Background: DeepFlu predicts symptomatic influenza onset from pre-exposure gene expression, but the contribution of individual genes remains unclear.Methods: We apply SHapley Additive exPlanations (SHAP), an explainable AI (XAI) method, to evaluate gene-level contributions in DeepFlu models trained on H1N1 and H3N2 influenza gene expression data (GSE52428 and GSE73072 datasets).Results: Five key genes—HLA-DQA1, HLA-DQB1, XIST, RPS4Y1, and KDM5D—were shared across influenza subtypes. These genes are associated with immune response and sex-linked pathways, with elevated XIST and reduced RPS4Y1/KDM5D expression suggesting higher susceptibility in females. Conclusion: Integrating SHAP with DeepFlu enhances the interpretability of influenza risk predictions, offering actionable biological insights for early intervention.
This paper investigates diffusion models for data augmentation in fetal ultrasound image classification, addressing the critical challenge of limited medical image datasets for rare conditions. We fine-tuned Stable Diffusion with Low-Rank Adaptation (LoRA) to generate synthetic ultrasound images and systematically evaluated their impact on classification performance across multiple experimental conditions. Our results demonstrate that augmenting training data with synthetic images significantly improves classification accuracy and effectively addresses class imbalance issues. Notably, models trained with a combination of real and synthetic data showed enhanced generalization capabilities and improved performance on minority classes. These findings establish diffusion-based augmentation as a promising approach to overcome data scarcity constraints in medical imaging applications, with potential implications for clinical decision support systems.
Stroke (Cerebrovascular Accident, CVA) is one of the diseases with the greatest impact on human health, particularly due to the neurological dysfunctions that often follow, such as upper or lower limb paralysis, foot drop, muscle weakness, and cognitive impairment, which significantly affect patients' daily lives. Existing research indicates that the first three months after a stroke is the "golden period" for treatment, during which rehabilitation training can greatly improve a patient's quality of life and promote neuroplasticity. However, some stroke patients are unable to engage in active rehabilitation training due to physical limitations and can only rely on passive rehabilitation methods. To address this issue, this study proposes a closed-loop brain-controlled exoskeleton system based on brain-computer interface (BCI) technology to enhance rehabilitation in post-stroke patients. This study designed a brain-controlled exoskeleton system consisting of open-loop control and closed-loop feedback. In the open-loop control stage, patients control the exoskeleton's standing, walking, and sitting commands through electroencephalogram (EEG) signals; in the closed-loop feedback stage, the rehabilitation effect is evaluated through brainwave analysis and clinical evaluation. Experimental results demonstrate that the system can effectively assist patients in regaining motor control during training and elicit significant neurophysiological responses in the preparation and execution stages of movement. Additionally, clinical evaluation results showed that patients made notable progress in muscle strength and functional recovery. Although the study sample size was small (n=3) and the treatment period was short, the study highlights the potential application of brain-controlled exoskeleton systems in post-stroke rehabilitation. Future studies should expand the sample size and include long-term follow-up to further assess its clinical feasibility and long-term effects.
This research investigates the complex biochemical mechanisms underlying aging by analyzing primary human fibroblasts using a longitudinal multi-omics dataset. This dataset includes cytology, DNA methylation and epigenetic clocks, bioenergetics, and cytokine profiling. Key findings indicate that mitochondrial efficiency declines with age, while glycolysis becomes more prevalent to compensate for energy demands. Epigenetic clocks, such as Hannum and PhenoAge, showed strong correlations with biological age (rho > 0.650, p < 1e-6), validating the experimental setup and confirming that the cultured fibroblasts were aging appropriately. Fibroblasts with SURF1 mutations exhibited accelerated aging, marked by bioenergetic deficits, increased cell volume, and reduced proliferative capacity, underscoring the pivotal role of mitochondrial dysfunction in cellular senescence. Novel insights were gained from analyzing cytokines like IL-18 and PCSK9, some of which were linked to age-related diseases such as Alzheimer's and cardiovascular disorders. Experimental treatments revealed distinct effects on cellular aging. Dexamethasone reduced inflammation but also increased DNA methylation, induced metabolic inefficiencies, and shortened cellular lifespan. By uncovering connections between mitochondrial dysfunction, epigenetic biomarkers, and immune dysregulation, this research identifies potential therapeutic targets for age-related diseases.
Rapid identification of antibiotic-resistant infections is crucial, as antimicrobial resistance is a global health crisis. Yet, conventional antibiotic susceptibility tests (AST) often require days to yield results. Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) has emerged as a rapid, cost-effective tool for bacterial identification and shows promise for resistance profiling by detecting spectral biomarkers. In this study, we harness MALDI-TOF MS with machine learning and deep learning to predict ciprofloxacin resistance across four Gram-negative bacteria, Escherichia coli, Klebsiella pneumoniae, Acinetobacter baumannii, and Acinetobacter nosocomialis, using a cross-species "basket-wise" approach. We extracted features from mass spectra using kernel density estimation-based peak detection and m/z binning, then trained a random forest (RF) classifier and a convolutional neural network (CNN) to distinguish ciprofloxacin-resistant and susceptible isolates. To interpret the models, we employed dual feature importance analyses: gradient-weighted class activation mapping (Grad-CAM) for the CNN to highlight critical m/z regions and an ensemble RF-based method to identify significant peak features. The CNN achieved higher overall accuracy than the RF, especially in three of four species, while the ensemble RF approach identified interpretable sets of around 20 important m/z peaks per organism. Several informative peaks overlapped between species, indicating some common resistance-associated spectral signatures. However, no single universal marker was found across all species. These findings demonstrate an AI-enhanced MALDI-TOF MS framework for rapid AMR detection, yielding accurate predictions and interpretable spectral markers. The approach highlights clinical potential to guide effective therapy and bolster antimicrobial stewardship, particularly for underrepresented pathogens such as A. nosocomialis.
Motivation: Alzheimer's disease (AD) is the most common type of dementia. Given the lack of a cure, early identification of high-risk populations is crucial for timely prevention. While several studies have focused on AD risk prediction, single feature (e.g., age) may dominate model performance, limiting the discovery of other potential risk factors. This study incorporates key features identified in previous research and applies propensity score matching for age and sex, aiming to improve the predictive performance of AD risk models for older adults.Methods: This study utilized data from the UK Biobank to integrate genetic and clinical data and developed 5-year and 10- year AD risk prediction models for older adults, respectively. The workflow included genome-wide association studies (GWAS) on 433,589 participants were conducted to identify significant Single nucleotide polymorphisms (SNPs) under three p-value thresholds, followed by polygenic risk score (PRS) calculation for 13,282 participants using PRSice-2 and Lassosum, and the integration of multiple features to construct prediction models. Clinical features, PRS, and significant SNPs were then incorporated into four machine learning models: Logistic Regression, LightGBM, XGBoost, and Multi-Layer Perceptron (MLP) for prediction and performance comparison.Results: For the 5-year risk prediction, the MLP model demonstrated the best performance, achieving an AUC of 0.88 based on 37 clinical features and 206 significant SNPs. For the 10-year risk prediction, the MLP model also demonstrated the best performance, achieving an AUC of 0.89 based on 37 clinical features, 206 significant SNPs, and PRS based on these SNPs. SHAP analysis revealed that key contributors across both models included ApoE genotype, urinary tract infection (N390), disorientation, depressive symptoms, and pairs matching time. The 5-year model emphasized immediate clinical and cognitive indicators such as reaction time and number of medications taken, whereas the 10-year model highlighted long-term risk factors including BMI, diabetes, and peak expiratory flow.Conclusion: This study demonstrates that integrating clinical features with PRS can effectively enhance the accuracy of AD risk prediction models for older adults. However, to further validate the utility of PRS, future research should involve collaborations across diverse populations and databases. Additionally, further exploration of other potential risk factors is needed to enhance the clinical applicability of these models.
Image classification tasks in scientific research often struggle with limited data, particularly when few images are available per class, necessitating effective data augmentation strategies. For snakes and other string-shaped objects, traditional augmentation methods such as flipping, adding noise, or altering colour prove inadequate in enhancing classification accuracy. This paper introduces KGOT (Keypoint-Guided Ophidian Transformation), a novel image augmentation technique designed to generate realistic, non-linear transformations of string-shaped objects. KGOT first straightens the snake image, then warps it according to a reference curve defined by key points, effectively inverting the straightening process. By utilizing advanced curve interpolation, pixel mapping, and post-processing techniques, KGOT ensures smooth, biologically plausible deformations while preserving important individual-specific features. Our experiments demonstrate that KGOT-augmented data increase accuracy compared to the traditional augmentation method using geographic transformation techniques. Although some limitations persist in scenarios where snake bodies overlap, the success of KGOT augmentation provides valuable information on factors that improve classification performance. This approach holds promise not only for snake individual identification but also for a broader range of applications involving string-shaped object classification tasks.
Alzheimer's disease (AD) is the most common form of dementia worldwide, but most existing risk prediction models are based on European populations and lack generalizability to other ethnic groups. This study combined genetic and clinical data from both the UK Biobank and a Taiwanese population to develop a model more broadly applicable across populations. By selecting SNPs with similar minor allele frequencies (MAF) between groups and performing genotype imputation, models were built using polygenic risk scores (PRS), genotype, and clinical data. Logistic regression, XGBoost, and multilayer perceptron (MLP) were used for comparison. All models performed well on the UKB validation set (AUC = 0.81), and the logistic regression and MLP models showed improved performance (AUC = 0.87) in the Taiwanese test set. Key predictive features included PRSice2_PRS, lassosum_PRS, and age. The results highlight the potential of integrating genetic and clinical data to improve risk prediction for AD across populations, offering insights into AD pathogenesis and aiding the development of precision medicine strategies.
Long non-coding RNAs have gained significant attention due to their crucial roles in the pathogenesis of complex human diseases, such as neurological diseases, cardiovascular diseases, AIDS, diabetes, and various types of cancer. In the machine learning literature, lncRNA-disease association (LDA) has been widely investigated as a binary classification problem, where each lncRNA-disease pair is seen as an independent instance. This approach presents drawbacks as it does not exploit the correlation among the diseases, aggravates the already imbalanced dataset, and substantially increases the execution time. Furthermore, the literature focuses on the transductive setting where new disease associations are predicted in lncRNAs already seen by the model, which naturally restricts its application to already seen lncRNAs. As a solution, we propose to address LDA prediction as a structured output prediction problem, namely (hierarchical) multi-label classification, where all LDAs are predicted at once for a given lncRNA. We compared several LDA methods and their structured output variants with recent (hierarchical) multi-label classification methods in an inductive setting, e.g., disease associations are predicted in unseen lncRNAs. Our experiments reveal that approaching LDA prediction with structured output prediction leads to superior or competitive results while drastically reducing the running time.
Accurate immune cell composition profiling is crucial for understanding immunological dynamics and disease mechanisms. Bulk RNA sequencing (bulk RNA-seq) is widely employed due to its cost-effectiveness and scalability; however, it lacks the resolution to identify cell-specific gene expression. To address this limitation, we propose a self-attention enhanced deep learning model designed for precise immune cell deconvolution from bulk RNA-seq data. We systematically annotated immune cell types from four single-cell RNA-seq (scRNA-seq) peripheral blood mononuclear cell (PBMC) datasets and validated these annotations against established automated identification tools (SingleR, Seurat, scPred, ScType). Leveraging these annotations, we generated realistic pseudo-bulk RNA-seq training samples using Dirichlet-distribution-based composition sampling, significantly enhancing the model’s performance, particularly for rare cell populations. Comparative evaluations demonstrated that our self-attention enhanced deep learning model consistently outperformed existing approaches, including CIBERSORTx and Scaden, achieving lower prediction errors and higher correlations on benchmark PBMC datasets. Integrating multi-head self-attention allowed the model to dynamically capture intricate dependencies among gene expression features, substantially improving deconvolution accuracy for specific cell subsets. While demonstrating robust performance on PBMC datasets, we acknowledge that broader validation is essential due to potential limitations in generalizability across different tissue types and conditions. Our study highlights the potential of self-attention mechanisms and realistic training data generation strategies to enhance computational deconvolution techniques, providing valuable tools for clinical diagnostics and translational immunology research.
Failed labor induction presents a significant challenge, posing considerable risks for maternal health. The development of accurate predictive models is essential for assisting healthcare providers in determining the most suitable delivery methods. This study explores the application of machine learning (ML) models to predict failed labor induction, utilizing data from the Indonesian National Health Insurance (INHI). A retrospective cohort of 9-month period of pregnancy was established using INHI Sample Data, comprising 27,953 pregnancy cases. Failed labor induction was identified through ICD-10 codes (O61). The features considered included demographic data (age, insurance class, region) and binary-encoded diagnoses. Various ML models were assessed for their predictive performance. The ensemble model demonstrated a high area under the curve (AUC) of 0.759 (95% CI, with a balanced trade-off between sensitivity (0.816) and specificity (0.589). SHAP analysis was used to interpret feature contributions, highlighting both clinical and demographic factors influencing model predictions. ML models showed strong potential in predicting failed labor induction using administrative data. While the ensemble model achieved the best overall performance, simpler models such as logistic regression or XGBoost may offer greater practicality for clinical integration in Indonesia.
To diagnose Helicobacter pylori (H. pylori) infection from endoscopic images, conventional methods require a time-consuming labeling process to annotate individual endoscopic images based on pathological findings. In this paper, we aim to diagnose H. pylori infection for each patient from a group of endoscopic images captured during endoscopy, where only patient-level labels are available for each patient. To achieve the goal, we propose a multi-class token-based multiple instance learning method which consists of the feature extractor, the multi-class token selector module, and the aggregator. By using learnable positive class tokens and negative class tokens with transformer encoders, the multi-class token selector module selects the proper class tokens to improve the H. pylori infection prediction performance of the aggregator. Compared with supervised methods and state-of-the-art multiple instance learning methods, the proposed method achieves the best results.
Single-cell RNA sequencing (scRNA-seq) has profoundly reshaped our understanding of cellular diversity and functionality; however, accurate cell-type annotation is required for biological interpretation. Current annotation methods, which are predominantly reliant on gene expression alone or manual curation, suffer from subjectivity and a limited biological context. Here, we introduce a novel approach that integrates textual biological knowledge via gene embeddings, derived from fine-tuning Large Language Models (LLMs), with gene counts to enrich the input space for supervised models in automatic cell-type classification. In particular, we trained an XGBoost model and a multi-layer perceptron (MLP) to automatically classify the cell populations. We demonstrate that combining Modern-BERT embeddings and raw counts enhances the performance of MLPs, particularly in complex classification scenarios that involve subtle cell-subtype distinctions. Our results also show that ModernBERT generated better embeddings than smaller LLM architectures, underlining the value of enriched, biologically informed embeddings. By embedding prior knowledge from curated biological databases and literature, our approach enhances the MLP’s ability to distinguish sub-cell populations and biological signals. This work provides a scalable framework for integrating broader biological context into scRNA-seq analyses, offering new opportunities for downstream tasks such as gene regulatory network inference and cross-species annotation.
Viral titer measurement is vital in virology, offering insights into viral dynamics, disease severity, and treatment efficacy. This article underscores its importance across research domains and introduces BacCal (Baculovirus Calculator), a specialized software for streamlining viral titer experiments. BacCal automates titration using the Reed-Muench method, providing a user-friendly interface for inputting parameters and analyzing results. It enables precise quantification of viral concentrations, aiding in understanding virus-host interactions and replication kinetics. BacCal enhances quality control in virology and diagnostic assay production, ensuring consistency. Analysis involves preparing plates, inoculating cells, and assessing effects, with TCID50 calculation and optional PFU conversion. BacCal features localized data storage, preserving conditions for reproducibility. By integrating automated titration and robust data management, BacCal advances virology research, facilitating efficient experimentation and safeguarding critical data.
Optical Coherence Tomography (OCT) is a widely used non-invasive imaging technique that provides detailed three-dimensional views of the retina, which are essential for the early and accurate diagnosis of ocular diseases. Consequently, OCT image analysis and processing have emerged as key research areas in biomedical imaging. However, acquiring paired datasets of clean and real-world noisy OCT images for supervised denoising models remains a formidable challenge due to intrinsic speckle noise and practical constraints in clinical imaging environments.To address these issues, we propose SDPA++: A General Framework for Self-Supervised Denoising with Patch Aggregation. Our novel approach leverages only noisy OCT images by first generating pseudo-ground-truth images through self-fusion and self-supervised denoising. These refined images then serve as targets to train an ensemble of denoising models using a patch-based strategy that effectively enhances image clarity. Performance improvements are validated via metrics such as Contrast-to-Noise Ratio (CNR), Mean Square Ratio (MSR), Texture Preservation (TP), and Edge Preservation (EP) on the real-world dataset from the IEEE SPS Video and Image Processing Cup. Notably, the VIP Cup dataset contains only real-world noisy OCT images without clean references, highlighting our method’s potential for improving image quality and diagnostic outcomes in clinical practice.
Gastrointestinal cancer is one of the deadliest diseases all over the world. In medical imaging technology, cancer diagnosis has been evolving, especially endoscopes and currently applying artificial intelligence and deep learning on improving endoscopist’s imaging abilities. By using acquisition of images of tissues and organs, physicians are able to diagnose cancer in gastrointestinal track, besides of some limitations as depending on quality of endoscopy, experience of physicians or time consuming. With the development of technology, applications of deep learning are more effective in improving imaging techniques and diagnosis. Due to the state-of-the-art of deep learning, we applied VGG16, DenseNet201, HarDNet MSEG and HarDNet MSEG with attention to obtain some better results. Our results provide the comparison with mDice, mIoU and inference time on some of our trained models on the CVC-Clinic DB and Kvasir-SEG dataset. The best obtained mDice and mIoU are over 0.9 and 0.8, respectively.
Obstructive sleep apnea (OSA) remains significantly underdiagnosed due to limitations associated with traditional polysomnography (PSG), including complexity, high cost, and patient inconvenience. To overcome these challenges, this study introduces a novel event-based detection method using the YOLOv8 Nano object detection model applied to electrocardiogram (ECG) spectrograms. ECG features were generated via Continuous Wavelet Transform ( CWT) with the Morlet wavelet, enabling precise identification and temporal localization of individual apnea and hypopnea events within the signal. Model performance was rigorously validated using datasets comprising 50 subjects from clinical source: clinical ECG data collected at the National Cheng Kung University Hospital Sleep Center (NCKUHSC). Evaluation involved both Subject-Wise 5-Fold Cross- Validation and Leave-One-Subject-Out (LOSO) cross- validation methods to assess generalization. Experimental results demonstrated that the YOLOv8 Nano model achieved by providing accurate temporal granularity. Additionally, apnea-hypopnea index (AHI) estimation, derived from counting precisely localized events, strongly correlated with clinical ground-truth values (Pearson correlation coefficient R = 0.86, RMSE = 12.16 events/hour for detection on 135-sec ECG spectrogram segments), highlighting clinical accuracy and reliability. The compact model size (similar to 6 MB, 3.2M parameters) facilitates seamless integration with Home Sleep Apnea Tests (HSAT) and edge computing platforms, enabling accessible, cost-effective, and patient-friendly diagnostic solutions. Furthermore, the event-based nature of the results allows for extraction of richer, event-specific clinical parameters beyond AHI. Future research directions include enhancing feature extraction, improving model generalization with larger datasets, exploring these extended clinical parameters, and optimizing embedded computing implementations for real-world clinical use.
This study presents a comprehensive evaluation of vision-language models (VLMs) for skin lesion classification, comparing them with conventional convolutional neural networks (CNNs) and Vision Transformer (ViT) architectures. Using the HAM10000 dataset, we evaluated six models in four key dimensions: classification accuracy, robustness to visual perturbations, zero-shot generalization to unseen lesion types, and the quality of semantic explanations. Although traditional vision-only models achieve higher accuracy in clean images, their performance degrades significantly under various types of image distortion. In contrast, VLMs—particularly Qwen2.5—demonstrate stronger robustness and generate more coherent and clinically relevant explanations. However, they still fail to achieve overall classification performance and exhibit limited generalization in zero-shot settings. These findings highlight the trade-offs between task-specific accuracy and multimodal adaptability, offering practical insights into the current capabilities and limitations of VLMs in dermatology-focused artificial intelligence applications.
MicroRNAs (miRNAs) play crucial regulatory roles in cancer biology, but accurately quantifying their expression remains a significant unmet challenge in single-cell and spatial transcriptomics. Current sequencing technologies predominantly capture polyadenylated messenger RNAs (mRNAs), rendering them incapable of directly profiling miRNAs, which lack poly(A) tails. To address this gap, we thus propose miSAM, a novel computational framework for estimation of miRNA expressions based on a bi-objective combinatorial genetic algorithm in conjunction with support vector regression. The miSAM jointly optimizes the selection of a minimal subset of mRNAs, called signatures, while maximizing the Spearman correlation coefficient (SCC) between inferred and actual miRNA expression levels. Evaluation on The Cancer Genome Atlas (TCGA)-BRCA dataset, comprising 1,095 breast cancer samples with expression profiles of 1,881 miRNAs and 19,937 mRNAs, demonstrates the effectiveness of miSAM. Using the top five prognostic miRNA biomarkers for breast cancer, miSAM achieved a mean SCC of 0.633 on the test set while utilizing only 32.6 mRNAs in average, significantly outperforming baseline approaches including: (1) differential expression filtering-based SVR using 887 mRNAs (SCC = 0.562), and (2) LASSO-based SVR using 147.4 mRNAs (SCC = 0.470). miSAM also outperformed XGBoost, which yielded a SCC of approximately 0.45-0.50 across various cancer types. Furthermore, the identified mRNAs signatures offer explainable insights into the regulatory associations between miRNAs and their corresponding mRNA targets. These results underscore miSAM's potential as a robust, interpretable, and scalable tool for miRNA inference in spatial transcriptomics, single-cell sequencing analyses, and precision oncology.
RNA inverse design is an essential part of many synthetic biology applications. The current machine learning approaches can effectively predict sequence of the RNA from its 3D atomic-scale information and even generate hypothetical sequences predicted to have similar structural behavior to a desired template. Natural riboswitches are RNA molecules capable of switching their conformation upon sensing an environmental signal and are great candidates for synthetic RNA devices. Design of a novel riboswitch, however, is a formidable challenge, due to inherent plasticity of certain RNA structures and also given that there is little structural similarity among the already discovered riboswitch classes making it difficult to combine structural information. In this work, we explore the capabilities of the GNN-based structure-informed RNA inverse design tool, to inverse design the Thiamine Pyrophosphate (TPP) riboswitch from its backbone geometry. Results show that sequence recovery of TPP riboswitches is higher for models containing solo RNA structures compared to those containing other types of riboswitches as well as those containing RNA-Protein complexes. Design of a bound state required further modeling of ligand binding as well as impact of metal ions on structure.
