Diabetes Mellitus Type 2 prevalence is still increasing around the world, and there is growing evidence of the role of environmental co-exposures to heavy metals-cadmium, lead, arsenic, mercury, and chromium-in glycaemia impairment. Contrarily to studies on single metal toxicology, real-life situations consist in exposure to mixtures of heavy metals through environmental, occupational and dietary pathways, especially in places like the Indo-Gangetic plain of India, characterized by the presence of geological arsenic pollution in association with metals exposure. This paper provides a narrative review on the current state of art in the domain of three key issues connected to risk stratification in Type 2 Diabetes Mellitus: methods for biomonitoring metals, inflammatory and molecular pathways involved in diabetes onset and progression related to metal exposures and statistical approaches for modeling joint effects of mixtures. Firstly, discuss on the biomonitoring methods, with particular emphasis to blood, urine and hair and nail as classical matrices and new emerging matrices (saliva, buccal cells). The analytical technique (AAS vs ICP-MS) is considered as well. It will focus on the heterogeneity between different methodologies, including matrix-specific exposure window and inter-laboratory variability. Regarding mechanistic aspects, there is an agreement in scientific literature on several molecular pathways involved in glycaemia alterations, including oxidative stress, NF-κB inflammation pathway, mitochondria dysfunction, epigenetics alteration and insulin signaling impairment, with possible synergistic rather than additive action of heavy metals in the impairment of beta-cell function and insulin resistance. Finally, it discusses some recent advances in quantifying joint metal effects on diabetes through specific statistics as WQS regression and Bayesian Kernel Machine regression (BKMR) and quantile g-computation. Even though important methodological advances have been made, a translational gap remains: no validated procedure currently exists that integrates exposure metrics, inflammatory biomarkers and clinical variables for individual risk stratification. We therefore outline a research-stage framework that could guide future development of such tools.

Accelerating magnetic resonance imaging (MRI) remains challenging, particularly under realistic acquisition noise. While diffusion models have recently shown promise for reconstructing undersampled MRI data, many approaches lack an explicit link to the underlying MRI physics, and their parameters are sensitive to measurement noise, limiting their reliability in practice. We introduce Implicit-MAP (ImMAP), a diffusion-based reconstruction framework that integrates the acquisition noise model directly into a maximum a posteriori (MAP) formulation. Specifically, we build on the stochastic ascent method of Kadkhodaie et al. and generalize it to handle MRI encoding operators and realistic measurement noise. Across both simulated and real noisy datasets, ImMAP consistently outperforms state-of-the-art deep learning (LPDSNet) and diffusion-based (DDS) methods. By clarifying the practical behavior and limitations of diffusion models under realistic noise conditions, ImMAP establishes a more reliable and interpretable baseline for diffusion-based accelerated MRI reconstruction.

Ultrasound imaging is a widely used diagnostic tool, and beamforming techniques are integral to its performance. Traditional methods such as delay-and-sum (DAS) and delay-multiply-and sum (DMAS) have limitations in terms of image quality and computational efficiency. Recently, deep learning approaches have shown significant promise in improving ultrasound beamforming, offering enhanced image resolution and quality, alongside potential for real-time processing. This review is aimed at providing a comprehensive overview of the current state-of-the-art in deep learning-based ultrasound beamforming. It focuses on analyzing various deep learning models, their applications, challenges, and the integration of hardware optimization strategies to enhance performance and efficiency. The review examines a range of deep learning models applied to ultrasound beamforming, including convolutional neural networks (CNNs), generative adversarial networks (GANs), transformer-based models, and hybrid approaches. It also discusses the challenges associated with dataset limitations, model interpretability, and the risk of overfitting. Furthermore, the review explores hardware acceleration using field-programmable gate arrays (FPGAs), along with cloud-edge frameworks for real-time inference. The findings highlight the potential of deep learning models to outperform traditional methods in terms of image quality and resolution. However, challenges such as the need for large and diverse datasets, the black box nature of deep learning models, and the risk of overfitting remain. Hardware optimization through FPGAs has proven effective in enabling real-time processing, but the integration of edge computing with cloud-based solutions offers promising avenues for balancing performance, latency, and energy efficiency. Deep learning-based ultrasound beamforming has great potential to advance medical imaging. Future efforts should enhance model generalizability with diverse and synthetic data, optimize hardware for real-time use, and establish standardized validation protocols to support clinical adoption.

Globally, significant progress toward ending the HIV/AIDS epidemic includes advances in HIV treatment, converting a fatal infection to a chronic disease. Adolescents living with HIV (AWHIV) continue to underachievein HIV viral suppression (VS) needed to remain healthy. This cross-sectional study aimed to examine the role of self-management and related factors in VS among AWHIV aged 15-19 years in Dar es Salaam, Tanzania. A total of 658 AWHIV from ten large HIV care and treatment sites were assessed on HIV-related self-management, depression, HIV stigma, physical quality of life and social support. HIV viral load was abstracted from clinical records. Generalized linear models were used to examine unadjusted and adjusted mean differences in self-management scores between the two groups. Over half of the AWHIV, 403(78%), had achieved VS. Overall, sub-scale self-management scores did not differ significantly between groups (p > .05); however, three self-management items significantly differed by VS status: health self-efficacy (p = 0.049), life ambitions/goals (p = 0.031) and ART administration knowledge (p = 0.044). Adjusted models attenuated self-efficacy and ambition differences while ART knowledge differences persisted (p = 0.005). Physical quality of life, social support and stigma significantly moderated relationships between self-management items and VS. These findings suggest that focusing on targeted self-management behaviors may provide a more accurate understanding of their influence on VS than aggregated self-management scores.

Ultrafast plane-wave imaging (PWI) trades image quality for frame rate under sparse acquisitions. Existing deep learning-based PWI reconstruction methods are limited by the lack of physics-guided degradation modeling, leading to distortions in anatomical structures and speckle statistics. In addition, an inherent low-frequency spectral bias attenuates high-frequency components, further exacerbating detail loss and amplifying reconstruction artifacts. In this study, we propose DiffPWI, a degradation-aware reconstruction framework. DiffPWI employs a latent diffusion model (LDM) to learn the distribution of physically informed degradations in a compressed latent space, providing a physics-informed prior for reconstruction. Furthermore, we introduce a frequency decoupling attention module (FDAM) that adaptively decomposes feature representations into hierarchical frequency sub-bands, enabling preservation of high-frequency anatomical details while suppressing acquisition artifacts. Extensive evaluations on phantom, multi-organ in vivo, and zero-shot datasets demonstrate that DiffPWI achieves improved spatial resolution, contrast recovery, and speckle fidelity compared with state-of-the-art methods, advancing high-fidelity ultrafast ultrasound imaging.

Patient-generated text in internet hospitals represents a massive but unstructured source of clinical demand. Misalignment between rigid administrative menus and actual patient intent leads to triage inefficiency and resource mismatches. We developed a hybrid artificial intelligence framework to decode patient needs and optimize service allocation using real-world data from a tertiary internet hospital, with external validation on an independent multihospital dataset (N = 497). In this study, we analyzed 28,851 inquiries from a tertiary internet hospital using a 2-stage approach: (a) unsupervised K-Means clustering to discover intrinsic intent taxonomies, and (b) supervised classification using a fine-tuned MacBERT model. We benchmarked this specialized small model against traditional baselines and state-of-the-art large language models (LLMs; Qwen2.5 and HuatuoGPT-II) on a gold-standard dataset of 4,814 annotated records. Unsupervised analysis identified 8 distinct patient intents, revealing substantial semantic overlap between rehabilitation and follow-up requests. The fine-tuned MacBERT model achieved a weighted F1-score of 0.680, outperforming both the medical-specific LLM (F1 = 0.595) and the general LLM (F1 = 0.585). External validation on an independent nationwide platform dataset (N = 497) confirmed cross-platform transferability, with the hybrid learning framework achieving a weighted F1 of 0.578 versus 0.389 (Qwen2.5-7B) and 0.190 (HuatuoGPT-II). Departmental analysis exposed distinct "demand fingerprints": 39% of Neurology inquiries focused on prescription renewals, whereas 68% of Dermatology inquiries involved initial symptom consultations. These findings indicate that, within the studied Chinese hospital setting, fine-tuned task-specific models may offer superior utility over general LLMs for clinical triage, with the performance gap widening under domain shift-though multicenter prospective validation is required to establish generalizability. Mapping intent fingerprints across specialties provides a quantifiable basis for restructuring digital hospital operations.

Deep vein thrombosis (DVT) is characterized by the formation of blood clots in the deep veins, most commonly in the lower limbs, obstructing blood flow. In about 50% of cases, the clot dislodges and travels to the lungs, causing potentially fatal pulmonary embolism. Diagnosing DVT is challenging, as up to half of cases are asymptomatic. Early and accurate diagnosis is critical, but the current state of the art in ultrasound and plethysmography is insufficient for continuous, operator-independent, point-of-care monitoring. This paper introduces the concept and advances of the ThrombUS+ Horizon Europe project, which aims to innovate toward modular and wearable system for continuous, point-of-care DVT monitoring. The system leverages a multimodal approach, incorporating compression ultrasonography, electrical impedance plethysmography, and light reflection rheography, by re-engineering these clinical technologies for seamless integration into a wearable architecture. Manual compression of an ultrasound transducer is replaced by a sophisticated actuator, enabling precise pressure control via a piezoelectric micropump with feedback signals. The same actuator supports plethysmography modules, providing additional data for DVT assessment. These inputs, combined with physical movement monitoring, are processed by an artificial intelligence-driven central intelligence unit for DVT detection, risk assessment, and prevention. Preliminary laboratory testing of the individual modules highlights the system's potential for autonomous DVT monitoring and detection. The ThrombUS+ uses big datasets for machine learning training collected in the project via 3 large-scale clinical studies. The entire system will be validated in the clinical setting via an early feasibility study and a multicenter clinical trial.

Accurate prediction of drug-target interaction (DTI) facilitates the translation of lead compounds into clinical therapeutics. Despite progress in deep learning, two challenges remain: the representation gap between heterogeneous drug and protein modalities and insufficient use of complementary knowledge from pretrained encoders. To address these issues, we propose SF-DTI, a unified framework that integrates multi-encoder protein fusion, structure-semantic drug alignment, and progressive cross-modal refinement for DTI prediction. Specifically, SF-DTI captures global protein semantics from multiple pretrained protein language models, aligns graph-based structural features with pretrained semantic embeddings for drugs under subspace constraints, and progressively refines drug-protein interaction representations through cascaded cross-modal signal propagation. Experiments on six benchmark datasets under standard, cold-start, and cross-domain settings demonstrate that SF-DTI achieves competitive performance relative to seven state-of-the-art baselines across evaluation scenarios. These results suggest that SF-DTI is a promising framework for DTI prediction and may support virtual screening, drug repositioning, and drug discovery.