Machine learning models (ML) often require localization to perform optimally in local populations. We hypothesize that smaller community healthcare centers may not have the necessary patient volume to facilitate localization based on statistical guidelines. This work investigates the ability for community medical centers to localize ML and performs a simulation study to evaluate synthetic data generation (SDG) to augment local data for recalibration. We conducted an experiment using data from a real network of hospitals (two rural, one urban academic medical center) to predict 30-day unplanned hospital readmission and using data from a multi-site ICU dataset to simulate using synthetic data generation (SDG) in a network of hospitals of various sizes. We also performed a simulation study using data from a multi-site ICU dataset to evaluate the utility of SDG to augment local data volumes. In the real-world evaluation, the urban medical center met the guidelines for the number of samples for recalibration (Required: 14,224, Available: 42,303) and had the best calibrated model using local data (α=0.1,β=1.05; best: α=0,β=1). For the smaller sites, neither site had the samples required for recalibration (Site 1: Required: 16461, Available: 3187; Site 2: Required: 15299, Available: 905). In the simulation study, deep learning-based SDG was most effective at improving calibration performance. Connections to large medical centers are not enough to promote accurate ML at all sites within a healthcare system. Data augmentation and SDG may provide the necessary data volumes to enable local recalibration at smaller facilities.

CRB-NCE: An adaptable cohesion rule-based approach to number of clusters estimation

Assessing the quality and performance of synthetic data augmentation to identify stigmatizing language in obstetric clinical notes.

Learning to ask and answer in specialized documents: Exemplifying through modular integrated construction regulatory documents

• Synthetic data from generative AI preserves privacy in healthcare modeling. • Generative AI adapts dynamically, surpassing static traditional CEA models. • Enhanced scenario simulations by generative AI aid robust decision-making. • Generative AI integrates real-world evidence, refining predictive accuracy. • Non-linear modeling in AI captures complex healthcare cost-outcome relations. Healthcare economic evaluation increasingly relies on predictive modeling to inform resource allocation decisions. Traditional cost-effectiveness analysis (CEA) methodologies face significant challenges when processing complex, heterogeneous healthcare datasets and accommodating dynamic system variables. This review examines how generative artificial intelligence technologies may transform predictive modeling frameworks in healthcare economics, specifically focusing on potential improvements in accuracy, adaptability, and efficiency in cost-effectiveness analyses. A literature search was conducted across PubMed, Scopus, Web of Science, and IEEE Xplore between October 2024 and January 2025, examining publications from 2018-2024. Critically, we identified a near absence of empirical studies that directly apply and validate generative AI technologies within formal health economic modeling or health technology assessment contexts. Most identified literature addresses general AI/ML applications in healthcare or synthetic data generation in adjacent domains, rather than demonstrating validated use in cost-effectiveness analysis. Generative AI demonstrates promising theoretical capabilities in handling non-linear healthcare relationships, generating privacy-preserving synthetic datasets, and enabling dynamic scenario exploration based on performance in related fields. However, direct empirical evidence comparing generative AI to traditional CEA approaches in real-world health technology assessment remains virtually non-existent. Potential advantages include automated model support, enhanced integration of real-world evidence, and improved handling of missing data scenarios. Technologies such as Generative Adversarial Networks and Variational Autoencoders show early-stage promise in addressing traditional modeling limitations in adjacent applications. Generative AI represents a conceptually significant potential advancement in healthcare economic modeling. However, claims presented are predominantly forward-looking and conceptual rather than empirically validated. Implementation challenges including model interpretability, regulatory frameworks, validation requirements, and ethical considerations require substantial empirical research before successful integration into healthcare decision-making processes.

Enhancing automated shoeprint comparison via synthetic data generation and deep segmentation

In this work, we address the problem of estimating full-field planar strain when a reference image of the undeformed configuration is not available. Classical Digital Image Correlation (DIC) and related image-based methods rely on the comparison between undeformed and deformed states and therefore cannot be applied in such scenarios. To overcome this limitation, we propose a deep learning-based approach that estimates the in-plane strain state directly from a single deformed image. A Convolutional Neural Network is trained to regress the two principal strains and their orientation by learning deformation-induced texture features, formulating strain estimation as a data-driven regression problem. The method is validated using synthetically deformed images generated from both classical DIC speckle patterns and steel microstructure images acquired by optical microscopy. A systematic sensitivity analysis is performed to assess the influence of subset size and training dataset size. On synthetic data, the proposed approach achieves coefficients of determination exceeding 0.96 and average strain errors on the order of 0.01 m/m. Preliminary validation on real experimental images of deformed patterns demonstrates that the method can capture meaningful strain distributions, although with lower precision than reference-based DIC, as expected from a statistical, reference-free formulation. The proposed approach is therefore not intended to replace classical DIC, but to enable strain estimation in experimental situations where reference configurations are unavailable.

Hydraulic models are critical tools for planning and managing water distribution networks (WDNs), yet inaccuracies in nodal demand estimates remain a key source of modeling errors. Demand calibration is essential to enhance model reliability but is challenging due to the large number of unknown demands and limited sensor availability, resulting in an underdetermined problem. This study introduces a novel hierarchical grouping methodology that integrates sensitivity analysis with graph-theoretic shortest-path to cluster nodes into demand groups, effectively reducing the dimensionality of the calibration problem. Building on this, a two-stage calibration framework is proposed: Stage 1 performs conventional calibration across demand groups, while Stage 2 targets sensors with high residual errors by hierarchically subgrouping nodes in their associated groups for focused recalibration. The approach is validated on two looped WDNs and one tree type WDN using synthetic data under multiple sensor configurations. Results demonstrate that the proposed methodology achieves up to 34% reduction in nodal demand estimation errors compared to conventional single-stage calibration, particularly in scenarios with limited sensor coverage. These findings highlight the effectiveness of hierarchical node grouping and two-stage calibration in improving demand estimation accuracy, providing a scalable and practical solution for utilities facing data constraints in WDN modeling and operation.

Twisted convolutional networks (TCNs): Enhancing feature interactions for non-spatial data classification.

Quantification of local dune dimensions from bathymetric data obtained in large sand-bed rivers is important for navigation assessments and hydraulic roughness estimation. This study developed, calibrated, and tested a method for wavelet-based quantification of local dune dimensions (i.e., length and height) from linear bed elevation profiles (BEPs). The method uses the continuous wavelet transform (CWT) and leverages strengths of both the energy-normalized and amplitude-normalized CWT definitions. The method was calibrated using synthetic BEPs that mimic data obtained on the Missouri River to obtain median values and 95% confidence intervals for the method’s dune length and height estimation coefficients. Analysis of results for synthetic BEP data suggests that the wavelet-based method provides locally averaged dune dimensions. When using median coefficient values, root-mean-square errors (RMSEs) for predictions were less than 32% of mean values for the synthetic data. The method was validated by comparisons with reference dune dimension data for the Missouri and Paraná Rivers, for which the RMSE prediction errors were less than 59% and 27% of the mean values, respectively.

Synthetic data and deep neural networks enable prediction of sensory perception attributes in texture-modified plant-based smoothies

Air pollution interpolation is crucial for civil management: it is used to transform limited sensor data into comprehensive pollution maps. Various methods, including deterministic, geostatistical, and Machine Learning (ML)-based techniques have been utilized for this purpose. Deterministic methods rely on mathematical rules for estimation, whereas geostatistical techniques are based on spatial correlations. ML leverages historical data for predictions. Each method has limitations: deterministic methods do not adequately model environmental complexity, geostatistical methods struggle with small-scale areas, and ML depends heavily on data availability. On the other hand, air pollution simulators based on physicochemical dispersion models capture intricate pollution dispersion patterns, yet running them at high resolution for continuous interpolation is often computationally demanding. Recent advancements in ML offer a complementary pathway by leveraging simulated data to enable rapid, real-time interpolation. This study introduces an air pollution interpolation approach combining simulated air-dispersion patterns through an educated machine that includes machine learning and environmental modeling that considers boundaries, pollution sources, obstacles, wind dynamics, and topography. Specifically, we present a combined ML-based linear regression model designed to infer dense concentration maps from a sensor array using state-of-the-art simulation methods. We dub this the Ridiculously Simple data-driven air pollution Interpolation Method (RSIM). The RSIM method was evaluated on both synthetic and real-life-based simulation models. The synthetic scenarios included an industrial area with point pollution sources and an urban road surrounded by buildings simulating traffic-related pollution. The real-world environment consisted of sensor data and simulations from Antwerp, Belgium. The results indicate that this method outperforms standard techniques for reconstructing dense pollution maps from sparse sensing, and demonstrates significant promise for other real-world applications. • We propose a novel method for interpolating air pollution using machine learning. • The approach is validated on both synthetic data and real data from Antwerp, Belgium. • Our method outperforms classical techniques in generating dense pollution maps. • The workflow integrates simulated dispersion fields with real-world sensor readings. • The methodology supports scalable, high-resolution mapping for urban air quality.

• Introduced a novel WGAN-AAE framework for real-synthetic data fusion • Achieved 94.89% accuracy in multi-label pipeline condition classification • Demonstrated high sensitivity of FBG sensors to pressure and flow changes • Identified optimal sensor positions for improved signal and detection • Proposed a robust 2D-CNN architecture tailored for leakage classification Pipelines for water and oil transport are vulnerable to ageing, operational loads, and harsh environments, necessitating reliable condition monitoring. This study presents a hybrid sensing and data-driven framework for accurate leak detection and localisation in steel pipelines instrumented with Fibre Bragg Grating (FBG) sensors. For the research, a 6 m steel pipe was instrumented with FBG sensors placed longitudinally and circumferentially to quantify strain responses under various leakage and operating conditions. The study evaluates the sensitivities of the FBG strain recordings to variations in leak size, leak location, pressure and flow. A hybrid WGAN–AAE data fusion methodology is proposed to analyse limited and noisy strain data for accurate classification of leakages, flow rate and pressure. The framework integrates a Wasserstein GAN (WGAN) that generates synthetic FBG signals tuned to the operating envelope; an adversarial autoencoder (AAE), which provides domain-aware latent regularisation and learned feature-level fusion of real and synthetic data; and a 2D-Convolutional Neural Network classifier that operates on the fused representations. Models are trained on 70% real data augmented with WGAN samples and evaluated on the remaining 30% real data. The study addresses both single-label (pressure, flow, leak size, leak location) and multi-label classification based on a Binary Relevance approach. The classifiers achieved a 94.89% ± 1.13% accuracy for multi-label classification on the test set based on real data. Additionally, single classification tasks show high accuracy rates, with 91.04% ± 0.90% for flow rate classification, over 98.69% ± 0.93% for pressure classification, 100% ± 0.00% for leakage size classification, and 91.22% ± 0.98% for leakage localization. A comparative analysis of leakage-location classification across the ten best sensor layouts shows that WGAN+AAE outperforms GAN+AAE and WGAN-concatenation baselines, supporting the benefits of Wasserstein synthesis, latent regularisation, and learned fusion. These results demonstrate that the proposed hybrid sensing and data-fusion approach enables accurate and robust leak detection and localisation in pipeline systems, even with limited datasets and under noisy, variable operating conditions.

As a basic machine learning task, Multi-View Classification (MVC) has garnered considerable attention and achieved great success. However, the existing MVC methods, especially late fusion style ones still suffer from some problems: 1) hidden valuable information is not well exploited; 2) a lack of interaction before decision making. To address these problems, we propose a novel framework named ”TrashtoTreasure” that leverages mutual information to effectively exploit hidden valuable information. Specifically, the framework explicitly disentangles multi-view information into ”useful” components and ”trash” (noisy) components, and further extracts potentially valuable ”treasure” information from the ”trash”components of all views. Additionally, we design a tailored objective function that facilitates the effective separation of ”useful” and ”trash” components, as well as the synergistic extraction of ”treasure” information. This function guides model optimization through triple mutual information constraints. Experimental results on synthetic data and several real-world data sets verified the effectiveness and superiority of the proposed method. The fresh perspective offered by this article may inspire more interesting exploration in this direction. The codes are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jiezhang054/TrashToTreasure</uri>.

Physics-informed neural networks enable patient-specific tumor microenvironment modeling: Identifying parameters in combined immune therapy with integer- and fractional-order dynamics

Corrosion is one of the big problems in material science, and options to avoid and analyse it represent an active research area. Among the corrosion assessment methods and techniques, Electrochemical Impedance Spectroscopy (EIS) has an important role in determining and explaining the electrochemical mechanisms which occur in applications such as batteries, pipelines, structures, and so on. Its analysis is very specialised, and it can require a certain amount of time and sensibility to achieve high reliability. On the other hand, machine learning (ML) can represent a powerful tool to optimise how the data is analysed and processed. Recent works have brought ML to surface engineering, and the present work is aligned with it. This study presents a proof-of-concept machine learning workflow for the inversion of Electrochemical Impedance Spectroscopy (EIS) data, which specifically refers to the inverse mapping from synthetic EIS spectra to circuit parameters within a known-forward-model framework, using five predefined equivalent electrical circuits (EECs). The proposed approach employs a modular two-stage convolutional neural network (CNN) architecture (ML tandem): one establishes an initial guess (rough fit) and, sequentially, performs a fine-tuning fitting. To isolate methodological performance, the models are trained and evaluated on large, controlled synthetic datasets generated from the same forward EEC equations across five corrosion-relevant circuit topologies. Under these conditions, the architecture reveals a marked circuit-dependent difference in invertibility, which some EECs were identified accurately, whereas others exhibited systematic ambiguities due to overlapping spectral features within explored parameter ranges. Misclassification and inversion failures are also discussed. Overall, this work aims to evaluate the feasibility and limitations of a modular ML-based inversion strategy using controlled conditions and provides a transparent basis for future extensions to experimental EIS applications. • A brand-new ML architecture is proposed to analyse synthetic EIS data. • CNNs were configured in tandem to get an initial guess and fit the EIS data. • Model failure sources in classification and inversion are discussed.

SyntheITS: Synthetic industrial time-series data with prior knowledge and deep generative models for equipment anomaly detection under small samples

DDTracking: A diffusion model-based deep generative framework with local-global spatiotemporal modeling for diffusion MRI tractography.

Application-driven multi-modal depth completion in fringe projection profilometry

Graph neural network-based surrogate modeling for fast and scalable simulations of meshed district heating networks