Search Results (1,257)

Background: Metabolic dysfunction-associated steatotic liver disease (MASLD) is a significant and escalating global health concern, with an estimated prevalence of 30%. Current assessments of hepatic steatosis, a hallmark of MASLD, rely on semi-quantitative grading by pathologists, which is inherently limited by inter-observer variability. Objective: To address this limitation, we developed a novel deep learning pipeline, named SteatoStat, to standardize and enhance the quantification of hepatic steatosis in patients with MASLD. Method: The SteatoStat pipeline employs and integrates multiple components such as file format standardization, rule-based cell filtering, and multiple segmentation models across various liver structures, resulting in an output of a continuous quantitative measure of steatosis percentage and translated into steatosis grades. Results: We report a high degree of accuracy and reliability with SteatoStat achieving the following performance metrics (DICE score = 0.8955, AUROC = 0.9928, F1 score = 0.8990). When benchmarked against expert pathologists, the weighted Kappa coefficient is 0.837. Furthermore, in comparison with an existing, well-established model, SteatoStat demonstrated a weighted Kappa coefficient = 0.765. Conclusions: These robust findings underscore its potential clinical utility in providing a standardized objective quantification of hepatic steatosis. Future directions include enhancing the model’s generalizability and its clinical integration through validation on independent, multi-institutional datasets. Full article

To address the issues of strong inter-well interference during multi-well fracturing in shale reservoirs, low efficiency of conventional numerical simulation, and the tendency of machine learning models to overfit and lack interpretability under small-sample conditions, this paper constructs an explainable ensemble learning framework for predicting hydraulic fracture asymmetry. A geology–engineering integrated numerical simulation is adopted to quantify the fracture asymmetry index η as an interference metric, and an initial dataset is constructed comprising natural fracture orientation, well spacing, and injection rate. Subsequently, Jensen–Shannon (JS) divergence-constrained Gaussian data augmentation and second-order interaction features are introduced, and the GBRT model parameters are optimized using particle swarm optimization (PSO). Furthermore, random forest and ridge regression are incorporated, and ensemble weights are determined via cross-validation to build a weighted ensemble prediction model. The results show that the proposed model achieves good predictive performance in repeated validation, with an average coefficient of determination R² of 0.8484 and a 95% confidence interval of 0.8179–0.8790, while also demonstrating favorable overall accuracy in multiple baseline model comparisons and regularization-controlled experiments. Through leave-one-simulation-scenario validation, prediction interval analysis, and interpretability robustness testing, the model’s generalization boundary, prediction uncertainty, and explanation reliability under small-sample conditions are further evaluated. SHAP analysis and grouped permutation importance results indicate that the natural fracture angle is the dominant factor controlling asymmetric fracture response, while the interaction between well spacing and the natural fracture angle also significantly affects the predictions, suggesting that asymmetric fracture propagation is primarily governed by the combined effects of natural fracture steering and inter-well stress interference. The proposed framework can serve as a fast surrogate model for evaluating inter-well interference and screening fracturing designs within a given simulation parameter space, providing an interpretable data-driven approach for fracturing design optimization in shale reservoirs under small-sample conditions. Full article

Non-contact body-temperature prediction from facial thermography is affected by pose, occlusion, missing measurements, and inter-subject variation. This study proposes a graph–tabular latent-representation fusion framework for predicting body temperature from thermal facial landmark profiles. A Pearson correlation coefficient (PCC)-guided landmark graph models landmark-to-landmark thermal dependencies. At the same time, the same landmark-temperature signal is retained as a tabular representation to preserve global temperature-pattern interactions. The graph and tabular branches are encoded independently, fused at the latent level, and trained for target-landmark temperature regression with auxiliary reconstruction losses. Experiments were conducted on TFD68 under complete, missing completely at random (MCAR), and structured missing not at random (MNAR) conditions. The structured MNAR simulation combines 3D head-pose visibility modelling, accessory-driven occlusion, validation against real TFD68 occlusion annotations, and graph-construction sensitivity analyses. Results show that selected fused configurations improve over strong stand-alone graph and tabular baselines, particularly under MNAR-imputed evaluation, with the best selected configuration reducing prediction error by approximately $6\%$. Statistical testing further confirms significant improvements in most MNAR fused–baseline comparisons. Accuracy–efficiency analysis shows that fusion improves robustness at the cost of additional inference time, providing a flexible design space for thermal landmark-based body-temperature prediction. Full article

Background: The growing complexity and cost of oncohematological treatments has created an urgent need for standardized methodologies capable of enabling inter-institutional comparisons of healthcare expenditure within homogeneous patient groups. Cancer-related pharmaceutical costs vary substantially depending on tumour type, disease stage, and therapeutic approach, making cross-institutional benchmarking challenging due to heterogeneity in patient populations and clinical practice patterns. Therefore, integrating cost analysis with clinically meaningful patient stratification is essential to improve resource allocation and outcome evaluation. Methods: A multicentre working group comprising four tertiary hospitals in Madrid (Spain) was established to develop and preliminarily evaluate a novel classification system for adult oncohematological patients. A standardized methodology was designed to stratify patients into homogeneous groups (PATONCO categories) based on tumor location, therapeutic objective, and clinically relevant biomarkers. A cost indicator was defined as the average cost per patient per month for each PATONCO category. Data were extracted from pharmacy dispensing systems and analyzed using descriptive and inferential statistics, including Kruskal–Wallis and post hoc Dunn tests. Results: A total of 3659 patients were included (3168 oncology; 491 hematology), distributed across 62 programmes (54 oncology; 8 hematology). The PATONCOS tool enabled the identification and validation of a cost indicator (average cost/patient/month per category), allowing inter-hospital comparison. Significant differences in costs were observed across most high-prevalence categories, reflecting variability in drug selection within homogeneous patient groups, as documented by the differential use of specific therapeutic agents across centers. The model demonstrated its capacity to detect intra-group homogeneity and inter-group variability, improving the identification of high-cost patient subgroups and supporting benchmarking across centers. Conclusions: The PATONCOS tool provides a novel, clinically oriented stratification methodology that integrates pharmacotherapy, biomarkers, and disease stage with economic evaluation. This approach enables more accurate comparisons of oncology treatment costs between institutions and may support data-driven decision-making in resource allocation. Its implementation may contribute to more sustainable healthcare systems by aligning clinical practice with economic outcomes. Full article

Climate change threatens nitrogen cycling in agricultural ecosystems. Optimizing sowing dates and nitrogen management for maize–soybean intercropping is critical for sustainable production in the North China Plain (NCP). Using a calibrated Agricultural Production Systems Simulator (APSIM) model driven by three representative global climate models (GCMs) selected from 20 Coupled Model Intercomparison Project Phase 6 (CMIP6) GCMs, we evaluated management strategies under two Shared Socioeconomic Pathway scenarios (SSP2-4.5 and SSP5-8.5) across three climatic zones for near-term (2030–2059) and long-term (2070–2099) periods. Under SSP5-8.5, warming was 1.8–2.2 times greater than under SSP2-4.5, nitrate nitrogen (NO₃⁻-N) leaching increased by 12.1%, and nitrate storage in the 100–150 cm soil layer rose by 53.4% in Zone III. Biological nitrogen fixation contributed 20.1–29.1% of soybean nitrogen uptake under low nitrogen and 14.9–23.4% under medium nitrogen. Optimal strategies were identified: sowing on 7 June (S3) with medium nitrogen (220.8 kg N ha⁻¹) under SSP2-4.5, and advancing sowing to 28 May (S2) with medium nitrogen under SSP5-8.5 to alleviate heat stress. This study reveals a climate-driven “earlier supply–shortened demand–concentrated leaching” mismatch, providing adaptive management guidance for maize–soybean intercropping systems in the NCP. Full article

Gully erosion is a major driver of irreversible soil loss in Northeast China’s Mollisol belt, a region that supplies roughly one-quarter of the national grain output. Existing susceptibility assessments in this region have rarely combined multi-model comparison with spatially explicit cross-validation, and the predictive contribution of composite anthropogenic indicators such as the Human Footprint Index (HFI) has not been quantitatively benchmarked against conventional topographic variables. This study addresses these gaps for the Heihe region by combining an inventory of 4020 gully polygons supported by field checks in Xunke County, 16 VIF-screened environmental factors, three tree-based ensemble models and a logistic regression baseline. Under stratified random splitting, XGBoost achieved the highest discrimination (AUC = 0.95, $\kappa$ = 0.74); under leave-one-district-out spatial cross-validation all tree-based models retained AUC above 0.83, confirming that random-split metrics overestimate discrimination by approximately 0.11 AUC units due to spatial autocorrelation and inter-district covariate shift. SHAP analysis identified LULC and HFI as the dominant predictors, exceeding all topographic variables, while slope gradient contributed least—consistent with the low-relief, intensively cultivated character of the study area. Susceptibility was highest in the southwestern agricultural lowlands. A one-factor sensitivity test in which only NDVI was increased by 20% suggested a reduction in modelled high-susceptibility area of approximately 12%, although co-occurring land-cover and hydrological changes were not simulated. The multi-model framework, integrating spatial cross-validation and post hoc interpretability, provides an explicit estimate of conventional evaluation optimism and supports spatially differentiated erosion management. Full article

Purpose: To develop a fully automated 2D nnU-Net pipeline for multi-class skeletal muscle segmentation (psoas, paraspinal, and abdominal wall) at the third lumbar (L3) vertebral level, and to quantitatively evaluate its diagnostic performance and reliability compared to manual segmentation. Materials and Methods: A 2D nnU-Net was trained on 164 axial L3 CT slices from the multi-institutional AMOS22 dataset, spanning diverse abdominal pathologies and multivendor imaging. To assess generalizability under severe anatomical distortion, independent external validation was performed in 50 consecutive patients with advanced liver disease from a single institution (January–December 2025; mean age, 63 ± 15 years; 32 women, 18 men), of whom 88% had moderate-to-severe ascites. Model stability was examined by comparing a five-fold ensemble with the best-performing single-fold model. Intra-observer reliability of the manual reference standard was evaluated in a random subset of 30 cases. Inter-observer agreement was additionally assessed using an independent second reader. Performance metrics included the Dice Similarity Coefficient (DSC), Pearson correlation coefficient (r), and Bland–Altman analysis for cross-sectional areas and mean attenuation. The inference workflow was deployed via a custom Streamlit-based graphical user interface (GUI). Results: In this anatomically complex external validation cohort, the 5-fold ensemble 2D nnU-Net achieved an overall mean DSC of 0.937 ± 0.043 (95% CI, 0.925–0.950), with 80% of cases achieving a mean DSC ≥ 0.90. While the mean DSC was statistically comparable to the best single-fold model (0.937, [95% CI, 0.921–0.952], p = 0.736), the ensemble strategy increased the minimum observed DSC (worst-case performance) from 0.720 to 0.822. Class-specific external validation performance for the 5-fold ensemble was highest for the paraspinal muscles (DSC: 0.960; 95% CI, 0.952–0.967), followed by the psoas muscles (DSC: 0.941; 95% CI, 0.927–0.956), and lowest for the anatomically complex abdominal wall muscles (DSC: 0.911; 95% CI, 0.893–0.929). Comparison between the ensemble model and manual segmentation yielded a Pearson correlation of r = 0.955 (p < 0.001) for total skeletal muscle area, with a mean bias of +7.17 cm². Intra- and inter-observer agreements for the manual reference standard demonstrated correlation coefficients of r = 0.995 and 0.090 for total areas, respectively. The automated pipeline required 3–5 s per case for inference and quantitative reporting, compared to 3–5 min for manual segmentation. Conclusions: In patients with advanced liver disease and substantial anatomical distortion from ascites, an ensemble-based 2D nnU-Net provides high quantitative agreement with manual L3 skeletal muscle segmentation, while mitigating lower-bound (worst-case) errors relative to single-fold models. Integration with a dedicated GUI enables substantial time savings and supports scalable quantitative body composition measurement. Full article

Driven by China’s “dual carbon” (carbon peak and carbon neutrality) goals and the national strategy of new-type urbanization, prefabricated construction has emerged as a pivotal pathway toward industrialized and sustainable development in the construction sector—leveraging its distinctive advantages in construction efficiency, cost optimization, environmental performance, and design adaptability. Nevertheless, the inherently sequential and interdependent nature of the full construction process—encompassing off-site component manufacturing, logistics transportation, and on-site assembly—introduces pronounced cross-stage risk transmission mechanisms, with prefabricated components serving as critical risk carriers. Such transmission dynamics significantly impede the scalable and safe deployment of prefabricated construction. To date, scholarly efforts on construction safety in prefabricated buildings have predominantly addressed isolated, stage-specific risks, falling short in quantitatively modeling the coupled propagation of risks across stages, accommodating epistemic uncertainties and latent (i.e., unknown or unobserved) risks, and informing targeted, evidence-based mitigation strategies. To bridge this gap, this study develops a rigorous quantitative framework for assessing cross-stage risk transmission in prefabricated construction safety. Specifically, it aims to (i) uncover the structural patterns and driving mechanisms underlying inter-stage risk propagation; (ii) reduce the likelihood of safety incidents throughout the construction life cycle; and (iii) deliver actionable theoretical insights and methodological guidance for practitioners and policymakers. Methodologically, we first conduct a systematic identification of safety-critical risk factors and establish a hierarchical risk indicator system comprising three first-level dimensions and twenty second-level indicators. Second, using the Decision-Making Trial and Evaluation Laboratory (DEMATEL) method, causal relationships among risk factors are clarified, while incorporating the Leaky Noisy-or Gate (LNOG) extended model to account for unknown risks. Risk data are processed using triangular fuzzy functions, and a Bayesian network (BN) topology diagram is constructed via the GeNIe 5.0 platform, forming a DEMATEL-LNOG-BN-based model for assessing cross-phase risk transmission. Finally, applying the model to an actual project—”a prefabricated construction project in Shanghai”—the study conducts a cross-phase risk transmission analysis. Through forward probability inference, backward causality tracing, sensitivity analysis, and pathway decomposition, sensitivity comparisons are performed under different LNOG unknown risk parameters. Results are compared with those from the traditional DEMATEL-BN model to validate the stability and consistency of high-sensitivity risk factor identification, comprehensively verifying the applicability and predictive reliability of the proposed DEMATEL-LNOG-BN model. The study quantitatively reveals the progressive diffusion and amplification mechanisms of risks across the production–transportation–assembly process, providing scientific support and practical reference for precise safety risk prevention, critical node control, and the optimization of management systems in prefabricated construction sites. Full article

The permafrost in the Tibetan Plateau is extremely sensitive to climate warming, which poses challenges to regional sustainability. Predicting the evolution of permafrost on the Tibetan Plateau in the future could provide a reference for future engineering, construction, and resource management on the Tibetan Plateau. In this study, the Random Forest regression model and the temperature at the top of permafrost (TTOP) model are combined. The Random Forest regression model is used to simulate the long-term series of land surface temperatures. The multiple climate model data sets in the Coupled Model Intercomparison Project Phase 6 (CMIP6) and TTOP model are used to simulate the historical (1979–2018) and predict the future (2019–2100) distribution of permafrost on the Tibetan Plateau. The results show that since 1979, due to climate warming, more than 20% of the permafrost in the Tibetan Plateau has disappeared. The permafrost will degrade at different rates under each of four Shared Socioeconomic Pathways (SSPs), namely SSP1–2.6, SSP2–4.5, SSP3–7.0, and SSP5–8.5. The degradation rate under SSP1–2.6 is the slowest, indicating that about 20.1% of the permafrost will disappear by 2100. The degradation rate under the SSP5–8.5 is the fastest, predicting that about 82.4% of the permafrost will disappear by 2100. Under SSP2–4.5 and SSP3–7.0, 37.57% and 69.1% of the permafrost will disappear by 2100, respectively. The above results can provide a reference for sustainable engineering construction, infrastructure planning, and climate adaptation strategies on the Tibetan Plateau. Full article

Aims: This retrospective clinical MRI study aimed to evaluate regional differences in normalized signal intensity (SI) of the native anterior cruciate ligament (ACL) and to assess the influence of demographic and clinical factors on SI distribution. Methods: MRI scans obtained on a 3T scanner from 84 patients were analyzed. Regions of interest were defined in the proximal, middle, and distal portions of the native ACL, with normalization to the posterior cruciate ligament (PCL) tibial insertion. SI was expressed as the anterior–posterior native cruciate ligament ratio (APRn). A linear mixed-effects model with a random intercept for the patient was used to account for repeated measurements, and pairwise regional comparisons were performed with Tukey adjustment. Statistical significance was set at p < 0.05. Results: A significant regional effect was observed (p = 0.0001), with the distal portion demonstrating the highest SI, followed by the proximal and middle regions. Compared with the distal region, both the middle (β = −0.557, p = 0.0001) and proximal (β = −0.568, p = 0.0002) portions showed significantly lower SI values. The ACL signal was approximately 1.65–2.45 times higher than that of the PCL. Post hoc Tukey comparisons confirmed significant differences between all anatomical regions. The log-transformed model demonstrated improved fit (marginal R² = 0.24; conditional R² = 0.26). Inter- and intra-observer reproducibility demonstrated excellent agreement (ICC = 0.88–0.91). Conclusions: Native ACL signal intensity demonstrates a location-specific pattern, with the distal region showing the highest values. These findings provide reference data for regional native ACL signal distribution in a clinical cohort. Full article

Accurately assessing and predicting vegetation dynamics is of great significance for evaluating regional hydrological and ecological environments. This study focuses on the climate-sensitive Qinghai-Tibetan Plateau (QTP), aiming to reveal the spatiotemporal patterns, underlying driving mechanisms, and future trends of vegetation dynamics. The historical turning points of greening trends were identified using the running slope difference method, and the SHapley Additive exPlanations (SHAP) method was employed to analyze the key driving factors. An Xtreme Gradient Boosting (XGBoost) prediction model was constructed and validated, and then coupled with Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model ensemble data to project seasonal vegetation changes under different Shared Socioeconomic Pathways (SSP). The main conclusions are as follows: (1) Vegetation on the QTP showed an overall greening trend with significant spatial heterogeneity. Approximately 47.25% of the area exhibited no trend shift (NS), while 29.42% experienced a shift from greening to browning (GB), with most shifts occurring between 1990 and 2010. (2) Soil moisture and precipitation were the dominant driving factors, with contributions significantly higher than those of temperature, wind speed, and other variables, and they exhibited nonlinear interactive effects with the Normalized Difference Vegetation Index (NDVI). (3) In the future, vegetation is projected to show an overall increasing trend, with stronger responses in spring and autumn. The regional average rate of change is highest in spring, especially under the SSP5-8.5 scenario (17.8% for 2030–2060 and 26.4% for 2061–2100); in autumn, although the regional average rate of change is small, the internal spatial variability is significant. The humid regions in the eastern and southeastern parts of the QTP demonstrated more active greening across all seasons except winter, and high-emission scenarios are expected to exacerbate regional and seasonal differences. This study systematically reveals the adaptive dynamics and future scenarios of vegetation dynamics on the QTP, providing scientific support for the adaptation of alpine ecosystems to global change and the management of regional ecological security barriers. Full article

The accurate estimation of State of Charge (SOC) and Remaining Useful Life (RUL) is a key challenge in lithium-ion battery management systems, due to the nonlinear, time-varying, and multi-physics nature of battery dynamics. This work presents a systematic comparative study for SOC and RUL estimation based on the analysis of the NASA battery dataset, characterized by significant heterogeneity in operating conditions, temperature regimes, and cycle durations. The study combines a physically informed feature engineering process with machine learning models, including tree-based ensembles, kernel methods, and neural networks. The dataset is analyzed from an electrochemical, thermal, and impedance perspective, highlighting the role of internal resistance evolution, SOC–voltage characteristics, and temperature dynamics as indicators of battery degradation. Based on these observations, two regression problems are formulated: a local window-based representation for SOC estimation and a cycle-level representation for RUL prediction. Particular attention is devoted to the impact of dataset heterogeneity, feature construction, and target representation on the predictive behavior of the considered models. In addition, the work investigates the effect of normalized RUL representations and provides an interpretability-oriented comparison of the learned regressors through feature-importance analysis and parity plots. Experimental results show that SOC estimation is a comparatively well-conditioned problem, achieving high accuracy across nonlinear models, although the dominant role of temporal and current-derived features highlights the strong dependence of the prediction task on the structure of the experimental protocol. In contrast, RUL prediction exhibits significantly higher complexity due to long-term degradation uncertainty and inter-battery variability. The introduction of a normalized RUL representation substantially improves prediction accuracy and stability, particularly for ensemble-based approaches. Feature importance analysis confirms that capacity-related variables dominate RUL estimation, while voltage, temporal, and current-derived features play a central role in SOC prediction. Overall, the results show that physically interpretable feature construction combined with ensemble learning methods provides an effective framework for battery state estimation and degradation analysis under heterogeneous operating conditions. Full article

Accurate prediction of pore pressure is crucial for ensuring drilling safety and improving the efficiency of oil and gas development. However, regarding feature selection, most existing studies focus primarily on the correlations among features, with little consideration given to inter-well distribution differences; this may result in insufficient model generalization capabilities for cross-well prediction tasks. To improve cross-well prediction accuracy, this paper introduces the DASP (Domain Adaptation with SHAP-guided Particle Swarm Optimization) method for feature selection. Using 20 logging and rock mechanical parameters as input features, a four-stage selection process reduces the number of features to 12, achieving approximately 40% feature dimensionality reduction. In terms of model performance comparison, models such as RF (Random Forest), XGB (eXtreme Gradient Boosting), LGB (Light Gradient Boosting Machine), and CNN (Convolutional Neural Network) were constructed for comparative analysis. The results indicate that LGB performs better during the validation phase, while in the cross-well testing phase, RF achieves the best prediction accuracy among the base models, demonstrating strong generalization stability. Addressing the issue that CNNs still have room for further optimization in cross-well prediction, this study further refines their architecture by introducing an attention mechanism to enhance the model’s adaptive weighting capability for key features and combining it with a physical constraint mechanism to suppress non-physical prediction fluctuations, thereby improving the model’s stability and geological plausibility. Experimental results demonstrate that the improved model shows significant improvements in both error metrics and fitting capabilities. The combination of the DASP-based feature selection method and the improved CNN model effectively enhances the accuracy of cross-well pore pressure prediction, providing new technical insights and data support for intelligent pore pressure prediction and drilling safety decision-making. Full article

Accurate mapping of winter rapeseed cultivation areas is crucial for food security assessment and agricultural resource management, yet remains a persistent challenge in mountainous regions characterized by complex topography and highly fragmented field parcels. To address these challenges, this study develops a progressive deep learning framework using single growing-season data from the Hanzhong Basin. We conducted a structured comparison of remote sensing indices, machine learning, and deep learning approaches for rapeseed identification in heterogeneous landscapes. First, sensitivity analysis of the Flowering Index for Rapeseed was performed to identify the optimal parameterization, yielding high inter-class separability (ND = 0.959) during peak flowering and a threshold-based overall accuracy (OA) of 94.41%. Second, a multidimensional feature space was constructed by integrating Sentinel-2 spectral bands, image texture metrics, and topographic variables; Random Forest-based feature importance selection subsequently enhanced Support Vector Machine classification performance to an OA of 90.70%. Third, we proposed an innovative three-stage progressive UNet++ architecture: Stage1 focuses on binary rapeseed/non-rapeseed classification to establish spatial priors; Stage2 refines discrimination among spectrally similar vegetation classes (rapeseed and other vegetation); and Stage3 achieves comprehensive seven-class semantic segmentation. A weighted focal loss function combined with a weight inheritance mechanism was employed to mitigate class imbalance and facilitate inter-stage knowledge transfer. The final model attained an OA of 98.65% and a mean intersection over union of 95.29%, while effectively suppressing salt-and-pepper noise artifacts in geometrically fragmented parcels. Our findings demonstrate the substantial advantages of progressive deep learning strategies for crop monitoring in topographically constrained environments. Full article

Background: Accurate identification and positioning of thoracic devices on chest radiographs is critical for patient safety in intensive care. Multimodal large language models (LLMs) offer potentially generalizable automated evaluation, but their performance in this domain is underexplored. Methods: Three multimodal LLMs (GPT-4o, gpt-4o-2024-08-06; Gemini 3.1 Flash Lite Preview; Claude Sonnet 4.6) were evaluated on 4813 chest radiographs from the RANZCR CLiP dataset for device presence and positioning of ETT, NGT, CVC, and Swan–Ganz catheters. Performance was quantified with 95% Wilson confidence intervals, balanced accuracy, MCC, Cochran’s Q, Bonferroni-corrected McNemar, and Cohen’s/Fleiss’ kappa. Six additional analyses were performed: a blinded paired reader study (n = 377; two board-certified radiologists, blinded to ground truth and to all LLM outputs), external validation on PadChest (n = 200, device-presence detection only—PadChest lacks granular position labels), three-variant prompt-sensitivity analysis (n = 103), repeat-inference stability across three runs (n = 50), systematic error taxonomy, and a failure-case analysis. Results: Device-presence performance varied widely across models; abnormal-position sensitivity was uniformly poor (MCC ≤ 0.028; balanced accuracy 0.41–0.53). Inter-model agreement was poor to slight (Fleiss’ κ: 0.005–0.383 for presence; −0.280 to −0.025 for classification). Radiologists numerically outperformed all three LLMs in 42/42 paired comparisons; the superiority was statistically significant after Bonferroni correction in 33/42 (32/42 at p < 0.001). PadChest replicated the negative finding for device-presence detection (malposition not externally validated). Prompts and inference stochasticity introduced 2–3× sensitivity swings and run-to-run κ from 0.20 to 0.85. Case failures concentrated systematically in multi-device cases (p < 0.0001) but not in abnormal-position cases (p = 0.14). Conclusions: Current general-purpose multimodal LLMs are not yet reliable for autonomous thoracic-device assessment; their failure patterns are structurally characterizable across models, prompts, and case types and support, at most a circumscribed role, as adjunct device-presence screening tools. The findings do not generalize to purpose-built, regulator-approved clinical AI systems. Full article