Tuning an engineering controller or a structural design normally means running a differential-equation model inside an outer search loop: the objective is available only pointwise, carries no usable gradient, and mixes smooth tracking terms with non-smooth actuator saturation and constraint penalties. The search algorithm is therefore itself a design decision, yet it is almost always fixed once, on a single plant, without evidence that the choice transfers to any other. This paper asks whether it does, and when a metaheuristic is needed at all. We answer both questions with a unified differential-equation (DE) and operations-research (OR) framework in which six representative plants—a separately excited DC servomotor, a mass–spring–damper stage, a stirred heat exchanger, a continuously stirred tank reactor (CSTR), a series RLC circuit, and an inverted pendulum on a cart—are wrapped in a single proportional–integral–derivative (PID) evaluation harness, alongside three constrained mechanical-design benchmarks and the ten-dimensional Rastrigin function. Seven long-established metaheuristics (GA, PSO, DE, SA, GWO, WOA, SCA) are compared over fifteen seeded runs at a budget of exactly 3000 objective evaluations, enforced by an evaluation counter rather than by nominal population and iteration settings, so that trajectory- and population-based solvers are treated identically. Four non-metaheuristic baselines, a hyperparameter sensitivity sweep and two ablation studies test the framework itself. The solver ranking reverses across problem classes: differential evolution attains the best Friedman rank on the PID landscapes (1.52 of 7) where the sine–cosine algorithm is worst (6.44), whereas on Rastrigin that ordering inverts (SCA 1.20, WOA 2.07, DE 5.60), and GWO and DE dominate the constrained designs. The baselines explain why. On the smooth PID surfaces a multi-start Nelder–Mead search matches every metaheuristic to five decimal places, so the metaheuristic buys robustness rather than accuracy; classical Ziegler–Nichols tuning, by contrast, is 8.9 to 19.8 times worse, and no local method remains competitive once constraints or multimodality are present. The framework therefore does more than rank solvers—it identifies which problems justify a metaheuristic—and is released as a fully specified, reproducible pipeline.
Early identification of respiratory and cardiovascular problems depends on the accurate characterization of cardiopulmonary sounds. Nevertheless, the majority of current research concentrates on the independent analysis of either heart or lung sounds and frequently employs a single feature representation. In response to these shortcomings, this paper presents XCardioPulmoNet, an image-based framework that develops an approach to classifying heart and lung sounds. Through the utilization of Mel spectrograms and Wavelet transforms, the proposed method converts audio inputs into representations that are based on time and frequency. The subsequent step involves the reconstruction of a hybrid image by merging the information that is retrieved from both representations, which are complementary to one another. To improve multiclass classification performance and discriminative feature learning, the proposed EAGR_Net, a residual attention-guided convolutional neural network, is used to categorize the produced hybrid images. Furthermore, Gradient-Weighted Class Activation Mapping (Grad-CAM) is used to depict the discriminative regions that influence the model’s predictions and improve interpretability. Five-fold cross-validation is used to test the baseline CNN and Attention CNN models. With a ROC-AUC of 0.9851 and PR-AUC of 0.9558 on the heart sound dataset and a ROC-AUC of 0.9978 and PR-AUC of 0.9917 on the lung sound dataset, respectively, the experimental findings show that EAGR_Net outperforms the baseline models. These findings demonstrate the framework’s robustness, interpretability, and diagnostic potential for automated multiclass cardiopulmonary sound classification and disease detection.
Reproducibility is not only a cornerstone of the advancement of science, but also of scientific credibility. However, it remains a persistent challenge across many disciplines, and it is now referred to as the “reproducibility crisis.” This issue is particularly relevant in computational modeling. In this paper, we want to know if the usage of different tools and implementation choices can lead to divergent outcomes. We investigate the replicability of a SEIRS epidemiological model implemented via agent-based modeling (ABM) across seven different platforms and programming languages, namely C++, Julia, Python, NetLogo, GAMA, Cormas, and PythonPDEVS. Each implementation was based on a common formal specification using the Overview, Design concepts and Details (ODD) protocol and developed independently by different modelers. Our results show that while all implementations qualitatively reproduce the expected dynamics of the model, a sharp infection peak followed by damped oscillations, they produce statistically significant differences in key metrics such as peak amplitude and timing. These differences are also large in magnitude: the variation between implementations is about 7.6 times larger than the stochastic variation within an implementation (across its 30 replications). Because each implementation was produced by a different modeler on a different platform, this variance reflects the combined effect of the tool and the modeler, which are confounded in the present design and cannot be separated. A complementary analysis of several independent developers working in a single language (C++ or Java) shows that the modeler alone induces statistically significant and large differences, confirming that the individual developer is an important source of divergence; the specific contribution of the technology therefore remains open. We conclude that even when a model is formally specified, independently developed ABM implementations can yield divergent numerical results. A controlled ablation traces much of this divergence to two under-specified execution details, namely how the residence times are discretised into whole days and whether the state-transition test is strict; specifying them in the ODD would remove much of the divergence. Our findings highlight the need for rigorous reproducibility and replicability practices, and for model specifications precise enough to pin down such implementation choices.
High discriminative performance alone does not ensure reliable clinical decision support; well-calibrated confidence estimates are also essential. Standard deterministic deep learning classifiers produce probability predictions without explicitly quantifying predictive uncertainty, limiting confidence assessment for safety-critical clinical applications. We propose an uncertainty-aware evidential deep learning framework for binary stroke prediction based on the Dirichlet distribution. This framework estimates predictive confidence in a single pass, with explicit decomposition of epistemic and aleatoric uncertainty, characterizing model and data uncertainty respectively. The model achieved strong discriminative performance, with an area under the receiver operating characteristic curve approaching 0.99 and high average precision in distinct experimental repeats. At the operating point selected using Youden’s J statistic, the model achieved a sensitivity of 0.968, a specificity of 0.949 and a negative predictive value of 0.980. The calibration analysis demonstrates that predicted probabilities are closer to the observed outcomes, while the uncertainty analysis indicates that ambiguous and misclassified cases exhibit higher epistemic uncertainty, conveniently identifying them as low-confidence predictions. Through comprehensive ablation studies, the proposed full evidential formulation consistently outperforms the baseline convolutional networks, focal loss and Monte Carlo dropout. Computational evaluation showed that evidential inference required a single forward pass, whereas MC Dropout used 10 stochastic forward passes per prediction. The results show that evidential deep learning correctly identifies and calibrates uncertainty, providing clinically useful stroke detection at a lower computational cost, which is ideal for any clinical decision support system.
Multi-organ segmentation is critical for clinical diagnosis and treatment planning, yet real-world deployment is hindered by cross-center imaging heterogeneity, scarce annotations, limited computational resources, and strict privacy constraints. Vision foundation models such as SAM provide strong representations but are impractical for on-site fine-tuning due to their scale and communication cost. We propose FedMed-LoRA, a federated and parameter-efficient fine-tuning framework for adapting foundation models to medical imaging. FedMed-LoRA operates on a fully frozen SAM encoder and introduces dual low-rank adaptations in both attention and MLP pathways, enabling effective medical-domain transfer with only a small set of trainable parameters. INT4-quantized frozen weights combined with BF16 low-rank training substantially reduce memory requirements, allowing large models to be fine-tuned on hospital-level hardware. Within a federated workflow, only low-rank updates are shared, preserving data privacy and model intellectual property. Experiments demonstrate that FedMed-LoRA matches full-parameter fine-tuning on FLARE22 and achieves superior cross-domain performance on Synapse, outperforming MedSAM, Med-SA, and SAMed. Ablations confirm the complementary roles of the two LoRA paths. FedMed-LoRA offers an efficient, privacy-preserving, and clinically deployable solution for medical multi-organ segmentation.
