Emotion recognition (ER) aims to identify affective states from facial expressions, speech, text, and physiological signals. The context-dependent and multimodal nature of emotion presents challenges for systems that rely on statistical associations alone. Knowledge graphs (KGs) organize affective concepts and their relations, providing explicit knowledge for contextual modeling, multimodal integration, and reasoning. This survey reviews KG-based emotion recognition and related affective tasks, covering knowledge extraction and graph construction, unimodal and multimodal systems, representation learning, reasoning, and integration with large language models (LLMs). The synthesis draws on 49 core studies and uses additional literature for theoretical and methodological context. It distinguishes semantic KGs from computational graphs and characterizes graph construction along five dimensions: knowledge scope, semantic foundation, construction mechanism, human involvement, and multimodal representation. Selected within-study comparisons report gains from enabling knowledge-related components, while differences in tasks, protocols, and comparison designs limit conclusions across systems. The survey also examines evaluation datasets, knowledge quality, and the computational and grounding challenges of LLM and KG integration. It identifies research priorities in knowledge validation, cross-domain generalization, privacy, temporal modeling, and multimodal alignment, with evaluation procedures for assessing their contribution to reliable affective systems.
Decision-making in healthcare in the Intensive Care Unit (ICU) is a very complex and multi-dimensional process. Sepsis management requires well-coordinated, simultaneous decisions across vasopressor titration, fluid restoration, antibiotic supervision, and haemodynamic monitoring, each operating on different physiological timescales. Despite advances in reinforcement-learning-based clinical decision support, existing systems treat these as a single flat action space and cannot capture the multi-level structure of real clinical reasoning. Moreover, most systems lack explainability mechanisms, making them unsuitable for accountable clinical practice. This study proposes the Hierarchical Multi-Agent Clinical Decision Support System (HM-CDSS), a four-phase hierarchical multi-agent reinforcement learning framework designed to address these limitations. Using a cohort of 5,282 adult ICU sepsis patients from the Medical Information Mart for Intensive Care (MIMIC-IV) database, approximately 78,200 offline trajectory transitions are constructed. A single-agent Dueling DQN baseline achieves a mean test reward of 4.385 but exhibits conservative policy collapse, reaching only 10.3
Medical image segmentation plays a pivotal role in clinical diagnosis assistance, lesion detection, and computer-aided medicine. However, factors such as complex organ boundary morphologies, multi-scale structural variations, and low contrast continue to pose challenges for accurate segmentation. To address this, this paper proposes an efficient segmentation network based on the fusion of anatomical prior and features—Anatomical-prior-guided Fusion Mamba-UNet (PF-Mamba-UNet). Firstly, the CR-VSS module is designed. Within the VSS module, a convolutional refinement branch is constructed to enhance local detail modeling. An integrated CSS2D module enables multi-directional scanning along horizontal, vertical, diagonal, and anti-diagonal paths, boosting the model’s multi-directional perception and feature capture capabilities for complex geometries. Secondly, the Coordinated Spatial-Channel Fusion Attention (CSCFA) module is designed and embedded in the skip-connection layers of the encoder and decoder. Through a spatial-first mechanism, it synergistically enhances spatial and channel features, thereby highlighting critical anatomical regions while suppressing redundant background noise. Finally, an anatomical prior regularization is constructed to balance boundary smoothing and regional consistency. The former ensures continuity and smoothness of segmentation edges, while the latter guarantees structural consistency within regions. Collectively, they ensure the anatomical plausibility and structural continuity of segmentation results. Medical image segmentation results demonstrate that PF-Mamba-UNet exhibits significant performance advantages over mainstream models including U-Net, SwinUNet, Mamba-UNet, Mamba-Sea, KM-Unet and CSWin-Unet.
Abstract Borderline Personality Disorder (BPD), the prototypic Cluster B condition, is characterised by profound instability in affect, self-image, and interpersonal relationships, including rapid oscillation between idealisation and devaluation. Despite decades of neurobiological and psychotherapeutic research, no unified mechanistic theory accounts for the simultaneous presence of interoceptive dysregulation, interpersonal hypersensitivity, unstable identity, and splitting. Here we (i) review the hot research topics in Cluster B personality disorders in the recent literature, including triple-network neuroimaging dysconnectivity, allostatic interoceptive overload, and computational psychiatry approaches; (ii) identify the critical gap — the absence of a formal mechanistic account coupling interoceptive and social predictive hierarchies; (iii) propose the Hierarchical Interoceptive–Social Inference Dysregulation (HISID) framework, positing that BPD arises from a precision dissociation in which interoceptive precision is chronically suppressed while social-threat precision is pathologically elevated, leaving the self-model anchored exclusively to volatile social-affective inputs; and (iv) present a Level-1 computational simulation implementing two active inference agents — a healthy control (HC) and a BPD agent characterised by suppressed interoceptive precision, asymmetric social-threat amplification, a bimodal (splitting) social prior, and a hysteresis commitment mechanism — embedded in a 300-trial social-threat paradigm. While the HISID framework is introduced within the broader context of Cluster B personality disorders, the formal theory, parameterisation, simulations, and empirical predictions presented in this paper are specific to BPD as the prototypic Cluster B condition. Extension of the precision-dissociation formalism to narcissistic, antisocial, and histrionic personality disorders — each of which is predicted to occupy a distinct parametric region within the same generative architecture — is deferred to subsequent work. Across $$n=40$$ agents per group, the BPD agent produces a 4.22 $$\times $$ increase in self-model variance, a 10.33 $$\times $$ increase in per-agent median threat-recovery time, and a 4.45 $$\times $$ increase in unweighted cumulative prediction-error accumulation (alongside a 2.61 $$\times $$ increase in precision-weighted free energy; the unweighted measure is reported as a control on precision-induced mechanical scaling). A factorial control study crossing precision profile with generative architecture decomposes the BPD phenotype into distinct contributions: identity-instability variance is driven primarily by the bimodal-prior architecture, whereas threat-recovery slowing is driven primarily by the asymmetric threat-conditional precision, with a super-additive interaction in the full agent. These differences are large and internally consistent within the simulation and constitute demonstrations of model sufficiency, not empirical theory validation. A sensitivity analysis across a $$3\times 3$$ precision grid confirms qualitative robustness. A key architectural finding is that the bimodal prior requires Bayesian mixture belief updating with hysteresis commitment — not gradient descent — providing a candidate formalisation of splitting as discrete pole-occupancy inference. The HISID framework, grounded in the Free Energy Principle, generates five operationalised empirical predictions testable with existing neuroimaging and computational phenotyping methods.
Traditional picture fuzzy sets (PFSs) cannot capture temporal dynamics in dynamic uncertainty scenarios. This paper proposes dynamic picture fuzzy sets (DPFSs) and corresponding similarity measures. DPFSs integrate a discrete time set into the PFS framework, retaining the four-dimensional membership structure (positive, neutral, negative, refusal) while modeling time-dependent variations, and their basic operations and key properties are rigorously proven. Six types of dynamic similarity measures (Hamming, Euclidean, Minkowski, Chebyshev, Lambert, absolute sum-difference) are constructed, which are divided into Type-I (excluding refusal degree) and Type-II (including refusal degree), and time weights with a decay factor are incorporated to emphasize the cumulative effect of recent data. Two practical applications validate the effectiveness of the proposed methods: time-aware supplier selection (a multi-criteria decision-making scenario) can consistently identify the optimal alternative by amplifying the advantages of recent performance, and type 2 diabetes risk assessment (a medical diagnosis scenario) can capture dynamic symptom progression to provide clinically meaningful risk classifications, with both applications showing high consistency across different measures.
