Contactless respiratory monitoring using radio frequency (RF) sensing and Wi-Fi channel state information (CSI) has emerged as a promising approach for unobtrusive physiological assessment. However, reported classification performance may be substantially inflated when augmented datasets contain duplicated or source-related samples that are inadvertently shared between training and test sets. This study presents a leakage-aware benchmark audit using a publicly available IEEE DataPort dataset comprising ESP32 Wi-Fi CSI respiration recordings and exhaled-breath RF response curves labeled at respiratory rates of 12, 20, and 28 breaths per minute. The exhaled-breath RF subset was evaluated exclusively for breaths-per-minute class discrimination because no independently verified clinical hydration reference was provided. As parent-recording identifiers, session metadata, and acquisition provenance were unavailable in the public release, the strictest feasible proxy protocol was established by combining SHA-256 duplicate detection with unique source-file identification, similarity-family grouping, near-duplicate analysis, grouped cross-validation, feature-group explainability analysis, perturbation-based robustness assessment, and fully reproducible reporting. Among the respiration recordings, 45 released files were reduced to 23 unique source hashes, whereas the exhaled-breath RF subset contained 40 unique hashes. Respiration classification remained highly discriminative after deduplication but exhibited substantially greater uncertainty under source-level validation. Logistic regression achieved a balanced accuracy of 1.000, whereas random forest achieved 0.700 with limited recall for the 28 breaths-per-minute class. For the exhaled-breath RF subset, complete separability of breaths-per-minute classes was preserved under source-level evaluation. Explainability analysis further indicated that respiration classification was predominantly driven by distributed CSI subcarrier variability rather than by a single dominant spectral component. Although breaths-per-minute discrimination remained robust following per-file normalization, balanced accuracy decreased to 0.906–0.925 after retraining with the upper-frequency quartile removed, while simulated upper-quartile dropout or spectral smoothing reduced classification performance to approximately chance level. Overall, the released dataset contains strong discriminative structure for breathing-rate classification but also substantial risks of data leakage and feature dependence. The primary contribution is a reproducible benchmark validity audit rather than evidence for clinical hydration assessment, respiratory disease diagnosis, or deployment readiness.
This work developed a machine learning (ML) model for predicting the minimum film thickness ($h_0$) in plain journal bearings. A reference dataset comprising 376,349 datapoints was generated from the approximated pressure journal design method developed by Reason and Narang. The inputs in the dataset included geometric and operational features: Society of Automative Engineers (SAE) oil grades, speed (500–10,000 rpm), bearing load (1,500–9,500 N), bearing length/diameter (0.01–0.08 m), clearance (0.00001–0.00006 m), and initial oil temperature (40–70 ℃). Four supervised learning models, i.e., Artificial Neural Network ($h_0$_ANN), Gaussian Process Regression ($h_0$_GPR), Ordinary Least Squares Regression ($h_0$_OLS), and Ridge Regression ($h_0$_Ridge) were developed and evaluated. The results indicated that Artificial Neural Network (ANN) provided the most accurate predictions via achieving an $R^2$ value of 0.99. The GPR model obtained an $R^2$ value of 0.98 while those of the OLS and Ridge Regression achieved $R^2$ of 0.95 and 0.93, respectively. Having shown superior predictive capabilities, the ANN model was selected for further statistical evaluation. Analysis from Local Interpretable Model-agnostic Explanations (LIME), Shapley Additive exPlanations (SHAP), and radar plots all proposed that the ANN model offered stable predictions within the considered parametric space. The model was presented in an explicit mathematical form in an Excel spreadsheet; its computational efficiency was recommended for the design of plain journal bearings.
Graph-based representations provide a useful systems-level framework for modelling interactions among structure, dynamics, and behaviour. This paper proposes a dual-graph framework for modelling indoor movements and activities. The first layer is a location graph that represents feasible movement through the spatial connectivity of an indoor environment. The second layer is a mixed causal/contextual activity graph that combines directed activity dependencies with undirected contextual associations. The two layers are coupled through an activity-to-location mapping, yielding a probability-preserving dynamical model in which spatial occupancy is jointly influenced by graph-constrained movement and activity-driven spatial expectations. Two features distinguish the proposed framework from conventional dual-graph models. First, the activity layer is explicitly constructed as a mixed directed/undirected network and second, a cross layer coupled mismatch residual framework is proposed to detect inconsistencies between semantic activity evolution and observed movement. The paper also establishes the probabilistic properties of the movement operator, discusses manual and data-driven construction of the interlayer mapping and introduces an optional reverse-coupling extension. Simulations in a six-location living environment examine the effects of the activity-mixture parameter, the mapping matrix, and the coupling gain. The results support the framework as an interpretable basis for indoor behaviour modelling and also highlight some of its limitations for future studies.
Obesity represents a substantial public health burden in Saudi Arabia, yet the predictive contribution of anthropometric and lifestyle characteristics beyond body mass index remains insufficiently characterized. An interpretable machine-learning framework was developed to classify obesity among Saudi adults using anthropometric, demographic, health, and lifestyle variables and to assess whether predictive performance was retained after excluding variables directly related to the body mass index-defined outcome. Of 294 survey responses, 279 were retained after consent and data-completeness criteria were applied. Numerical variables were median-imputed, categorical variables were mode-imputed and one-hot encoded, and obesity was defined as a body mass index $\geq$30 kg/m$^2$. Random forest performance was evaluated using stratified five-fold cross-validation with fixed tuned hyperparameters. Sensitivity analyses excluded body mass index alone and body mass index, height, and weight simultaneously. Model interpretability was assessed using Shapley additive explanations. With body mass index included, all evaluated performance metrics reached 1.000 ± 0.000, reflecting target leakage because body mass index directly defined the outcome. After body mass index exclusion, accuracy was 0.911 ± 0.038, recall 0.624 ± 0.168, F1-score 0.701 ± 0.125, and area under the receiver operating characteristic curve 0.972 ± 0.023. After simultaneous exclusion of body mass index, height, and weight, accuracy decreased to 0.842 ± 0.023 and area under the receiver operating characteristic curve to 0.815 ± 0.056. Waist circumference, hip circumference, waist-to-hip ratio, age, and selected health and lifestyle characteristics retained predictive information, although sensitivity to obesity decreased substantially after removal of the defining anthropometric variables. Shapley additive explanation analyses clarified feature contributions to individual predictions. These findings demonstrate that the exceptional performance of the complete model was predominantly attributable to target leakage. Complementary characteristics retained meaningful discriminatory information, but reduced sensitivity warrants cautious interpretation. External validation in larger, representative cohorts with independently measured anthropometric data is required before clinical or population-level screening applications are considered.
Agentic data pipelines, in which large language models select and invoke tools through the Model Context Protocol, consume tool outputs, and iteratively execute multi-step analytical or operational workflows, are increasingly being deployed in production environments. However, the observability infrastructure required to diagnose failures in such systems remains underdeveloped. Conventional distributed tracing effectively captures service-to-service execution but often represents large language model invocations as opaque spans and fails to preserve causal relationships across large language model-tool boundaries. Consequently, incident diagnosis can require an agent's execution trajectory to be reconstructed manually from chronologically ordered spans. To address this limitation, causal span linking across large language model and tool invocations was defined as a first-class observability primitive for agentic data pipelines. Hops-to-root-cause was introduced as the directed acyclic graph distance between a symptom span—the earliest span tagged error=true—and the identified causal span, with deterministic tie-breaking applied. The proposed approach was evaluated on a synthetic corpus comprising 20 incidents generated using a fully disclosed construction protocol. Compared with a flat-span baseline, causally linked tracing reduced the mean hops-to-root-cause from 6.4 to 1.5, corresponding to a reduction by a factor of 4.3. The greatest improvements were observed for incidents involving multi-hop tool chains. Causal span linking was further complemented by deterministic replay and by a human-artificial intelligence collaborative diagnostic workflow. Together, causal span linking, deterministic replay, and human-artificial intelligence collaborative diagnosis were established as complementary observability primitives for improving the reproducibility, interpretability, and efficiency of root-cause analysis in agentic data pipelines.
Open Access
Journal Metrics
Journal Overview
Associated Journals