Search Results (369)

High-resolution remote sensing object detection is a fundamental task in Earth observation. However, the storage, transmission, and downstream processing of high-resolution images impose substantial bandwidth, memory, and computational burdens. Compressed sensing (CS) samples and compresses the signals simultaneously, thereby reducing data transmission and storage overhead. Unfortunately, most existing CS-based pipelines require explicit image reconstruction before downstream inference, leading to heavy computational overheads and poor scalability for high-resolution remote sensing images (RSIs). This work focuses on post-acquisition image compression and explores algorithmically, rather than from a physical hardware implementation perspective, whether explicit image reconstruction is an indispensable intermediate step prior to object detection. To this end, we propose HR²SIOD-CL, an end-to-end compressed learning (CL) framework that performs object detection directly on CS measurements of high-resolution RSIs without explicit image reconstruction. HR²SIOD-CL integrates an entropy-driven content-aware adaptive sampling strategy and a measurement-domain detection backbone for multi-scale feature extraction. Although jointly optimized during training, the adaptive sampling and detection modules can be decoupled for flexible deployment. For fair comparison, an extra lightweight reconstruction network equipped with a single-step data-consistency correction is constructed as the baseline. Extensive experiments on the NWPU VHR-10 and DIOR datasets show that across various sampling ratios, HR²SIOD-CL surpasses the reconstruction-based detection method when integrated into two-stage detectors, and achieves comparable or superior detection performance to the reconstruction-based counterparts when integrated into single-stage detectors. Meanwhile, its computational overhead and GPU memory consumption are merely 6.97% and 35.74% of those of the reconstruction-based counterpart, respectively. These results indicate that CS measurements can function as an effective intermediate representation for object detection, and explicit image reconstruction is not a prerequisite when detection is the primary objective. Full article

Sustainable energy management in grid-connected AC microgrids is investigated through a comparative assessment of centralized, distributed, and decentralized multi-agent-system-based load management control architectures integrating photovoltaic (PV) generation and battery energy storage system. A sustainability-oriented rule-based load scheduling strategy is implemented to efficiently utilize available renewable energy while maintaining grid power consumption within the prescribed demand limits and ensuring priority support for critical loads. Multi-agent-system (MAS)-based load agents coordinate centralized, distributed, and decentralized load management operations. The control architectures are evaluated under identical load profiles, PV generation patterns, and demand-limit constraints to ensure a fair comparison of sustainability-oriented energy management performance. Their resilience is further assessed under agent failures, communication losses, and delays. The comparative analysis employs sustainability-oriented performance metrics, including load served percentage, load curtailment percentage, demand-limit violation duration, and PV utilisation, to assess reliable energy delivery, demand-side efficiency, grid compliance, and effective renewable-energy utilisation. The results reveal distinct architectural trade-offs in sustainable energy management: decentralized control offers greater resilience to agent failures and achieves the highest average load-served percentage (64.83%) and lowest demand-limit violation duration (40.61 ms). Distributed control provides enhanced coordination and priority-based load management, with intermediate performance (64.10%, 41.17 ms), whereas centralized control is constrained by single-point failures and scalability, exhibiting the lowest performance (58.52%, 43.33 ms). PV utilisation remains approximately 99% across all architectures, indicating near-complete utilisation of available solar generation for load supply and battery charging with minimal curtailment. Full article

Feature engineering has become a common step in AI-based HVAC forecasting, often involving variables calculated from raw building management system (BMS) measurements, such as temperature differences, setpoint tracking deviations, airflow balance indicators, rolling statistics, and temporal or seasonal descriptors. Accurate short-term forecasting can provide a baseline of expected operation for anomaly and fault detection and can support control optimization and operator decision making. However, real-world deployment is complicated due to differences in BMS sensor availability and data quality, as well as the preprocessing and maintenance burden associated with complex feature sets. The actual contribution of these features to the performance of AI forecasting remains underexplored, particularly for short-term prediction of air handling unit (AHU) operation. This study evaluates the impact of features on short-term AHU forecasting using three deep learning (DL) architectures: Temporal Convolutional Networks (TCNs), Long Short-Term Memory (LSTM) networks, and a hybrid CNN–LSTM model. An actual operational AHU dataset from a BMS was used to predict key operational variables, including supply and extract air temperatures, supply and extract fan operating signals, and supply air temperature setpoint-tracking error. Fan signal balance was additionally evaluated as a derived indicator calculated from the two predicted fan signals. Four input configurations were evaluated: (i) full (74 inputs), containing raw BMS measurements, short-cycle temporal variables, engineered and dynamic features, and annual-calendar information; (ii) no annual calendar (68 inputs), identical to full but excluding annual-calendar variables; (iii) raw + short-cycle temporal (20 inputs); and (iv) raw-only (12 inputs). The models used a 60-min input history to forecast the following 30-min at one-minute resolution. Persistence and Ridge models were included as reference baselines. All models were trained and tested on identical data splits and forecasting horizons to ensure a fair comparison. Each DL experiment was repeated across five independent runs, and performance was evaluated using MAE, RMSE, and R². The TCN showed the strongest overall DL performance. Raw-only achieved the highest mean R² in 11 of 15 architecture–target comparisons using just 12 inputs. The best mean DL R² ranged from 0.916 for the fan signals to 0.993 for extract air temperature. Annual-calendar features improved the TCN results but provided no consistent benefit for the LSTM or CNN–LSTM. Ridge slightly outperformed the best DL configurations for temperature-related targets, reflecting the strong short-term continuity of these signals. These findings show that recent raw BMS measurements contain most of the information needed for accurate 30-min AHU forecasting, while explicit seasonal and engineered features provide limited additional value. The resulting simpler models may in the future be used as forecasting components in predictive control and fault detection systems. However, their control and energy-saving benefits must be tested separately. Full article

Sinogram images represent projection-domain data acquired during digital breast tomosynthesis (DBT), preserving angular information prior to reconstruction. However, their low contrast and noise-related degradation limit their direct use in downstream tasks. This work proposes a budget-constrained hybrid biomimetic optimization framework for projection-domain contrast enhancement based on the integration of Particle Swarm Optimization (PSO) and the Nelder–Mead (NM) simplex method. The approach combines global exploration with local refinement under a fixed number of objective function evaluations (NFE), enabling fair and computationally efficient comparisons with standalone optimizers. Experiments were conducted on 222 labeled mammographic images (136 benign and 86 malignant), which were transformed into sinograms via the Radon transform. The proposed method achieves competitive performance in terms of PSNR, SSIM, and FSIM, while exhibiting faster convergence and reduced computational cost compared to several state-of-the-art swarm-based approaches. Additionally, the impact of the enhanced sinograms was evaluated through a downstream classification task using a Vision Transformer-based knowledge distillation scheme, demonstrating improved discrimination between benign and malignant cases. These results demonstrate that the proposed budget-constrained hybrid biomimetic strategy provides an effective and computationally efficient solution for contrast enhancement in projection-domain medical imaging. Full article

Remote-sensing-derived spectral indices and in situ measurements provide complementary information for aquatic monitoring, but their practical integration is complicated by asynchronous observations, heterogeneous tables, missing acquisitions, and non-reproducible preprocessing. This study presents AI-Aquatica-RS, a modular Python framework for spectral-index calculation, station-aware nearest-neighbor temporal alignment, feature-set construction, regression benchmarking, command-line execution, and structured result export. The software was evaluated using a fully synthetic controlled benchmark; no real satellite scenes or field-monitoring measurements were used. The benchmark comprised 1080 daily in situ-like observations from six stations and 181 unique remote-sensing-like acquisitions generated as water-like surface-reflectance proxies. A ±3-day alignment tolerance produced a shared complete-case cohort of 954 records. To ensure a fair comparison, the in situ-only, spectral-index-only, and fused configurations were evaluated on exactly the same 667 training and 287 validation records. The fused configuration achieved the best performance using ridge regression (RMSE = 3.481 NTU, MAE = 2.768 NTU, R² = 0.729), compared with RMSE values of 5.086 NTU for the in situ-only configuration, and 5.538 NTU for the spectral-index-only configuration. The benchmark demonstrates reproducible execution and recovery of complementary information under controlled conditions; it does not constitute environmental validation. AI-Aquatica-RS provides an extensible software layer for future studies using real satellite products, monitoring networks, sensor-specific preprocessing, and spatially blocked validation. Full article

As regulatory requirements increasingly shape automated lending decisions, fairness remains a critical challenge in high-stakes domains, particularly credit scoring. Although artificial intelligence models can achieve strong predictive performance, they may also reproduce biased outcomes that reduce financial inclusion or transfer harm to overlooked protected groups. Existing fairness interventions commonly operate at a single stage of the decision-making pipeline, despite bias often propagating across representational and decision layers. This study proposes iCert-Fair, a two-layer framework for technical fairness assessment and harm recovery in credit scoring. The first layer adopts a fairness-through-explainability paradigm, using SHAP-based explanations to identify direct and proxy dependence on protected attributes and guide structural dataset repair, while the second layer applies targeted threshold-policy adjustments to recover residual harm while preserving decision utility. Experiments on the German and Taiwanese credit datasets show that fairness gains are model- and dataset-specific and may be collective, concentrated, transferred, or recovered unevenly across protected attributes. The direct comparison with representative pre-processing, in-processing, and post-processing methods revealed that baseline methods targeting one protected attribute at a time frequently transferred residual harm to other monitored attributes. In contrast, the fairness-focused recommendations generated by iCert-Fair achieved larger collective fairness improvements across all considered protected attributes while avoiding residual harm. These gains were obtained while preserving predictive utility on the German dataset and with utility degradation remaining below 5% across the evaluated performance metrics on the Taiwanese dataset, alongside consistently lower false-negative risk. The empirical findings support the use of complementary structural and policy-level interventions and demonstrate the importance of jointly evaluating aggregate disparity, worst-case attribute-level harm, cross-attribute transfer, and predictive utility. Full article

Reliable Predictive Maintenance (PdM) in Industrial Internet of Things (IIoT) environments is challenged by severe class imbalance, heterogeneous sensor variables, inconsistent experimental protocols, and deployment constraints. This study proposes a Physics-Informed Benchmarking Framework that integrates engineering-guided feature construction, Mutual Information (MI)-based feature relevance analysis, standardized model development, and deployment-oriented evaluation within a unified and reproducible workflow. Using the AI4I 2020 Predictive Maintenance Dataset, Logistic Regression, Isolation Forest, Random Forest, and Extreme Gradient Boosting (XGBoost) were evaluated using identical feature representations, train–test partitions, preprocessing procedures, and imbalance-handling strategies. The engineered feature space incorporates thermal, mechanical, interaction, and degradation-related information derived from the original sensor measurements. The results show that tree-based ensembles provide the strongest overall performance under severe class imbalance. Random Forest achieved an accuracy of 0.986, an F1-score of 0.722, and a ROC-AUC of 0.983, providing the best balance between failure detection and false-alarm control. XGBoost achieved an accuracy of 0.978, a recall of 0.853, and the lowest inference latency of 0.35 ms, indicating its suitability for latency-sensitive IIoT deployment. These findings demonstrate that combining engineering-guided feature representation with a standardized evaluation protocol enables fair comparison of representative learning paradigms while preserving engineering interpretability and deployment relevance. Full article

Educational platforms increasingly personalize which AI learning-path features (adaptive homework, learner choice) each student receives. The natural correlational baseline ranks students by predicted performance—deliver the feature to those expected to do well—a heuristic that need not identify who actually benefits. We formalize feature recommendation as a causal targeting-policy problem: rank students by the estimated conditional average treatment effect (CATE) of a feature and recommend to the top of the ranking. We prove three results: (i) causal top-CATE targeting maximizes policy value at any budget and weakly dominates predictive (outcome-based) targeting, strictly when the two rankings disagree; (ii) a split-sample doubly robust evaluation of targeting quality is leakage-free (null-exact in finite samples), whereas the naive in-sample version is optimistically biased; and (iii) greedily targeting by CATE traces the optimal cost–benefit (Qini) frontier, with the deployment rule “recommend when $\widehat{\tau }>0$.” We validate the method on 17 randomized embedded experiments from the ASSISTments platform. Because the 40 held-out splits re-partition the same students, we do not treat them as independent replicates: we calibrate every headline comparison against a within-experiment permutation null and an experiment-clustered bootstrap. Under that calibrated inference, causal targeting outperforms predictive targeting for adaptive homework at every budget (permutation $p\le 0.005$, the resolution floor of 200 replicates; Holm-corrected $p\le 0.040$), while for learner choice the same contrast is directionally consistent but not statistically significant (permutation $p=0.23$–$0.38$; clustered $p=0.42$). The direction is stable in both families: no leave-one-experiment-out refit reverses its sign. Predictive targeting is nonetheless the one rule that is reliably worse than the alternatives, because it recommends the feature to high-performing students who benefit least—realized benefit falls monotonically across predicted-performance deciles (from $+0.087$ in the lowest to $-0.030$ in the highest). Against a fuller baseline suite, causal (CATE) targeting does not beat random, a simple risk-based rule (target low performers), or treating everyone. Indeed, the estimated benefit ranking is close to noise—split-half rank agreement is $\rho \approx 0.002$–$0.008$ and its calibration slope is $0.018$, far below the ideal of 1—so the gain over predictive targeting comes from avoiding an actively harmful ordering rather than from recovering individual benefit. A fairness analysis shows why this matters: predictive targeting is regressive, concentrating feature access on high-ability students, whereas causal and risk-based targeting reverse that gradient in this corpus; no policy differentiates by neighborhood opportunity zone. Group-conditional policy values, however, are not individually distinguishable from zero once dependence across students and experiments is accounted for; what survives resampling is the allocation itself—predictive targeting directs $0.33$ fewer of its recommendations to low-ability than to high-ability students (95% CI $[-0.46,-0.01]$, experiment-clustered)—so we frame the fairness result as improved access, not established equity gains. The actionable finding is therefore narrow and specific: outcome-based targeting systematically mis-allocates learning-path features and should be replaced by some benefit-aware rule; whether that rule needs to be a learned CATE model, rather than a simple risk-based heuristic, is not established by this corpus. Full article

Background/Objectives: Meniscal tears frequently occur during athletic activities and are often associated with anterior cruciate ligament injuries. The purpose of this study was to compare the posterior tibial slope (PTS) in young adults undergoing arthroscopic surgery for isolated meniscal tears with that of a control group, and to investigate its association with meniscal tears. Methods: This retrospective, single-centre, controlled study included 190 patients aged 18–30 years who presented with unilateral knee pain and underwent knee magnetic resonance imaging (MRI) between 2015 and 2022. Patients with an isolated medial meniscal tear (Group A, n = 67) or an isolated lateral meniscal tear (Group B, n = 31), confirmed on MRI and directly with arthroscopy, were compared with a symptomatic control group who presented with knee pain but had no meniscal or anterior cruciate ligament (ACL) tear on MRI (Group C, n = 92). Medial and lateral PTS angles were measured on pre-existing standing lateral knee radiographs. Interobserver and intraobserver reliability were quantified using intraclass correlation coefficients (ICCs). Group comparisons, receiver operating characteristic (ROC) analysis with the Youden index, and multivariable binary logistic regression adjusted for age, sex and side were performed. Results: Both medial and lateral PTS angles were significantly higher in Groups A and B than in the control group (p < 0.05). ROC analysis identified a cut-off of 9.2° for the medial PTS in the medial-tear group (AUC: 0.756, 95% CI: 0.68–0.83; sensitivity: 64.18%, specificity: 80.43%) and 9.9° for the lateral PTS in the lateral-tear group (AUC: 0.709, 95% CI: 0.60–0.82; sensitivity: 58.06%, specificity: 81.52%). The association persisted after adjustment for age, sex and side (medial slope in medial tears adjusted OR: 1.36 per 1°, 95% CI: 1.19–1.55, p < 0.001; lateral slope in lateral tears adjusted OR: 1.25 per 1°, 95% CI: 1.08–1.44, p = 0.003). Conclusions: In this retrospective cohort of young adults, an increased posterior tibial slope was associated with isolated meniscal tears. A medial PTS greater than 9.2° and a lateral PTS greater than 9.9° were associated with medial and lateral meniscal tears, respectively. Given the modest sensitivity and the observational design, PTS should be regarded as a supportive morphological marker rather than a stand-alone diagnostic metric, and a causal relationship cannot be inferred based on the results of this study. Because discrimination was only fair (AUC 0.693–0.756) and sensitivity limited (54.8–64.2%), these cut-offs are not clinically applicable decision thresholds and should be used only in conjunction with clinical assessment and MRI. Full article

Our first objective was to quantify the associations between health-related events (HRE) before insemination, the relative increase in estrus intensity (REI) at insemination, and the probability of cow-level pregnancy per artificial insemination (P/AI) in organic Holstein dairy cows. Quantifying these associations may aid on-farm decision-making, such as setting the voluntary waiting period, choice of type of semen, do-not-breed and culling decisions. A second objective was to develop predictive models to estimate P/AI based on readily available data, and present common goodness-of-fit results also used in the machine learning community. All data were collected from a certified organic dairy farm in the western USA from 2019 to 2021. Health-related and reproduction data were obtained through Dairy Records Management Systems (DRMS; Raleigh, NC, USA). Activity data were collected using pedometers (IceRobotics, Stirling, UK) mounted on the rear legs. The REI, defined as walking steps per hour before insemination divided by the cow’s baseline steps per hour, was available for 17,238 inseminations from 4759 cows. The REI was categorized as ≤200%, >200–400%, >400–600%, or >600%. The HRE were available for 65,684 inseminations from 13,365 cows. The HRE were categorized as mastitis, metabolic disease (i.e., hypocalcemia, ketosis, displaced abomasum, digestive problems), reproductive disease (i.e, metritis, endometritis, pyometra, retained fetal membranes), lameness, 2 different diseases, ≥3 different diseases, or as healthy (none of these diseases prior to insemination). Combinations (COMBO) between REI categories and 0, 1, or ≥2 HRE were also created. Data were split into training and test sets. The training data were used to fit three logistic regression models that included either HRE, or REI, or COMBO. Each of the three models also included the covariates of 3-mo herd-average P/AI prior to insemination, days in milk, and the fixed effects of parity, insemination season, days after the previous insemination or days to 1st insemination. A random effect accounted for repeated inseminations within cow. Parameter estimates, odds ratios, and the estimated marginal means of the estimated P/AI of the fixed effects were obtained from the logistic regression models. The models’ estimates were applied to the test datasets, and discrimination and calibration statistics were calculated to judge goodness-of-fit. Unadjusted mean P/AI were 0.31, 0.28 and 0.28 for the HRE, REI and COMBO training datasets. For the HRE model, estimated P/AI ranged from 0.20 (≥3 different HRE) to 0.30 (healthy). The estimated P/AI associated with four REI categories were not different from 0.27 in the REI model. The estimated P/AI associated with the combinations of HRE and REI in the COMBO model varied from 0.18 after ≥2 HRE and >200–400% REI, to 0.30 when inseminations were in healthy cows with REI >600%. Inseminations in older cows, in the spring, and outside 18–24 d after the previous insemination were also associated with lower estimated P/AI. The area underneath the Receiver Operating Characteristic curve ranged from 0.57 (COMBO) to 0.60 (HRE) for the test data, indicating fair discrimination ability of the models. Calibration plots showed that the prediction models produced unbiased predicted P/AI. In conclusion, the results showed no conclusive evidence of greater estimated P/AI related to greater REI as a measure of estrus activity. More HRE were associated with lower estimated P/AI. Combinations of low REI and more HRE were associated with notably decreased estimated P/AI. The logistic regression models produced unbiased predicted P/AI. We found no evidence that the strength of the relationship between REI and P/AI depended on the HRE category. The applications of the results are as follows. First, these predictive models may help inform insemination decisions in organic dairy cows, although further external validation is recommended, and the discriminatory performance is weak. Second, a variety of goodness-of-fit statistics were calculated to allow comparisons of the current logistic regression analyses with future analyses made by other machine learning techniques. Full article

Interconnector geometry strongly affects gas transport, polarization loss, and pressure drop in planar solid oxide fuel cells (SOFCs). In this study, four interconnector configurations were investigated for an anode-supported planar SOFC, including one conventional straight-rib interconnector and three connected-rib interconnectors, namely circular-rib (CI), rectangular-rib (RI), and triangular-rib (TI) designs. A three-dimensional multi-physics model coupling electric field, flow field, species transport, and temperature field was established and validated against experimental polarization data of the conventional straight-rib cell. To ensure a fair comparison, all interconnectors were designed with the same interconnector–electrode contact area. The effects of rib configuration on electrical performance, overpotential components, reactant distribution, velocity distribution, and pressure drop were systematically analyzed. At 800 °C, the peak power densities of CI-SOFC, RI-SOFC, and TI-SOFC increased by 4.9%, 9.7%, and 11.7%, respectively, compared with SI-SOFC. The connected-rib interconnectors mainly reduced cathode-side activation and concentration overpotentials by improving oxygen redistribution beneath the ribs. Among the four configurations, the TI-SOFC showed the highest power density and the strongest under-rib transport enhancement, while the RI-SOFC provided a better compromise between flow uniformity and pressure drop. Full article

Energy scarcity and subsequent global fuel market shocks have become a significant concern for the United States. Process heating in industry accounts for over half of all industrial energy usage and is almost entirely (>95%) supplied by natural gas, coal, and byproduct fuels. Many existing alternatives, technologies, and strategies can reduce dependency on this fossil fuel usage to promote energy security and competitiveness. It is expected that adoption of alternatives is limited by capital investment. A working group exploratory exercise was performed to evaluate the current barriers to alternative process heat technology adoption for the manufacturing sector. Although the working group’s findings emphasize electro-technologies due to their emergence as the central topic of discussion, we have contextualized these results by providing a fair and consistent comparison against several alternative thermal technology options. The most reported issue was the lack of financial incentives, with the second key issue for engineers and manufacturing sector decision-makers being the lack of awareness or understanding of available alternatives. This paper aims to provide an analysis of the levelized cost of heating for a variety of alternatives as part of addressing concerns from the emergent patterns reported in the exploratory exercise, as well as provide guidance on barriers to further adoption. Levelized cost of heating analysis considered the capital investment, operations and maintenance, lifespan, and fuel stream costs of various systems delivering heat to the product or process as a generalized cost per megawatt-hour delivered. The analysis indicates that biomass burners, industrial open-loop heat pumps, and central-receiver heliostat fields are at cost parity under average 2024 United States natural gas and electricity prices against well-optimized natural gas burners, though further cost reductions are necessary for consistent adoption. Full article

As organizations lean more heavily on their IT systems, managing cyber risk is gaining increasing importance. Organizations are often challenged to determine which cybersecurity risk framework they should adopt. Choosing the right framework can have a significant impact on the quality of governance, operational resilience, and assurance in risk reporting. However, most prevalent cybersecurity risk frameworks vary significantly in their intent, design, and analytical approach. This makes it difficult for organizations to understand how each framework may meet their business needs. This study presents an AI-enhanced multi-criteria decision support approach for evaluating cybersecurity risk frameworks. The model incorporates machine learning-driven risk scoring as a conceptual input layer, enhancing the objectivity and analytical rigor of the comparison without executing new predictive algorithms. The methodology includes a hybrid approach of literature review, document analysis, and multi-criteria decision analysis (MCDA) to compare and rank NIST CSF, ISO 27001, FAIR, OCTAVE, and CRAMM based on eight criteria that are designed to represent modern requirements for risk frameworks, including governance, scalability, quantitative focus, and interoperability. These criteria also reflect differences in security metrics supported by each framework to provide an organized means to compare qualitative versus quantitative measurement methodologies. The results indicate that NIST CSF performs the best overall in agility, business alignment, and interoperability. ISO 27001 outperforms all others in established governance and compliance. FAIR outperforms all others in quantitative risk analysis and provides superior analytical depth that other frameworks do not offer. OCTAVE and CRAMM function well in legacy systems but lack scalability and are not well-suited for modern distributed systems. Robustness analysis shows that the ranking of NIST CSF, ISO 27001, and FAIR is consistent under different weighting combinations and industry types. The result of this research demonstrates that a combined or hybrid approach to cybersecurity risk framework selection, such as using NIST CSF with FAIR, can give organizations a more well-rounded foundation for applying machine learning-enabled risk analytics with cyber controls. This research also offers a reusable decision support tool that organizations can leverage when aligning their risk priorities to the features of cybersecurity risk frameworks. Full article

Background/Objectives: This scoping review examined recent evidence on the use of transformer-based language models, encompassing encoder-only architectures (e.g., BERT and its clinical variants) and generative large language models (LLMs; e.g., GPT-4 and Llama), to support clinical decision making from unstructured clinical notes, with implications for behavioral-health services where narrative documentation is central. Methods: Following PRISMA-ScR guidelines, PubMed, PsycINFO, and Web of Science were searched for peer-reviewed studies published between 1 January 2023, and 5 August 2025. Studies applying transformer-based language models to clinical narratives for healthcare tasks and reporting evaluative outcomes were included. We extracted data on clinical tasks, model architectures, enhancement strategies, and evaluation metrics; mapped each study by primary purpose, care setting, and primary model approach; and charted reported validation design, direct human comparison, fairness assessment, workflow evaluation, and clinical deployment. Results: Thirty-six studies were included. Information extraction/de-identification and classification/prediction predominated, whereas summarization/generation was less commonly represented. Model approaches appeared to align with task characteristics: encoder-only and decoder-only systems were frequently used for extraction, encoder–decoder systems for generation, and hybrid or pipeline-based approaches for classification and prediction. Standard task-specific metrics (e.g., F1 and AUROC) predominated, whereas evidence beyond retrospective task performance, including direct human comparison, fairness assessment, workflow evaluation, clinical deployment, and temporal or external validation, was rare. No included study evaluated a transformer-based language model application within a behavioral-health service or behavioral-health workflow. Conclusions: Transformer-based language models have been applied across diverse clinical-note tasks, but the evidence base more strongly supports retrospective task feasibility than transportability, equitable performance, workflow benefit, or safe clinical deployment. Future research should prioritize transparent reference standards, external and prospective validation, clinically meaningful human comparison, and equity-focused evaluation, including direct studies in behavioral-health services. Full article

Cloud and quality screening removes approximately 65% of daily TROPOMI tropospheric NO₂ pixels, creating structured data gaps that coincide with meteorological conditions driving pollution extremes. Standard gap-filling methods—kriging, Random Forests and other machine learning methods—act as statistical smoothers that systematically suppress extreme concentrations and ignore the Missing Not At Random (MNAR) character of cloud-induced missingness. Here we present a physically informed framework that treats urban NO₂ as a forced nonlinear dynamical system and reconstructs missing satellite observations through geometric navigation on a shadow manifold rather than statistical interpolation. The framework integrates five components: (i) Multivariate State-Space Reconstruction (MSSR) using multiview embeddings of continuous ground-based NO₂, O₃, and ERA5 meteorology, grounded in Stark’s forced-system embedding theorem; (ii) Short-Time Regime-Conditioned Convergent Cross Mapping (ST-RC-CCM) with a spatial-mismatch negative control for falsifiable causal validation; (iii) Inverse Probability Weighting (IPW) to correct the clear-sky sampling bias; (iv) trajectory-matrix denoising via Singular Spectrum Analysis (SSA) and Robust PCA; (v) topology-inspired fidelity metrics—Manifold Overlap Ratio (MOR) and Dynamic Trend Capture (DTC)—that penalize smoothing artefacts. The physical basis for this coupling is the shared dynamical history of surface and column NO₂: tropospheric NO₂ has a photochemical lifetime of 1–4 h near urban emission sources, comparable to the boundary layer mixing timescale, ensuring that surface and column concentrations are jointly governed by the same emission–photolysis–transport attractor. The planetary boundary layer height (PBLH), solar zenith angle (SZA), and surface O₃—all included as MSSR coordinates—are the dominant physical drivers of the instantaneous surface-to-column scaling, and their joint trajectory in state space constitutes the physically grounded basis for analogue selection. The framework is validated on a synthetic forced Lorenz-96 system, then applied to five European primary cities spanning contrasting regimes (Sofia, Milano, Stuttgart, Kraków, Hamburg) plus five N1 spatial-mismatch control stations (Plovdiv, Genova, Frankfurt, Warszawa, Berlin)—ten urban-background stations across four countries—with structured ablations (A0-A4V-A4K). Across >3600 evaluations, MOR_ext distributions for EDM and non-EDM methods are non-overlapping by a factor exceeding 5× (EDM minimum 0.59 vs. non-EDM maximum 0.10; median non-EDM MOR_ext ≤ 0.05 at every city × mask combination), while EDM achieves MOR_ext up to 0.915 (Milano Po Valley). Under a fair-comparison benchmark that withholds ground-level NO₂ from Random Forest, EDM’s RMSE advantage remains robust at a median of 3.9× (RF_FULL) and increases to 4.2× (RF_METEO), confirming that the performance gap is physical rather than an information artefact. A three-level temporal validation—within-window pseudo-cloud masking, cross-year transfer (full 2022 holdout and DJF 2023/24), and a COVID-19 out-of-distribution test—demonstrates robustness beyond standard train/test splits, with CCM library-length convergence confirmed for 60/60 ablations (p < 0.001) across all ten stations. Spatial-mismatch tests confirm local dynamical specificity at all five primary–control pairs (Δρ = 0.090–0.210), with seasonal modulation driven by orographic and synoptic mechanisms. These results establish manifold-based gap filling as a dynamically informative complement to statistical approaches, particularly in topographically confined, stagnation-prone basins where preserving extreme-event geometry is essential for exposure assessment. Full article