensure transparent, verifiable coordination among participating nodes. Furthermore, deep vulnerability detection using optimized averaged stochastic gradient descent weight-dropped long short-term memory models may further enhance the framework’s security, enabling threat identification during decentralized data exchanges. Experimental results show that the proposed hierarchical FL model achieves 94.23% accuracy on the modified National Institute of Standards and Technology dataset, outperforming federated averaging (92.66%) under the same environments. In addition, communication analysis proved that the overall transmission is minimized by collecting updates at local servers before sending them to central servers. Therefore, it is nearly a future-ready technology that can be implemented without many geopolitical issues, even in the case of hypersensitive global situations.

Healthcare services generate and store large quantities of data, requiring significant resources to manually analyze and gain meaningful insights. Recent advancements in automation tools—such as generative artificial intelligence (GenAI)—provide new opportunities to reduce human labor. This study explores the potential utilization of GenAI for a healthcare data analysis task—specifically, the conversion of clinical data from one diagnostic classification system to another (i.e., the Australian extension of the Systematized Nomenclature of Medicine Clinical Terms to the International Classification of Diseases, 10th Revision, Clinical Modification)—and examines the time and cost benefits of performing this using GenAI compared to a human rater. Conversions were completed using three methods: manual conversion using the National Library of Medicine’s I-MAGIC tool, ChatGPT-4o, and Claude 3.5 Sonnet. The accuracy of the GenAI tools was mapped against the manually extracted codes and examined in terms of a perfect, partial, or incorrect match. Task completion time was recorded and extrapolated to calculate and compare the cost associated with each method. When compared to the manually extracted codes, Claude 3.5 Sonnet yielded the highest level of agreement over ChatGPT-4o, whilst being the most time- and cost-effective. GenAI tools have greater utility than they have currently been given credit for. The automation of big data healthcare analytics, whilst still the domain of humans, is increasingly capable of being undertaken using automation tools with low barriers to entry. The further development of GenAI’s capabilities, alongside the capability of the healthcare system to use it appropriately, has the potential to result in significant resource savings.

Heart rate variability (HRV) is a critical non-invasive marker of autonomic nervous system regulation and plays an essential role in cardiovascular health. Individual differences in autonomic function necessitate the development of personalized health strategies. This study aimed to develop and validate a method that integrates principal component analysis (PCA) and K-means clustering to identify distinct patterns of autonomic regulation in healthy men using HRV data. A total of 80 young, healthy men (22.0 ± 2.8 years old, 65.2 ± 6.9 kg, and 171.0 ± 6.5 cm) were recruited, and their HRV data were analyzed using time-domain and frequency-domain parameters. PCA was applied to reduce the dimensionality of the HRV data, while K-means clustering was employed to identify distinct autonomic profiles. Silhouette index values were 0.397 for one cluster, 0.481 for two clusters, and 0.556 for three clusters, indicating that the three-cluster solution provided the best fit. Three statistically distinct and physiologically meaningful clusters were identified. Cluster 3 (n = 19) demonstrated significantly higher HRV parameters than cluster 1 (n = 33) and cluster 2 (n = 28) (p = 0.001). Post hoc analysis further confirms that cluster 1 differed significantly from both cluster 2 and cluster 3 (p = 0.001). Based on HRV characteristics, the clusters were characterized as “high vagal tone,” “intermediate vagal tone,” and “low vagal tone.” The “high vagal tone” cluster exhibited the strongest parasympathetic activity, while the “low vagal tone” cluster showed evidence of sympathetic predominance. This study demonstrates a robust approach for stratifying autonomic profiles, highlighting the potential of machine learning in advancing personalized cardiovascular health assessment.

The Segment Anything Model (SAM), originally built on a two-dimensional vision transformer, excels at capturing global patterns in two-dimensional natural images but faces challenges when applied to three-dimensional (3D) medical imaging modalities such as computed tomography and magnetic resonance imaging. These modalities require capturing spatial information in volumetric space for tasks such as organ segmentation and tumor quantification. To address this challenge, we introduce RefSAM3D, an adaptation of SAM for 3D medical imaging by incorporating a 3D image adapter and cross-modal reference prompt generation. Our approach modifies the visual encoder to handle 3D inputs and enhances the mask decoder for direct 3D mask generation. We also integrate textual prompts to improve segmentation accuracy and consistency in complex anatomical scenarios. By employing a hierarchical attention mechanism, our model effectively captures and integrates information across different scales. Extensive evaluations on multiple medical imaging datasets demonstrate that RefSAM3D outperforms state-of-the-art methods. Our work thus advances the application of SAM in accurately segmenting complex anatomical structures in medical imaging.

External ventricular drain (EVD) placement is a critical neurosurgical procedure traditionally performed freehand, with inherent risks of malposition, infection, and hemorrhage. Recent advances in artificial intelligence (AI), particularly in medical imaging and real-time computer vision, have enabled the development of portable navigation tools that may enhance accuracy, safety, and bedside accessibility. This study evaluated whether iOS devices equipped with a TrueDepth camera could perform real-time object and facial recognition, tracking, and semantic segmentation of computed tomography (CT) scans for non-immobilized heads to guide EVD placement via a custom AI-driven application. A custom iOS application was developed to provide a complete, real-time surgical navigation experience on an iPhone or iPad Pro. Three AI models were trained, tuned, and validated: a semantic segmentation model for brain anatomy, a semantic segmentation model for facial features, and an object detection model for a custom EVD stylet attachment. GPU programming accelerated on-device real-time, continuous registration while optimizing power consumption. A UNet convolutional neural network trained on eight 1 mm head CTs achieved 98.3% testing and 98.2% validation accuracy using a 50/50 test-validation split, segmenting a thin-cut CT in 3 s on an iPhone 12 Pro. Point cloud merging of patient anatomy took 4 seconds with an initial depth scan of 30,000 points, updating in real time with a cumulative error of 1 × 10^-8 cm. Transfer learning-powered EVD tracking, trained for 1,000 epochs, achieved an intersection over union of 1.0 and 0.98 for the detection model, with inference times of 800 μs on Apple’s Neural Engine. This feasibility study demonstrates that iOS devices with TrueDepth cameras can enable real-time, continuous surgical navigation for EVD stylets.

Medical residency training faces persistent challenges in delivering individualized learning experiences. While flipped classroom models promote engagement, they often lack real-time, personalized feedback. Artificial intelligence (AI)-driven platforms offer a promising solution by dynamically adapting content to residents’ evolving needs. This study evaluated the feasibility and effectiveness of integrating adaptive AI beings into a flipped classroom model for internal medicine residents. The AI-powered platform, edYOU, incorporated a personalized ingestion engine to customize learning content and an intelligent curation engine to ensure content integrity. Residents interacted with AI beings capable of adjusting real-time content delivery based on performance and progress. Learning outcomes were assessed using platform engagement metrics, simulation-based quiz results, and resident feedback. Among eligible residents, 92% actively used the platform, spending an average of 32.3 h (a few minutes to 148 h). A significant positive correlation was observed between time spent on the platform and quiz performance (r = 0.63, p<0.001), with 82.6% of educational topics engaged. Learners spent more time on difficult content areas, highlighting the system’s ability to adapt to individual challenges. Integrating AI into the flipped classroom proved feasible and was associated with improved engagement, learning efficiency, and academic performance. These results support using AI-enhanced educational tools to foster tailored, learner-centered experiences in graduate medical education. Further research is warranted to optimize implementation strategies and evaluate the long-term impact of AI-driven learning environments on resident development and competency outcomes.