The gait parameters, i.e., gait speed, stride length, step length, and cadence, serve as key indicators for patients with mobility impairments. Monitoring these parameters can significantly aid in assessing recovery progress, especially in stroke patients. Compared to traditional clinical visits, wearable sensors, or motion capture systems, millimeter-wave (mmWave) device offers unique advantages: low-cost, contactless, light-free, and no privacy invasion, making frequent gait monitoring feasible for home deployment. We introduce mmGait, an mmWave-based system that extracts foot features from reflections and learns walking patterns to accurately predict the 3D locations of both feet, from which we can derive the gait parameters automatically without requiring subjects to wear any sensors or be recorded by cameras. Across multiple stroke patients, mmGait achieves low median errors in estimating key gait parameters: less than 0.03 m/s for gait speed, ∼3 cm for stride length, 7.4 cm for step length, and below 10 steps/min for cadence. Agreement analysis further demonstrates good agreement with the OptiTrack reference for gait speed and stride length, supporting the feasibility of the proposed contactless approach for coarse spatiotemporal gait assessment.

Retinopathy of Prematurity (ROP) is among the major causes of preventable childhood blindness. Automated screening remains challenging, primarily due to limited data availability and the complexity of the condition involving both structural staging and microvascular abnormalities. Current deep learning models depend heavily on large private datasets and passive multimodal fusion, which commonly fail to generalize to small, imbalanced public cohorts. We thus propose the Context-Aware Asymmetric Ensemble Model (CAA Ensemble) that simulates clinical reasoning through two specialized streams. First, the Multi-Scale Active Query Network (MS-AQNet) serves as a structure specialist, utilizing clinical contexts as dynamic query vectors to spatially control visual feature extraction for localization of the fibrovascular ridge. Secondly, VascuMIL encodes Vascular Topology Maps (VMAP) within a gated Multiple Instance Learning (MIL) network to precisely identify vascular tortuosity. A synergistic meta-learner aggregates these orthogonal signals to resolve diagnostic discordance across multiple objectives. Tested on a highly imbalanced cohort of 188 infants (6004 images), the framework attained state-of-the-art performance on two distinct clinical tasks: achieving a Macro F1-Score of 0.93 for Broad ROP staging and an AUC of 0.996 for Plus Disease detection. Crucially, the system features 'Glass Box' transparency through counterfactual attention heatmaps and vascular threat maps, proving that clinical metadata dictates the model's visual search. Additionally, this study demonstrates that architectural inductive bias can serve as an effective bridge for the medical AI data gap.

Respiration affects the electrocardiogram (ECG) through heart motion and changes in thoracic conductivity, leading to the development of ECG-derived respiration (EDR). Here, we propose a dipole-based EDR approach that tracks the heart's beat-to-beat rigid-body motion from standard 12-lead ECG recordings and encodes each beat as a six-dimensional log-SE(3) descriptor. An unsupervised Rayleigh-quotient formulation extracts a respiratory surrogate by maximizing spectral energy in the respiratory band (0.1-0.4 Hz) without requiring a reference signal. A cohort of 80 participants was analyzed, including healthy volunteers (n=51, 49% women) and patients with ischemic heart disease (n=29, 21% women). Performance was evaluated against amplitude-based methods (QRSpk and QRS integral) and a VCG-based approach using Pearson correlation (CC) and breathing rate error (MAE). The proposed method achieved median CC=0.87 and MAE=0.7 BPM, comparable to VCG (CC=0.85, MAE=0.89 BPM) and superior to amplitude-based methods (CC=0.75-0.78, MAE=0.93-1.24 BPM; p<0.05). Performance was significantly reduced for all methods in ischemic patients (p<0.0001), and a significant sex-pathology interaction was observed (p<0.01). However, the dipole method remained the most accurate. A dedicated 128-electrode pipeline revealed a dominant cranio-caudal cardiac displacement of 7.6 mm consistent with imaging literature, and a sequential activation of all six degrees of freedom across the respiratory cycle. This framework provides a robust, physiologically grounded approach for respiratory estimation from ECG, with potential applications in clinical monitoring and non-invasive characterization of cardio-thoracic dynamics.

For more than a century, non-typhoidal Salmonella has been recognised as a major foodborne pathogen, responsible for human illness underscoring its continued importance to public health. In the current study, we employed molecular dynamic simulation and Surface Plasmon Resonance (SPR) biosensing based experimental approaches to assess the interaction between the ssDNA aptamers (DFRM13, DFRM110, DFRM29, DFRM33, and DFRM39) and Salmonella FimD usher protein. The docking studies revealed superior binding performance for DFRM13-FimD complex based on its low RMSD (0.7 ± 0.4 nm) and z-score (-1.9) values followed by DFRM39-FimD complex (RMSD: 1.1 ± 0.6 nm, z-score: -1.6). The RMSD, RMSF, Rg, SASA analysis and high number of H-bonds with the ssDNA aptamer (∼18 bonds) observed during 50 ns molecular dynamics simulations revealed that DFRM29-FimD complex also exhibited stable binding interactions. SPR assay optimized with biotinylated aptamers on streptavidin sensors exhibited relatively lower dissociation constant (KD) or higher affinity for native FimD protein from Salmonella lysates (KD10[-7] to 10[-8] M) than rFimD (KD10[-6] to 10[-7] M) due to the cross-over SELEX approach adopted for enriching the aptamers. The DFRM13 demonstrated lowest KD toward FimD in Salmonella thermal lysates (KD 1.24 × 10[-8] M), followed by DFRM39 (KD 4.17 × 10[-8] M) indicating highest binding affinity among other aptamers as determined by SPR interaction kinetics. The sensorgrams with curvatures characteristic of FimD detection was not observed when thermal lysates of non-Salmonella bacteria were examined in the SPR assay, thus confirming the specificity of aptamers for Salmonella detection. Overall, the computational and experimental findings from our study demonstrated potential for possible utilization of the aptamers DFRM13 and DFRM39 in Salmonella diagnostic platforms.

Public tabular datasets are widely reused to benchmark machine learning for health-related screening, but their evaluation validity is rarely audited. Using the widely cited, synthetic Sleep Health and Lifestyle Dataset (374 records, three classes) as a case study, we show that the very high accuracies reported for it reflect evaluation optimism rather than a property of the prediction task: several published studies report accuracies approaching 98%, whereas our reproduction under conventional cross-validation reaches approximately 91%, and confining duplicate records to a single fold reduces this further. Our central finding is duplicate-record leakage: only 132 of the 374 rows are distinct, 64.7% of rows are exact copies of another row, and the conventional cross-validation used in prior work allows identical rows to fall on both sides of the train/test split. When duplicate records are confined to a single fold using group-aware cross-validation, a gradient-boosted model drops from 91.2% to 72.6%, and on fully de-duplicated data to 61.7%: standard cross-validation overestimates performance by 18.6 percentage points relative to duplicate-group-aware evaluation, and performance falls by a further 10.9 percentage points after full de-duplication, reflecting the reduced effective sample size and conflicting-label structure. Under the corrected protocol we benchmark seven classifiers with correlation-corrected (Nadeau-Bengio) confidence intervals, nested cross-validation for hyperparameter selection, and effect-size-oriented significance testing. Simple models (logistic regression, support vector machine, multilayer perceptron) and CatBoost remain robust (≈87-91%), whereas the untuned random forest, XGBoost, and LightGBM show substantial variability, with very wide intervals; after nested tuning they partially recover, underscoring that model rankings on this dataset are not reliable. As a secondary, deliberately negative result, a feature-graph attention network we designed for the task (with a learnable sparse graph and a residual gate) does not outperform a plain multilayer perceptron (full model 88.2%±4.2 vs 89.3%±3.4; paired Wilcoxon p=0.03), indicating that graph inductive biases do not provide measurable benefit for this small static tabular dataset. We also report a near-deterministic association between the Occupation field and the label (Cramér's V=0.75; an 85% no-learning lookup), which we treat as suggestive of a data-construction artefact rather than as proven leakage. We conclude, in line with recent leakage-aware work in medical imaging, that this dataset is unsuitable for strong claims about sleep-disorder screening unless duplicate structure and synthetic-construction artefacts are handled explicitly; the contribution is a reusable, duplicate-aware benchmarking protocol rather than a clinical result.

Alzheimer's disease staging from structural magnetic resonance imaging (MRI) is challenging when clinically adjacent stages and severe class imbalance are considered simultaneously. This study evaluates six-class classification of cognitively normal, subjective memory complaint, early mild cognitive impairment, mild cognitive impairment, late mild cognitive impairment, and Alzheimer's disease using T1-weighted MRI from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The final cohort included 3592 MRI volumes from 904 participants, with subjective memory complaint representing 3.79% of the data. We developed a parameter-compact three-dimensional convolutional neural network (3D CNN) with multi-scale volumetric feature extraction and imbalance-aware training using class weighting, focal loss with label smoothing, weighted sampling, and minority-class intensity perturbation. In the primary evaluation, partitioning was performed at the MRI ImageID (scan) level: a fixed test set of 719 MRI volumes was excluded from model development, while the remaining 2873 vol were used in five stratified training and validation folds. ImageIDs were mutually exclusive across partitions; however, participants were not grouped, and longitudinal scans from the same participant could therefore occur in different partitions. The proposed model achieved 0.9018 ± 0.0452 accuracy, 0.8804 ± 0.0393 macro F1, 0.8762 ± 0.0544 Matthews correlation coefficient (MCC), and 0.9463 ± 0.0056 macro area under the receiver operating characteristic curve (AUC) under this scan-level protocol. A same-data EfficientNet-B0 baseline achieved stronger aggregate performance, while the proposed model used 43.7% fewer parameters. Paired ablation analysis showed that weighted sampling did not improve aggregate metrics but increased subjective memory complaint precision, recall, and F1, with all corresponding 95% confidence intervals excluding zero. These findings support class-specific evaluation under severe class imbalance while distinguishing scan-level discrimination from participant-independent generalization.

Muscle mass and muscle quality, characterized by intramuscular adiposity and fiber composition, are clinically established biomarkers significantly associated with obesity, sarcopenia, frailty and cardiometabolic disorders. Current muscle quantification efforts are predominantly based on CT, often focusing on single anatomical region such as the L3 vertebral level, due it its relatively standardized acquisition protocols and the overall simplicity of developing muscle segmentation algorithms. However, MRI offers superior soft-tissue contrast for visualizing intramuscular fat and enables muscle assessment across a wider range of body regions. In this study, we developed SegmentAnyMuscle, a model capable of segmenting muscles across diverse anatomical regions and various MRI sequences, addressing the barriers to clinical muscle assessment. By developing upon a robust, expert-curated dataset of 316 MRI exams from 160 patients covering 11 anatomical regions, SegmentAnyMuscle demonstrated high segmentation performance with an averaged Dice Similarity Coefficient (DSC) of 88.45% on routinely acquired MRI sequences and maintained strong performance (DSC of 86.21%) under clinical challenging cases, including less common sequences, abnormalities such as muscular atrophy, implant-associated artifacts, and significant noise. Furthermore, automatic skeletal muscle mass measurement provided by SegmentAnyMuscle showed excellent correlation with expert manual assessments (Pearson correlation coefficient >0.99, p < 0.001). By providing the first reliable, automatic muscle segmentation tool across diverse anatomical regions and MRI acquisition settings, SegmentAnyMuscle can open new opportunities for MRI-based muscle health assessment and encourages broader indicators of metabolic and functional health. We made codes and model publicly available at: https://github.com/mazurowski-lab/SegmentAnyMuscle.

Detecting lung cancer at an early stage improves survival outcomes, whereas accurate segmentation and classification of lung nodules are fundamental to dependable diagnosis. Classical approaches encounter difficulties, like limited accuracy, data imbalance, and complex feature extraction, emphasizing the need for enhanced automated techniques to increase timely lung cancer diagnosis. Hence, a Deep Sequential Convolutional Neural Harmonic Network (DSeqCH-Net) is devised for lung nodule classification from Computed Tomography (CT) scans. Lung CT images are initially subjected to adaptive median filtering to remove unwanted noise. The enhanced images are then forwarded to the segmentation phase, where Lung Nodule Detection Transformer (LN-DETR) extracts the lung nodule regions using the proposed FusionLoss-LFC loss, formed by combining Lovász-Softmax, Cross-Entropy, and Focal losses. The segmented images are augmented using rotation, flipping, and resizing techniques to increase image variability. Feature extraction is then carried out to obtain the most discriminative image characteristics, which are subsequently used by a Sequential Convolutional Neural Network (SCNN) for lung cancer detection. Lastly, lung cancer classification is performed employing DSeqCH-Net. Additionally, DSeqCH-Net achieved the highest True Negative Rate (TNR), accuracy, and True Positive Rate (TPR) of 95.813%, 96.745%, and 97.643%.

The microscopic detection of Mycobacterium tuberculosis bacilli in Ziehl-Neelsen (ZN)-stained sputum smears remains essential for tuberculosis (TB) diagnosis, but it is limited by inter-observer variability, high operational workload, and a shortage of qualified specialists. Deep learning (DL) has shown strong potential to support this task; however, progress is constrained by the limited availability of well-annotated datasets and the lack of standardized evaluation protocols. This study presents a comprehensive investigation of DL-based object detection methods for TB bacillus identification, comprising (i) a systematic comparison of seven detectors, (ii) the introduction of a new annotated ZN-stained dataset containing 500 images, and (iii) an image-partitioning strategy designed to mitigate the performance degradation caused by heterogeneous image resolutions. Three public datasets and the proposed dataset were evaluated using cross-validation and cross-dataset experiments. Among the seven methods tested, Faster R-CNN consistently achieved the highest overall precision, sensitivity, F1-score, and average precision across datasets. The proposed partitioning strategy significantly improved detection performance, particularly for large images. Cross-dataset experiments revealed limited generalization across laboratories, underscoring the importance of standardized datasets and context-aware training. Overall, the proposed dataset and experimental findings offer practical guidance for future research toward robust, clinically deployable AI-assisted TB diagnostic workflows.

Ensuring diver safety remains a major challenge in breath-hold diving due to physiological stressors such as hypoxia and increased hydrostatic pressure. Assessing physiological parameters before and after diving enables characterization of baseline and post-dive variations associated with these stressors. This study presents an imaging photoplethysmography (iPPG)-based method to estimate vital parameters, aiming to develop a non-invasive monitoring framework for diver safety. The proposed framework is based on an Attention U-Net architecture for segmenting regions of interest in video recordings of nine breath-hold divers acquired before and after a 30-meter dive in the controlled thermal pool at Y-40 The Deep Joy, Italy. The segmentation model was trained on the Face and Skin Detection (FSD) dataset. iPPG signals were derived from pulsatile skin regions to estimate heart rate (HR) and blood oxygen saturation (SpO2). The pipeline was first validated on the Pulse Rate Detection (PURE) dataset and then evaluated on the acquired diver data. The segmentation model achieved an accuracy of 97% and an intersection over union of 89% on test data. HR and SpO2 were estimated with mean absolute errors (MAEs) of 5 bpm and 4%, respectively, on the PURE dataset. The proposed pipeline further estimated pre-dive MAEs of 6 bpm for HR and 3% for SpO2, while post-dive MAEs were 7 bpm for HR and 3% for SpO2, indicating increased physiological variability following breath-hold diving. This study demonstrates the effects of breath-hold diving on physiological responses and establishes the feasibility of iPPG-based non-contact monitoring for detecting stress-related changes in divers.

Self-supervised pretraining is promising for large-scale neuroimaging representation learning, yet the impact of region-aware masking and hybrid sequence modeling remains underexplored. In this work, we introduce Rhamba, a region-aware pretraining framework that integrates anatomically guided masking with hybrid Attention-Mamba architectures for resting state functional magnetic resonance imaging (fMRI) analysis. Models were pretrained on the ABIDE dataset using region-aligned patch embeddings and three masking strategies with increasing spatial specificity. We evaluated four architectural variants: a Mamba-dominant model, an Alternate architecture with interleaved Mamba and Attention blocks, and two hybrid encoder-decoder configurations (Attention-Mamba (AM) and Mamba-Attention (MA)). The pretrained models were fine-tuned on downstream classification tasks using the COBRE and ADHD-200 datasets for schizophrenia and attention-deficit/hyperactivity disorder discrimination. We employed Integrated Gradients, an explainable AI method, to identify the brain regions contributing to model predictions. Our masking strategy strongly influenced reconstruction behavior, with reconstruction loss following a consistent ordering (Any>Majority>Pure). However, this trend did not directly translate into downstream performance, where differences were modest and dataset-dependent. The hybrid architecture with the MA configuration achieved the highest average AUROC across both datasets. Rhamba demonstrated competitive performance against state-of-the-art methods. Region-wise analysis showed that peak performance depends on the interaction between masking strategy and architecture rather than a single dominant configuration. The hybrid architectures underscore the importance of combining global context modeling with efficient sequence dynamics. Overall, Rhamba offers a flexible framework for balancing interpretability, scalability, and performance in large-scale fMRI representation learning.

Understanding systemic determinants of breast tumor immunity requires bridging transcriptomically distinct tissue compartments that cannot be sampled simultaneously in a single patient. We developed an MMD-regularized Domain Adaptation Autoencoder (DAA) to align unpaired RNA-seq profiles from GTEx neuroendocrine tissues (n=189) and TCGA-BRCA tumors (n=1391) within a shared 128-dimensional latent space, enabling the first cross-tissue transcriptomic interrogation of the neuroendocrine-breast tumor immune interface. The dominant cross-tissue axis was identified by Pearson correlation and rigorously validated by permutation testing (n=1000 iterations), then independently assessed in METABRIC microarray (n=1980) and SCAN-B RNA-seq (n=3273) cohorts via a strict gene-intersection protocol that eliminated zero-padding artefacts. The DAA achieved stable cross-domain alignment (mixing score =26.58%), and Latent Dimension 31 emerged as a significant systemic immune-inflammatory axis (p=0.001; aggregate correlation 13.9× above the permutation null), driven by T-cell receptor variable chains, immunoglobulin genes, and the tolerogenic phospholipase PLA2G2D. METABRIC validation recovered a mechanistically concordant acute-phase secretory signature (LBP, SAA1, PLA2G2A), while SCAN-B confirmed PLA2G2D and CCL18 on a unified cross-platform latent axis. The latent score significantly stratified overall survival (p=0.0036) and relapse-free survival (p=0.0084), and precisely reproduced the established breast cancer immune topology across all six molecular subtypes (Kruskal-Wallis H=139.4, p<0.0001). External validation in the independent neoadjuvant GEO cohort GSE25066 (n=508; Affymetrix GPL96) via a Strict Intersection Protocol (604-gene intersection, zero-padding eliminated) confirmed axis recovery (Latent Dimension 115; PLA2G2D |r|=0.266), significant distant relapse-free survival stratification (log-rank p=0.022), and non-significant pathological complete response to chemotherapy (p=0.241), establishing the axis as a prognostic but not predictive biomarker. Functional annotation in GSE25066 revealed significant correlation with all 12 curated immune cell signatures (Spearman ρ=0.10-0.42; all padj<0.05), and GSEA pre-ranked analysis across 13,236 genes identified 34 significantly enriched Hallmark pathways (FDR < 0.25), led by Interferon Gamma Response (NES =2.90) and opposed by Estrogen Response Early (NES =-2.71). These findings, validated across 7152 patients in four independent cohorts, provide a computational transcriptomic framework linking systemic neuroendocrine regulation to breast tumor immunobiology and nominate PLA2G2D, SAA1, and LBP as candidate circulating biomarkers warranting prospective proteomic validation.

Electrocardiographic imaging (ECGI) is a promising non-invasive technique that reconstructs epicardial potentials by combining high-density body-surface recordings with patient-specific 3D geometries. To systematically compare the performance of the main ECGI regularisation methods, an experimental setup was developed using isolated Langendorff-perfused rabbit hearts. Panoramic optical mapping, epicardial electrograms, and torso-tank signals were acquired simultaneously during atrial sinus rhythm and ventricular tachycardia. A tailored pre-processing pipeline was applied prior to inverse reconstruction using multiple methods, including Tikhonov (orders 0-2), truncated singular value decomposition (TSVD), damped singular value decomposition (DSVD), generalised minimal residual (GMRES), and Bayesian approaches. Results showed that no single method was universally optimal, with performance strongly dependent on cardiac region, rhythm, and evaluation metric. Tikhonov regularisations achieved the highest waveform similarity, reaching mean cross-correlation (CC) values up to 0.84 in the right atrium during sinus and 0.83 during ventricular tachycardia, though performance varied across orders and regions. In contrast, TSVD- and DSVD-based approaches yielded lower correlations (typically 0.62-0.78). Second-order Tikhonov achieved the lowest localisation error during ventricular tachycardia (6.77 ± 3.65 mm), while GMRES offered a competitive balance between spatial precision (7.04 ± 2.36 mm) and temporal accuracy. Bayes showed the highest CC variability across electrodes. Despite these differences, all methods consistently preserved dominant activation frequencies found in the measured signals (≈1.7 Hz in sinus rhythm and ≈4.8 Hz in tachycardia).