Deformable image registration aims to achieve nonlinear alignment of image spaces by estimating dense displacement fields. It is widely used in clinical tasks such as surgical planning, assisted diagnosis, and surgical navigation. While efficient, deep learning registration methods often struggle with large, complex displacements. Pyramid-based approaches address this with a coarse-to-fine strategy, but their single-feature processing can lead to error accumulation. In this paper, we introduce a dense Mixture of Experts (MoE) pyramid registration model, using routing schemes and multiple heterogeneous experts to increase the width and flexibility of feature processing within a single layer. The collaboration among heterogeneous experts enables the model to retain more precise details and maintain greater feature freedom when dealing with complex displacements. We use only deformation fields as the information transmission paradigm between different levels, with deformation field interactions between layers, which encourages the model to focus on the feature location matching process and perform registration in the correct direction. We do not utilize any complex mechanisms such as attention or ViT, keeping the model at its simplest form. The powerful deformable capability allows the model to perform volume registration directly and accurately without the need for affine registration. Experimental results show that the model achieves outstanding performance across four public datasets, including brain registration, lung registration, and abdominal multi-modal registration. The code will be published at https://github.com/Darlinglinlinlin/MOE_Morph.

Network slicing is a cornerstone of 5G/6G vertical services, yet practical deployments require mobile network operators (MNOs) to adjust slice service level agreement (SLA) weights based on quality of experience (QoE), causing rapid non-stationary objective changes that can destabilize deep reinforcement learning (DRL) slicing policies and necessitate retraining. This paper proposes Continual Mixture of Experts (CoMEx) for fast policy adaptation. CoMEx pre-trains and freezes multiple expert policies under diverse SLA preferences, explicitly appends the SLA weight vector to observations, and trains a DRL-based gating network to fuse expert actions at the step level for fast adaptation to unseen SLA configurations. To broaden coverage without degrading existing experts, CoMEx further incorporates a masked expert expansion mechanism that incrementally adds new experts and fine-tunes the gate. Step-level DRL gating demonstrates superior generalization in RAN slicing, attaining a mean score of 78.95 under unseen SLA weights—surpassing episode-level and supervised gating by 2.40% and 27.67%, respectively. Moreover, CoMEx’s extensibility is highlighted by a 7.08% performance boost (reaching 84.54) upon the addition of a fourth expert. Such results confirm the framework’s capacity for timely and robust policy adaptation in non-stationary SLA environments.

Sleep-predominant network hyperexcitability is increasingly recognized as a potential disease-accelerating comorbidity in Alzheimer's disease (AD). However, its prevalence and risk-factors remain debated, largely due to cohort-specific and methodological differences across studies. In this prospective case-control study, we investigated potential ways of improving detection, from translational approaches focusing on rapid eye movement (REM)-sleep to refined electroencephalogram (EEG) setups and added clinical questionnaires. We recruited 30 patients with early-stage AD without a history of epilepsy and 30 age-matched controls. Participants underwent overnight polysomnography with video-EEG. Interictal epileptic discharges (IEDs) were identified through a structured 3-step review by multiple independent experts using recommended criteria. Neuroanatomic patterns and sleep-related abnormalities were investigated as potential risk factors. Clinical symptoms in favor of epileptic seizures were evaluated through a tailored questionnaire at follow-up. IEDs were detected in 3 patients (10%) and 1 control (3.33%), a difference not reaching statistical significance (p = 0.612). Most events occurred during non-REM (NREM) sleep. Eight patients (26.67%) reported symptoms compatible with epileptic seizures-one of whom also presented with IEDs. Patients with IEDs or reported symptoms suggestive of potential seizures exhibited more severe sleep-disordered breathing and reduced precuneus volume compared with those without. Despite efforts to optimize detection accuracy, our findings reveal a lower-than-expected percentage of patients with AD with IEDs, yet support previous findings suggesting that sleep-disordered breathing and specific atrophy patterns could flag at-risk patients, guiding screening in clinical settings. Our findings also favor validation efforts of questionnaires to support the diagnostic process. Finally, we highlight methodological issues in IED detection and call for the re-evaluation and standardization of diagnostic methods and criteria in this population to improve patient care. ANN NEUROL 2026.

IntroductionLarge language models (LLMs) offer potential as clinical decision support systems (CDSS) for detecting drug-related problems (DRPs), yet their real-world performance compared to clinical pharmacists (CPs) remains unclear, especially in complex hematology care. We aimed to evaluate the concordance between a clinical pharmacist and three LLMs in identifying DRPs within a Bone Marrow Transplantation unit.MethodsThis prospective observational study evaluated the concordance between a CP and three LLMs (ChatGPT-4o, Grok-3, DeepSeek-v3) in a Bone Marrow Transplantation unit. Eighty-three anonymized patient cases encompassing 210 CP-identified DRPs, classified via the PCNE v9.1 system, were presented using a standardized CDSS-simulating prompt. Performance was assessed based on direct detection, prompted detection after structured follow-up, and the clinical relevance of AI-generated therapeutic recommendations against the CP's gold-standard assessments.ResultsDirect detection of intervention-requiring DRPs was limited (51.4%-60.5% across models), with nearly half missed initially. Guided prompting significantly improved overall detection rates to 93.8%-98.1%, with ChatGPT achieving the highest accuracy. All models produced hallucinations. Recommendation concordance with the CP exceeded 70% in most DRP categories. DeepSeek and ChatGPT showed more consistent performance in context-dependent evaluations, whereas Grok demonstrated higher direct detection but lower recommendation alignment. LLMs demonstrate meaningful potential to assist in DRP detection but are not sufficiently reliable as standalone tools. Expert-guided interaction substantially enhanced their performance, underscoring the critical value of hybrid pharmacist-AI workflows.ConclusionFuture research should validate these findings across broader populations with multiple expert evaluators and integrate next-generation AI architectures for safer CDSS implementation.

Sequential recommenders aim to enhance prediction accuracy by leveraging user interaction sequences, with transformer-based models showing particularly strong performance. Among them, cold-start sequential recommenders are particularly challenging because these models typically require extensive historical data to perform optimally. Some works attempt to address this issue by enhancing the adaptive ability of the sequence recommenders with meta-learning approaches. However, they are unsuitable for enhancing the popular Transformer-based sequence recommenders: MAML-based models cannot adapt the large number of parameters of Transformers, while transition-based and metric-based meta-learning models rely on unique architectures that are incompatible with Transformer-based frameworks. Also, they usually lack mechanisms to recognize and cater to multiple interests within short interaction sequences. To address these limitations, we propose MESA, a meta-modulation plugin module specifically designed to enhance the cold-start recommendation of Transformer-based sequential recommender systems. (1) We design a meta-modulation method to directly modulate the parameters in Transformer-based sequence encoders, thus enabling the model to adapt more effectively to new users in cold-start scenarios. (2) Additionally, MESA integrates the Mixture of Experts (MoE) mechanism, which refines sequence representations by utilizing multiple experts, each focusing on different aspects of user interests. This structure enhances the personalization of the recommendation by effectively handling diverse user interests within the sequences. Experiments demonstrate the effectiveness of MESA in cold-start scenarios. Our codes are available here : https://github.com/Mushroom-cat/MESA .

ABSTRACTHippocampal subfields differentially develop and age, and they vary in vulnerability to neurodegenerative diseases. Innovation in high‐resolution imaging has accelerated clinical research on human hippocampal subfields, but substantial differences in segmentation protocols impede comparisons of results across laboratories. The Hippocampal Subfields Group (HSG) is an international organization seeking to address this issue by developing a histologically valid, reliable, and freely available segmentation protocol for high‐resolution T2‐weighted 3 T MRI (http://www.hippocampalsubfields.com). Here, we report the first portion of the protocol focused on subfields in the hippocampal body; protocols for the head and tail are in development. The body protocol includes definitions of the internal boundaries between subiculum, Cornu Ammonis (CA) 1–3 subfields, and dentate gyrus, in addition to the external boundaries of the hippocampus apart from surrounding white matter and cerebrospinal fluid. The segmentation protocol is based on a novel histological reference dataset labeled by multiple expert neuroanatomists. With broad participation of the research community, we voted on the segmentation protocol via an online survey, which included detailed protocol information, feasibility testing, demonstration videos, example segmentations, and labeled histology. All boundary definitions were rated as having high clarity and reached consensus agreement by Delphi procedure. The harmonized body protocol yielded high inter‐ and intra‐rater reliability. In the present paper we report the procedures to develop and test the protocol, as well as the detailed procedures for manual segmentation using the harmonized protocol. The harmonized protocol will significantly facilitate cross‐study comparisons and provide increased insight into the structure and function of hippocampal subfields across the lifespan and in neurodegenerative diseases.

CellApop: A knowledge-guided decoupled distillation framework for label-efficient apoptotic cell segmentation and dynamic analysis in brightfield microscopy.

Low reliability has consistently been a challenge in the application of deep learning models for high-risk decision-making scenarios. In medical image segmentation, multiple expert annotations can be consulted to reduce subjective bias and reach a consensus, thereby enhancing the segmentation accuracy and reliability. To develop a reliable lesion segmentation model, we propose CalDiff, a novel framework that can leverage the uncertainty from multiple annotations, capture real-world diagnostic variability and provide more informative predictions. To harness the superior generative ability of diffusion models, a dual step-wise and sequence-aware calibration mechanism is proposed on the basis of the sequential nature of diffusion models. We evaluate the calibrated model through a comprehensive quantitative and visual analysis, addressing the previously overlooked challenge of assessing uncertainty calibration and model reliability in scenarios with multiple annotations and multiple predictions. Experimental results on two lesion segmentation datasets demonstrate that CalDiff produces uncertainty maps that can reflect low confidence areas, further indicating the false predictions made by the model. By calibrating the uncertainty in the training phase, the uncertain areas produced by our model are closely correlated with areas where the model has made errors in the inference. In summary, the uncertainty captured by CalDiff can serve as a powerful indicator, which can help mitigate the risks of adopting model's outputs, allowing clinicians to prioritize reviewing areas or slices with higher uncertainty and enhancing the model's reliability and trustworthiness in clinical practice.

A recent survey in the Philippines, PRESYON-4, showed increasing prevalence of hypertension from 22% in the 1990s to 37% in 2021, of which only 52% were aware of their diagnosis. While rates of treatment and adherence were 68% and 86%, respectively, the rate of BP control was low at 37%. Furthermore, there remained a high degree of unawareness regarding hypertension, its role in CV morbidity and mortality, and how it can be optimally managed. In particular, there is a knowledge gap in the diagnostic approach and management of severe acute elevations in blood pressure. In response to this, the Philippine Society of Hypertension, Philippine Heart Association, and multiple experts from various sectors worked together to develop the 2024 Clinical Practice Guideline on the Diagnosis and Management of Severe Blood Pressure Elevation. The CPG provides eleven (11) recommendations and four (4) best practice statements addressing key clinical questions on the diagnosis and management of severe BP elevation. The guideline development process adhered to the GRADE approach through the Evidence to Decision (EtD2) framework, including the identification of critical questions and outcomes, retrieval of current evidence, appraisal and synthesis of the evidence, and formulation of draft recommendations. A multisectoral consensus panel (CP) was convened to discuss values, preferences, and socioeconomic impact and finalize the strength of the recommendations. The CPG is intended to be used by general practitioners, specialists, family physicians, allied health professionals, emergency medical personnel, and healthcare workers who may encounter adult patients with hypertension, whether in the inpatient or outpatient setting.

Although digital health technologies have the potential to improve cardiac patient health outcomes, there are significant digital health inequities experienced by immigrant communities. It is important to understand the barriers to digital health equity within cardiovascular healthcare for immigrants. The purpose of this paper is to apply the Health Equity Impact Assessment, Digital Health Supplement within the context of immigrant communities with an intersectional lens. The Health Equity Impact Assessment, Digital Health Supplement can be used to ensure health equity remains central in digital health technologies used within cardiovascular healthcare. The instrument includes 5 steps: (1) scoping, (2) potential impacts, (3) mitigation, (4) monitoring, and (5) dissemination. Social determinants of health, intersectional factors, and patient and family involvement are necessary to understand digital health inequities experienced among immigrants. Access is one of the potential impacts highlighted by the framework. Access can be promoted through funding and tailored digital health technologies. Immigrants need to be active partners in the design and development of the digital health technologies. It is important that equity remains a central outcome. Multiple experts are needed to analyze the results in a fair manner. Findings should be disseminated within various avenues. Future research is necessary to strengthen the evidence base for applying the Health Equity Impact Assessment, Digital Health Supplement among immigrant populations. Since digital health equity research requires an intersectional lens, the diversity dimensions can serve as a foundational framework in future studies.

We analyze a varying-coefficient spatial autoregressive model with spatial fixed effects. One salient feature of the model is the incorporation of multiple spatial weight matrices through their linear combinations with varying coefficients, which help solve the problem of choosing the most “correct” one for applied econometricians who often face the availability of multiple expert spatial weight matrices. We estimate and make inferences on the model coefficients and coefficients in basis expansions of the varying coefficients through penalized estimations, establishing the oracle properties of the estimators and the consistency of the overall estimated spatial weight matrix, which can be time-dependent. We further consider two applications of our model in change point detections in spatial autoregressive models, providing theoretical justifications in consistent change point locations estimation and practical implementations. Simulation experiments demonstrate the performance of our proposed methodology, and real data analyses are also carried out.

This work reviews the critical improvements needed to enhance high temperature (HT) geothermal instrumentation towards reliable characterization of deep subsurface geothermal wells and reservoirs. To close the current technology gap R&D efforts must be rationally prioritized; Sandia National Laboratories interviewed multiple experts from geothermal industries who have firsthand experience with evaluating the geothermal subsurface, including Original Equipment Manufacturer (OEM) geothermal instrumentation developers and service providers for well evaluation. To support the rigorous advancement of geothermal energy production as a global renewable energy resource, we identify four priority R&D needs: HT embedded electronic system, HT high pressure harsh fluid sensing hardware, HT means for communications, and HT energy storage.

The teaching quality evaluation of university music programs constitutes multi-attribute group decision-making because it involves synthesizing judgments across diverse assessment dimensions from multiple experts. Probabilistic linguistic term sets have emerged as a robust tool for capturing the uncertainty and subjectivity in such evaluations, effectively quantifying ambiguous expert opinions. Recent advancements in the multi-attribute group decision-making method have seen the application of the exponential tomada de decisao interativa multicriterio (ExpTODIM) method, which offers unique advantages in handling complex trade-offs and decision-makers' psychological tendencies. (This Portuguese term means interactive multicriteria decision-making in English.) To address the specific demands of music program teaching quality evaluation under uncertain conditions, the authors of this study developed a novel probabilistic linguistic ExpTODIM approach tailored to integrate probabilistic linguistic term sets with the ExpTODIM framework. The authors present a numerical study focused on the teaching quality evaluation of university music programs. This practical application validates the feasibility and effectiveness of the proposed probabilistic linguistic ExpTODIM method.

Due to its multi-factor mechanism, variable opioid response, and high-risk adverse reactions, cancer pain remains a major challenge in oncology. To overcome these obstacles, we have developed a collaboration framework based on large language models (LLMs): OncoPainBot. This framework can simulate the reasoning and decision-making of multiple clinical experts to conduct comprehensive cancer pain assessment and management. Our OncoPainBot integrates four specialized agents: Pain-Extraction, Pain-Mechanism Reasoning, Treatment-Planning, and Safety-Check, each corresponding to a unique clinical role. In this paper, we compare seven LLMs and three Retrieval-Augmented Generation(RAG) strategies to determine the optimal model configuration. The final framework was verified on 516 real-world electronic medical records of cancer pain collected. We tested our solution through multiple dimensions. Ultimately, Claude-4 combined with RAG achieved the best overall performance, demonstrating outstanding semantic consistency and evidence-based reasoning in multiple metrics. In clinical validation, OncoPainBot achieved a high degree of consistency between the generated reports and actual clinical documents, while maintaining a high decision-making accuracy (0.841) in the analgesic recommendation task. At the same time, our error analysis shows that most of the differences are caused by patient-specific factors and monitoring recommendations rather than incorrect drug selection, which demonstrates the reliability of our framework. OncoPainBot has demonstrated the feasibility of a cancer pain management system based on LLMs, providing a transparent, evidence-based, and clinical-based framework for personalized analgesic care.

Abstract Forensic expert opinions profoundly influence legal outcomes, yet how judges, jurors, and lawyers should interpret these opinions and what information helps them do so remains underexplored. We highlight the Bayesian solution provided by Morris (1971, 1974, 1977). Rather than adopting expert assessments at face value, recipients assign weight to expert opinions through their own uncertainties regarding what opinions experts would provide under each considered proposition. This “performance uncertainty” is distinct from recipients’ uncertainty about which proposition is true. Validation data reduce performance uncertainty, enabling recipients to weight expert opinions based on demonstrated performance. We illustrate this framework through examples spanning categorical conclusions, likelihood ratios, ranges, and multiple experts, demonstrating how it accommodates case-specific factors, incomplete information, and varying recipient beliefs. Though recipients will not conduct explicit computations, the key implication of Bayesian reasoning remains that, regardless of an expert’s opinion scale, judicial stakeholders require access to detailed performance data to make scientifically defensible interpretations of expert opinions. Restricted access to validation data prevents recipients from updating their performance beliefs with empirical evidence, leaving interpretations dependent on whatever initial assumptions each recipient brings. Science in forensics requires not just generating validation data but ensuring meaningful access for those interpreting expert opinions.

Federated Domain Generalization (FedDG) aims to train a globally generalizable model on data from decentralized, heterogeneous clients. While recent work has adapted vision-language models for FedDG using prompt learning, the prevailing "one-prompt-fits-all" paradigm struggles with sample diversity, causing a marked performance decline on personalized samples. The Mixture of Experts (MoE) architecture offers a promising solution for specialization. However, existing MoE-based prompt learning methods suffer from two key limitations: coarse image-level expert assignment and high communication costs from parameterized routers. To address these limitations, we propose TRIP, a Token-level pRompt mIxture with Parameter-free routing framework for FedDG. TRIP treats prompts as multiple experts, and assigns individual tokens within an image to distinct experts, facilitating the capture of fine-grained visual patterns. To ensure communication efficiency, TRIP introduces a parameter-free routing mechanism based on capacity-aware clustering and Optimal Transport (OT). First, tokens are grouped into capacity-aware clusters to ensure balanced workloads. These clusters are then assigned to experts via OT, stabilized by mapping cluster centroids to static, non-learnable keys. The final instance-specific prompt is synthesized by aggregating experts, weighted by the number of tokens assigned to each. Extensive experiments across four benchmarks demonstrate that TRIP achieves optimal generalization results, with communicating as few as 1K parameters. Our code is available at https://github.com/GongShuai8210/TRIP.

CaliDiff: Multi-rater annotation calibrating diffusion probabilistic model towards medical image segmentation.

Mixture-of-Experts (MoE) has emerged as an effective and efficient scaling mechanism for large language models (LLMs) and vision-language models (VLMs). By expanding a single feed-forward network into multiple expert branches, MoE increases model capacity while maintaining efficiency through sparse activation. However, despite this sparsity, the need to preload all experts into memory and activate multiple experts per input introduces significant computational and memory overhead. The expert module becomes the dominant contributor to model size and inference cost, posing a major challenge for deployment. To address this, we propose MC# (Mixture-Compressor-sharp), a unified framework that combines static quantization and dynamic expert pruning by leveraging the significance of both experts and tokens to achieve aggressive compression of MoE-LLMs/VLMs. To reduce storage and loading overhead, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates adaptive bit allocation as a linear programming problem. The objective function jointly considers expert importance and quantization error, producing a Pareto-optimal trade-off between model size and performance. To reduce runtime computation, we further introduce Online Top-any Pruning (OTP), which models expert activation per token as a learnable distribution via Gumbel-Softmax sampling. During inference, OTP dynamically selects a subset of experts for each token, allowing fine-grained control over activation. By combining PMQ's static bit-width optimization with OTP's dynamic routing, MC# achieves extreme compression with minimal accuracy degradation. On DeepSeek-VL2, MC# achieves a 6.2× weight reduction at an average of 2.57 bits, with only a 1.7% drop across five multimodal benchmarks compared to the 16-bit baseline. Moreover, OTP further reduces expert activation by 20% with less than 1% performance loss, demonstrating strong potential for efficient deployment of MoE-based models.

Knowledge graph embeddings (KGE) are effective for representing factual data for numerous applications. However, real-world facts continually evolve, necessitating ongoing updates to knowledge graphs as new information emerges. Under these circumstances, existing KGE models in transductive, inductive, and continual learning settings are prone to catastrophic forgetting or require costly retraining to integrate new information. To address these challenges, we propose a novel model called the Context-aware Adaptive learning model for Knowledge Graph Embeddings (CAKGE). Our model first identifies semantic-relevant entities and uncovers latent relational paths to facilitate the acquisition of new knowledge. To ensure the paths are semantically aligned with the query, we employ a context-aware fusion module, which leverages multiple specialized expert networks to assess and integrate the relevance of these relational paths. Building on this, we introduce an adaptive message aggregation module that incorporates a knowledge replay strategy, enabling the model to integrate both new and existing knowledge efficiently, without retraining the knowledge graph. Additionally, to mitigate catastrophic forgetting, we reformulate the challenge of aligning new with existing knowledge as a graph-matching task using the Fused Gromov-Wasserstein distance, enabling the alignment of old and new knowledge from both semantic and topological perspectives. Furthermore, we provide theoretical guarantees for the expressiveness and reasoning ability of CAKGE, showing that it is the first unified framework tackling transductive, inductive, and continual settings. Extensive experiments show that CAKGE achieves state-of-the-art performance, demonstrating its effectiveness in dynamic KGE modeling.

ABSTRACT Predicting learning outcomes is an important problem in educational research, as it enables timely intervention for learners and supports educators in implementing personalized strategies to improve training quality. This task is especially important in online and blended learning environments, where students must independently manage their learning progress. Deep learning (DL) models, particularly Long Short‐Term Memory (LSTM) networks, have shown promise in improving prediction accuracy with time‐series data. However, single‐network models often face high computational costs and struggle to capture the diversity of learner behaviors, resulting in limited performance. To address this, we propose a novel architecture called MoELA to solve the pass/fail classification problem, based on the Mixture of Experts (MoE) framework, which uses multiple expert networks to flexibly model different behavior groups. Each expert includes an LSTM layer integrated with an attention mechanism to effectively process time‐series data and emphasize critical learning moments. The number of experts is determined by clustering learner behavior using the Fuzzy C‐Means (FCM) algorithm. Experiments were conducted on the Open University Learning Analytics Dataset (OULAD). To mitigate class imbalance in the Pass/Fail prediction task, we introduced SMOTERN, a data balancing method that combines SMOTE with noise removal. Results show that MoELA models significantly outperform the traditional four‐layer stacked LSTM. The best performance was achieved by the MoELA4 model trained on SMOTERN‐balanced data, with Accuracy, AUC, and F1 scores of 0.9238, 0.9568, and 0.8637, respectively. Additionally, the proposed architecture requires fewer parameters and achieves faster training and prediction times compared to the stacked LSTM model.