Abstract: As the information environment evolves, social media has become the primary channel through which the public accesses and shares information, and perceived credibility has emerged as a critical influence on how users evaluate the credibility of information. Existing research suggests that heuristic credibility cues can enhance users’ perceived credibility, yet the findings remain inconsistent. Consequently, it is necessary for researchers to systematically examine whether heuristic credibility cues can effectively enhance perceived credibility. This study employed a meta-analysis to analyse 18 studies meeting the selection criteria, involving a total sample size of 14,188 participants. The aim was to assess the overall effect of social media heuristic credibility cues on perceived credibility and to explore the influence of potential moderating mechanisms on perceived credibility. The results indicate that manipulating source cues and social cues which serve as heuristic credibility cues on social media significantly increased perceived credibility (g = 0.307, p = 0.000). Effect sizes varied across moderating variables such as the type of heuristic credibility cue, participant type, method of measuring perceived credibility, experimental design, sample size, and year of publication. Among these, the type of heuristic cue and participant type as significant moderators, specifically, authoritative sources were more effective than other types of information sources in enhancing perceived credibility; the impact of different types of social cues on perceived credibility was also significant to varying degrees. Furthermore, student groups were more susceptible to the influence of heuristic credibility cues than non-student groups. These findings provide theoretical and practical insights for the design of information dissemination and the construction of perceived credibility on social media. It should be noted that, given the limited number of studies included in this meta-analysis and the restricted range of moderator variables, the above conclusions require further empirical research to be tested and confirmed.

Abstract: Bicycling is an important mode of sustainable and active transportation, but bicycle safety remains a major concern in urban areas, especially at intersections where cyclists interact with turning vehicles, crossing traffic, and complex roadway conditions. This study assesses the effect of roadway, environmental, and infrastructure-related factors on bicycle safety at intersections in Baltimore City. Crash data from 2022 to 2024 were obtained from the Maryland crash data records and analyzed for bicycle-involved crashes at intersections. The study used descriptive statistics, GIS-based spatial analysis, visualizations, and exploratory regression models, including linear regression, binary logistic regression, and multinomial logistic regression. The results showed that Baltimore City had 180 bicycle-involved crashes at intersections during the study period, most of which resulted in injury. Spatial analysis indicated that crashes were concentrated mainly in downtown Baltimore. Descriptive results showed that many crashes occurred during daylight, clear weather, and dry surface conditions, which may reflect higher bicycle activity during these periods. The Sankey diagram suggested that severe crash outcomes were more common in locations without bike lanes. However, the regression models did not identify statistically significant relationships between the selected variables and crash severity. The findings highlight the need for better bicycle exposure data, more complete infrastructure variables, and improved intersection-level safety planning in Baltimore City.

Abstract: Purpose: Sustaining high-functioning interprofessional teamwork in intensive care settings is essential for patient safety, workforce well-being, and reliable care delivery. However, evidence regarding the role of structured team-training interventions in already high-performing critical care teams remains limited. The purpose of this study was to assess the impact of the Team Training and Clinical Excellence Academy (TTrACE), a structured interprofessional team training program, on collaboration, psychological safety, need satisfaction, and team effectiveness in a critical care setting. Methods: We conducted a prospective pre-post evaluation of an interprofessional critical care team-training program in a medical/surgical ICU at a large academic hospital. The evaluation used previously published measures of collaborative environments and psychological need satisfaction, along with exploratory team-functioning items assessing shared mental models, psychological safety and communication, and perceived team effectiveness. Results: Of the 35 TTrACE enrollees, 28 (80%) consented to participate in the evaluation study. Baseline scores across collaboration, psychological need satisfaction, and team functioning were high and remained stable at one-month follow-up, suggesting preservation of strong team functioning over time. Descriptive improvements were observed in perceived teamness and shared mental models. Most respondents reported having an opportunity to implement TTrACE learnings in practice (11/13, 84.6%; 1 missing response). All respondents who answered the recommendation item indicated they were likely or very likely to recommend TTrACE to colleagues. Conclusions: In high-performing critical care environments, structured interprofessional training may contribute less to large measurable performance gains and more to reinforcing and sustaining relational, communicative, and psychological processes essential for safe patient care. These findings suggest that TTrACE may support the maintenance of high-functioning team environments in critical care, while future longitudinal and comparative studies are needed to examine longer-term impacts on team functioning, workforce well-being, and patient safety outcomes.

Abstract: Personal environmental exposure monitoring increasingly relies on low-cost multi-pollutant sensors, yet existing devices keep all perceptual intelligence on remote servers. As a result the device cannot self-check sensor degradation in real time, onboarding a new sensor takes days of firmware work, and quality control fails whenever connectivity is lost. We present Zhiwei, a personal-exposure monitoring system that internalizes the reasoning loop on the device. Built on a Raspberry Pi 5, it combines reference-free on-device multi-sensor self-diagnosis, a five-layer declarative skill-package mechanism with a capability-association graph for plug-and-play sensor onboarding, and a three-tier resilient reasoning architecture that sustains quality control offline. Over a 30-day indoor deployment in Beijing comprising 1,896,789 records at 99.9% completeness, the device autonomously graded a nominal oxidizing-gas channel as untrustworthy for ozone from three complementary physical-consistency checks: a temperature dependence of −7.9 ppb per °C, a cross-interference of the wrong sign with NO2, and an apparent drift that became non-significant once confounders were removed. It also confirmed the relative consistency of the PM2.5 and NO2 channels against a nearby reference station, with correlations of 0.90 and 0.86. Zhiwei establishes the feasibility of fully on‑device autonomous sensor quality assurance for trustworthy personal‑exposure monitoring.

Abstract: The increasing complexity of environmental analysis requires new approaches for real-time simulation across in-door and urban spaces. While computational fluid dynamics (CFD) models provide detailed representations of gas dispersion and aerosol transport, they are often computationally intensive and difficult to integrate into emerging AI-native digital twins and agentic AI systems. This pilot study presents a GPU-accelerated voxel simulation framework for modeling three-dimensional gas dispersion and aerosol transport using structured voxel representations derived from BIM, LiDAR, GIS, and digital twin environments. The framework provides physically informed, CFD-inspired simulation at sub-meter to meter-scale spatial resolutions while maintaining interactive runtime performance suitable for building management, ventilation analysis, environmental monitoring, hazard assessment, and emergency response applications. Transport dynamics are modeled using a discretized advection–diffusion formulation incorporating airflow-driven advection, diffusion, source emissions, and voxel-level sink mechanisms. A key contribution is the development of a voxel-native GPU-parallel computational architecture implemented in Python using Taichi kernels. Prototype simulations demonstrate stable transport behavior, browser-based three-dimensional visualization, and efficient execution on commodity GPU hardware. Experimental scenarios include a voxelized three-story Industry Foundation Classes (IFC) building model comprising approximately 34.5 million active voxels (582 × 382 × 155 voxels) and an urban-scale 3D city model spanning approximately 300 × 300 × 150 m and containing up to 13.9 million active voxels. Simulations containing tens of millions of voxels were completed within minutes on a single consumer-grade GPU, demonstrating the scalability of the framework. These results establish a practical foundation for AI-native digital twins and agentic AI-assisted environmental simulation applications.

Abstract: Background: The incidence of placenta accreta spectrum (PAS) with placenta previa has been previously reported. However, the incidence varies across reports, suggesting that unknown risk factors may be involved. This study aimed to reevaluate the risk of PAS in patients with placenta previa. Methods: This multicenter retrospective study was conducted from 2015 to 2024 in patients with placenta previa who underwent cesarean section (CS). PAS was defined based on pathological or clinical findings, such as manual removal of the placenta or obvious retention of the placenta if a hysterectomy was not performed. The incidence of PAS and associated risk factors were analyzed using multivariable logistic regression. Results: PAS was observed in 26% of women with placenta previa. The incidence of PAS increased significantly with the number of prior CSs: 13.2% in women with no prior CS, 41.9% in those with one prior CS, and 66.7% in those with two or more prior cesarean sections. Multivariate analysis identified major placenta previa (aOR 2.69, 95% CI 1.11–6.54), including complete and partial placenta previa, and number of prior CSs (one prior: aOR 4.35, 95% CI 1.94-9.73; two or more prior: aOR 9.48, 95% CI 2.55-35.2) as independent risk factors. Conclusions: The incidence of PAS with placenta previa was higher than that previously reported, and major placenta previa was shown to be an independent risk factor, regardless of prior CS history. Comprehensive evaluation, including prior CS and placenta previa classification, is crucial for accurate risk stratification and perinatal management.

Abstract: Background/Objectives: Suboptimal dietary patterns are among the leading modifiable contributors to global morbidity and mortality, particularly in cardiovascular disease, type 2 diabetes mellitus (T2DM), obesity, metabolic syndrome, and hypertension. Digital nutrition platforms have emerged to improve adherence to evidence-based dietary strategies; however, many systems lack structured optimization, processing-aware nutrient profiling, and explainable artificial intelligence (AI) mechanisms. The integration of large language models (LLMs) into digital health introduces conversational personalization but also risks hallucination and unsafe outputs without constraint enforcement. This study aimed to describe the system development, architecture, database infrastructure, optimization algorithms, explainability enforcement, and digital health implications of NutriSteppe-AI, a chatbot-first LLM-driven system for personalized health menu generation constrained by deterministic nutrient logic and processing-aware scoring. Methods: NutriSteppe-AI integrates: (1) a multi-source structured nutrient database of 20,000 food products with up to 130 tracked nutrients; (2) energy requirement estimation using the revised Harris-Benedict equation; (3) linear programming-based multi-objective optimization; (4) a Healthy Food Index (HFI; 0.5–5.0 scale) incorporating NOVA processing classification penalties; (5) traffic-light nutrient gating; and (6) a constrained LLM orchestration layer governed by structured API contracts. Algorithmic validation was performed using 10,000 simulated user profiles spanning diverse age, anthropometric, activity, dietary exclusion, and budget parameters. Results: The system achieved 96.8% full constraint satisfaction with macronutrient mean absolute errors of 11.60% (energy), 18.86% (protein), 16.26% (fat), and 20.91% (carbohydrates). Incorporating NOVA processing penalties reduced ultra-processed food HFI scores by 0.73 points (P < 0.001). Median optimized menu HFI improved from 3.6 to 4.3. Median system latency was 1.8 s. Explainability validation confirmed 100% deterministic alignment with zero hallucinated numeric claims. Conclusions: NutriSteppe-AI demonstrates that LLM-driven nutrition chatbots can achieve deterministic, explainable, and clinically aligned performance when governed by structured optimization, processing-aware scoring, and explainability enforcement. This architecture provides scalable digital health infrastructure for cardiometabolic disease prevention in diverse populations.

Abstract: The increasing complexity and volume of technical documentation, includ-ing requirements specifications, patents, and engineering reports, creates significant challenges for manual analysis and knowledge extraction. This paper includes a systematic review of methods for semantic content analy-sis of technical documents, with a particular focus on Natural Language Processing (NLP) techniques and Transformer-based models. The study formalizes the task of structured information extraction and provides a mathematical description of Named Entity Recognition (NER) as a core subtask. A practical case study demonstrates an end-to-end NER pipeline for Russian-language technical requirements, leveraging ruRoberta-large via spaCy-transformers. The results highlight both the potential and limitations of current approaches, emphasizing the critical role of annotation con-sistency and document format normalization. This work contributes to the development of intelligent systems for engineering documentation analysis and outlines key directions for future research.

Abstract: This study examines how accounting information on environmental protection expenditure can support fiscal transparency and sustainability risk management in the public sector. Using harmonised Environmental Protection Expenditure Accounts (EPEA) data for EU Member States, together with GDP and population indicators, the paper develops a comparative framework for analysing public-sector environmental expenditure. The study constructs scaled indicators, including expenditure per capita and expenditure as a percentage of GDP, and examines the functional composition of expenditure through the Classification of Environmental Protection Activities and Expenditure (CEPA). Exploratory clustering and panel regression diagnostics are used to identify cross-country expenditure profiles and descriptive associations with macroeconomic indicators. The findings show substantial variation among Member States and confirm that environmental expenditure should not be interpreted as a direct measure of environmental ambition or performance. Instead, differences reflect accounting scope, institutional arrangements, service-delivery models and infrastructure needs. The paper contributes to sustainability accounting, public financial management and sustainable finance by demonstrating how harmonised accounting information can improve comparability, auditability and decision usefulness in public-sector environmental reporting. It also highlights the relevance of environmental expenditure information for identifying fiscal exposure, infrastructure priorities and sustainability-related risks.

Abstract: Calpains constitute an ancient, extensive family of calcium-dependent cysteine proteases found in some bacteria and most eukaryotes. They are involved in a wide variety of developmental and cellular processes and are implicated in major human diseases, but whether they share an ancestral or broadly conserved cellular role remains unclear. Beyond their core CysPc catalytic domain, calpains contain diverse domain combinations and can be either cytosolic or membrane bound. Here, we develop the hypothesis that both cytosolic and transmembrane calpains may contribute to cytokinesis through positional anchoring and organization of microtubules (MTs). We propose that during plant cell division, the singular transmembrane calpain DEK1 play a role in localizing and organizing the array of cortical MTs from the microtubule organizing center (MTOC) and may thereby position the cell division plane, potentially affecting preprophase band placement and subsequent cell plate formation. Similarly, during cell division in animals, their cytosolic calpains may be involved in setting the point of membrane invagination via their association with membrane-bound proteins. We discuss this novel model for calpain activity in the context of data from the animal and plant literature, as well as of our discovery of putative calpain sequences in both brown and red algal genomes. These findings are consistent with the view that calpains were present early in eukaryotic evolution and diversified alongside distinct modes of cell division. Finally, we consider the possibility that early calpain functions may have been linked to the formation and function of MT arrays in flagella and cilia, from which later roles in cytokinesis might have evolved. This model is intended as a testable framework for future studies of calpain function across eukaryotes.

Abstract: Cell types are fundamental biological units and partially independent evolutionary units, shaped by individualised gene regulatory networks and developmental lineages. Despite the recent explosion in single-cell sequencing and increased attention on cell characterization, we still lack a unified and consistent naming system for cell types that works across species. Since cell types are the products of evolutionary diversification, we propose that cell-type names should explicitly reflect evolutionary history, and suggest a naming system with a phylogenetic representation prefix as a simple, informative and intuitive way to do this. The key to this is establishing the evolutionary/taxonomic level of comparison, coupled with understanding homology and innovation in cell-type evolution. Put simply, it can apply to both individual cell types and their clades. We illustrate this approach using two case studies: chordate macroglia and more explicitly on vertebrate photoreceptors. The long-term goal is to stimulate progress towards a more coherent and informative language for cell-type identity and comparative analyses that is evolutionarily extendable as single-cell research proliferates across the tree of life.

Abstract: Left ventricular noncompaction (LVNC) is a cardiomyopathy characterized by excessive trabeculation and deep intertrabecular recesses, yet its molecular mechanisms remain poorly understood. Here, we identify Bcl11b as a novel regulator of cardiomyocyte (CM) growth and ventricular wall maturation. CM-specific deletion of Bcl11b in mice recapitulates key LVNC features, including increased noncompact-ed-to-compacted ratio, impaired compact layer expansion, reduced CM proliferation and size, and systolic dysfunction. Mechanistically, Bcl11b deficiency leads to marked upregulation of Pou3f2, a transcriptional repressor that further suppresses Titin (TTN) expression. Loss of Bcl11b disrupts sarcomere integrity and reduces TTN protein levels, while forced Pou3f2 overexpression similarly represses TTN. Notably, heterozygous loss of Pou3f2 rescues the LVNC phenotype in Bcl11b-deficient hearts, restoring CM growth and TTN expression. Our findings establish a critical relationship among Bcl11b, Pou3f2 and TTN that governs CM proliferation and hypertrophic maturation during cardiac development. Dysregulation of this regulatory network impairs ventricular compaction and contributes to the development of LVNC, providing new insights into disease pathogenesis and potential therapeutic targets.

Abstract: Background: This study applies Buysse’s (2014) sleep health framework to examine sleep in children and adolescents with neurofibromatosis type 1 (NF1). By examining sleep timing, daytime sleepiness, sleep quality, sleep behavior, sleep duration and sleep effi-ciency together, this framework captures the multidimensional nature of sleep and its relationship with biopsychosocial factors and health related quality of life (HR-QoL) in NF1. Methods: This multi-site, prospective, cross-sectional study included 131 children and adolescents with NF1 and 71 typically developing (TD) controls aged 6 to 16 years. A sleep health composite was derived from carer rating scales and seven days of actigra-phy. A biopsychosocial framework was used to examine factors associated with sleep health in NF1, including socio-demographic, cognitive, psychopathology, and biological variables. Independent predictors of QoL were examined, to assess the unique contri-butions of sleep quality, sleep duration and previously established predictors of HR-QoL in NF1. Results: Poorer sleep health was evident in children with NF1. Compared to TD controls, children with NF1 were five times as likely to have poor sleep quality with almost 78% demonstrating impaired sleep efficiency and nearly half not obtaining suf-ficient sleep at night. The strongest risk factors were being male, elevated pain, and having greater levels of ADHD and autism spectrum disorder traits. Conclusions: Findings suggest sleep health in NF1 is interconnected to multiple biopsychosocial fac-tors. A better understanding of these relationships will help identify early risk markers, improve prediction of clinical trajectories, and guide the development of targeted mul-timodal interventions for sleep disruption in NF1.

Abstract: Neodymium (Nd) and dysprosium (Dy) are two critical rare earth elements for fabricating NdFeB permanent magnets, which have crucial applications in modern technologies. The increasing global demand for Nd and Dy emphasizes new efficient processes for their recovery and purification, which are technologically challenging due to their close physical and chemical properties. Through systematic exploration, it was found that Lewatit VP OC 1026 resin impregnated with di-(2-ethylhexyl) phosphoric acid (D2EHPA) had a strong adsorption preference for Dy³⁺ over Nd³⁺, which is highly suitable for Dy-Nd separation from their mixed solutions under optimized conditions. The loaded resin could be eluted using dilute sulfuric solutions for recycling to the adsorption process. By employing a multistage adsorption-elution process analogous to distillation, efficient Dy-Nd separation and purification were realized from their mixed solution, with a prospective purity of 99.13% and recovery of 97.45% for Dy and a prospective purity over 99.96% and recovery of above 99.90% for Nd, despite the large concentration disparity between Dy and Nd where Nd concentration is over 26 times of that of Dy. This research demonstrates that efficient recovery and purification of metals from aqueous solutions can be achieved using selective resin adsorption processes analogous to distillation, despite large concentration differences of the metals in the solutions, which presents new alternative approaches.

Abstract: We derive a Riesz-Frechet representation for bounded linear functionals defined on the padic Hilbert spaces introduced by Kalisch [Ann. of Math. (2), 1947]. We also notice the surprising difference between the Archimedean case and the non-Archimedean case (exact non-Archimedean version of Riesz-Frechet representation fails).

Abstract: This research applies the principles of human resource accounting (HRA) and intangible asset valuation frameworks under IAS 38/IFRS to examine the relationship between the quality of player performance metrics, human capital metrics, and the quality of their financial reporting on the market valuation of football players and the financial performance of the leading football clubs in Europe. Based on a dual-level database, composed by 20 leading European clubs (club-level) and by 120 players (player-level) in the season 2023/24, the study constructs a performance-adjusted valuation model for estimating the interconnection between on-field statistics (goals, assists, expected goals, defensive actions, and performance indices imposed on a composite measure) and accounting or financial number (transfer fees, amortization charges, intangible asset values, book values) and financial results (ROA, club market valuation). The Outcome of multiple OLS Regression Models using Robust Standard Errors shows that Performance Index is the most important predictor of player market value (max 0.497, p < 0.01) whereas club revenue is the most important predictor of club market valuation (max 0.009, p < 0.01, R2 = 0.879). The market to book ratio analysis shows systematic difference between economic value and accounting book value based on player age, duration of contract signed, and performance indicators (Adj. R² = 0.363). The Moderated regression shows existence of positive moderating relationship between IFRS compliance and Big4 audit quality with on-field performance and financial outcomes. The findings add to the intersection of sports finance, accounting and human capital theory, stressing the inadequacy of current IAS 38 provisions in capturing the true economic value of football players as human capital assets.

Abstract: The loess gully region is characterized by complex terrain with crisscrossing gullies,where coal mining can readily induce surface subsidence and slope deformation. Such deformation often leads to geological hazards and ecological issues,including collapses,landslides, soil erosion, vegetation dry up,and land degradation.Therefore,understanding the deformation behavior of mining‑induced slopes is essential for the restoration and management of mine geological environments.This study focuses on five slopes within working faces 50205 and 50206 of the Zhen’er Coal Mine in Fugu County.Using a combination of 3DEC numerical simulations and orthophoto-based fracture identification, we systematically investigated mining-induced slope deformation under the complex topographic conditions of the loess gully region.The goal is to answer three key questions: where mining-induced slope deformation primarily occurs,how it evolves over time, and what the main controlling factors are.Spatially,the primary deformation zones and their propagation paths vary significantly among the five slopes.The largest deformation occurs in the slope body directly above the main section of the working face,gradually decreasing toward the edges of the working face. Temporally, mining-induced slope deformation exhibits a time lag, meaning that surface responses lag behind underground mining activities and continue to develop even after the working face is fully extracted.In the loess gully region, slope deformation induced by mining is controlled not only by mining activities but also by topographic factors such as slope shape, aspect,gradient, and height. The spatiotemporal evolution of deformation becomes even more complex for slopes that span multiple working faces. These findings provide a scientific basis for monitoring mining-induced slope deformation and preventing geological disasters in the loess gully region,while also offering practical guidance for safe mining operations and hazard control in similar settings.

Abstract: Improving soil fertility in the context of climate change is of paramount importance. This study addresses this challenge in Kinshasa (DR Congo) where the combined effect of charcoal waste and Tithonia diversifolia biomass was evaluated in an alley cropping trial, with two successive maize crops. The objective was to sustain optimum maize yields, and to derive insights into sub-Saharan Africa (SSA). Three treatments were applied: a control (T0) plots; and two other plots receiving 5 t ha^-1 of charcoal prior to cultivation combined with alley cropping using T. diversifolia pruned in situ at 50 cm (T1) or at 100 cm (T2)) and applied as mulch. The results showed that Tithonia biomass production reached approximately 100 t DM ha^-1 year-1. Maize grain yields in the first season were higher in the amended plots (2.7 to 2.9 t ha^-1) compared to the control (1.6 t ha^-1). The yields obtained in the second season were similar for all plots, but they declined significantly for T0 compared to the first season. While yields stabilized with amendments, they stayed below SSA self-sufficiency targets (4.45 t ha-1). Improving crop N absorption and use efficiency, which were low in this study, is key to closing the yield gap.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a central paradigm for post-training large language models, yet group-relative methods often suffer from zero advantage failures, where identical rollout rewards erase the policy-gradient signal. A growing body of work addresses this bottleneck by intervening in rollout-group construction to restore learnable contrasts. Among these efforts, methods that introduce external textual signals beyond the model’s own distribution, such as reference trajectories, abstract scaffolds, and reusable experience, have emerged as a key branch, as they can restore learnable contrasts while expanding the model’s capability boundary. This survey provides the first systematic survey of this branch: we introduce Hint as a unifying concept for such external textual signals and organize hint-based RL methods into sample-level hints, covering trajectory-based and scaffold-based guidance, and task-level hints, covering static and evolving experience bases. Beyond taxonomy, we further clarify the boundaries, cross-level analysis of construction and utilization, and future directions. We maintain an up-to-date resource list at https://github.com/WYRipple/Awesome-Hint-Based-RL

Abstract: Background/Objectives: Sepsis is responsible for approximately 270,000 deaths annually in the United States. Conventional scoring systems, such as SOFA and qSOFA, are largely reactive and do not effectively leverage longitudinal ICU data for early prediction. This study aimed to develop a deep learning framework capable of predicting sepsis onset up to 6 hours before Sepsis-3 criteria are met, while also providing clinically interpretable temporal explanations. Methods: The PhysioNet/CinC 2019 Challenge dataset, comprising 1,552,210 patient-hours from 40,336 ICU patients, was utilized. A Temporal Transformer Encoder (TTE) was trained using 12-hour look-back windows with 92 engineered features. Severe class imbalance (2.6% positive rate) was addressed through weighted random sampling and focal loss. Five-fold patient-level cross-validation was employed to prevent temporal leakage. Platt scaling was applied for probability calibration. Grad-CAM was adapted for temporal explainability, while SHAP was used for feature-level attribution. BiLSTM-Attention and XGBoost models served as baseline comparators. Results: The TTE model achieved a cross-validated AUROC of 0.8320±0.0032 and an AUPRC of 0.1505±0.0148, significantly outperforming BiLSTM Attention (AUROC: 0.7859) and XGBoost (AUROC: 0.7731; DeLong p < 0.0001). Platt scaling reduced the Expected Calibration Error from 0.3154 to 0.0017. The median alert lead time was 46.5 hours (IQR: 21–84 h), with 95.3% of septic patients receiving alerts at least 3 hours before onset. Grad-CAM analysis identified timesteps t − 10 and t −9 as the most predictive. However, high-severity patients (SOFA proxy ≥ 3)demonstrated substantially reduced performance (AUROC: 0.257). Conclusions: The proposed TTE framework demonstrates strong and well-calibrated early sepsis prediction with substantial clinical lead time. The concentration of predictive signals 10–11 hours prior to alert generation supports the feasibility of continuous automated ICU monitoring from admission onward. Reduced performance in high-severity patients highlights the need for severity-stratified modelling in future research.