Estimating causal effects in quasi-experiments with spatio-temporal panel data often requires adjusting for unmeasured confounding that varies across space and time. Gaussian processes (GPs) offer a flexible, nonparametric modeling approach that can account for such complex dependencies through carefully chosen covariance kernels. In this paper, we provide a practical and interpretable framework for applying GPs to causal inference in panel data settings. We demonstrate how GPs generalize popular methods such as synthetic control and vertical regression, and we show that the GP posterior mean can be represented as a weighted average of observed outcomes, where the weights reflect spatial and temporal similarity. To support applied use, we explore how different kernel choices impact both estimation performance and interpretability, offering guidance for selecting between separable and nonseparable kernels. Through simulations and application to Hurricane Katrina mortality data, we illustrate how GP models can be used to estimate counterfactual outcomes and quantify treatment effects. All code and materials are made publicly available to support reproducibility and encourage adoption. Our results suggest that GPs are a promising and interpretable tool for addressing unmeasured spatio-temporal confounding in quasi-experimental studies.

ABSTRACT Geographic Artificial Intelligence supports smart city land management, where modeling complex inter-parcel relationships and extracting effective features remain key challenges for accurate land use classification. Urban areas exhibit diverse relationships including spatial similarity between adjacent blocks, configurational similarity between non-adjacent blocks, and heterogeneous relationships among functional zones. However, existing research lacks comprehensive frameworks to fully describe these complex interaction systems. We propose a graph neural network framework based on higher-order Markov inference that integrates three types of complex relationships for urban land use identification. The framework utilizes social media check-in data to construct a third-order transition matrix, explicitly modeling population mobility’s chain influence mechanism. It employs hypergraph structures to fuse point-of-interest semantic features with remote sensing visual features, capturing similarities among spatially distant but functionally homogeneous areas. Finally, it integrates multi-source feature embeddings and block adjacency relationships through distance-weighted graph attention networks. Empirical studies using real data demonstrate superior performance compared to traditional machine learning methods. Higher-order activity type inference performs optimally in areas with high population density, monofunctional land use, and heterogeneous destination land use patterns for inter-regional travel. This model provides scientific modeling approaches and analytical tools for urban land use planning and smart city management.

Brain tumors impair brain function both locally and across distant regions by disrupting network connectivity, contributing to cognitive deficits and triggering compensatory plasticity. Traditional resting-state fMRI methods average brain activity over time, missing transient, dynamic events critical to cognition. Co-Activation Pattern (CAP) analysis captures these brief brain states, enabling quantification of state engagement and duration. To investigate alterations in transient brain states in patients with brain tumors using CAP analysis of resting-state fMRI, and to assess whether these changes reflect modified engagement of cognitive states compared to healthy controls. This retrospective cross-sectional study included 106 patients with left-hemispheric brain tumors (72 high-grade gliomas, 19 low-grade gliomas, 15 metastases; mean age 61.15 ± 8.95 years; 43 women) and 100 age-matched healthy controls. Resting-state fMRI data were analyzed using a seed-free clustering method (TbCAPs toolbox) to extract CAPs. CAPs were first identified in controls and then matched to patients via spatial similarity. Each CAP was assigned to a canonical brain network using the GIFT toolbox. Dynamic metrics computed included: persistence (duration of a CAP), transitions (switching frequency), in-degree, and out-degree. Group comparisons used two-tailed t-tests with Benjamini-Hochberg correction (p < 0.05). Six CAPs were identified. Compared to controls, patients showed significantly increased transitions, in-degree, and out-degree, and decreased persistence in CAPs linked to the default mode and executive control networks (all p < 0.01), suggesting more frequent but less stable engagement. Visuospatial network CAPs demonstrated lower transition, in/out-degree and persistence in patients (p < 0.05). No significant differences were observed among tumor types. Patients with brain tumors display altered CAP dynamics involving higher-order cognitive networks and perceptual networks. These alterations may reflect the combined effects of tumor-related network damage and potential adaptive reorganization, with potential implications for functional preoperative planning and prognosis. However, in the absence of direct clinical or neuropsychological correlations, these interpretations remain hypothesis-generating. The findings are also limited by the retrospective cross-sectional design preventing causal or longitudinal interpretation, and the restriction to left-hemispheric tumors. Future studies integrating CAP dynamics with cognitive and clinical outcomes will be necessary to determine whether these changes reflect compensatory functional reorganization in brain tumor patients.

Spatial transcriptomics (ST) enables mapping gene expression in native tissue context to resolve architecture and cellular interactions, but current analytical workflows rely on separate algorithms for distinct tasks. We present RGAST (Relational Graph Attention network for ST analysis), a framework that builds upon and extends our earlier HERGAST model (specifically designed for large-scale ST data analysis) for diverse downstream analysis. By introducing a relational graph attention auto-encoder, RGAST jointly models spatial proximity and gene expression similarity to capture both local and global structures in ST data. This design enables a wide range of downstream tasks within a single framework. Through comprehensive benchmarking, RGAST demonstrates superior performance in spatial domain identification across multiple platforms, improving adjusted rand index by ~10% compared to the second-best model in the dorsolateral prefrontal cortex dataset. RGAST accurately reconstructs known neuroglial interaction patterns in the mouse hypothalamus, including long-range signaling pathways that are often missed by distance-constrained methods. Moreover, RGAST also excels in boosting spatially variable gene identification accuracy, delivering more precise inference of developmental trajectories in the human cortex, and robust reconstruction of 3D tissue architectures from serial sections. Collectively, these results establish RGAST as a powerful tool for providing coherent solution to advance ST data analysis across multiple research scenarios.

AI automated grid placement in the OMERACT knee inflammation MRI scoring system (KIMRISS) for bone marrow lesion assessment: A multi-reader exercise.

Beijing’s urban development has moved from the era of expansion to renewal. The exploration of industrial spatial evolution laws aims to tackle the major theoretical and practical problems in Beijing’s high-quality urban development. Few studies have analyzed the spatial and temporal patterns of industrial spatial aggregation based on long time-series data and fine-grained spatial scale industrial data sets. Based on the multitype industry data set from 2006 to 2020, this study uses a spatial similarity model and a circle analysis method to deeply analyze the rule and circle differentiation characteristics of Beijing’s industrial dominant types and aggregation levels. The results show that (1) the spatial similarity index of Beijing industry increases, showing the general characteristics of the continuous enhancement of the degree of aggregation. There exists a high aggregation degree and an obvious aggregation trend within the service industry and a low aggregation degree between the service and manufacturing industries. (2) The spatial aggregation of producer and living services is always high, and the aggregation of public services or cultural and creative industries with other services increases significantly. (3) The differential distribution of spatial circles and the slow process of aggregation are the main reasons for the spatial separation of the service and manufacturing industries. (4) Beijing’s industrial spatial aggregation shows a significant ring pattern, and the dominant types and aggregation levels of industrial evolution vary in different circles. The research conclusions scientifically support the optimization of the capital’s spatial structure and its high-quality spatial development.

Focal epilepsy is increasingly conceptualized as a network disorder, yet the extent to which network dysfunction reflects a shared phenotype remains unknown. Spatially conserved patterns of network dysfunction may implicate a centralized mechanism underlying widespread impairment. Here, we investigate whether network connectivity disruptions are spatially similar across temporal lobe and extra-temporal lobe epilepsy cohorts and whether shared dysfunction aligns with thalamic connectivity profiles. We retrospectively analyzed resting-state magnetoencephalographic imaging from 71 individuals with nonlesional, drug-resistant focal epilepsy (n = 45 temporal, n = 26 extratemporal), collected between 2014 and 2023, and healthy controls (n = 18). Source reconstructed time series were bandpass filtered, and long-range functional connectivity was quantified using imaginary coherence. Network disturbance maps were computed as T-score maps, comparing functional connectivity in epilepsy cohorts to controls, across topographical parcels and frequency bands. Spatial similarity of temporal and extratemporal network dysfunction maps were assessed using Pearson correlations. To infer thalamic involvement, shared network dysfunction maps were correlated with normative functional magnetic resonance imaging-derived thalamocortical connectivity profiles. Extra-temporal lobe epilepsy demonstrated reduced global network connectivity relative to controls in the delta (p = .012), alpha (p = .034), and gamma (p < .001) frequency bands. Across all frequencies, the spatial patterns of network disturbances between temporal and extratemporal cohorts were significantly correlated (r = .287-.717, all p < .001), indicating a shared network dysfunction. Shared spatial maps of network dysfunction correlated with normative thalamocortical connectivity profiles, with significant correlations in the anterior, pulvinar, and dorsomedial thalamus. Nonlesional focal epilepsy exhibits a common, frequency-dependent pattern of cortical network dysfunction that is spatially aligned with thalamic connectivity, supporting a thalamic hub contribution to widespread network impairment.

Spatiotemporal clustering of state-level opioid, opioid-stimulant, and opioid-benzodiazepine polysubstance deaths across the U.S.: Hierarchical agglomerative clustering analysis.

Understanding the mechanisms of brain function and dysfunction is at the core of the neuroscience mission. However, the field's grasp of causal relationships between brain properties has been hindered by a focus on single modalities that neglects the complex interplay between the features found at different neural scales. Progress in neuroinformatics and the increasing availability of open datasets have helped overcome this limitation by facilitating the contextualization of brain maps against cellular, metabolic and network features. Despite the rapid uptake of data contextualization methods proposing that quantification of spatial similarity between brain maps may shed light on pathways of structure-function coupling, development and disease, their potential pitfalls have received little attention. In the context of neuroimaging research, these limitations include reliance on often small-sample and non-representative reference datasets, repeated use of the same brain maps across studies, and problems with intermodal and interindividual alignment. Applying data contextualization without considering these limitations can lead to circular reasoning, overfitting and correlational overreach, and limits the interpretation of findings to the properties of the source data. Here we provide a Roadmap of practical guidelines operating at the level of study design, analysis pipelines and interpretation of findings to encourage the development of best practices in data contextualization. A more informed use of brain map correlation approaches will improve mechanistic investigations and our understanding of causal relationships between brain properties.

PurposeDeep learning-based 3D dose prediction boosts radiotherapy planning efficiency and consistency, yet most models rely solely on anatomical data and assume homogeneous beam configurations, impairing their robustness in esophageal cancer intensity-modulated radiotherapy (IMRT) with heterogeneous beam arrangements. This study explored whether explicit beam geometry modeling enhances voxel-level dose prediction accuracy, robustness in rare beam configurations, and clinical workflow efficiency.MethodsA retrospective analysis was performed on 751 esophageal cancer IMRT patients. Two models with the identical AS-NeSt backbone were constructed: an anatomy-only dose prediction model (ADP) and an anatomy-and-angle-based model (AADP) integrating ray-tracing-derived beam geometry representations (normalized beam coverage and overlap maps) accessible in pre-planning. Model performance was assessed on an independent test set (n = 100), a rare-beam configuration cohort (n = 33) and a clinical validation cohort (n = 42), against clinical plans as the reference baseline. Evaluation was based on dosimetric metrics and isodose spatial similarity, along with an analysis of its robustness on unseen beam configurations and impact in a crossover clinical workflow study.ResultsCompared with ADP, AADP significantly reduced prediction errors for most targets and organs at risk, cutting average dosimetric error from 2.88% to 2.02%, with prominent improvements in low-to-intermediate dose regions (lung and heart). The mean Dice similarity coefficient of isodose volumes rose from 0.90 to 0.93. For rare beam configurations, AADP exhibited superior robustness (average error: 3.08% vs. 4.52%). Notably, AADP-assisted planning shortened total planning time by up to 65% (from 111.4 to 38.5 min for junior physicists), reduced iterations, and improved intra- and inter-physicist dose consistency.ConclusionsThis study confirms explicit ray-tracing-based beam geometry modeling enhances the accuracy, robustness, and clinical utility of 3D dose prediction for esophageal cancer IMRT, supporting beam geometry as a critical component of clinically deployable, physics-informed dose prediction models.

Computerized diagnosis of brain tumor using graph based CNN classification.

MultiSP deciphers tissue structure and multicellular communication from spatial multi-omics data.

Accurate molecular annotation is essential for deciphering biochemical processes in spatial biology. Here, we present a scalable and broadly applicable molecular annotation tool for tandem mass spectrometry imaging (MS2I). Our workflow includes parallel image acquisition (PIA) for parallel MS2I and an open-access computational framework for spatial similarity networking (SSN) that enables molecular annotation of MS2I data with isomeric specificity. The PIA enables simultaneous untargeted MSI and targeted MS2I ensuring structure-specific imaging of hundreds of molecules in a single experiment. The SSN increases annotation confidence through graph-based spatial correlation of product ion distributions, opening up new avenues for data investigation and annotation from both MSI and MS2I data. By integrating PIA and SSN into a single workflow, we visualize and annotate 134 phospholipid isomers and isobars in mouse brain tissue. Furthermore, we demonstrate the biological utility of the platform by mapping cholesterol metabolism in human multiple sclerosis brain tissue, revealing lesion-associated cholesterol oxidation pathways. Finally, we propose annotation confidence levels for structural annotation in MSI. Overall, PIA and SSN together provide large-scale, structure-specific MSI, expanding the scope for spatial metabolomics, lipidomics, and chemical pathology through molecular annotation beyond current capabilities.

Tau protein aggregation in the brain is a hallmark of Alzheimer’s disease (AD). Positron emission tomography (PET) is the only in vivo method to visualize tau pathology and estimate both its burden and regional distribution, but the use of tau-PET is constrained by high cost and limited accessibility. Here, we develop a deep learning model to synthesize tau-PET scans from more accessible data: structural magnetic resonance imaging (MRI), demographics, and when available, blood biomarkers. We included 5,191 participants across the AD continuum or with another neurological disorder from 13 cohorts (mean age 70 years, 51% female) and optimized a 3D U-Net neural network with residual and attention units for this task. In held-out test data, synthetic tau-PET reliably modeled tau burden, with correlations of R=0.77–0.86 with true tau-PET across individuals in common AD regions of interest. Spatial similarity between synthetic and true tau-PET was likewise high, with mean regional correlation of R=0.75. Synthetic scans also captured clinically meaningful prognostic information comparable to true tau-PET, including distinction between early (HR=12, p<0.001) and late (HR=45, p<0.001) stages of tau accumulation. These findings demonstrate that clinically informative synthetic tau-PET scans can be generated from widely available modalities using deep learning, potentially offering a scalable and cost-effective approach for estimating tau AD pathology in the brain.

Background/Objectives: Abdominal aortic aneurysm (AAA) surveillance is based largely on monitoring the maximum diameter, a single scalar metric that obscures regional remodeling and offers limited information on the location and time dependency of the growth rate. The present work addresses this limitation with a geometry-based patient-specific framework that learns local, linear evolution from longitudinal clinical imaging, yielding 3D forecasts of AAA geometry at arbitrary future times. Methods: Lumen and outer wall surfaces are represented on a centerline-anchored cylindrical grid, with subsequent implementation of individualized linear mixed-effects models. The model is explicitly interpretable as the fixed effects predict global trends and the random effects represent regional heterogeneity. In a multicenter cohort of 79 patients, we evaluated forecasts using spatial similarity (with the 95th percentile of the Hausdorff distance—HD95) and clinically relevant global geometric scalars such as maximum diameter and volume. Results: When forecasting a future AAA geometry, the model achieved sub-millimetric HD95 spatial errors and less than 6% error for the aforementioned global scalars. The model was deployed in an interactive application named the Aneurysm Forecasting Studio, which allows a user to visualize the AAA in an explorable forecast space. Conclusions: During typical clinical surveillance intervals, AAA geometric remodeling is reasonably approximated as locally linear in time, enabling transparent, fast forecasts that support surveillance optimization, threshold timing, and digital twin-based interventional planning.

Contrasting land-use sources of microplastic and endocrine-disrupting compound pollution in a major coastal river network.

Abstract Variations in the thermal forcing of the Qinghai–Tibet Plateau (QTP) constitute an important precursor for downstream climate prediction. Especially for its downstream population, the Sichuan Chongqing region with dense economy and frequent extreme high temperatures in summer has a significant impact. Based on meteorological observations from the QTP and the Sichuan–Chongqing Region (SCR) during 1981–2023, this study investigates the spatiotemporal characteristics of the winter Land–Air Temperature Difference (ΔT) over the QTP and its predictive relevance for summer high temperatures in the SCR. The results show that winter ΔT over the QTP has exhibited a significant upward trend over the past 43 years, with an interdecadal shift occurring in the early 21st century. The spatial enhancement is mainly concentrated in the eastern and southern parts of the plateau. Singular value decomposition (SVD) analysis reveals a dominant coupled mode between pre-winter ΔT over the QTP and summer high temperatures in the SCR, explaining 91.74% of the covariance and passing the 99% confidence level in Monte Carlo tests. The corresponding time coefficients are significantly positively correlated, indicating a stable cross-seasonal statistical linkage between winter thermal anomalies over the plateau and downstream summer heat. Composite circulation analysis further supports this relationship. On this basis, a statistical prediction framework using only pre-winter ΔT over the QTP as a predictor is constructed. Hindcast experiments for summer high temperatures during 2021–2023 demonstrate that the method effectively reproduces the primary spatial patterns of summer heat in the SCR, with high spatial similarity between predicted and observed fields. These results suggest that pre-winter ΔT over the QTP can serve as an important precursor signal for summer high temperatures in the SCR and has potential value for cross-seasonal heat prediction.&amp;#xD;

Accurate prediction of student performance is crucial for enabling timely interventions and supporting data-driven instructional strategies. Beyond single-term forecasts, it also facilitates early alerts and targeted feedback for learners and instructors. However, many existing models handle spatial relationships and temporal sequences independently, which limits their ability to capture the intricate dependencies embedded in academic data. In this study, we propose a unified spatial-temporal graph aggregation (USTGA) framework that embeds spatial similarity directly into temporal dependency modeling, forming a coherent and dynamically evolving graph representation. Unlike conventional architectures that treat spatial structure as static or auxiliary, the unified graph enables spatial relations to actively modulate temporal information propagation. To further enhance representation learning, an attention-guided aggregation strategy is employed to adaptively highlight informative spatial-temporal neighbors, while stacked aggregation layers progressively refine node representations across multiple scales. By jointly capturing fine-grained spatial-temporal interactions, the proposed framework overcomes limitations inherent in independent modeling strategies. Experiments on an educational dataset demonstrate that USTGA consistently outperforms competitive baselines, achieving reductions in multiple error metrics under repeated evaluations. The results confirm the framework’s effectiveness and robustness, and its design offers high scalability and adaptability, making it well-suited for deployment in diverse educational analytics systems such as early-warning dashboards and program-level advising workflows.

ABSTRACT Aim It is well‐documented that the tropical forest biota is vertically stratified, and ecological theories from studies of the latitudinal gradient have been applied to predict and understand how communities vary across vertical strata. In butterflies, differences in abiotic conditions between the canopy and the understorey promote the evolution of distinct flight morphologies and physiologies. However, how these distinct morphologies relate to differences in dispersal ability is poorly explored and we lack a general understanding of how and why vertical stratification influences community turnover in tropical forests. Here, we explored how vertical stratification influences diversity, horizontal spatial similarity of assemblages and distance‐decay patterns in understory and canopy butterflies across multiple ecosystems. Location Seven forest ecosystems in Ecuador, South America. Taxon Butterflies (Lepidoptera: Papilionoidea). Methods We assessed patterns of diversity by employing data from standardized butterfly monitoring programmes during the years 2011–2019 across seven sites in Ecuador (37,370 records from 1099 species), and a framework based on metacommunity theory and Jost's diversity estimates. Results Our results suggest three vertical patterns for neotropical butterfly communities: (a) a strong partitioning of canopy and understorey subcommunities, with distinct resulting diversity profiles; (b) greater spatial similarity for the canopy assemblages compared to the understorey (both locally and regionally); and (c) steeper distance‐decay patterns for understorey assemblages compared to the canopy. Main Conclusions Our study shows the generality of vertical stratification diversity patterns across multiple Neotropical ecosystems, including previously unstudied montane cloud forests. It also shows that horizontal variation in community composition depends on the vertical position of taxa within Neotropical forests and is in general consistent with predictions based on species ecology and morphology.

Multicomponent approach to optimize the performance of existing air quality networks. Practical applications