Accurate prediction of ultra-high-performance concrete (UHPC) compressive strength is essential for optimizing mixture design and reducing experimental iterations. Existing machine learning approaches suffer from limited algorithm diversity, insufficient statistical validation, and inadequate uncertainty quantification. This study presents a comprehensive framework through systematic evaluation of 20 algorithms across seven categories on 863 experimental observations. Six physically meaningful composite features (such as water-cement ratio, total binder content, and fiber aspect ratio) are engineered to capture intrinsic material relationships, with the Boruta algorithm employed for feature selection. Statistical robustness is ensured through 30 repeated experiments analyzed using both frequentist (p-value, effect size, 95% CI) and Bayesian approaches. CatBoost achieves optimal performance (R² = 0.8979 ± 0.0239, RMSE = 10.58 ± 1.45 MPa), with curing age, sand content, and steel fiber volume identified as dominant predictors through multi-perspective interpretability analysis integrating SHAP, ALE, permutation importance, and LIME. External validation on 810 independent samples yields R² = 0.5923 (RMSE = 25.68 MPa) under significant cross-dataset conditions, with performance reduction attributed to feature availability differences and distribution shift. Comprehensive uncertainty quantification yields prediction uncertainty of 3.48%, substantially below previously reported thresholds. The proposed framework offers practitioners a reliable tool for UHPC mixture screening while maintaining prediction confidence for structural engineering applications.

Flexural strength (FS) governs the structural performance of 3D-printed concrete (3DPC) under bending loads yet remains difficult to predict owing to the coupled influence of binder composition, supplementary cementitious materials, water-to-binder ratio, and fiber reinforcement geometry on interlayer fracture behavior. Much of this compositional diversity stems from supplementary cementitious materials, industrial by-products whose reuse as partial cement replacement lowers the embodied carbon of printable mixes. A machine learning framework was trained on 209 FS records covering OPC- and SAC-based systems (FS: 3.65–45.0 MPa; W/B: 0.15–0.65). Six composite features were constructed from physical principles, two of them specific to bending: Fiber_Pullout_Index encoding post-crack pullout energy and Binder_Efficiency capturing cement quality per unit water content at the fiber–matrix interface; Lasso regularization with the one-standard-error rule reduced the 19-variable space to 14 active predictors. Twenty regression algorithms spanning eight families were benchmarked under 30 independent partitions; the Friedman test rejected equal performance (${\chi }^2=337.02$, $p=4.88\times {10}^{-60}$) and all 19 pairwise Wilcoxon comparisons against CatBoost were Holm-significant. CatBoost ranked first (mean rank of 18.17/20; 30-run $R^2=0.9302\pm 0.0722$; seed-42 partition: $R^2=0.9557$; RMSE = 1.761 MPa; MAPE = 11.20%). n(W/B) emerged as the primary driver across SHAP, ALE and LIME, with a monotonic ALE profile spanning 8.99 MPa and no inflection over the full printable window; Binder_Efficiency ranked second (PDP range: 5.21 MPa), isolating cement grade and paste dilution as independent strength levers. Cross-conformal prediction provided finite-sample coverage guarantees without distributional assumptions (empirical coverage: 95.24%; conformity quantile: 3.83 MPa); bootstrap analysis put the epistemic component at a mean predictive SD of 0.946 MPa, a quarter of that quantile. External validation yielded $R^2=0.769$ (Pearson $R=0.923$, RMSE = 1.91 MPa), with 19 of 20 predictions (95.0%) within Bland–Altman 95% limits of agreement, confirming transfer to a source study withheld from model training. A graphical user interface packaging the 14-feature CatBoost pipeline supports mix design queries without programming.

Compact malaria detectors for microcontrollers are almost always benchmarked on the NIH Malaria dataset with a per-cell random split. This leaks slide identity because the cells come from only about 200 slides and a random split mixes same-slide cells across training and testing. The leakage also distorts architectural conclusions: under a leakage-free slide-disjoint protocol, per-module ablation gains collapse to seed noise and an apparent cross-site robustness variant loses most of its advantage. Headline accuracy falls from 97.1% to 95.6%, a gap that sits within the cross-seed noise, and all eight tested architectures move the same way. The evidence is this unanimous direction, not the size of any single gap. This benchmarking finding is our main contribution. Two results survive. First, MalariaNet, our 21 K-parameter detector, reaches about 95.6% accuracy at 23.5 KB of INT8 weights, with a numerically faithful on-chip forward on an STM32H743 at a 1.2 FPS triage rate. Second, it is among the most interference-robust of the eight networks and the most robust microcontroller-deployable model. Scope is limited to single P. falciparum thin-smear cells. Slide-disjoint evaluation should become standard, and we provide MalariaNet as the first leakage-free, on-device-validated point-of-care malaria reference.

The Controller Area Network (CAN) protocol lacks native authentication mechanisms, exposing modern vehicles to critical security threats. While deep learning-based intrusion detection systems show promise, existing solutions require computational resources far exceeding automotive-grade microcontroller constraints, hindering practical embedded deployment. This paper proposes LCSMC-Net, an ultra-lightweight neural architecture for resource-constrained CAN intrusion detection. The framework integrates three innovations: (1) Separable Multiscale Convolution Lite (SMC-Lite) blocks capturing multitemporal attack patterns with minimal parameters; (2) Lightweight Channel-Temporal Attention (LCTA) achieving linear $\mathcal{O}\left(N\right)$ complexity through adaptive pruning; and (3) 6-dimensional CAN-optimized features exploiting protocol-specific characteristics for aggressive compression. The framework employs Bayesian hyperparameter optimization and knowledge distillation for systematic model compression. Extensive experiments on CAN and CAN-FD datasets demonstrate that LCSMC-Net achieves 99.89% accuracy with only 9401 parameters and 2.84M FLOPs, outperforming existing solutions while meeting real-time constraints of automotive embedded systems, providing a viable edge AI deployment solution.

Compact malaria detectors for microcontrollers are almost always benchmarked on the NIH Malaria dataset with a per-cell random split. This leaks slide identity, because the cells come from only about 200 slides and a random split mixes same-slide cells across training and test. Under a leakage-free slide-disjoint protocol, the leakage fabricates architectural conclusions. Per-module ablation gains collapse to seed noise, and an apparent cross-site robustness variant loses most of its advantage. Headline accuracy falls from 97.1% to 95.6%, a gap that sits within the cross-seed noise, and all eight tested architectures move the same way. The evidence is this unanimous direction, not the size of any single gap. This self-critical finding is our main contribution. Two results still survive. First, MalariaNet, our 21K-parameter detector, reaches about 95.6% accuracy at 23.5KB of INT8 weights, with a numerically faithful on-chip forward on an STM32H743 at a 1.2FPS triage rate. Second, it is the second most interference-robust of the eight networks and the most robust microcontroller-deployable model. Scope is limited to P. falciparum thin-smear single cells. Slide-disjoint evaluation should become standard, and we provide MalariaNet as the first leakage-free, on-device-validated point-of-care malaria reference.

Chronic kidney disease (CKD) impacts more than 850 million people globally, yet existing machine learning methodologies for risk stratification encounter substantial challenges: computationally intensive hyperparameter tuning, model opacity that conflicts with clinical interpretability standards, and class imbalance leading to systematic prediction bias. We constructed an integrated architecture that combines XGBoost with Optuna-driven Bayesian optimization, evaluated against 19 competing hyperparameter tuning approaches and tested on CKD patients using dual-paradigm statistical validation. The architecture delivered 93.43% accuracy, 93.13% F1-score, and 97.59% ROC-AUC—representing gains of 6.22 percentage points beyond conventional XGBoost and 7.0–26.8 percentage points compared to 20 baseline algorithms. Tree-structured Parzen Estimator optimization necessitated merely 50 trials compared to 540 for grid search and 1069 for FLAML, whereas Boruta feature selection accomplished 54.2% dimensionality reduction with no performance compromise. Over 30 independent replications, the model exhibited remarkable stability (cross-validation standard deviation: 0.0121, generalization gap: −1.13%) alongside convergent evidence between frequentist and Bayesian paradigms (all p < 0.001, mean CI-credible interval divergence < 0.001, effect sizes d = 0.665–5.433). Four separate explainability techniques (SHAP, LIME, accumulated local effects, Eli5) consistently identified CKD stage and albumin-creatinine ratio as principal predictors, aligning with KDIGO clinical guidelines. Clinical utility evaluation demonstrated 98.4% positive case detection at 50% screening threshold alongside near-optimal calibration (mean absolute error: 0.138), while structural equation modeling revealed hyperuricemia (β = −3.19, p < 0.01) as the most potent modifiable risk factor. This dual-validated architecture demonstrates that streamlined hyperparameter optimization combined with convergent multi-method interpretability enables precise CKD risk stratification with clinical guideline alignment, supporting evidence-informed screening protocols.