Misdeployment of balloon-expandable covered stents extending from the common iliac artery (CIA) to the internal iliac artery (IIA) is a rare but serious complication. Optimal bailout strategies remain limited. Herein, we report a case of endovascular bailout for inadvertent stent deployment from the CIA to the IIA. A 68-year-old man presented with severe claudication of both legs. The bilateral ankle-brachial index was low, and computed tomography showed chronic total occlusion of the left CIA. During revascularization, a balloon-expandable covered stent was inadvertently deployed from the CIA into the IIA. To salvage the external iliac artery (EIA), a modified-tip 0.035-inch guidewire was used to perforate the stent graft, creating a fenestration. A 0.014-inch guidewire was advanced through the newly created opening, followed by deployment of a self-expanding bare-nitinol stent across the fenestration to reconstruct the EIA flow. The final angiography indicated successful revascularization without complications. Therefore, percutaneous fenestration of a balloon-expandable covered stent is a potential therapeutic option for the treatment of aortoiliac lesions.
In scientific studies of human-AI interaction dynamics, researchers often need to present participants with opportunities to interact with live large-language models (LLMs). However, technical and practical challenges (from survey platform limitations and logging chat data to manipulating AI behaviors for experimental designs) often inhibit survey-based deployment of AI stimuli. We developed DiSCoKit-an open-source toolkit for deploying live LLM experiences (e.g., ones based on models delivered through Microsoft Azure portal) through JavaScript-enabled survey platforms (e.g., Qualtrics). We describe the toolkit's scientific motivation, architecture, and operation. We also offer an example of toolkit deployment and customization, along with discussing its possibilities and limitations. Altogether, DiSCoKit gives researchers a flexible, secure, scalable solution for deploying naturalistic LLM stimulus experiences through online surveys.
Immunization with antigens presented multivalently on virus-like particles elicits potent immune responses. The SpyTag/SpyCatcher system, which allows spontaneous covalent bond formation between two proteins, provides a simple route to functionalize virus-like particles for multivalent presentation. Termed SpyVLPs, these reactive nanoparticles allow modular plug-and-display of different antigens. SpyCatcher003-mi3 is a resilient SpyVLP that can be produced at high yield, allows efficient coupling, and can be applied to vaccines against a diverse range of pathogens, as well as against cancer and allergy. Here, we describe the procedure for the expression and purification of SpyCatcher003-mi3. We outline considerations related to antigen design, multivalent display, nanoparticle validation, and immunogen preparation that are critical for effective implementation of SpyCatcher003-mi3 and other SpyVLPs. The methods discussed are applicable beyond vaccines, extending to therapeutic targeting, diagnostics, and catalysis.
Fast and accurate industrial Load Station (LS) inspection is crucial in manufacturing environments. This study addresses the challenges of deploying a deep learning-based solution for LS inspection in semiconductor wafer handling, where the Load Station must be properly aligned and occlusion-free before pin pack assembly operations. A significant challenge in industrial settings is data scarcity, as collecting and annotating large amounts of datasets is often impractical due to operational constraints and the rarity of abnormal conditions. This study examines how training data size affects inspection reliability in a real-world smart manufacturing context. The experiments utilise YOLOv5 and YOLOv8 variants across five training set sizes to determine minimum data requirements for reliable deployment, evaluated under 5-fold cross-validation with multiple random seeds. Our results demonstrate that YOLOv8 achieves superior data efficiency, with only 40 samples per class (T-40), YOLOv8n achieves 0.981 ± 0.014 inspection accuracy and 0.842 ± 0.050 mean Average Precision (mAP@0.5) on a held-out real test set, while YOLOv5 requires substantially more training data to achieve comparable performance. Smaller variants (YOLOv8n, YOLOv8s) consistently outperform larger models in this data-scarce environment. An ablation study confirms that combining real and synthetic obstruction samples is essential, and the approach is validated on an operational semiconductor manufacturing dataset, providing practical, statistically grounded recommendations for deploying deep learning inspection systems when training data are limited.
Computer-integrated surgical navigation systems are often run in the OR with the assistance of a technician for controlling the user interface and advising on technical details of the system. In lower-resource healthcare settings, limited access to additional OR staff and technical training for operating navigation systems can represent a barrier to sustainably deploying a low-cost surgical navigation solution. Recent advancements in locally deployable large language models (LLMs) have improved their ability to answer technical questions based on source materials and safely perform limited tasks on behalf of a user. The objective of this paper is to explore the feasibility of using a network of local LLM-based agents to act as a natural language interface for a low-cost surgical navigation system, facilitating hands-free manipulation of the user interface and providing documentation-grounded technical guidance. We propose the navigation offline virtual agent (NOVA), an end-to-end architecture that integrates distinct local LLM-based agents for knowledge tasks, action tasks, and delegation between the agents. Two semisynthetic benchmark datasets were generated for ablation studies of individual agents, and a prototype was built which integrates these agents into NousNav, an open-source neuronavigation system. The agents based on local LLMs were found to perform comparably to closed-source, commercially hosted LLMs. In a user study with nine participants, NOVA facilitated hands-free patient registration with a mean end-to-end latency of 10.4 ± 2.4  s and an 81% command success rate. This work demonstrates that local LLM-based agents can be deployed to support users of low-cost navigation systems by providing a context-aware natural language interface, representing an important step toward reducing technician dependence in low-cost surgical navigation.
Radiology artificial intelligence (AI) is increasingly developed on large external datasets and deployed across institutions, but real-world model performance may vary substantially after implementation. Imaging AI interacts with a local ecosystem shaped by scanner hardware, acquisition protocols, reconstruction methods, technologist practices, disease prevalence, patient demographics, reporting conventions, and clinical workflow. These factors can produce domain shift, degrade calibration, alter false-positive and false-negative patterns, and affect clinical utility. In this article, we argue that radiology AI should move beyond a "plug-and-play" deployment paradigm toward institution-calibrated AI stewardship. We propose an institution-specific radiology AI performance profile, conceptually analogous to a local antibiogram, to summarize how AI tools perform within a specific clinical environment. Unlike prior MLOps and radiology AI governance frameworks, the proposed profile translates lifecycle management into a radiology-specific, locally maintainable artifact that captures technical context, clinical context, performance metrics, reference standards, equity checks, workflow effects, and governance triggers. The framework applies not only to diagnostic decision support, but also to research cohort generation and radiology education, where local reporting language, imaging protocols, and case mix may strongly influence model reliability. We emphasize that institution-specific calibration does not require every hospital to develop AI models de novo. Rather, externally developed, vendor-based, and foundation-model approaches should be paired with local validation, cautious threshold adjustment, calibration when applicable, surveillance for drift and protocol changes, and multidisciplinary governance. Responsible radiology AI deployment should therefore ask not only whether a model is accurate, but whether it remains accurate, relevant, equitable, and useful locally over time.
Healthcare systems increasingly depend on interoperable electronic health records, picture archiving and communication systems, cloud analytics, remote monitoring, and connected medical devices. These workflows create a long-lived confidentiality problem: clinical, genomic, pediatric, psychiatric, and family-linked data may remain sensitive for decades, whereas the public-key algorithms that protect current transport and identity systems are vulnerable to future cryptographically relevant quantum computers. Shor's algorithm threatens RSA, finite-field Diffie-Hellman, elliptic-curve Diffie-Hellman, and elliptic-curve signatures, while Grover's algorithm reduces the effective security margin of symmetric primitives. This review validates the post-quantum transition pathway for healthcare and distinguishes software-deployable post-quantum cryptography from hardware-specialized quantum key distribution. We summarize the standardized NIST primitives, correct the status of Falcon/FN-DSA as a FIPS 206 algorithm still in development, analyze deployment constraints in Internet of Medical Things environments, and map cryptographic migration to healthcare governance. A staged, hybrid, crypto-agile migration strategy is recommended: prioritize long-confidentiality data, inventory cryptographic dependencies, deploy hybrid key establishment at enterprise gateways, reserve QKD for selected high-value fixed links, and require vendor update pathways for clinical devices.
Speckle patterns in ultrasound images often obscure anatomical details, leading to diagnostic uncertainty. Recently, various deep learning-based techniques have been introduced to effectively suppress speckle; however, their high computational costs pose challenges for low-resource devices, such as portable ultrasound systems. To address this issue, we introduce Edge Speckle Reduction and Image Enhancement (EdgeSRIE), a lightweight hybrid deep learning framework for real-time speckle reduction and image enhancement in portable ultrasound imaging. The proposed framework consists of an unsupervised despeckling branch and a self-supervised deblurring branch trained separately in sequence with AdamW (learning rate 1 × 10- 4) and an L2 loss. The integrated model was converted to an 8-bit deployment model using post-training quantization on a low-resource system-on-chip. Training used 1779 B-mode images (BUSI and HC18), validation used 77 raw acquisitions (PICMUS and CUBDL), and external evaluation used 94 EdgeFlow UH-10 bladder images. EdgeSRIE was compared with OSRAD, OBNLM, DIAE, DUNet, BRUNet, and USNet using contrast-to-noise ratio (CNR), speckle signal-to-noise ratio (SSNR), average gradient magnitude (AGM), and structural similarity index measure (SSIM); paired inference used two-sided Wilcoxon signed-rank tests with Holm correction and rank-biserial and Hedges' g effect sizes. Across the four representative cases, post-training-quantized EdgeSRIE achieved the largest mean improvement in CNR (64.9 ± 21.0%) and the smallest mean decrease in AGM (-16.6 ± 33.2%). SSNR (109.8 ± 45.8%) was in the OBNLM-led range, whereas SSIM remained slightly lower than OSRAD and OBNLM on its native [0, 1] scale. The deployed model contained 17.67K parameters and reached 64.10 frames/s on the target hardware. In the case-matched analysis (n = 8), EdgeSRIE showed Holm-corrected significant advantages over all six baselines in CNR, over five of six in SSNR, and over four of six in AGM. For SSIM, EdgeSRIE was significantly higher than the four deep learning baselines but lower than OSRAD and OBNLM. Across the statistically significant comparisons, the magnitude-based effect sizes were large (Hedges' g ≥ 0.8). These results support the feasibility of EdgeSRIE as a compact, deployment-oriented framework that balances speckle suppression, structural preservation, and real-time execution for portable ultrasound imaging.
Orchestration of infrastructure and data resources for research on sensitive data in Trusted Research Environments (TREs) is complex and time-consuming, with fragmented processes for data governance, identity management, disclosure checks, and resource provisioning. Research Object Crates (RO-Crates) aim to make metadata and analytical outputs FAIR (findable, accessible, interoperable, reusable), but current approaches neither record infrastructure requirements in RO-Crates nor use them for automated provisioning. CR8TOR addresses this challenge by extending standardised metadata models (Five-Safes RO-Crate) to take an active role in creating project-dependent infrastructure. Built on the Kubernetes Operator pattern within K8TRE (a Kubernetes-native, cloud-agnostic TRE implementation), CR8TOR demonstrates how contextual project information modelled in RO-Crate-like schemas can provide machine-readable representations of research resource requirements. By extending the Kubernetes API with custom resource definitions (CRDs), these resource descriptions sit alongside traditional research object metadata while maintaining FAIR principles, acting as computable units that can be recreated when required. CR8TOR supports automated data ingress from external sources, project governance controls, data provenance tracking, and on-demand reconciliation of project resources including analytics workspaces, user accounts, and access/network policies. Deployed across cloud and on-premises infrastructure supporting multi-institution collaborations, the same declarative models can be shared across deployments, enabling reproducible and portable research settings. This work demonstrates how provisionable metadata may strengthen methodological foundations of federated research by allowing TRE infrastructure specifications-including workspace configuration, controlled access rules, and project-specific identities-to be described, shared, and reproduced with clarity, consistency, and automation, ultimately lowering operational costs while improving governance and reproducibility.
Velopharyngeal dysfunction (VPD) is an impaired ability to achieve adequate velopharyngeal closure during speech, often resulting in hypernasality and reduced intelligibility. VPD screening and diagnosis require specialized expertise and controlled recording conditions, limiting scalable access outside high-income countries.Key challenge: Speech-based machine learning models can perform extremely well under standardized clinical recording conditions. However, performance often deteriorates when deployed on consumer devices (e.g., phones or tablets) and in uncontrolled acoustic environments. This degradation is largely driven by domain shift arising from differences in recording conditions (e.g., device and channel characteristics, background noise, and room acoustics), which can cause models to rely on spurious recording artifacts rather than pathology-relevant cues. This study introduces a two-stage framework to improve robustness under realistic recording scenarios. Nasality representation pre-training employs a nasality-focused representation via supervised contrastive learning (SupCon) using an auxiliary dataset with phoneme alignments to form oral-context versus nasal-context supervision. During Frozen-encoder VPD screening, the encoder is frozen to perform VPD screening using lightweight classifiers on 0.5-second chunks with probability aggregation to produce recording-level decisions using a fixed decision threshold. Here, in-domain refers to standardized clinical recordings used for model development, and out-of-domain refers to heterogeneous public Internet recordings collected under uncontrolled conditions and evaluated without any adaptation. The proposed method is then compared against prior-study baselines, including MFCC features and large pretrained speech representations, using the same evaluation protocol. On the primary in-domain subject-disjoint held-out split of 82 subjects (60 train/22 test; 345 training recordings; 131 test recordings; multiple recordings per subject), the proposed approach reached ceiling recording-level screening performance under this standardized clinical protocol (macro-F1 = 1.000, accuracy = 1.000). To assess sensitivity to this fixed split, an additional subject-level nested 5-fold cross-validation analysis was performed on the full in-domain cohort (82 subjects, 476 recordings), with the encoder frozen and only the second-stage classifiers retrained; the best mean performance was obtained with SVM (macro-F1 = 0.981 ± 0.022, accuracy = 0.985 ± 0.016). On a separate out-of-domain set of 131 public Internet recordings, large pretrained speech representations degrade substantially, and MFCC is the strongest baseline (macro-F1 = 0.612, accuracy = 0.641). The proposed method achieves the best overall out-of-domain performance (macro-F1 = 0.679, accuracy = 0.695), improving over the strongest baseline by +0.067 macro-F1 and +0.054 accuracy (point-estimate improvements) under the same evaluation protocol and fixed threshold. Learning a nasality-focused representation prior to clinical classification can reduce sensitivity to recording artifacts and improve robustness when moving from the laboratory to real-world audio recording scenarios. This design supports practical deployment of VPD screening and motivates domain-robust evaluation protocols for deployable speech-based digital health tools.
Standard Retrieval-Augmented Generation (RAG) systems only use semantic similarity to retrieve information, and since this method is quite limiting, it may find clinically irrelevant evidence and produce outputs that are unsafe or hallucinated. This drawback is particularly important in respiratory care, where the diagnosis relies heavily on very accurate physiological indicators such as spirometry patterns and symptom profiles. We propose MMCRAG-Resp., a clinically grounded, physiology-aware corrective RAG framework for respiratory intelligence. The system harmonizes 17,516 patient records from three heterogeneous public datasets (Respiratory Sound Database, NHANES spirometry, and clinical guidelines) through a unified respiratory ontology. Multi-modal embeddings (acoustic, spirometric, and textual) are fused into 144-dimensional vectors and indexed in FAISS. A novel Respiratory Relevance Score (RRS) combining spirometric pattern agreement (weight 0.45), symptom overlap (0.30), and embedding similarity (0.25) gates evidence quality before LLM consumption via a three-path correction policy. A locally deployed small language model (Ollama; llama3.2, Mistral-7B-Instruct) generates evidence-constrained, fully traceable clinical explanations without GPU requirements. Evaluated on 200 held-out test queries drawn from the harmonized corpus, MMCRAG-Resp achieved a Clinical Precision@10 of 0.81, a Hallucination Rate of 0.11 (71% relative reduction over Vanilla RAG), a Spirometry Alignment Score of 0.79, and an Evidence Citation Rate of 0.87. Ablation studies confirmed that spirometry-first scoring contributes the largest single-component gain. All pairwise comparisons with baselines were statistically significant (p < 0.001, McNemar's test; Cohen's d = 1.42). MMCRAG-Resp is, to the best of our knowledge, the first system combining multi-modal respiratory evidence retrieval, physiologically-grounded corrective RAG, and privacy-preserving local LLM inference in a single deployable clinical reasoning framework. The approach advances safe, explainable AI-assisted respiratory decision support.
Continuous glucose monitoring (CGM), initially recommended primarily for people with type 1 diabetes, has rapidly expanded to patients with type 2 diabetes and, more recently, even to individuals without diabetes. This broadening use underscores the importance of critically examining the evidence supporting CGM across populations, clarifying where benefits are established, where they are modest, and where they remain uncertain. We reviewed systematic reviews and meta-analyses of randomized trials and clinical guidelines evaluating CGM use beyond type 1 diabetes, with particular attention to glycemic outcomes, hypoglycemia, patient-reported outcomes, and potential unintended consequences in nonpregnant adults. In adults with type 2 diabetes, evidence from randomized trials consistently demonstrated that CGM use was associated, on average, with a modest yet consistent reduction in hemoglobin A1c of approximately 0.3% compared with finger-stick monitoring or usual care; some patients benefited considerably more. Only limited and indirect evidence supported the adoption of CGM in people with type 2 diabetes not receiving glucose-lowering therapy or in those with prediabetes or obesity. CGM should be deployed in response to a specific patient problem rather than as a default intervention. Deliberate, equitable deployment of CGM that is aligned with patient goals, clinical context, and system capacity will be necessary to ensure that those most likely to benefit are not left behind, while avoiding overuse in populations for whom benefit remains unproven.
Recent years have witnessed a surge in FDA approved AI tools for healthcare applications. While this growth offers considerable potential benefits for clinical practice, it also introduces substantial challenges related to ethics, regulation, and patient safety. These challenges are further compounded by previously documented gaps in the regulatory approval pathway. These gaps include inconsistent pre-market evaluation practices, over-reliance on retrospective studies, and the limited systematic post-market surveillance of AI devices in real-world clinical settings. Using publicly available FDA data, we developed therefore an interactive web-based dashboard for assessing and predicting the performance of FDA-approved AI software, called PROACTIVE-AI, for the purpose of pro-viding the user with a structured guidance on the anticipated performance of AI-enabled medical devices in real-world clinical settings. The dashboard supports exploratory analysis by diverse stakeholders via knowledge graph visualization and longitudinal trend monitoring of performance indicators, including device recalls and safety-related issues. In addition, PROACTIVE-AI incorporates an AI-aided post-market surveillance risk assessment calculator, derived from historical recall data, to identify device characteristics and con-textual factors associated with elevated deployment risk. Our findings using the PROACTIVE-AI dashboard highlight some of the important challenges related to real-world monitoring and accountability of deployed AI medical devices. Furthermore, it illustrates the potential value of such dashboard in narrowing the trust gap surrounding AI in healthcare by providing quantitative metrics of expected clinical performance and recall-related risk factors.
Type II endoleak (T2EL) remains the most common endoleak after endovascular aneurysm repair. We evaluated whether the cuff-first technique (CFT), in which an aortic cuff is deployed at the branch-vessel level before main-body deployment, is associated with lower early T2EL rates. We performed a single-center retrospective comparative study of infrarenal endovascular aneurysm repair (October 2020-December 2024). Patients without contrast-enhanced computed tomography at 6 months and other predefined exclusions were excluded. The primary endpoint was T2EL on 6-month computed tomography (early and delayed phases). Secondary endpoints included aneurysm sac regression ≥5 mm at 6 months, reintervention for T2EL, operative time, contrast volume, and radiation exposure (air kerma). Freedom from reintervention was analyzed by Kaplan-Meier methods. We analyzed 100 patients (cuff-first, n = 33; noncuff-first, n = 67). Inferior mesenteric artery patency (93.9% vs 74.6%; P = .028) and diameter (2.4 ± 1.0 vs 1.8 ± 1.3 mm; P = .039) were higher in the cuff-first group, whereas lumbar artery counts were similar (5.5 ± 1.8 vs 5.4 ± 1.8; P = .827). At 6 months, T2EL was less frequent with the CFT (9.1% vs 31.3%; P = .014), and aneurysm sac regression ≥5 mm was more frequent (24.2% vs 7.5%; P = .027). Operative time, contrast volume, and radiation exposure did not differ significantly. Reintervention for T2EL occurred in one cuff-first patient and seven noncuff-first patients; freedom from reintervention at 36 months was 94.7% vs 85.5% (log-rank P = .251). The CFT was associated with lower 6-month T2EL rates and higher early sac regression without increasing procedural burden. Larger studies with longer follow-up are warranted.
The exponential adoption of artificial intelligence (AI), worsening climate disruptions, and new One Health approaches to global health policy, prompt research organizations to re-examine their responsibilities. AI offers powerful capabilities from precision medicine to ecosystem monitoring. Yet its deployment raises concerns related to high energy and water demands, hardware that depends on scarce resources, data governance, ethics and equity. There is further risk that technological solutions may distract from essential ecological and social action. A One Health lens, recognizing the interconnectedness of human, animal, and environmental health, encourages organizations to examine whether their AI tools and infrastructures truly support healthy ecosystems and communities. This raises critical questions: How should organizations balance scientific urgency with the responsibility to protect people, places, and data? How can meaningful interdisciplinary collaboration be structured? How can institutions uphold accountability to land, water, Indigenous communities, and future generations when deploying AI? Operational issues are central to this conversation. This moderated panel will explore how data-intensive research organizations can balance innovation with stewardship. Discussion topics include: What constitutes climate-resilient and environmentally responsible research infrastructure? What is the purpose of calculating organizational carbon footprints, and how can institutions meaningfully measure and manage the carbon and material impacts of AI? How can research cultures encourage the use of energy-efficient models, sustainable computational practices, and responsible procurement? Input from this panel will inform recommendations that help data-intensive organizations reimagine their responsibilities not only as the producers of knowledge but as institutions whose everyday operations shape a sustainable and ethical future.
The efficacy of convalescent plasma (CP) in COVID-19 has been contested, with large randomized trials producing conflicting results that led to recommendations against routine use. We performed a systematic review with design-stratified narrative synthesis (PRISMA 2020; SWiM; OSF: https://osf.io/tjcuq) to test whether these conflicting outcomes can be reconciled through three interacting determinants-timing, antibody titre, and host characteristics-using a standardized 91-field extraction matrix applied to each study. PubMed, Web of Science, and Scopus were searched through May 2026. Thirty-one studies (20,793 patients; 8 RCTs, 23 observational) were included. The synthesis indicates that CP is effective, but only under a specific convergence of conditions: when administered early in the disease course, with sufficient antibody titre, to seronegative or immunocompromised hosts who have not yet mounted an endogenous antibody response. Among RCTs analysing timing, 4 of 6 found benefit with earlier administration; for titre, 3 of 4 RCTs with explicit comparisons found benefit with higher-titre CP; and two large RCTs (combined n = 12,522) demonstrated preferential benefit in seronegative recipients. This interpretation is consistent with modelling estimates that CP, deployed largely to hospitalised patients, nonetheless saved tens of thousands of lives in the United States during the first pandemic year, an effect attributed to the subset treated early enough to modify disease progression. A second, previously unquantified finding concerns antibody class: characterisation of non-IgG immunoglobulins was nearly absent, with only 2 of 31 studies-both from our group-reporting IgA or IgM data, despite evidence that IgA and IgM are among the most potent neutralising classes against SARS-CoV-2. This hypothesis-generating observation identifies a systematic gap in CP product assessment. These findings reframe CP efficacy from a binary question to a conditional one and provide a template for the precision deployment of passive immunotherapy in future pandemics.
We report on the case of a pediatric patient with homozygous familial hypercholesterolemia (HoFH) who developed significant coronary artery disease (CAD). Percutaneous coronary intervention (PCI) guided by intravascular ultrasound (IVUS) was performed. A 6-year-old female patient was diagnosed with HoFH following the identification of cutaneous xanthomas. Genetic test identified a homozygous mutation in the LDLR gene. Cardiovascular assessment including a coronary CT angiography revealed a severe (70-90%) stenosis of the ostium of the left main coronary artery (LMCA) without significant lesions in other coronary branches. Given the severity of the lesion, PCI was indicated. The patient underwent PCI via the right femoral artery. The LMCA lesion was confirmed (70-90% stenosis, TIMI 3 flow) and a drug-eluting stent was deployed. Intravascular ultrasound (IVUS) played a crucial role in guiding the intervention, ensuring optimal stent expansion and apposition. Post-procedure, dual antiplatelet therapy (aspirin + clopidogrel) was initiated, clopidogrel was stopped after six months. A six-month follow-up CT angiogram demonstrated satisfactory stent positioning, despite metallic artifacts limiting full assessment. Two years after angioplasty, there have been no clinical, electrocardiographic, or echocardiographic signs of myocardial ischemia. This case highlights the feasibility and safety of LMCA stenting in pediatric patients with HoFH. Given the challenges of coronary interventions in pediatric patients, IVUS should be considered a critical adjunct to angiography in complex cases of early-onset coronary artery disease.
This study presents a Field-Programmable Gate Array (FPGA)-based convolutional neural network (CNN) accelerator for preliminary Parkinson's disease (PD) handwriting classification using hand-drawn circle images, with emphasis on arithmetic-level optimization through efficient multiplier architectures. Although optimized multipliers have been extensively studied for machine learning acceleration, their application-specific effects on inference consistency and hardware efficiency in healthcare-oriented FPGA implementations remain underexplored. To address this issue, a lightweight binary CNN classifier, trained on Google Colab, is deployed on FPGA hardware and evaluated with three multiplier architectures: standard multipliers, approximate logarithmic multipliers, and Karatsuba multipliers. The desktop CNN model was evaluated using both non-augmentation validation and standard augmentation strategies. The primary evaluation methodology used a non-augmentation validation approach, in which augmentation was applied exclusively to the training set, resulting in a software validation accuracy of 92.86%. The standard augmentation strategy achieved a validation accuracy of 97.83% and was used to compare the effects of augmentation before splitting. The trained model was quantized to Q4.12 fixed-point precision and implemented on FPGA hardware, where dense-layer computations were performed using different multiplier architectures. Hardware inference was validated on the NewHandPD hand-drawn circle dataset, and FPGA outputs were compared with software inference results via graphical analysis and Mean Absolute Deviation (MAD) to assess numerical consistency. Experimental results indicate that the FPGA-based implementation achieved classification behavior closely aligned with software inference while improving hardware efficiency. Under the non-augmentation validation approach, the FPGA implementation achieved 89.73% accuracy compared to 92.86% in software, whereas the standard augmentation strategy achieved 95.40% accuracy compared to 97.83% in software using the Approximate Logarithmic Multiplier. The results indicate the potential feasibility of lightweight CNN deployment with optimized multipliers for resource-efficient edge healthcare applications. However, due to the limited dataset size, the presented findings should be interpreted as a preliminary proof-of-concept study rather than definitive clinical validation.
The deployment of Federated Learning (FL) across the Internet of Medical Things (IoMT) is severely hindered by computational asymmetry and statistical heterogeneity. Traditional synchronous aggregation protocols suffer from severe straggler effects when deployed across devices with varying computational capacities, such as hospital servers vs. ambulatory wearables. In this study, we propose the Asynchronous Proximal Federated Aggregation (APFA) framework to address these dual bottlenecks. APFA integrates a local proximal regularizer with a server-side staleness dampening penalty, permitting continuous, uncoordinated model updates from edge devices. Evaluated on highly skewed partitions of the CheXpert and MIMIC-IV datasets, APFA reached an 80% diagnostic viability threshold in just 4.1 simulated hours, representing a 71% reduction in total wait time compared to standard synchronous baselines like FedProx. Our mathematical integration effectively mitigates weight divergence, indicating the robustness of asynchronous machine learning for scalable, privacy-preserving clinical diagnostics.
To systematically characterize intraoperative bleeding patterns during laparoscopic pancreaticoduodenectomy (LPD) and conduct an in-depth analysis of hemostatic behavior and its association with hemostasis quality. We retrospectively collected 210 LPD surgical videos from 15 centers across China (2018-2022), of which 190 were enrolled. Two surgeons manually annotated bleeding events, causes, surgical phases, anatomical locations, bleeding types (arterial, venous, lymph node-related), and hemostasis behavior assessments [good hemostasis (GH), poor hemostasis (PH)). A surgical instrument detection model deployed on the SurgSmart platform extracted per-frame instrument usage and duration data. Hemostatic behaviors-including phase duration, hemostatic instrument utilization patterns, hemostatic instrument switching frequency, and the hemostatic complexity index (HCI) based on Simpson's diversity index-were compared by bleeding type and hemostasis behavior assessment category. A total of 675 bleeding events were identified, occurring in 87.9% (167/190) of the surgical videos. The pancreatic uncinate process was recognized as the "high-risk bleeding zone," with the highest bleeding incidence observed in the superior mesenteric vein-portal vein axis (34.3%) and superior mesenteric artery-celiac trunk axis (17.4%). Venous bleeding was the most frequent type (n = 334, 49.5%), followed by arterial bleeding (n = 313, 46.4%) and lymph node-related bleeding (n = 28, 4.1%). Excessive tissue tension (41.6%) and incorrect anatomical plane dissection (28.1%) were the leading causes of bleeding. Regarding hemostatic behaviors stratified by bleeding type: venous bleeding exhibited the longest hemostasis duration (68.0 (30.0-148.0) s), and significant differences were observed in hemostatic instrument utilization patterns, instrument switching frequency, and the HCI across bleeding types (all P < 0.05). Comparisons between PH and GH groups revealed that PH groups had a significantly prolonged hemostasis duration (76.0 (29.0-171.5) s vs. 50.0 (24.0-118.0) s, P = 0.001), more frequent instrument switching (median 2.0 (0.5-6.0) vs. 1.0 (0.0-3.0), P < 0.001), and significantly distinct instrument utilization patterns (P = 0.004). Additionally, the HCI was significantly higher in PH groups than in GH groups (0.338 ± 0.233 vs. 0.286 ± 0.222, P = 0.008). Our study systematically characterized bleeding sites, causes, and surgical phases during LPD, identifying procedure-specific high-risk anatomical regions and distinct bleeding mechanisms. We demonstrated that hemostatic instrument utilization patterns vary by bleeding type and correlate with hemostasis quality, and proposed the instrument HCI as an objective instrument-based metric for summarizing hemostatic behavior. Clinicaltrials.gov, number NCT06128603.