Search Results (3,413)

(1) Background: Information diffusion models developed for graph-based platforms such as Reddit and broadcast architectures such as Telegram identify temporal features—particularly the timing of peak spread—as dominant predictors of coverage. Whether these predictors generalise to platforms where content is distributed through algorithmic recommendation rather than social-graph contagion remains an open question. (2) Methods: We analyse the YouNiverse dataset, comprising 133,364 English-language YouTube channels observed weekly from January 2015 to September 2019 (18.9 million observations). We derive channel-level diffusion features—including time-to-peak, post-peak decay rate, diffusion volatility, and upload frequency—and train three machine learning models (Linear Regression, Random Forest, and LightGBM) on two tasks: predicting peak weekly view growth (regression) and identifying viral channels (classification). A single-feature naive baseline (subscriber count alone) establishes the marginal contribution of the broader feature set beyond subscriber count alone, and a temporal split experiment (training on channels peaking before 2018, testing on 2018–2019) assesses cross-temporal stability. Because subscriber count and subscriber rank are measured at the October 2019 crawl, this is a retrospective characterisation rather than a strict real-time forecasting design. (3) Results: LightGBM achieves $R^2=0.776$ (5-fold CV: $0.778\pm 0.003$) compared with $R^2=0.548$ for the naive baseline, a net gain of $+0.228R^2$. Because subscriber rank and subscriber count are near-perfectly collinear, we interpret them jointly as a channel-size dimension (42.2% of total mean absolute SHAP attribution), rather than as independent effects. Time-to-peak ranks fourteenth (1.1%), in contrast to its dominant role on Reddit ($r=0.995$, rank #1). For virality classification, LightGBM achieves ROC-AUC $=0.967$. Under the temporal split, Random Forest ($R^2=0.703$) outperforms LightGBM ($R^2=0.683$), showing greater cross-temporal stability within this retrospective split. (4) Conclusions: Within the 2015–2019 data, the results are consistent with algorithmic recommendation weakening the relationship between temporal diffusion dynamics and coverage magnitude at the channel level. Time-to-peak is weakly informative in this setting, while generalisation to the current recommendation system requires validation on newer data. Full article

Early in ovo sex identification is important for improving hatchery resource utilization; however, sex-related optical differences are weak during early embryonic development and exhibit clear stage dependence. In this study, longitudinal visible–near-infrared transmission spectra were acquired from 1600 fertilized pink-shelled eggs of Bashang Long-tailed chickens across four independent incubation batches during incubation days 1–7. Before model development, 200 eggs were reserved as a fixed internal hold-out test set, while the remaining 1400 eggs were used for exploratory incubation-day and feature screening, five-fold cross-validation, and model comparison. Among all 21 dual-day combinations, D2 + D5 achieved the highest exploratory validation AUC of 0.951. The selected scheme required separate measurements on D2 and D5, with the final sex classification completed on D5. Further incorporation of the signed difference spectrum ΔS and relative change rate R increased the AUC to 0.965, demonstrating the incremental discriminative value of longitudinal developmental-change information. After the input scheme had been fixed, DSSF-Net achieved the numerically highest performance among the evaluated models under unified egg-level stratified five-fold cross-validation, with an accuracy of 0.915 ± 0.015, an AUC of 0.974 ± 0.012, and an F1-score of 0.915 ± 0.014. The final model achieved an accuracy of 0.915 and an empirical AUC of 0.966 (95% CI: 0.940–0.987) on the fixed internal hold-out test set. Complete four-round leave-one-batch-out validation yielded mean accuracy and AUC values of 0.846 ± 0.014 and 0.926 ± 0.012, respectively, indicating that the longitudinal spectral information from D2 + D5 maintained relatively stable discriminative performance across incubation batches within the investigated population. These results demonstrate the feasibility of longitudinal spectral modeling for early sex identification in fertilized pink-shelled eggs of Bashang Long-tailed chickens. Systematic offline validation was completed under laboratory conditions, including evaluation on the fixed internal hold-out test set and four-round leave-one-batch-out validation, supporting the effectiveness and batch-level stability of D2 + D5 longitudinal developmental spectra. Post hoc wavelength-level interpretation further identified 594–620 nm and 660–675 nm as two candidate spectral regions with relatively high predictive value. Collectively, these findings provide a reliable longitudinal spectral modeling framework for early non-destructive in ovo sex identification and establish an experimental and methodological foundation for wavelength selection, device development, and subsequent engineering validation of multispectral detection systems under hatchery conditions. Full article

Battery state-of-health (SOH) benchmarks often mix cycles from the same cell across training and test sets. We examined whether predictor structure changes the gap between within-cell interpolation and cross-cell transfer. Seven regressors and five predictor sets were evaluated in four NASA cells (636 cycles) using ten random 80:20 splits and leave-one-cell-out (LOCO) testing repeated over ten model seeds. Across-model median random split RMSE fell from 5.00 SOH points with age alone to 1.52 with diagnostics plus age, whereas median LOCO RMSE fell only from 5.83 to 4.61; the absolute protocol gap therefore increased from 0.83 to 3.09 points, and the median cell-paired LOCO/random ratio increased from 1.24 to 4.53. With all diagnostics, model-level ratios ranged from 1.65 to 5.16, and random-versus-LOCO ranks were uncorrelated (descriptive Spearman ρ = 0.00). Holding the number of training cycles constant, increasing the number of source cells from one to three reduced median LOCO RMSE from 5.25 to 4.73 points, but 32.1% of cell–model pairs did not improve. Standardised sliced Wasserstein shift was associated with the validation ratio (descriptive ρ = 0.73) but not with absolute LOCO RMSE (ρ = −0.01). Rated capacity normalisation preserved the predictor-dependent gradient. An eight-cell Oxford analysis was limited to an age-only boundary check and was not a feature-matched replication. These four-cell laboratory results show that richer discharge diagnostics can improve represented cell interpolation much more than unseen cell transfer; claims should therefore report absolute and relative errors with complete cell-held-out validation. Full article

Underwater point cloud registration using Multibeam Echosounder (MBES) data is fundamental to marine exploration and seafloor mapping. However, MBES point clouds present unique challenges compared to terrestrial Light Detection and Ranging (LiDAR): high noise levels, low overlap rates, and strongly anisotropic distributions caused by the strip-like sonar scanning pattern. These characteristics degrade existing registration algorithms, which predominantly assume locally isotropic point distributions. To address these challenges, this paper proposes Anisotropic Kernel Point Convolution (A-KPConv), a novel operator tailored to the strip-like structure of MBES point clouds. A-KPConv uses Principal Component Analysis (PCA) to estimate local geometric principal directions and constructs an affine transformation that adapts the convolution kernel shape and orientation to align with the local geometry, thereby shifting feature extraction from isotropic aggregation to structure-aware feature learning along the principal structural directions. Building upon this operator, we integrate A-KPConv into Predator—a framework for low-overlap registration—to develop A-Predator, in which the standard isotropic KPConv in the first three encoder layers is replaced with A-KPConv so that structure-aware feature learning is performed where geometric information is most salient. Extensive experiments on the public Dotson-east dataset and a self-collected LiQuan Lake (LQL) MBES dataset demonstrate that A-Predator achieves the highest registration recall among the evaluated methods. On Dotson-east, recall improves from 31.63% to 59.55% under 10% overlap, with consistently low translation and rotation errors. Ablation studies support the effectiveness of A-KPConv relative to the evaluated anisotropic operators, and cross-dataset experiments suggest more effective transfer than the evaluated baselines from Dotson-east to the rescaled LQL-MBES data without fine-tuning. Full article

AI-driven credit scoring is supervised as a high-risk application in banking and insurance, yet unfairness is rarely operationalized as a measurable category of model, conduct, legal, and reputational risk. Using 20,000 anonymized applications from a Southern European digital lender (15.2% twelve-month default rate), we estimate three model families—a regularized logistic regression, a gradient-boosting machine, and a multi-layer perceptron—under a fully crossed design in which each family is evaluated without mitigation and under pre-processing (reweighing), in-processing (an exponentiated-gradient reduction, applicable to any base learner, together with adversarial debiasing where gradient-based training permits it), and post-processing (reject-option) interventions, so that the mitigation effect is no longer confounded with the choice of estimator. No sensitive-group field enters any estimated specification; group membership is used exclusively for auditing. Predictive performance (AUC-ROC, Brier score and Brier skill score relative to the base-rate forecast, F1 on the default class, Gini, and the Kolmogorov–Smirnov statistic) is reported jointly with group fairness (demographic-parity and equal-opportunity differences, disparate-impact ratio, Theil index) and with group-conditional calibration, at an explicitly stated and economically justified decision threshold. Every fairness quantity is accompanied by stratified-bootstrap confidence intervals and, for stochastic learners, by seed-level dispersion. The interpretable benchmark attains an AUC of 0.780 and a Brier score of 0.104 against 0.129 for the constant base-rate forecast, and the high-capacity models improve on it by under one AUC point. Disparity is present but is located geographically rather than in the composite group label: the disparate-impact ratio is 0.724 [0.693, 0.754] for the lowest socio-economic neighborhood cluster, excluding the four-fifths screening value, against 0.809 [0.776, 0.840] for the ethno-socioeconomic proxy, whose interval contains it, and no measurable gender disparity. Group membership is recoverable from the neutral feature set at an AUC of 0.654, and 42% of the group gap in predicted risk travels through the bureau credit score alone, so feature deletion cannot close the channel. Feature attributions and an auxiliary group-recoverability test locate the proxy pathways through which disparity arises, and a misclassification-sensitivity analysis bounds the effect of error in the group proxy, which attenuates measured disparity toward parity. We map the results onto Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, the GDPR as interpreted in SCHUFA Holding, Directive (EU) 2023/2225, EBA loan-origination guidance, and Solvency II, EIOPA, and IAIS expectations, and propose fairness-risk controls organized around impact assessment, independent validation, and three lines of defense governance. Because the evidence comes from credit origination at a single lender, the insurance argument is developed at the level of regulatory and governance architecture rather than as an empirical transfer of estimates. Full article

Cross-event fire-severity mapping requires a model to delineate the burned footprint and grade severity within it across wildfires whose spectral expression varies with vegetation and observation conditions. Nominal multiclass models treat the classes as parallel alternatives and leave the order implicit. We introduce Co-Burn, a bi-temporal Siamese model that adds a pre-to-post dNBR channel to the post-fire branch as an NIR-SWIR change anchor and uses a conditional ordinal head to estimate burn presence before high-severity assignment. Ten methods were compared across 14 New South Wales wildfires against a Sentinel-2 FESM-derived three-class target, with 4 complete fires held out from model development and selection. Co-Burn ranked first under both pixel-pooled and event-mean aggregation, reaching 0.520 ± 0.017 and 0.524 ± 0.028 external burned mIoU. Event-level factorial contrasts showed that burned-mIoU effects varied among fires, while dNBR anchoring reduced false-burn rate on all four external events. At the fixed operating point, Co-Burn assigned 19.9% of reference-unburned pixels to burned classes, against 28.9% for the reflectance-only nominal variant. On the hardest held-out fire, limited false-burn expansion coexisted with a downward shift across the ordered severity classes. Co-Burn supports ordered three-class mapping of previously unseen forest fires before target-fire labels become available. Full article

Conceptual metaphors structure abstract domains through concrete, perceptual, affective, or bodily experience, but they are accessed through metaphorical expressions. Controlled experimental work requires normed materials characterizing expressions, not only the conceptual mappings they instantiate. This is particularly relevant for European Portuguese, where research on metaphors lacks multidimensional normative resources for everyday metaphorical expressions. We introduce ME6D-PT, a database of 213 European Portuguese metaphorical expressions instantiating 20 conceptual metaphors and accompanied by English counterparts. Portuguese native speakers rated the expressions on six dimensions relevant to language processing: familiarity, concreteness, valence, arousal, embodiment, and transparency. Items were presented individually without context, organized in seven lists in a within-subject design. Aggregated item-level reliability ranged from moderate to excellent across dimensions. Descriptive results showed that the expressions were generally familiar and transparent while varying substantially in concreteness, valence, arousal, and embodiment. Further analyses showed strong positive associations among familiarity, transparency, and concreteness, whereas embodiment was weakly related to experiential and semantic dimensions. Exploratory quadratic models revealed nonlinear patterns involving valence, including the expected U-shaped association with arousal, and a decelerating positive association between transparency and familiarity. ME6D-PT provides a controlled resource for stimulus selection and research on metaphors, abstract language, embodiment, and future cross-linguistic research. Full article

To investigate freeze–thaw damage evolution and mix proportion effects in PVA fiber–fly ash–slag powder composite concrete (PVA-FA-SPC), nine composite concrete mixtures were designed using an L₉(3³) orthogonal array at a water to binder ratio of 0.45, with plain concrete serving as the reference, and subjected to 200 rapid freeze–thaw cycles. Freeze–thaw resistance was evaluated using surface deterioration, the mass loss rate, and the relative dynamic elastic modulus. Within the investigated factor levels, PVA fiber volume content exhibited the strongest main effect trend, followed by the total mineral admixture replacement rate and the fly ash to slag powder mass ratio. T20R1:2P0.3, containing 20% total mineral admixture replacement, a fly ash to slag powder mass ratio of 1:2, and 0.3% PVA fiber, showed the best measured performance, retaining a relative dynamic elastic modulus of 79.18% after 200 cycles. A two-parameter Weibull function was used as a phenomenological description of the damage evolution; compared with the classical exponential model, the average RMSE decreased from 0.032 to 0.019 and the average MAPE decreased from 5.07% to 2.98%. Because only nine independent orthogonal mixtures were available for parameter mapping, a common shape parameter of α₀ = 2.1403 was adopted, and a parsimonious equation for the scale parameter β was selected using the small-sample-corrected Akaike Information Criterion (AICc) together with leave-one-mixture-out cross-validation (LOMO-CV). LOMO-CV yielded R² = 0.886, RMSE = 0.034, MAE = 0.025, and WMAPE = 19.98%; however, T20R1:2P0.3 exhibited a node WMAPE of 72.75%, indicating a local limitation of the reduced order mapping. Three non-orthogonal mixtures within the same material system yielded R² = 0.943, RMSE = 0.026, and WMAPE = 15.36%. The proposed model is therefore intended for local trend analysis and preliminary mix screening within the calibrated material system and parameter range rather than for universal service life prediction. Full article

Background: An optimal 25OHD level can prevent osteopenia of prematurity (bone health) and premature delivery (non-skeletal activity). Although maternal and fetal vitamin D (25OHD) levels correlate, data on preterm infants’ 25OHD status at birth are limited, with a well-recognized deficiency among term infants. Objectives: To evaluate maternal and neonatal 25OHD status at preterm birth and its determinants in a northern-latitude country, Poland. Design/Methods: Cross-sectional study between August 2015 and September 2016 investigating cord and maternal blood 25OHD levels at delivery in 69 pairs of mothers and newborns ≤ 32 6/7 weeks of gestation, and analyzing correlations between maternal and neonatal 25OHD levels and their influencing factors, with univariable and multivariable linear regression. Results: Median (IQR) 25OHD levels were 25 (16.6–34 ng/mL) in cord blood and 21.2 (15–29 ng/mL) in mothers and correlated positively (r = 0.69, p < 0.0001). 25OHD insufficiency was detected in 84.06% of mothers and 62.32% of newborns. Maternal 25OHD levels were higher in summertime delivery (p = 0.0008). The rate of vitamin D supplementation during pregnancy was 54.4%, increasing maternal (27 ng/mL vs. 16.7 ng/mL, p = 0.0006), cord blood (32.4 ng/mL vs. 17.4 ng/mL, p = 0.0005), and summer seasonal cord blood 25OHD levels (38.89 ng/mL vs. 25.5 ng/mL, p = 0.023), and the percentage of maternal (72.7% vs. 27.3%, p = 0.005) and neonatal 25OHD sufficiency (80% vs. 20%, p = 0.0021). Conclusions: Vitamin D insufficiency is common in mothers and their prematurely born infants at birth; vitamin D supplementation reverses the trend. To avoid inadequate supplementation in preterm newborns and in mothers at risk of preterm delivery, regular assessment of 25OHD levels should be considered. Full article

The Min pig, a Chinese indigenous breed, is valued for its excellent meat quality and stress tolerance. Nevertheless, its growth rate falls considerably short of commercial pig breeds, a pattern typical of most unselected indigenous populations, and marked individual variation further undermines its economic viability. To explore the genetic basis of this variation, we evaluated a series of nonlinear fixed-effects and mixed-effects models based on the Gompertz, Logistic, and von Bertalanffy functions using body weight records from 92 Min pigs. Model selection was based on AIC, BIC, and leave-one-out cross-validation. The best-fitting model was a Logistic nonlinear mixed-effects model with individual-level random effects on all three growth parameters (Asymptotic weight, timing parameter, and be parameter), which clearly outperformed models with simpler random-effect structures and fixed-effects models. From this model, we identified three characteristic growth transition points: the early transition point (Growth Rate Index, GRI) at 94.83 days (22.32 kg), the single inflection point (Maximum Growth Rate, MGR) at 166.06 days (52.80 kg), and the late transition point (Late Growth Rate Index, LGRI) at 237.29 days (83.29 kg), with a maximum absolute growth rate of 488.2 g/day. These points partitioned growth into initial acceleration, rapid growth, deceleration, and Asymptotic growth phases. Using individual fitted growth curves, we selected five fast-growing and five slow-growing pigs that reached approximately 90 kg during the plateau phase, defined as a predicted body weight of at least 95% of the individual Asymptotic weight. The fast-growing group reached 90 kg at 240.5 ± 20.42 days, whereas the slow-growing group reached the same weight at 286.67 ± 19.20 days. At the 90 kg slaughter weight, we collected longissimus dorsi muscle samples from these pigs during the plateau phase and performed RNA-seq. Transcriptome analysis revealed 864 differentially expressed genes between the two groups, with 540 upregulated and 324 downregulated in the fast-growing group. Pathway enrichment implicated the PI3K-Akt and TGF-β signaling pathways in muscle development, and differential expression of IGFN1, DCN, COL3A1, MYOC, COL1A2, COL1A1, and IGF2 may explain the growth variation between the groups. In summary, the Logistic mixed-effects model with individual-level random effects on all growth parameters effectively captures the growth pattern of Min pigs, and the PI3K-Akt and TGF-β pathways likely mediate growth differences in this breed. Full article

Long COVID, defined by the National Academies of Sciences as an infection-associated chronic condition persisting at least three months following SARS-CoV-2 infection, presents a growing public health challenge. Despite national prevalence estimates of 11 to 15 percent among U.S. adults who have had COVID-19, the general population estimates of approximately 5 to 7 percent population-level symptom burden in diverse metropolitan areas remain insufficiently characterized. A cross-sectional survey was administered to adults residing in Houston and Harris County. Participants self-reported infection history, persistent symptoms lasting at least 90 days, health status, access to care, and functional impacts. Descriptive statistics summarize infected and affected, symptom frequency, and functional limitations. Of 264 eligible respondents who tested positive for COVID and rated the worst symptoms, 92 (34.85 percent) met the operational definition of Long COVID, representing an estimated sample-based prevalence of 16.91 percent across all 544 survey respondents. Given the use of convenience sampling, this figure should be interpreted as descriptive of this sample rather than as a generalized population-level estimate. Fatigue was the most reported persistent symptom (69.32 percent), followed by brain fog (62.5 percent), shortness of breath (48.84 percent), headache or migraine (43.02 percent), and joint or muscle pain (41.38 percent). Cognitive and functional symptoms predominate over respiratory symptoms. These findings demonstrate a substantial burden of persistent, multisystem symptoms in a sample of community-dwelling urban adults, highlighting the need for longitudinal research and targeted public health strategies to address long-term consequences of SARS-CoV-2 infection in diverse metropolitan settings. Full article

Most computational work on emotional intelligence in education targets students; where the emotional states of teaching staff are modeled at all, the signal is typically a static questionnaire score rather than an input to an automated response. This study asks a narrower question: does routing a detected emotion through Self-Determination Theory—rather than naming the emotion, or using generic language—produce a message that university faculty themselves rate as better, and does that advantage survive once the underlying emotion classifier is realistically imperfect? Using two public emotion corpora (GoEmotions and ISEAR), a 75-item researcher-authored and independently unvalidated set of faculty-oriented stress vignettes, four classifier families spanning lexicon counting to large-language-model few-shot prompting, a four-level message-personalization ablation (generic, emotion-only, need-only, full situational context), and blind ratings from 30 higher-education faculty and researchers, two findings are reported. First, when classifiers are compared on a matched label space, the few-shot large-language-model classifier performs best on both out-of-domain evaluations without fine-tuning, including 0.893 accuracy on the synthetic faculty-oriented vignette set. Second, the human evaluation does not support a uniform monotonic benefit from progressively adding personalization. Need-only routing does not yield a consistent gain over emotion-only messages, whereas the full-context condition shows its clearest positive signal for relevance and motivational usefulness; these two contrasts are significant in the crossed mixed-effects analysis but are not uniformly significant under the rater-level paired tests. The automatic text-similarity proxy does not detect this contextualized-condition advantage. Because the faculty vignettes are synthetic and were not independently content-validated, the results should be interpreted as evidence of technical and methodological feasibility rather than ecological validity or deployment effectiveness. Full article

Background/Objectives: Traumatic intracranial hemorrhage (tICH) in older adults following low-energy falls (LEF) represents a common yet diagnostically challenging condition in the emergency department (ED), where predicting injury prior to computed tomography (CT) remains difficult. Machine learning (ML) has been proposed to support CT decision-making, but its feasibility using routinely available pre-CT clinical variables in this specific population remains unclear. This study presents a systematic exploratory benchmarking of ML pipeline configurations for pre-CT tICH prediction in a well-defined retrospective cohort of older emergency patients following LEF. Methods: We performed a secondary analysis from a retrospective observational bicentric study from two university hospital EDs, including 2250 patients aged ≥65 years presenting after an LEF and undergoing cranial CT. Clinical data were extracted manually from electronic health records (EHRs). Eighteen pre-CT clinical features retrieved from electronic health records were selected based on routine availability and ≤10% missingness. Overall, 1224 valid ML pipeline configurations, combining nine classification algorithms, six imputation strategies, four class-balancing approaches, and optional hyperparameter tuning, were evaluated using 10-fold stratified cross-validation on a training set. The 20 highest-ranked configurations by cross-validation AUC were then assessed on a previously inspected exploratory hold-out test set (n = 563); training-derived rule-out operating points were evaluable for 17 of these 20, as three tuned SVM configurations lacked stored out-of-fold predictions. Results: tICH prevalence was 7.0% (n = 158). Across the 20 highest-ranked configurations, hold-out AUC ranged from 0.517 to 0.585, with Matthews correlation coefficient near zero and balanced accuracy of approximately 50% throughout, indicating differences in operating point rather than in discriminative ability. Some of these top-ranked pipelines reached higher cross-validation AUC (up to 0.679) but detected no cases at the default 0.5 threshold—an effect of the decision threshold under class imbalance rather than of the models’ rank-order discrimination, which was itself limited (hold-out AUC of 0.517–0.585). Conclusions: Despite comprehensive exploratory benchmarking across 1224 ML pipelines, routinely available pre-CT clinical features did not provide sufficient discriminatory signal to develop a clinically useful tICH prediction model in this cohort of CT-imaged older adults following LEF. These findings indicate that none of the evaluated configurations produced clinically adequate performance; this near-chance result persisted across all pipelines and most plausibly reflects a combination of limited feature signal, low outcome prevalence, and a sample size below the level required for reliable model development at this event rate. Future studies should target substantially larger prospective multicenter cohorts and evaluate additional feature domains, including structured clinical examination findings, point-of-care biomarkers, and imaging features. Full article

Accurately identifying the risk of academic failure at the student-course level is fundamental to data-driven early warning and precision student management, which requires a subsequent decision-layer aggregation to inform student-level interventions. However, existing methods fail to jointly capture temporal grade evolution, course prerequisite dependencies, and cross-cohort distribution shifts, limiting their predictive efficacy. To address this, we propose a hybrid model integrating Long Short-Term Memory (LSTM) with a Transferable Multi-Scale Graph Convolutional Network (TMSGCN). Experiments on real records from 17 cohorts across five engineering majors at a Chinese university show that our model achieves strong predictive performance in internal cohort evaluations. In a temporal hold-out test on the 2023 cohort, it attains an AUC of 0.946 and a false-negative rate of 6.7%, highlighting cross-year generalizability. The model consistently outperforms the re-implemented baseline in both predictive accuracy and stability. Interpretability analyses further validate the rationality of its decision-making. The framework shows promising internal transferability within the studied institution. Full article

Machine learning has transformed consumer credit scoring, delivering substantial gains in predictive accuracy over traditional scorecards—but whether those gains come at a cost to fairness has remained contested. The dominant assumption in the literature is that more complex, accurate models amplify bias by encoding historical patterns of disadvantage more effectively. This paper challenges that assumption with direct empirical evidence. We evaluate four model families—logistic regression, random forest, XGBoost, and a multilayer perceptron—across two real-world datasets: the UCI Credit Card Default Dataset and the 2024 US Home Mortgage Disclosure Act national loan-level data, comprising over six million mortgage applications. Using repeated cross-validation, we report predictive performance alongside two primary fairness metrics—demographic parity difference and equalized odds difference—supplemented by false positive rate difference and calibration difference, with confidence intervals across 15 estimation folds. On the UCI data, where demographic disparities are modest, model choice has negligible effect on fairness outcomes. On the HMDA mortgage data, where racial disparities are large and legally consequential, the expected accuracy–fairness tradeoff does not hold; more accurate models produce significantly fairer outcomes on equalized odds within the models, data, and fairness criteria examined here, with logistic regression occupying the worst position simultaneously on all dimensions. Persistent demographic parity disparity among the more complex models is consistent with feature-level bias that no model architecture can resolve. The findings have direct implications for the less-discriminatory-alternatives framework under US fair lending law and for the high-risk classification of credit scoring AI under the EU AI Act. Full article