Search Results (8,594)

Distribution drift can substantially degrade the performance of deployed machine learning models; for example, accuracy on SST-2 can fall from $88\%$ to $58\%$. Detecting such degradation is fundamentally challenging because deployment provides inputs but not labels, so the detection itself must be reference-free. We introduce DecayBench, the first reference-free, calibrated benchmark for evaluating drift detectors. DecayBench measures detector trustworthiness along five axes (calibrated, valid, timely, no-regret, adaptive), and compares ten existing detectors across ten NLP, vision, and multimodal datasets using paired-bootstrap significance testing. Evaluation on DecayBench shows that no existing detector is uniformly optimal. Motivated by this observation, we propose Alert, a label-free aggregation rule for drift detection. Unlike all competing combiners, it uses a label-free self-configuring selection rule with a no-regret guarantee. Alert has three contributions: (i) a dilution analysis yielding a self-configuring detector selection rule; (ii) a finite-sample conformal guarantee that controls the false-alarm probability on clean data at any prescribed level (e.g., $5\%$) for arbitrary score distributions; and (iii) a no-regret result: when no single detector dominates (constituents of comparable effect size, a condition checkable offline), Alert matches or beats the best constituent, being never significantly worse and sometimes better by a large margin; this holds across NLP, NLI, and vision (ResNet), with the largest gains under multimodal drift, and the proof identifies a dominant single detector (MMD on CLIP) as the only dilution exception. We prove the no-regret property and, across the benchmark, report its empirical counterpart, non-dominance under a paired bootstrap (Alert is never significantly worse than the best constituent), which at some operating points is statistically inconclusive rather than a strict win. Because Alert combines only embedding- and logit-based detector scores, it directly transfers across NLP, vision, and multimodal models. Empirically, Alert strictly improves over single-modality monitoring, increasing AUC by up to 25 points under mixed-modality drift and by approximately 50 points under cross-modal mismatch, where individual modality-specific detectors perform near chance. Alert also matches or outperforms the Fisher, Simes, Bonferroni, and median combiners, performs best under low-severity drift, and matches or surpasses early fusion (Concat-MMD) in both multimodal settings. Full article

Background: Prognostic assessment after neoadjuvant chemotherapy and curative gastrectomy remains challenging in patients with gastric adenocarcinoma because postoperative outcomes are influenced by both pretreatment disease burden and pathological response. We aimed to develop and internally validate a clinicopathological prognostic model and a simple postoperative risk classification for predicting disease-free survival (DFS). Methods: This single-center retrospective cohort study included patients with gastric adenocarcinoma who underwent neoadjuvant chemotherapy followed by curative gastrectomy. Pretreatment clinicopathological variables, Becker tumor regression grade (TRG), and serum tumor markers were evaluated. Logistic regression was used to identify predictors of favorable pathological response, whereas Cox proportional hazards regression was performed to identify independent prognostic factors for disease-free survival (DFS). Sequential prognostic models were developed and internally validated using 1000 bootstrap resamples. A simplified postoperative clinicopathological risk classification based on pretreatment clinical N stage and Becker tumor regression grade was additionally developed to facilitate clinical interpretation and postoperative risk stratification. Results: A total of 109 patients were included. Favorable pathological response (Becker TRG1–2) was achieved in 68 patients (62.4%), whereas 41 patients (37.6%) had minimal or no pathological response (TRG3). In multivariable logistic regression analysis, pretreatment clinical T stage (cT4 vs. cT1–3) and clinical N stage (cN2–3 vs. cN0–1) were independently associated with a lower likelihood of achieving a favorable pathological response. For disease-free survival, pretreatment clinical N stage, Becker tumor regression grade, and log10-transformed CA19-9 remained independent prognostic factors in the multivariable Cox model. Sequential model development demonstrated progressive improvement in model discrimination, with the optimism-corrected Harrell’s C-index increasing from 0.697 for the clinical N stage model to 0.770 for the final model incorporating clinical N stage, Becker tumor regression grade, and CA19-9. Bootstrap internal validation demonstrated minimal optimism, and calibration analysis showed good agreement between predicted and observed disease-free survival. A simple postoperative clinicopathological risk classification successfully stratified patients into distinct prognostic groups. Conclusions: A clinicopathological prognostic model integrating pretreatment clinical N stage, Becker tumor regression grade, and serum CA19-9 demonstrated improved prognostic discrimination for disease-free survival compared with clinical N stage alone. The derived postoperative risk classification may provide a simple framework for postoperative risk stratification and could assist in individualizing postoperative surveillance. External validation is warranted before routine clinical implementation. Full article

This paper presents an analytical compact DC current model and a numerical gate capacitance model for p-type cylindrical gate-all-around (GAA) nanowire metal–oxide–semiconductor field-effect transistors (MOSFETs). The models are formulated within the Landauer transport framework, incorporating source-to-drain tunneling (SDT) and quantum statistical charge analysis. The proposed current model is validated against non-equilibrium Green’s function (NEGF) simulations for different channel lengths, nanowire radii, and bias conditions, showing good agreement with the NEGF results in the ballistic limit. The model parameters are separated into physical parameters obtained or calibrated from the NEGF simulations and a single set of global empirical fitting parameters. The latter is extracted once and remains unchanged across the investigated device geometries and bias conditions, allowing its transferability to be evaluated. The compact model is implemented in Verilog-A, and its SPICE compatibility is verified through DC simulations of PMOS inverter circuits. All NEGF comparisons in this work are performed with a zero channel backscattering coefficient corresponding to the ballistic transport limit; validation of the quasi-ballistic regime is left for future work. Full article

Magnitude-threshold classifiers trained on multiple station records from the same earthquake are vulnerable to event-level data leakage, and comparisons across magnitude boundaries can be misleading because the class definition and prevalence change. We evaluated ten machine-learning classifiers using the official SMD-TR Metadata.csv and IM_RotD50.csv files. Record metadata and orientation-independent horizontal RotD50 intensity measures were joined one-to-one by a waveform identifier (WFID), and earthquakes were grouped by the official earthquake identifier (EQID). The Mw-only cohort contained 35,198 records from 3377 events and 965 stations. Events, rather than records, were assigned to an 80% development subset and a locked 20% independent test subset. Hyperparameters, classifier, and decision-score cutoff were selected within five event-wise development folds at the prespecified Mw = 5.5 boundary using event-level F2. Linear Discriminant Analysis (LDA; lsqr solver with automatic shrinkage) was selected in development (mean F2 = 0.903, SD = 0.043; score cutoff = 0.703). On 676 independent test events, including 23 positives, LDA achieved F2 = 0.779 (95% event-bootstrap interval 0.625–0.896), recall = 0.826, precision = 0.633, balanced accuracy = 0.905, MCC = 0.712, and PR-AUC = 0.865 (TN = 642, FP = 11, FN = 4, TP = 19). An Extra Trees regression sensitivity analysis yielded event-level MAE = 0.171, RMSE = 0.229, and R² = 0.860. Boundaries from Mw 5.0 to 6.0 are reported as separate descriptive tasks, not as evidence for an optimal physical threshold. Because the predictors are full-record RotD50 peak parameters, the findings support offline calibration and audit rather than operational real-time early warning. Full article

This study presents a sensor-fused field investigation and simulation-based analysis of four horizontal and vertical traffic-calming devices—two raised speed tables, a speed hump, and a raised crosswalk—installed along a 5250-ft two-lane residential collector in Nashville, TN, USA. A dual-sensor architecture combining a Miovision Scout video-based vehicle counter and WAAS/EGNOS-augmented GPS probe-vehicle logging (5 m 3-D RMS horizontal accuracy, 1 Hz sampling) was used to reconstruct 30 quality-controlled free-flow vehicle trajectories and 12-h per-lane volume counts. A spatial kinematic transform (a = v·dv/dx) was applied to extract device-specific approach-deceleration and post-device recovery-acceleration rates, and a three-parameter log-logistic cumulative-distribution function was fitted to the field-observed desired-speed percentiles (root-mean-square error below 0.043 for both speed-table devices). The camera- and GPS-derived observations were used to calibrate and statistically validate a PTV VISSIM microsimulation replica of the corridor, achieving a mean-speed calibration error of 0.71% or better at every device, a GEH statistic below 1.5 at all four analysis turning movements, and independent travel-time validation errors of 5.7–12.1%, within the accepted 15% threshold. The validated model was then used to reconstruct device- and spacing-specific May–Keller macroscopic speed–density–flow relationships, calibrated against simulated capacities of 650–775 vehicles per hour per lane at 350-, 700-, and 1050-ft device spacing. Results show capacity reductions of 20–33% relative to free-flow conditions and yield kinematically derived maximum recommended spacings of 265–630 ft to maintain crossing speeds at or below 15 mph, depending on device geometry. The findings demonstrate a reproducible, low-cost sensor-fusion workflow for quantifying the safety–capacity trade-off of traffic-calming corridors and for informing the design of sensor-in-the-loop adaptive-calming infrastructure. Full article

This study investigates the influence of rainfall variability on runoff generation in the Upper Ing River Basin and inflow to the Mae Tum Reservoir in northern Thailand using the physically based Soil and Water Assessment Tool (SWAT) version 2012. In small mountainous watersheds, water supply reliability for irrigation and domestic use—particularly for unmonitored royal initiated projects like the Mae Tum Reservoir—has become a critical concern due to shifting climatic extremes. A SWAT model was developed using detailed spatial data on topography, land use, and soil characteristics together with long-term daily climate and streamflow records. The model performance at Station I.17 was evaluated through calibration and validation using the R², Nash–Sutcliffe Efficiency (NSE), and percent bias indices. Rainfall regimes were classified into dry, normal, and wet years based on the mean and standard deviation of 25-year gauge records to drive scenario simulations. The calibrated model reproduced seasonal runoff patterns satisfactorily (monthly NSE up to 0.685 and R² up to 0.712). The simulations demonstrated the strong sensitivity of both the runoff at Station I.17 and reservoir inflow to interannual rainfall differences, with the annual runoff ranging from 71.5 to 379.7 million m³ and the annual inflow to Mae Tum Reservoir ranging from 28.84 to 48.33 million m³. These findings demonstrate that physically based spatial modeling can effectively replace traditional empirical operating rules, providing a highly transferable framework for runoff forecasting, reservoir inflow assessment, and climate responsive water resources planning in data-scarce tropical mountainous basins. Full article

Sudden tailings dam breaches trigger large-scale heavy metal compound pollution in coupled surface water–groundwater systems, requiring systematic full-cycle ecological damage quantification tools applicable to diverse contamination types. This study constructs an integrated full-cycle ecological damage assessment framework for sudden water pollution accidents, integrating three core modules: multi-model pollutant migration prediction, multi-scale aquatic biological damage diagnosis, and three-dimensional ecological-economic loss accounting. The framework adopts a modular design that can potentially accommodate heavy metals (Cd, Cr, As, Pb) and organic pollutants such as polycyclic aromatic hydrocarbons (PAHs), with standardized molecular, individual, and population-level biological endpoints and corresponding pollutant dose–response templates reserved as reference calculation modules. However, applicability beyond this case has not been validated and requires case-specific calibration. To verify the operability and accuracy of the proposed integrated system, a typical tailings dam leakage incident dominated by hexavalent chromium (Cr(VI)) and arsenic (As) pollution was selected as the practical validation case; all field monitoring, pollutant simulation, and final economic loss quantification in this case exclusively rely on on-site measured Cr(VI) and As data, while Cd and PAH-related biological response curves and remediation cost formulas retained in the manuscript only serve as illustrative universal template components of the framework rather than case-measured results. For the Cr(VI)/As pollution case, the advection–diffusion model simulation revealed that the Cr(VI) contamination plume horizontally spread 250 m within 48 h and extended to 560 m after seven days, and anaerobic groundwater environments drove the transformation of toxic mobile trivalent arsenic (As(III)) from primary pentavalent arsenic. The calibrated SWAT model achieved Nash–Sutcliffe efficiency (NSE) coefficients of 0.75 for dissolved Cr(VI) and 0.68 for particulate As. The graph theory-based rapid prediction model cut computation duration down to minutes; when validated against independent field monitoring data, it yielded an average relative error of 14.2%, and its consistency with the SWAT model reached 10.5% relative deviation, satisfying the accuracy requirement for emergency early warning. Field biological monitoring demonstrated substantial ecological impairment: metallothionein (MT) expression in fish tissues was markedly elevated (the reported 6.2-fold induction value derives from standard Cd exposure template tests within the framework, with analogous MT upregulation also observed for field Cr(VI)/As co-stress), and benthic community Shannon diversity declined by over 50% in polluted river reaches. The standardized Ecological Damage Index (EDI) of the case was calculated as 480.2, indicating severe aquatic ecosystem damage, with total comprehensive ecological and economic losses reaching 17.25 million CNY. This study innovatively couples high-precision physical transport models with fast emergency prediction algorithms and establishes a complete multi-tier biological indicator chain linking molecular biomarkers to community integrity metrics; the three-dimensional loss accounting system integrating ecosystem service impairment, restoration expenditure, and post-pollution recovery loss realizes closed-loop full-cycle damage evaluation. The proposed framework, demonstrated for Cr(VI) and As pollution, has a modular design that may potentially be extended to other pollutants such as Cd and PAHs by adjusting model parameters, providing a quantitative reference for emergency disposal, pollution remediation, and ecological compensation of water contamination accidents, although further validation across different pollutants and hydrological settings is required. Full article

Large mineral flotation workshops in severe cold regions represent special industrial indoor environments characterized by the coexistence of limited ventilation and intense heat release. Such conditions generate pronounced spatial thermal stratification and localized heat accumulation within the workshop, leading to uneven worker thermal exposure and increased thermal discomfort and heat stress risk. However, conventional thermal comfort models were primarily developed for ordinary buildings with relatively stable thermal environments. Their applicability to large industrial workshops remains insufficiently validated. Nine representative monitoring points were arranged in the summer operating areas of the workshop, and thermal comfort surveys were conducted among 35 workers who had adapted to the local climate and working environment. The predicted mean vote (PMV) model was used as the baseline assessment framework, while an adaptive predicted mean vote (aPMV) model was further calibrated using field-based thermal sensation information. Monte Carlo simulation was employed to evaluate uncertainty propagation under field-data constraints, and Sobol sensitivity analysis was conducted to identify the dominant factors affecting thermal comfort predictions. The results demonstrated that the conventional PMV model exhibited a clear warm prediction bias under the investigated industrial conditions. After adaptive correction, the deviation from the field-based TSV was reduced by 82.93%, indicating improved agreement with workers’ actual thermal perception. Sensitivity analysis identified metabolic rate as the dominant contributor to aPMV output variance, with first-order and total-effect Sobol indices of 0.530 and 0.535. The proposed framework provides a scenario-specific approach for thermal comfort assessment in the investigated flotation workshop and offers preliminary methodological references for similar large-scale flotation workshops. Full article

Weather radar calibration is essential for ensuring data consistency and quantitative precipitation estimation in X-band radar networks. Existing absolute calibration methods (e.g., metal sphere, horn antenna) suffer from high cost, poor timeliness, and difficulty in automation due to meteorological conditions and airspace restrictions, while spatiotemporal matching methods based on volume scan data suffer from interpolation and matching inaccuracies. To address these issues, this study proposes a collaborative calibration method for X-band radar networks based on opposing Range–Height Indicator (RHI) scans. The method uses a rigorously calibrated reference radar as a benchmark and performs opposing RHI scans with the radar under calibration to obtain synchronized observations within the spatial overlap region. Precise spatial matching is achieved using the nearest-neighbor algorithm based on beam-broadening cross-coverage thresholds, and bias is extracted using both the midline 9-point averaging method (midline method) and spatially constrained regional Statistics method (regional method). Based on a total of 58 sets of opposing RHI scanning cases conducted under stratiform precipitation, scattered precipitation, and weak cloud conditions, the results show that under conditions where echo continuity is maintained near the midline of stratiform and scattered precipitation, both the midline method and the regional method can obtain stable matching data. The midline method achieves a median correlation coefficient (0.821–0.942) higher than that of the regional method (0.860–0.872), and its bias standard deviation remains relatively stable (midline method: 1.39–2.20 dB; regional method: 2.61–3.16 dB). Continuous RHI calibration tests confirm that within a 30-min window, the fluctuation of the data matching correlation coefficient is less than 0.05, and the fluctuation of the bias mean is controlled within ±0.3 dB. Under weak cloud conditions, although the midline method can still achieve a high correlation coefficient, the correctness of its results still requires auxiliary validation through other calibration means. This study provides a relatively efficient and effective technical approach for the automated collaborative calibration of dense X-band radar networks. Full article

Flexible pressure sensors serve as core sensing components for wearable health monitoring systems, electronic skins for soft robots, and flexible human–machine interaction devices. Benefiting from the interfacial electric double-layer polarization effect, iontronic sensing delivers far higher pressure response sensitivity than conventional parallel-plate capacitive sensors, endowing it with distinctive advantages in the detection of weak physiological signals. Nevertheless, current dense ionic thin-film dielectric layers suffer from limited deformation space under compression and poor low-pressure sensing capability. Mainstream high-precision micropillar arrays are fabricated via photolithography, 3D printing, and metal etching molds, which require costly equipment and complicated fabrication procedures, making large-area mass production unfeasible. Random frosted concave-convex microstructures feature disordered dimensions, leading to severe device hysteresis and narrow linear ranges, which fail to achieve ultrahigh sensitivity alongside a wide pressure detection range simultaneously. To address the aforementioned multiple bottlenecks, this paper proposes a low-cost resin template replication process to fabricate TPU-based ionic thin-film dielectric layers with ordered micropillar array microstructures. Combined with inkjet-printed silver conductive PI flexible electrodes, an iontronic flexible pressure sensor with a sandwich layered structure is constructed. Multi-dimensional investigations including microscopic morphology characterization, electromechanical sensing performance calibration, and human wearable application tests are systematically implemented to thoroughly elucidate the synergistic enhancement mechanism of the arrayed micropillars. Test results demonstrate that the effective pressure detection range of the sensor spans 0–1038 kPa, accommodating ultra-low pressures such as pulse signals as well as medium-to-high-pressure loads including joint bending. The sensitivity reaches 23.27 kPa⁻¹ within the low-pressure range of 0–200 kPa and remains stable at 3.52 kPa⁻¹ in the high-pressure range of 200–1038 kPa, with piecewise linear fitting correlation coefficients of 0.93 and 0.96 respectively. Both the response time and recovery time of the device are 40 ms, and the hysteresis error throughout the loading-unloading cycle is merely 2.62%. After 20,000 consecutive cyclic loading-unloading tests, the peak capacitance output only decays by 5.1%, verifying outstanding mechanical fatigue resistance and electrical stability. Validations in multi-scenario applications prove that the sensor can accurately capture human physiological and motion signals including radial artery pulses, laryngeal deformation induced by multi-syllable vocalization, and multi-angle bending of fingers and elbow joints, suitable for home-based health monitoring, quantitative rehabilitation training, flexible tactile interaction and other scenarios. The entire fabrication process eliminates high-precision micro-nano processing equipment such as photolithography systems, plasma etchers and 3D printers; only general chemical raw materials and conventional laboratory instruments are adopted. The reusable templates enable low manufacturing costs and large-area coating forming, offering a novel low-cost technical solution for the engineering implementation and industrialization of high-performance iontronic flexible pressure sensors. Full article

Accurate parameterization of the photovoltaic single-diode model is usually obtained from complete current–voltage (I-V) measurements. However, full I-V curve tracing is not always available in real monitoring environments, where the most accessible variables are irradiance and module temperature. This paper proposes a hybrid methodology for reconstructing the five parameters of the single-diode model from irradiance and temperature data. The method first estimates the maximum-power point and the remaining remarkable points of the I-V curve as well as the photocurrent ($I_{ph}$) through regression models calibrated on measured data. These predicted points are sufficient to solve the SDM equation. A numerical approach is then used to identify the five SDM parameters while enforcing physical admissibility constraints. The method is validated using NREL outdoor datasets from three locations and several photovoltaic technologies. The results show that the maximum-power current is estimated with very high reliability, with $R^2$ values close to unity in almost all cases. Voltage estimation is less stable and depends more strongly on technology and temperature sensor location. The reconstructed I-V curves are physically admissible for most crystalline silicon, HIT, and CdTe modules, whereas CIGS and amorphous silicon modules exhibit lower admissibility. The proposed method should therefore be understood as an environmental-data-driven reconstruction tool when complete I-V curves are unavailable, rather than as a replacement for direct full-curve fitting techniques such as TSLLS or Reduced Form. Full article

Shear-wave time-of-flight (TOF) measurement enables controlled assessment of frequency-dependent wave propagation, but its feasibility in biological tissues remains insufficiently established. This study evaluated whether a custom shear-wave TOF device could detect frequency- and direction-dependent responses in ex vivo tissues. Three porcine liver samples and three chicken breast samples were examined. Chicken breast was measured with propagation parallel and perpendicular to visible muscle fibers. One-cycle sinusoidal excitations were applied at 40–160 Hz, with 50 acquisitions ensemble-averaged per sample–frequency measurement. TOF was estimated using Tx threshold detection and cumulative-energy-based Rx onset detection, and TOF-derived apparent shear-wave propagation speed was calculated from the Tx–Rx distance and the measured TOF. Frequency-dependent data were fitted using the Kelvin–Voigt fractional derivative model to obtain model-dependent KVFD fit parameters. Signal quality was assessed, and a preliminary descriptive comparison with HISKY EQTouch UD3000 (Wuxi Hisky Medical Technologies Co., Ltd., Wuxi, China) SWE was performed. All 63 averaged sample–frequency measurements satisfied the predefined primary-detection criteria. Mean apparent shear-wave speed was 3.145 m/s in porcine liver, 6.133 m/s in chicken breast measured parallel to the fibers, and 5.914 m/s in chicken breast measured perpendicular to the fibers, giving a parallel-to-perpendicular speed ratio of 1.037. Mean post-averaging, post-processing SNR ranged from 24.47 to 31.52 dB. The UD3000 comparison showed the same tissue ranking. The device detected frequency- and direction-dependent responses in averaged ex vivo signals, supporting its feasibility as a controlled research platform. Claims of absolute stiffness accuracy and intrinsic muscle anisotropy require independent calibration and validation. Full article

Large classes in Chinese College English programmes make frequent analytic assessment of student writing difficult. Large language models (LLMs) may support more frequent formative assessment, but their scores may vary across queries and be systematically harsher or more lenient than local teacher ratings. Using a corpus-based, five-fold cross-validated comparative rater-evaluation design, this study examined whether statistical calibration could make LLM-assisted scores more interpretable for College English writing assessment and where their use should remain limited. Data comprised 414 timed argumentative essays written by Chinese non-English majors at one applied undergraduate institution. Two trained College English teachers independently rated the essays on a seven-dimension analytic rubric informed by China’s Standards of English Language Ability, providing the local reference standard. Three LLMs rated each essay–dimension pair on five occasions. Under five-fold cross-validation, uncalibrated scores were compared with location–scale correction, isotonic calibration, and equipercentile linking, using quadratic weighted kappa, Spearman correlation, mean absolute error, signed bias, and half-point tolerance accuracy. Agreement between models did not imply agreement with teachers: two models showed inter-model kappa values of 0.70–0.78 but an average kappa of only 0.15 with teacher ratings while rating the essays about one band more severely. Calibration removed most of this severity difference and raised pooled kappa to 0.61–0.70 depending on the method (0.63–0.64 under equipercentile linking), compared with a teacher–teacher agreement benchmark of 0.747. The three methods differed little, and the improvement mainly reflected closer alignment of score distributions rather than better judgement of writing quality. Agreement was higher for vocabulary, syntax, and grammar but remained low for cohesion and conventions. The findings suggest that LLM-assisted scoring may support low-stakes formative feedback when calibrated to local teacher standards and used under teacher supervision, while teachers retain responsibility for judging content, coherence, argumentation, and communicative quality. Full article

Objectives: Women with a history of breast cancer are at increased risk of developing subsequent breast cancer, including ipsilateral recurrence and contralateral new primary breast cancer. This study evaluated the discriminatory performance of a mammogram-based artificial intelligence (AI) risk model for predicting subsequent breast cancer within one year after a negative screening mammogram. Methods: This enriched retrospective case–control study included women with a prior history of breast cancer who underwent screening digital breast tomosynthesis between January 2018 and December 2023 at three affiliated academic breast imaging centers. Digital breast tomosynthesis examinations classified as BI-RADS 1 or 2 were retrospectively analyzed using the ProFound AI^® Risk model version 1.0 to estimate 1-year breast cancer risk. Patients were classified according to whether they developed subsequent breast cancer within one year of the index screening examination. Model discrimination was evaluated using receiver operating characteristic analysis. Sensitivity, specificity, positive predictive value, and negative predictive value were calculated at an exploratory cutoff selected by maximizing the Youden index. Results: The study included 96 women (mean age, 65.3 ± 8.7 years), of whom 32 developed subsequent breast cancer within one year, and 64 did not. The mean AI risk score was significantly higher in the subsequent breast cancer group than in the control group (1.18 ± 0.59 vs. 0.49 ± 0.41; p < 0.001). The AI model demonstrated an AUC of 0.824 (95% CI: 0.728–0.921). At an exploratory cutoff of 0.39, sensitivity was 81.3%, specificity was 76.6%, PPV was 63.4%, and NPV was 89.1%. In separate exploratory analyses, the AUC was 0.790 (95% CI: 0.641–0.939) for ipsilateral recurrence and 0.860 (95% CI: 0.752–0.974) for contralateral new primary breast cancer. AI risk scores were not significantly correlated with tumor size or age at subsequent breast cancer diagnosis. Conclusions: In this enriched retrospective case–control study, higher mammogram-based AI risk scores were associated with subsequent breast cancer within one year after a negative screening examination. The model demonstrated discriminatory performance for both ipsilateral recurrence and contralateral new primary breast cancer; however, these analyses were exploratory. Because the cohort was enriched for subsequent breast cancer events, the reported predictive values are specific to the study sample and should not be extrapolated to routine surveillance populations. Larger prospective cohorts are needed to validate discrimination, calibration, and clinical utility. Full article

(1) Background: Drug-eluting bead transarterial chemoembolization (DEB-TACE) is increasingly used for unresectable colorectal liver metastases (CRLM), yet individualized prognostic tools are lacking. We developed and internally validated nomograms predicting overall survival (OS) and hepatic progression-free survival (hPFS). (2) Methods: In this single-center retrospective cohort, reported per the TRIPOD guideline, OS and hPFS were estimated by Kaplan–Meier methods, and independent predictors from multivariable Cox regression were assembled into nomograms. Internal validation combined 1000-sample bootstrap optimism-corrected concordance indices (C-index), a uniform shrinkage factor, a bootstrap calibration slope, and a LASSO–Cox sensitivity analysis. (3) Results: Among 63 patients (44 deaths; 42 intrahepatic-progression events), median OS was 10.9 months and median hPFS was 5.8 months. Independent OS predictors were baseline CEA, high liver tumor burden (≥10 lesions), CEA decline (protective), and second-line-or-beyond interventional therapy (corrected C-index: 0.796). Independent hPFS predictors were high liver tumor burden, CEA decline, and age (corrected C-index: 0.718). Nomogram-defined high-risk groups had markedly shorter OS (5.3 vs. 22.8 months) and hPFS (3.3 vs. 8.6 months; both p < 0.001). Grade ≥3 toxicity occurred in 6%. (4) Conclusions: In real-world DEB-TACE-treated CRLM, liver tumor burden and CEA dynamics dominated prognosis; the internally validated nomograms provide individualized estimates and risk stratification, pending external validation. Full article