Antimicrobial resistance (AMR) remains a significant global issue, driven by the ability of bacteria to utilize specialized mechanisms to produce harmful proteins, adapt to their environment, release effectors, or exchange resistance traits. These mechanisms include Sec/Tat pathways and many specialized secretion systems that enable protein translocation, interbacterial antagonism, and horizontal gene transfer among a range of bacterial pathogens. However, their structural diversity, rapid evolutionary dynamics, and inconsistencies in genomic annotation hinder experimental characterization, necessitating computational approaches. Recent advancements in bioinformatics and artificial intelligence, exemplified by tools such as BastionHub, Sismis, MacSyFinder/TXSScan, DeepSecE, and Effectidor II, have substantially improved genome mining, effector prediction, and comparative genomics, enabling more systematic identification of secretion components and substrates within large datasets. Integrative multi-omics, machine-learning-based quantitative structure–activity relationship (QSAR) modelling, and AI-assisted virtual screening support antibacterial candidate prioritisation, but compounds identified using AI are not necessarily inhibitors of bacterial secretion systems. In contrast, secretion-system-targeted inhibitors specifically aim to disrupt secretion machinery or effector translocation; their target-specific effects require experimental validation, and most such approaches remain computational or preclinical. Combining bioinformatics, deep learning, and multi-omics provides a scalable framework for functional characterization of secretion systems and informs the development of future AMR intervention strategies.
Background Deep learning can accurately classify thyroid nodules on ultrasound, but discrimination alone does not establish clinical reliability. We developed a calibrated deep-ensemble framework incorporating uncertainty-aware selective prediction and conservative image-based triage. Methods Backbone selection used nested five-fold cross-validation with three-fold inner architecture selection. ConvNeXt-Tiny with a squeeze-and-excitation block was selected consistently and used to construct a deterministic five-member ensemble. Within each outer-development partition, images were separated into training, early-stopping, calibration, and threshold-selection subsets, while the outer-test data remained isolated. Mutual information (MI), predictive variance, and maximum-softmax-probability (MSP)-derived uncertainty were compared using risk–coverage and triage analyses. Predictive values were assessed across assumed malignancy prevalences of 5%–70%. External evaluation applied the frozen TN5000-derived ensemble and operating parameters to TN3K without recalibration. Results On TN5000, pooled out-of-fold AUC-ROC was 0.9377, average precision 0.9708, expected calibration error 0.0138, and Brier score 0.0827. Under the primary MI-guided policy, 8.86% of images received a No-FNA suggestion, 60.00% an FNA recommendation, and 31.14% radiologist review. Among 3574 malignant images, 14 (0.39%) received a No-FNA suggestion and 708 (19.81%) were assigned for review. Predictive variance and MSP-derived uncertainty showed better error-ranking performance than MI on AURC and AUGRC, although no measure consistently optimized all triage outcomes. Across malignancy prevalences of 5%–70%, NPV ranged from 99.95% to 97.93%, while PPV increased from 29.28% to 94.75%. On TN3K, the frozen policy achieved an AUC-ROC of 0.7870, average precision of 0.7254, expected calibration error of 0.1899, and an 83.7% radiologist-review rate. Conclusion The ensemble demonstrated strong internal discrimination and calibration with conservative selective prediction. However, performance and calibration deteriorated substantially on TN3K, indicating limited transportability under dataset shift. Multicentre validation, site-specific recalibration, threshold validation, and prospective workflow evaluation are needed before clinical implementation.
Ossiculoplasty requires decisions about residual anatomy, the middle-ear environment, reconstruction configuration, and durable hearing restoration. Artificial intelligence and patient-specific computational modeling offer potentially complementary approaches, but their evidential roles must be separated. This narrative review critically interprets selected clinical reconstruction studies, adjacent hearing-prediction studies, computational investigations, and methodological guidance. A targeted literature update supplemented the original bibliography; the review does not claim systematic coverage or provide pooled estimates. Direct ossiculoplasty evidence is distinguished from type I tympanoplasty and other adjacent procedures. Prognostic performance under observed care does not establish comparative benefit from an alternative prosthesis. Similarly, experimental vibration or anatomical fit does not establish improved hearing or long-term implant safety. We propose a decision-centered framework that establishes anatomical eligibility before treatment comparison, specifies the prediction timepoint, separates core clinical data from optional imaging, and defines hearing and durability outcomes. The framework includes a causal roadmap, structured recording of surgical rationale, evaluation of calibration and explanation quality, and staged prospective implementation with lifecycle governance. Treatment comparisons are restricted to clinically feasible alternatives with adequate overlap or evaluated prospectively where equipoise exists. The nearer-term objective is appropriately evaluated prognostic decision support, not autonomous implant selection. Patient-specific design remains a future research pathway requiring explicit links between mechanical predictions and clinical outcomes, device-safety assessment, and prospective evaluation of clinical utility.
Breast-cancer classification necessitates evaluation of the predictive discrimination, probability calibration and explanations of the model. This paper compares an uncertainty-conscious graph neural network to the Wisconsin Diagnostic Breast Cancer dataset, which consist of 569 patients and 30 morphological features. A 30-node feature graph is associated with each patient where edge estimates are only done based on training data. A two-layer graph attention encoder is tested using 50-pass Monte Carlo dropout, and six base cases. The five-fold cross-validation strategy is repeated five times and re-stratified with five repeats of cross-validation, including training-only scaling, graph construction and internal model selection. The proposed configuration achieved mean accuracy of 96.34% (corrected 95% interval 94.53 to 98.16%), AUC-ROC of 0.990, malignant-class F1-score of 0.951, ECE of 0.043 and Brier score of 0.031. Comparative inference is done by paired procedure with Holm adjustment which is dependence corrected. Expected-gradient SHAP and attention are evaluated over the identical trained encoder over all 30 descriptors; mean rank correlation between these two in a global manner was 0.820. Generalized deletion/insertion and repeated-fit ranking stability experiments are an assay of the capacity to explain cross-method independent behaviour. These findings characterize discrimination, calibration and explanation behaviour under internal validation on the WDBC dataset. Independent clinical evaluation is required to assess generalizability.
Background Gallbladder ultrasound classification is challenging because of irregular lesion morphology, acoustic shadowing, and overlapping imaging characteristics. This study proposes Ada-TransGBC, a lightweight framework integrating complementary attention and boundary-guided learning mechanisms for five-class gallbladder ultrasound classification. Methods Ada-TransGBC combines an ImageNet-pretrained EfficientNet-B0 backbone with Adaptive Deformable Strip Attention (ADSA), Shadow-Aware Gated Attention (SAGA), and Edge-Guided Boundary-Aware Auxiliary Loss (EGBAL). ADSA captures irregular directional structures, SAGA modulates spatial features in potentially shadow-affected regions, and EGBAL encourages boundary-sensitive learning using Canny-derived pseudo-boundaries. Class-balanced sampling, focal loss, MixUp, progressive fine-tuning, and test-time augmentation support model optimization. Results The proposed framework achieved 81.57% accuracy, 77.68% macro-precision, 80.03% macro-recall, 78.40% macro-F1, and 0.957 macro-ROC-AUC on 689 validation images. Compared with EfficientNet-B0, accuracy improved from 74.82% to 81.57%, while macro-F1 increased from 70.15% to 78.40%. Five-fold cross-validation yielded a mean macro-F1 of 78.19 ± 0.94%. The model required 6.13 million parameters and approximately 0.52 GFLOPs. Conclusion Ada-TransGBC uniquely integrates adaptive directional feature extraction, spatial feature modulation, and boundary-guided auxiliary supervision within a compact architecture. The findings demonstrate improved multiclass classification over the evaluated baseline. However, independent patient-level and external multicenter validation remains necessary before clinical applicability can be established.
