Smart hospitals deploy Internet of Medical Things (IoMT) sensors and MQTT brokers to stream clinical telemetry over resource-constrained edge gateways. Centralized network intrusion detection systems (NIDS) expose sensitive traffic traces and create single-point failures; federated learning (FL) avoids raw-data centralization but remains vulnerable to client-side model poisoning and server-side inspection of client updates. This study presents Trust-Weighted Federated Aggregation (TW-Fed), a reputation-aware defense framework for federated learning that integrates 1) cosine similarity-based trust scoring using low-dimensional update sketches and 2) secure aggregation of trust- and data size-weighted model updates through pairwise masking with dropout recovery. TW-Fed has been evaluated on the CICIoMT2024 benchmark (Wi-Fi and MQTT subsets) with a four-class NIDS task (Benign, MQTT-Connect-Flood, TCP-DDoS, ARP-Spoofing) and 20 edge clients emulating IoMT gateways. In the clean setting, TW-Fed achieves 99.21% accuracy and 98.74% macro-F1, exceeding FedAvg by 0.82 macro-F1 points. Under label flipping with 30% malicious clients, TW-Fed sustains 97.06% accuracy versus 82.14% (FedAvg), 85.33% (FedProx), and 92.07% (Krum). Under backdoor injection, TW-Fed reduces attack success rate from 91.5% (FedAvg) to 6.8%. On Raspberry Pi 4 clients, TW-Fed adds 0.09 s local overhead per round while reducing server aggregation time by 34% relative to Krum. Secure aggregation increases uplink volume by 9.7% while preventing per-client update disclosure.
Suspended particle pollution (PM10) represents a serious public health problem in Metropolitan Lima, since its high levels are related to an increase in respiratory and heart diseases. This work aims to anticipate PM10 concentrations in seven districts of Lima using a multiple linear regression (OLS) model with the Ridge and Lasso regularization techniques. PM2.5 and NO2 were selected as predictors because they were the only pollutant variables available in the SENAMHI monitoring dataset; hourly records collected between 2015 and 2024 were used, yielding 169,308 observations after removing missing values and outliers. The data were separated into 80% for training and 20% for validation, applying cross-validation to determine the most appropriate regularization parameter. The final model reached a coefficient of determination (R2) of 0.23 and a mean squared error (MSE) of 521 in the test set. The findings indicate that PM2.5 is the main factor that predicts PM10, while NO2 exerts a secondary impact. Although the predictive capacity of the models is limited, the use of both Ridge and Lasso helped stabilize the coefficients and minimize overfitting. Given this limited explanatory power, the model should be regarded as an initial, interpretable baseline for PM10 monitoring in Metropolitan Lima rather than an operational forecasting tool. It is suggested to include meteorological variables and investigate non-linear models in future studies to enhance the accuracy and usefulness of the model in air quality management.
Production AI systems typically apply governance—compliance checking, auditing, and explanation—as an external control layer that observes outputs after they are produced. In this paper, embedded governance is proposed: a governance gate that forms part of the system architecture itself, positioned between inference and release, so that no output leaves the system without a measurable governance assessment. For each candidate output, a governance score is computed from three factors that are checked mechanically rather than judged by a model: intent conformance, formulation traceability, and verification. On the basis of this score, each output is released, remediated, or blocked. Governance efficiency η, the ratio of intercepted error cost to governance cost, is further defined, by which the claim that governance pays for itself becomes falsifiable rather than asserted. Weights and thresholds are calibrated by maximizing η on a validation split and reported on held-out data. Evaluation uses the full credit-card fraud detection dataset of 284,807 transactions under a temporal split, so that the evaluation period follows the training period and distribution drift arises naturally. The calibrated gate intercepts 99% (95% CI 94.6–99.8) of severely shifted requests at a mean overhead of 3.74 ms, and η reaches 4.52 when releasing unreliable outputs carries cost, while remaining below break-even under accuracy-only cost models. A severity sweep locates the detection boundary and shows interception falling to 2% under mild shift. An ablation identifies distributional checks as the load-bearing components and reveals that η can be inflated by a gate that governs less, so efficiency must be read alongside coverage. Under identical workload, external monitoring and post-hoc assessment detect the same errors but prevent none, whereas inline placement reduces released error cost by 82.
In the aviation industry, flight delays represent a major challenge due to their economic impact, operational disruptions, and adverse effects on passenger satisfaction and transportation efficiency. This study proposes an integrated feature engineering framework for flight delay prediction that combines three groups of engineered features, namely Departure Disruption Features (DDFs), Airline-Route Stability Features (ARSFs), and Schedule Intensity Features (SIFs), with the original flight delay dataset. These feature groups are designed to capture complementary operational, historical, and scheduling characteristics that are not directly represented in the original data. Five machine learning classifiers were evaluated across five experimental scenarios. These included the original feature set, the original feature set combined separately with each engineered feature group (DDFs, ARSFs, and SIFs), and the original feature set combined with all three engineered feature groups. This design enabled assessment of the individual and complementary contributions of the proposed features to flight delay prediction. The results indicated that the engineered features improved prediction performance across the evaluated configurations. In particular, combining all three engineered feature groups with the original feature set outperformed the baseline feature set and each individually combined feature-set configuration, supporting the complementary predictive value of the proposed engineered features. The complete feature set with LightGBM and SMOTE-Tomek achieved the best performance, with an accuracy of 0.9820 and an AUC of 0.9884, compared with 0.9487 and 0.9663, respectively, for the baseline feature set.
Based on Sommerville’s robust software reliability theory and eight design principles based on best practices (DPG), this study conducted a comparative assessment of two Dexcom software platforms: the Dexcom Clarity web portal and the Dexcom ONE+ iPhone application. The first phase of the study employed a “black-box” assessment framework, combining off-the-shelf security analysis tools with application-layer behavioral testing workflows developed specifically for this project. The initial two phases focused on the web portal, analyzing communication encryption and HTTP security header configurations. Subsequent phases involved a more in-depth implementation assessment, focusing on actual interactive features, data types, and automated responses triggered by glucose sensor alerts within the application. However, this approach also revealed that certain system elements required access to the source code for verification. The resulting workflow enables the assessment of all externally visible security attributes while identifying system components that necessitate source code access for validation. Of the eight assessment criteria, four were found to have issues, either in whole or in part. Key issues identified include: session cookies missing the SameSite attribute (G1); acceptance of clinically abnormal carbohydrate values without rejection or warning (G2); inclusion of the unsafe-inline directive in the Content Security Policy (CSP), increasing the risk of Cross-Site Scripting (XSS) attacks (G4); and the absence of a dedicated alert mechanism for Wi-Fi connectivity loss during critical medical functions (G7). Due to the requirement for source code access, this “black-box” approach could not evaluate the remaining four criteria (G3, G5, G6, and G8). Finally, this paper presents a re-producible evaluation procedure and offers applicable, standards-compliant mitigation recommendations for each identified issue.
Transformer health index (THI) prediction supports condition-based maintenance by mapping dissolved-gas, oil-quality, and furan indicators to an interpretable asset-condition score. This study develops a supervised benchmarking framework using 3,392 transformer oil diagnostic records from 510 transform-ers. Thirteen diagnostic features from dissolved gas analysis, oil quality analysis, and furan analysis are evaluated across three modelling groups: static machine learning, static deep learning, and variable-length temporal deep learning. Transformer-level splitting is applied to reduce data leakage from repeated trans-former histories. Static gradient-boosted tree models achieved the strongest results. XGBoost obtained the lowest test root mean square error of 3.4254 with a coefficient of determination of 0.9780 and health-index class accuracy of 88.20 percent. The best temporal model, the liquid time-constant neural network (LTC-LNN), achieved a test root mean square error of 4.8691 and a coefficient of determination of 0.9567. Permutation feature importance identified 2FAL, C2H2, and dielectric break-down as the most influential predictors. For the present dataset, static gradient-boosted tree models produced the strongest ob-served point-estimate performance, while LTC-LNN remained a promising temporal-learning alternative for repeated diagnostic histories. However, the uncertainty analysis indicates that small numerical differences among the leading static models should not be interpreted as definitive evidence of model superiority.
Three-dimensional condominium information systems require more than visual building models: spatial objects must be constructed from authoritative plans and linked consistently to cadastral, registration, building, unit, parcel, and juristic-person records. This study presents CondoMaps, a database-driven architecture for constructing, organizing, and serving 3D condominium information in Thailand. The workflow transforms verified floor plans and cadastral references into georeferenced LOD2 building and unit geometries, assigns persistent project–building–floor–unit identifiers, stores administrative and legal attributes in a normalized relational database, and publishes the integrated objects through controlled map, scene, query, and REST services. The accepted production snapshot contains 7,150 condominium projects, more than 180,000 units, over 650 LOD2 buildings, at least 6,500 contextual LOD1 buildings, 194,304 spatial features, and 457,936 linked attribute records. The combined database and file repository occupies approximately 6 TB and supports 48 published services across 22 land-office branches. Quality checks on the accepted snapshot found complete mandatory documentation and fields, no duplicate normalized keys, and no orphan records in the reported integrity tests. Linkage rates of 100% apply only to the specified sets of 7,150 evaluated records or objects and are not estimates for all units. Horizontal RMSE values of 0.0775 m and 0.031674 m were obtained in two four-control-point plan rectification cases; they are case-level results, not evidence of repository-wide positional accuracy. A browser implementation demonstrates retrieval of associated unit, juristic-person, coordinate, title-deed, and parcel-survey information from a selected 3D building. The contribution is the traceable end-to-end linkage among source documents, normalized records, identifiable 3D objects, and governed web services rather than visualization alone. Performance and formal security benchmarks remain future work.
Human-guided quadruped locomotion requires re-liable intent inference and safe authority allocation under un-certain interaction and uneven terrain. This study proposes an uncertainty-adaptive shared-control framework with multimodal intent fusion for a Unitree Go2 quadruped. Interaction force, relative human–robot motion, operator-motion cues, and terrain context are fused through a temporal model to estimate desired motion and maneuver mode. Intent confidence is evaluated using posterior entropy and prediction variance, while locomotion risk is computed from traversability, slope, obstacle proximity, interaction load, and stability margin. The human authority factor is then adjusted online to blend the inferred human command with a robot-safe command. The shared command is realized through gait scheduling, terrain-aware foothold optimization, whole-body control, and interaction-compliant base regulation. Experiments across five guidance scenarios show that the proposed method improves intent accuracy, interaction load, slip resistance, stability margin, and task success compared with force-only guidance, fixed-authority shared control, and ablated variants.
Drug-drug interactions (DDIs) are a leading cause of preventable adverse drug events, and the standard benchmark corpus for extracting them from biomedical text is severely im-balanced towards non-interacting pairs, making rare interaction types hard to learn. We asked whether combining focal loss, entity-aware transformer encoding, and feedback-guided feature fusion in a single architecture could improve minority-class DDI extraction under this imbalance while requiring less training data than prior methods. We propose FoLT-DMCNN-FBVL, which integrates a BiomedBERT backbone with entity marker injection, a multi-branch neural classifier, and feedback-based validation learning, trained on a small balanced subset of the SemEval- 2013 DDIExtraction corpus and evaluated on the full blind test set. FoLT-DMCNN-FBVL matched the most recent state-of-the-art result (statistically indistinguishable; nominal absolute+0.33%) and significantly outperformed the widely cited CNN-DDI baseline (absolute +4.32%), while using less than one-fifth of the available training data. These findings show that a doubly imbalance-aware transformer architecture can match or exceed current state-of-the-art DDI extraction performance with substantially greater data efficiency. Such architectures could support more scalable and cost-effective pharmacovigilance and clinical NLP pipelines in settings where annotated data for rare interaction types is limited.
As society’s digital infrastructure grows, so do the threats. We cannot deny internet access, so we need to safeguard ourselves against cyber threats. Every day, new threats emerge. The most conventional systems classify cyberattacks in a closed set; that is, training and testing use the same set of attack behaviors. In the real world, we cannot train our system on all attack patterns because attack behavior evolves daily. To address this issue, this study proposes open-set cyberattack classification, in which we train our system on a limited set of labeled attack types and, during testing, use a larger unlabeled data sample that includes both known and unknown attack patterns. Our framework integrates a Deep Autoencoder (DAE)-based latent representation, an Explainable Deep Q-Network (Explainable DQN) for feature selection, and a Multi-Head Residual Attention Deep Neural Network (MHRA-DNN) for final classification. The DAE transforms high-dimensional network traffic into a low-dimensional latent feature representation. This reduces redundancy while preserving important feature characteristics. The Explainable DQN selects the features that contribute most to the model’s performance. It also improves explainability by highlighting the most influential features. The selected features are then processed by the MHRA-DNN. It captures multiple feature combinations and strengthens learning through attention mechanisms and residual connections. We evaluate the proposed framework on the benchmark cybersecurity datasets UNSW-NB15, APA-DDoS, and CICIDS2017. We set up an open-set environment that involves known and unknown attack classes. The proposed model performs strongly across accuracy, precision, recall, F1-score, specificity, MCC, and NPV. In addition, parameter tuning improves model stability, and SHAP-based analysis provides evidence of feature contributions to model performance. Overall, the proposed model offers an adaptive, explainable, and robust solution to cyberattack classification in an open-set environment.
Semi-supervised anomaly detection is widely pro-posed for Internet of Things water quality monitoring in aquaculture. However, reported high scores often result from definitional label leakage. This occurs when an observation is labelled anomalous based on a physicochemical threshold, and that same variable is used as an input feature. The benchmark then measures threshold recovery rather than true anomaly detection. To quantify this artefact, we introduce two trivial reference detectors: an all-positive predictor and a per-channel range rule. Using a rigorous leakage-aware protocol, we evaluated Isolation Forest, one-class SVM, a Liquid State Machine, a convolutional autoencoder, and an LSTM autoencoder on two corpora: the three ponds of a public aquaponics dataset that survive a sensor-integrity screen fixed before modelling, eight of eleven ponds having been rejected, and a six-pond dataset from East Java in which all positive test windows originate from two ponds. One-class SVM proved the strongest detector on both corpora, achieving a Matthews correlation coefficient up to 0.77, although with two corpora the shared detector ordering is indicative rather than established. Crucially, on the public corpus, both autoencoders and the trivial range rule performed identically poorly, with recall capped near twenty-one percent. This collapse stems from normal-class contamination caused by duration-based labelling, which inflates the percentile threshold for reconstruction-based models. Furthermore, we demonstrate that the F1 score of a zero-information all-positive predictor exceeds several trained detectors, rendering F1 unsafe as a primary metric at these prevalences. Because the labels are derived from husbandry thresholds rather than expert annotation, and because our multi-seed statistics quantify reproducibility on fixed partitions rather than generalisation across farms or seasons, these results bound the interpretation of the benchmark and not field performance. We conclude by proposing a strict reporting protocol for aquaculture anomaly detection studies, with guidance for applying it retroactively to already-published corpora.
Large Language Models (LLMs) offer strong capabilities for Natural Language Processing, yet their inherent uncertainty often produces hallucinations, confident but incorrect statements, which is critical in domains requiring precise knowledge representation. Retrieval-Augmented Generation (RAG) reduces this risk through information retrieval, but standard pipelines still suffer from fragmented context and weak alignment between queries and legal provisions, limiting trustworthy knowledge extraction. This study proposes a Metadata-Aware RAG architecture to improve grounding in large legal corpora. It integrates: 1) Sub-chunking with Legal Metadata Inheritance, which transforms unstructured legal PDFs into granular, metadata-rich fragments; and 2) an Adaptive Filter Creator, a pipeline that extracts structured constraints and compiles optimized hybrid retrieval queries. These components enhance semantic alignment, reduce uncertainty-driven hallucinations, and strengthen neural information retrieval. Using a curated Peruvian labor law corpus and 150 manually validated question–answer pairs, the system was evaluated across three LLMs (Llama-3.1-8B, GPT-OSS-20B, Gemma-3-27B). The proposed architecture achieves double-digit improvements over a Naive RAG baseline across all four RAGAS metrics—Context Precision, Context Recall, Factual Correctness, and Faithfulness—with gains ranging from 13.10% to 28.15%; notably, Faithfulness surpasses 0.90 for Gemma-3-27B. Statistical analysis confirms significance (t(11) = 15.49, p = 4.06 × 10−9) with an extremely large effect size (Cohen’s d = 4.47). Regression results show minimal influence of model size (slope < 0.005), indicating that retrieval design has a stronger influence than parameter count in the evaluated setting.
Code smells are structural indicators of poor soft-ware design that have been empirically associated with higher defect rates and increased maintenance effort. Although research on smell detection, software quality risk, and maintenance effort estimation has grown significantly, these areas have largely been studied separately. In addition, the role of smell severity in predictive modeling remains insufficiently synthesized. This Systematic Literature Review (SLR) investigates empirical studies published between 2010 and 2025, while also including selected foundational studies where appropriate. A total of 38 studies were selected for analysis through a structured search and screening process. The review addresses two research questions: 1) how smell type and severity influence software quality risk and maintenance effort, and 2) the extent to which current predictive approaches incorporate smell-related information in software defect prediction. The findings show that design-level smells are more consistently associated with fault-proneness than method-level smells. The review also finds that severity-aware models generally outperform binary detection models. However, the independent effect of code smells on directly measured maintenance effort remains inconsistent when confounding factors such as code size and code churn are taken into account. Furthermore, fully unified models that jointly predict defect risk and maintenance effort using smell severity as structured input remain largely absent. Overall, these findings highlight the need for standardized severity operationalization, integrated datasets, and interpretable unified modeling frameworks to advance software quality analysis.
Cell segmentation continues to be a significant challenge in medical image processing, as traditional segmentation methods fail to provide precise and thorough segmentation of complex structures such as touching and overlapping cells. The geometric data points of these cells are inadequate for accurate contour estimation to yield a correct segmentation. Adaptive shape priors have demonstrated efficacy as a promising solution in these challenging circumstances. They can be employed to restrict segmentation models in medical imaging and enhance segmentation robustness in the presence of a complex structure of touched and overlapped cells. In this study, Adaptive Shape Prior Model (ASPM) is introduced and evaluated using three cervical cell datasets, where all their images have a varied number of cells with different properties, such as texture and degrees of overlap. The segmentation performance of ASPM demonstrates that the contour/shape based deformation procedure yields good segmentation accuracy and can be successfully adapted to numerous cytology images. Compared to the other segmentation models for overlapping cells in the literature, the proposed ASPM offers quicker with more precise segmentation for highly overlapped cells in different cervical images without extensive modifications. These results imply that the proposed ASPM approach is a promising choice for integration into a fully automated cervical cancer screening system.
Real-world optimization landscapes are typically dynamic, high-dimensional, and uncertain, and a single meta-heuristic with fixed control parameters rarely sustains strong performance across such environments, as formalized by the No Free Lunch theorem. Existing adaptive frameworks attempt to address this through online operator or algorithm selection, but they suffer from two persistent limitations: coarse feed-back mechanisms that reward the frequency rather than the magnitude of improvements, and cumulative memory bias that allows early-performing algorithms to monopolize selection long after their advantage has faded. This work proposes a problem-agnostic adaptive framework that integrates the Relative Improvement Metric (RIM), a proportional reward quantifying the magnitude of each solver’s contribution, with a Sliding Window (SW) policy that discounts older rewards temporally so that algorithmic influence remains contingent on recent effectiveness. The framework orchestrates a heterogeneous portfolio of eight metaheuristics (GA, PSO, GWO, ACO, SSA, ABC, WOA, FA) through probabilistic selection driven by SW-RIM weights, with a small base probability that prevents any solver from being permanently excluded. Empirical evaluation across 23 standard benchmarks and the CEC2020 suite shows that the proposed mechanism reduces maximum solver participation from above 70% in the baseline configuration to under 40%, improves mean fitness on the majority of functions in both groups, and achieves statistically significant gains over both a baseline portfolio and a sliding-window-only variant (Wilcoxon p < 0.001; Vargha-Delaney A12 between 0.72 and 0.85, large effect across all comparisons). The two exceptions are functions with deceptive or ill-conditioned landscapes (a narrow-valley Rosenbrock-type function and a highly multimodal Schwefel-type function), where all three configurations perform comparably, indicating that SW-RIM’s benefit is contingent on the portfolio containing at least one solver structurally suited to the current landscape rather than on the selection strategy alone. The results support SW-RIM as a lightweight, general-purpose mechanism for sustaining diversity and impact-sensitive adaptation in complex continuous optimization, without the training cost of reinforcement-learning-based selectors.
Artificial intelligence has become a central paradigm in personalised learning. However, it remains unclear whether the two principles that responsible AI treats as foundational, explainability and equity, are integrated into the field’s intellectual architecture or are only loosely associated with it. This question is addressed through a bibliometric analysis of 1371 publications indexed in Scopus and Web of Science between 2015 and 2025. Performance analysis and science-mapping techniques were applied using bibliometrix and VOSviewer, with cluster structure quantified through modularity and inter-cluster density metrics, and uncertainty assessed through 1000-iteration edge-resampling bootstrap. Keyword co-occurrence analysis showed that ethics-equity vocabulary connects to the applied-AI cluster at densities comparable to those of technical vocabulary (0.20 vs 0.21), but the ethics-equity and technical streams interact directly only at half that rate (0.11; bootstrap 95% CIs non-overlapping), engaging primarily through the applied-AI mediator cluster rather than with each other. Co-citation analysis showed that the foundational responsible-AI references in education do not register as cluster-forming nodes. No specialist responsible-AI venue appears in the field’s Bradford core, and Sub-Saharan Africa and Latin America remain almost absent from the country-level collaboration network. To account for this pattern of vocabulary diffusion without infrastructural consolidation, we propose the Structural Integration Model (SIM). The SIM is a three-proposition framework that specifies the conditions under which explainability and equity can become structurally embedded in the field’s foundation rather than only thematically present within it.
Sleep is essential for maintaining overall physical and mental health, yet analyzing sleep patterns manually is a com-plex and time-consuming process that requires expert knowledge and is often prone to subjectivity. In this study, an automated and efficient deep learning-based approach is developed to classify sleep stages using the Sleep-EDF dataset. Physiological signals such as Electroencephalogram (EEG), Electrooculogram (EOG), and Electromyogram (EMG) are pre-processed and segmented into fixed-length epochs for analysis. These segments are then fed into models including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM architecture, which effectively capture both spatial and temporal features of the signals. The system classifies sleep into stages such as Wake, N1, N2, N3, and REM accurately and robustly. Among the implemented models, the proposed hybrid CNN-LSTM model achieved the best performance with an accuracy of 94.67%. A basic analysis of sleep stage patterns is also performed to observe possible irregularities. Overall, this work presents a scalable solution for sleep stage classification using deep learning techniques.
Although software systems increasingly shape energy consumption, economic output, and societal welfare, most development methods still prioritise schedule, cost, and functionality. This paper presents an approach to software development that prioritises sustainability by embedding environmental, economic, and social objectives into the SDLC from the very beginning. The framework represents sustainability as a set of quantifiable variables that are combined into a Global Sustainability Index (GSI). These metrics include operational energy and carbon footprint, total cost of ownership (TCO), maintainability index, defect density, and a normalised Social Impact Score (SIS). An empirical measurement architecture gathers runtime and process data for continuous improvement, while phase-level “sustainability budgets” direct trade-offs across requirements, design, implementation, testing, and operation. The framework is evaluated in a repeated-measures industrial study of six production software systems (758 KLOC in total, 58 engineers, six application domains), in which every system is observed over four counterbalanced release cycles governed respectively by the proposed framework and by three established approaches: GREENSOFT, GreenSDLC, and the Sustainability Quality Model (SQM). The proposed framework attains the highest GSI (0.86 ± 0.03), a statistically significant improvement of 10–19% over the competing frameworks (paired t-tests, all Holm-adjusted p < 0.002, Cohen’s dz > 2.5). Relative to current models, energy usage and carbon emissions are cut by 10–20%, and they are decreased by approximately 30% when compared to a no-framework baseline. Normalised maintainability and defect density both improve over a five-year timeframe, and total cost of ownership drops 4–9%. Consistently higher levels of social impact and stakeholder satisfaction are observed, particularly for user groups who are marginalised. Ninety-five percent confidence intervals and effect sizes are reported for every headline comparison, and the principal limitations of the framework are stated explicitly together with mitigation strategies. These results show that all three dimensions can be improved with explicit quantitative sustainability integration without a rise in long-term costs.
Review activity is the standard public proxy of guest engagement on peer-to-peer accommodation platforms, yet is modeled almost exclusively as count regression. This study presents a three-level engagement classification of peer-to-peer home-sharing in Latin America: 76,624 listings from Buenos Aires, Santiago, and Mexico City, with tercile labels from training folds only. A pre-specified, amendment-logged evidence ladder of fifteen attempts plus three contingency campaigns measures which interventions affect macro-averaged F1. A tuned LightGBM over engineered features reaches average macro F1 0.8249, 23.0 points above the best cold-start result (0.5945); a feature-level temporal audit lowers the deployable cold-start reference to 0.5301, widening the gap to 29.5 points. Three findings emerge. First, every full-scenario result lands in a narrow band (0.81–0.83) that doubled budgets do not move: an observed performance plateau within the evaluated feature and model space. Second, Spanish text helps in the cold-start scenario (+2.4 points nominally, +5.3 over the audited strict set) but adds little once review-derived features are present, and sparse lexical vectors match dense multilingual E5 embeddings at far lower cost. Third, an untuned tabular foundation model reaches 0.8205 under a pre-specified 10,000-row context subsample. A clearly labeled post-hoc re-evaluation at full context, outside the pre-specified protocol, reaches 0.8266, the best average observed, which an exploratory multi-resampling confirmation upholds in all twelve paired comparisons. Ordinal metrics, per-class breakdowns, and a Frank–Hall baseline locate the residual error in the middle tercile; repeated outer resampling bounds selection optimism at 0.2–0.3 points.
Medical institutions increasingly hold data that is relational rather than tabular: patients linked by medical history, diagnostics by dependency, and doctors by consultations. Graph neural networks (GNNs) are a natural fit for such data, but hospitals cannot pool it directly, since records remain confined by law to the collecting hospital. Federated learning lets a central server train a shared model by aggregating parameter updates instead of raw graphs. GNN parameters, however, define a graph-dependent message-passing operator, so when hospitals differ in topology, degree distribution, and homophily, standard update-averaging no longer produces a coherent update for any hospital’s actual operator, a mismatch accuracy metrics can mask, hitting structurally atypical hospitals hardest. We introduce the Global Geometric Reference Structure (GGRS), a server-side layer that regulates updates before aggregation: directional soft-weighting down-weights updates that disagree with the hospital consensus, subspace projection keeps only shared directions, and sensitivity clipping stops any one hospital from dominating the aggregate. Across six benchmark graphs, two architectures, and four federated optimizers, GGRS improves accuracy by up to 5.1% for structurally under-represented hospitals, with the best-performing combination, SCAFFOLD paired with GGRS, also reaching target accuracy up to 12 rounds faster than unregulated aggregation. These results suggest geometry-aware aggregation helps federated GNNs serve heterogeneous hospitals more equitably, without requiring any hospital to change how it trains locally.
