Predicting seizure outcome is essential for tailoring epilepsy treatment. However, accurate prediction remains challenging with traditional methods, particularly in diverse patient populations. This study presents a graph-based deep learning framework for predicting seizure outcomes using stereo-electroencephalography (sEEG) data in pediatric patients with drug-resistant epilepsy and deep thalamic involvement. We analyzed 105 ictal events from sEEG recordings of 10 pediatric patients with documented thalamic seizure networks and evaluated our model in three different cross-validation strategies: seizurewise, windowed segmentation, and patient-wise analysis. Our graph neural network-based model represents each sEEG channel as a node with power spectral density features, while edges capture inter-channel correlations. The windowed segmentation approach, which divides seizure recordings into non-overlapping 2-second temporal windows, demonstrated superior performance across all metrics. This data augmentation technique achieved 93.9% accuracy, significantly outperforming both seizure-wise (82.%) and patient-wise (77.0%) approaches using complete seizure recordings. Network analysis revealed distinct thalamo-cortical connectivity patterns with denser network topology in sample patient with poor outcomes (<50% seizure reduction) as compared to sample patient with favorable outcomes (>50% seizure reduction). These findings demonstrate the potential of connectivity-based deep learning models for enhancing seizure outcome prediction in pediatric epilepsy, particularly in cases involving complex thalamo-cortical networks. This framework advances our understanding of thalamic seizure propagation and offers promise for AI-assisted personalized epilepsy treatment planning.
Protecting healthcare data from inference attacks, where adversaries deduce sensitive information from de-identified data, is critical. This study examines the vulnerability of such datasets, focusing on Tennessee facilities serving predominantly African American populations, while also incorporating analyses based on the MIMIC-III dataset representing Massachusetts. We apply differential privacy with varying epsilon values to assess its impact on statistical integrity and predictive model accuracy. Results show a clear trade-off: lower epsilon enhances privacy but degrades performance, while higher epsilon preserves utility at the cost of increased leakage risk. These findings underscore the importance of carefully balancing privacy and utility when allocating the privacy budget in clinical prediction tasks.
Simultaneous electrocardiography (ECG) and phonocardiogram (PCG) offer a multimodal view of cardiac function by capturing electrical and mechanical activity, respectively. However, their shared and unique information and potential for mutual reconstruction remain poorly understood across different physiological states and individuals.This study analyzes the EPHNOGRAM dataset of simultaneous ECG-PCG recordings during rest and exercise, using linear and nonlinear models—including a non-causal neural network—to reconstruct one modality from the other. Nonlinear models, especially non-causal neural network, outperform others, with ECG reconstruction from PCG proving more feasible. In the within-subject scenario, the non-causal neural network achieved a signal-to-noise ratio (SNR) of 6.5±5.2 dB and a cross-correlation of 0.78 ± 0.19 for PCG-based ECG reconstruction.These findings provide quantitative insight into the electromechanical relationship between cardiac signals and support the development of multimodal cardiac monitoring tools.
Lung adenocarcinoma (LUAD) represents a major global health challenge requiring more accessible and noninvasive screening methods. Traditional diagnostic approaches such as computed tomography or biopsies are effective but costly, resource-intensive, and carry associated risks. This study leverages gut microbiome data and machine learning techniques to develop a non-invasive pre-screening tool for LUAD. Using a dataset of 107 fecal samples (43 LUAD and 64 healthy controls), we explored the performance of nine machine learning algorithms and four distinct feature sets generated through feature selection methods to identify informative microbial biomarkers and construct accurate classification models. Our results demonstrate that feature selection significantly enhances model performance compared to baseline approaches. A Random Forest model combined with Correlation-based Feature Selection achieved an Area Under the Curve of 0.9967. Key taxa including Prevotella, Coprococcus, Phascolarctobacterium, Bilophila, Blautia, Enterococcus, and Bacteroides emerged as potential biomarkers. Functional predictions using PICRUSt2 revealed significant alterations in folate metabolism, methylation cycles, and photosynthetic bacterial activity, highlighting disrupted gut microbiome function in LUAD patients. These findings align with previous studies and suggest promising directions for non-invasive and cost-effective screening methods.
Diabetic foot ulcers, a life-threatening complication of diabetes, take a disproportionate toll on communities of color; however, these communities are currently underrepresented in dermatologic and wound image datasets. Further, many of these datasets were collected under controlled conditions, limiting the transferability of ulcer recognition models to naturalistic settings. In support of more equitable and generalizable computational modeling, we detail our two-year effort to create the first repository of diabetic foot ulcer images collected predominantly from patients of color in naturalistic settings. We conduct an evaluation of state-of-the-art foot ulcer segmentation and classification methods using our dataset of 3,362 foot images collected from 252 patients, and provide evidence that current ulcer recognition models result in insufficient performance: the best performing baseline model (Mask R-CNN) has been previously reported to achieve a Dice score of 90.2%, but achieves only 39.5% on our more naturalistic dataset from patients of color. We propose and evaluate a new pipeline which improves segmentation performance, including an ulcer detection model and a foundational segmentation model (Segment Anything 2 UNet) tailored to communities of color and specifically aiming for naturalistic assessment scenarios. We release our image dataset to support the development of larger, more diverse datasets, and ultimately more equitable models for diabetic foot care.
The increasing incidence of drug resistance and the spread of fungal diseases underscore the urgent need to investigate resistance mechanisms in Valley fever, a fungal infection caused by Coccidioides spp. that has increased sharply in recent years and mirrors broader antifungal resistance trends. ATP-binding cassette (ABC) transporters, which are shown to efflux drugs in well-studied fungi, remain structurally uncharacterized in Coccidioides immitis. This study proposes the first structure-guided framework for systematic binding pocket assessment and inhibitor testing across five C. immitis ABC transporters. High-confidence protein structures (mean pLDDT > 95) were predicted using AlphaFold2, and predicted pockets were identified using PrankWeb. AutoDock Vina docked five chemically diverse ligands, generating 520 protein-ligand-pocket complexes. Static filtering (based on docking scores and pocket probability) and reference protein alignment created a shortlist of 26 complexes, which underwent short explicit-solvent MD simulations to assess binding persistence. Ligand center-of-mass drift was used to evaluate binding retention, and 17 pockets with minimal drift were given the initial classification as stable within the 2 ns MD window. Extended 20 ns simulations on a representative subset confirmed that early-screened stable pockets generally persisted, validating the use of 2 ns MD as a pocket prioritization strategy. This work provides the first structural dynamics dataset for C. immitis ABC transporters, identifies promising binding pockets, and highlights milbemycin oxime as a consistent binder. The presented framework enables early-stage screening and filtering for pocket prioritization in fungal resistance-mediating ABC transporters, supporting precision antifungal development through a structure-guided analysis of transporter pockets.
Early recognition of clinical deterioration is crucial for timely intervention, especially during Emergency Medical Services (EMS) encounters. Early Warning Scores (EWS) translate raw vital signs into a clinically transparent risk scale. However, research on prehospital EWS applications is limited and often focuses on in-hospital outcomes and single snapshots, neglecting short-term risk trajectories. This paper explores whether EWS trends, captured just before initial EMS intervention, convey additional information and can predict the return of spontaneous circulation (ROSC) during out-of-hospital cardiac arrest encounters. In a retrospective study of 4,394 cardiac arrest encounters from the 2021-2023 National EMS Information System (NEMSIS), we applied eight different EWS models at every documented vital sign measurement and derived time-normalized preintervention features, including slope, mean, area under the EWS curve (AUC), and exponentially weighted average (EWA). Informational value was quantified with nonparametric tests and L1-regularized logistic regression models targeting prehospital ROSC. Our findings demonstrate that short-term EWS dynamics encode measurable patterns of clinical deterioration, achieving moderate predictive performance (AUROC: 0.665) and advancing the current understanding of prehospital risk assessment. These results highlight the potential of incorporating vital sign trajectories into real-time, data-driven decision-support tools for EMS and motivate further exploration with more flexible, AI-based modeling approaches.
Noninvasive neural recording methods like electroencephalography (EEG) offer high temporal resolution for capturing neural activity. However, interpreting EEG data is challenging scalp-recorded signals (sensor space) reflect complex, integrated activity from multiple cortical regions (source space), complicating the reconstruction of underlying neural dynamics. Traditional approaches like minimum norm estimation require extensive subject-specific data, including MRI scans, precise electrode placement, and detailed anatomical atlases. To address these limitations, we propose a two-part framework: (1) an unsupervised biLSTM autoencoder that reveals clustering patterns in EEG electrode activations and their temporal dynamics during auditory stimulus processing; and (2) a deep learning architecture to predict temporally evoked neural features in sensor space EEG from source representations using a dual-path network with independent stimulus processing and dilated convolutional layers.The clustering identifies evolving spatiotemporal co-activation patterns between stimulus onset and gaps, revealing functional reorganization. The reconstruction network reduces input dimensionality and integrates features via convolutional blocks with residual connections, trained using a hybrid loss that combines feature-based and spectral terms. Our results demonstrate accurate reconstruction of stimulus-related neural correlates and reveal topographical patterns consistent with the clustering findings. The model generalizes well across subjects. By analyzing both functional organization in sensor signals and source-to-sensor mappings, our framework enhances understanding of EEG transformations. This has significant implications for brain-computer interfaces, neuroimaging, and EEG processing where accurate reconstruction and interpretation are essential.
Detecting stroke risk remains a major challenge in preventive medicine. In this work, we introduce a novel computational approach for modeling the effect of aging to identify patients at risk of stroke by analyzing the intricate relationship between brain and heart dynamics during sleep. We analyzed whole-night Polysomnography (PSG) data focusing on sleep stage transitions, to capture changes in cortical and autonomic functions. Using an attention-based model tuned for age estimation, we identify patients at risk of stroke. The model has been tested on 782 patients and a systematic ablation study was performed to evaluate predictive performance across different signal modality configurations and sleep stages.Results from this study indicate that the patients at risk of stroke show pronounced aging effects, suggesting that Brain-Heart Interaction (BHI) during sleep may be applied on a population level as a novel biomarker to identify patients at risk of stroke.
Lung imaging lacks a standardized reference space, hindering the large-scale, voxel-wise analyses that are routine in neuroimaging. To address this gap, we developed a high-resolution, open-source 3-D lung template and probabilistic lobar atlas from a cohort of 30 subjects from the National Lung Screening Trial (NLST). Created using a fully automatic pipeline based on the Advanced Normalization Tools (ANTs) ecosystem, this template reached convergence (dice similarity coefficient of 0.992 between consecutive iterations) after 11 iterations. We demonstrated its utility by registering 60 subjects with varying emphysema severity, finding that voxel-wise Jacobian analysis could distinguish disease-specific deformation patterns. This work provides a foundational, open resource for standardizing anatomical localization, enabling robust group-level studies in lung cancer screening research.
Automated histopathological subtyping of lung cancer from stained whole-slide images (WSIs) remains a pivotal yet challenging task due to pronounced tumor heterogeneity, complex cellular morphology, and severe class imbalance in existing datasets. Deep learning models vary in their ability to capture pathomic diversity, and their diagnostic performance is closely tied to the quality of tissue patches extracted from WSIs. To address these challenges, we propose a novel ensemble deep learning framework augmented with fuzzy-weighted patch quality assessment to optimize the selection and weighting of informative regions. High-quality patches are identified using a fuzzy scoring mechanism and processed through multiple pre-trained convolutional neural networks (CNNs) and vision transformer (ViT) models to extract diverse feature representations. These are integrated via latent embeddings, with fuzzy scores incorporated both as auxiliary inputs and as weights in the loss function, reinforcing attention to clinically relevant regions. Our method outperformed current state-of-the-art models by 1.5% and 1.4% (CI: 95%), achieving accuracies of 96.1% on BMIRDS-LUAD and 93.0% on WSSS4LUAD, demonstrating enhanced robustness in subtype classification and strong potential for clinical integration.
Caregivers of people living with dementia (PwD) are highly susceptible to depression due to the substantial care burden they experience. While caregivers often neglect their own mental health and rarely seek necessary medical services, they often communicate their perceived burdens and depressive symptoms to social workers, who serve as critical points of contact for their loved ones. Thus, accurately estimating the risk of depression and care burden through these conversational interactions may facilitate early screening and intervention. This feasibility study explored the effectiveness of using caregivers’ demographic information and their narrative descriptions of caregiving experiences to estimate depression risk and caregiver burden. Utilizing Natural Language Processing (NLP) and machine learning (ML) techniques, we trained estimation models based on data from 65 caregivers, using clinical screening measures—the Patient Health Questionnaire-8 (PHQ-8) and Zarit Burden Interview (ZBI)—as reference standards. The best-performing models achieved F1 Scores of 0.74 and 0.78 for depression and burden estimation, respectively. These results demonstrate the potential of leveraging demographic and conversational data to enable early identification of caregiver distress, facilitating timely interventions that could ultimately enhance caregiver well-being and improve the quality of care provided to PwD.
Managing diabetes requires careful monitoring of food intake, yet manual logging is burdensome and error-prone. Prior research has shown that the macronutrient composition of a meal (e.g., carbohydrates, protein, fat, and fiber) can be inferred from its postprandial glucose response (PPGR). However, this is a challenging problem given the large inter-individual differences in PPGRs, and the complex interaction between macronutrients in mixed meals. To address these issues, we propose RankPPGR, a rank-learning framework that analyzes within-subject differences in pairwise PPGRs from meals with varying macronutrient composition, and learns a non-linear embedding of meal macronutrients that reflects their joint impact to glycemic responses. We also propose a few-shot regression module that uses outputs from RankPPGR to infer macronutrient composition using a limited number of labeled meals per individual. We evaluate the model on an experimental dataset containing PPGRs to mixed meals from 45 participants. RankPPGR significantly improves both pairwise classification and macronutrient inference performance over a sample-based regression baseline.
Answering complex medical questions requires both reliable information retrieval and the ability to generate responses that are medically accurate and contextually appropriate. In this paper, we present HemaRAG, a Retrieval-Augmented Generation (RAG) system designed specifically for hematologic malignancies. Our system combines a dense retriever enhanced with biomedical ontologies and a fine-tuned large language model (Gemma 3), trained locally on domain-specific literature and question-answer pairs. To build a robust retrieval base, we enriched PubMed abstracts and curated datasets such as BioASQ and PubMedQA using synonym mappings from MeSH, NCIT, DOID, and UMLS. We used a local vector database to support high-speed semantic search without sharing data externally. Evaluation across both BioASQ and long-form PubMedQA benchmarks showed high semantic accuracy (BERTScore: 87-89%), strong lexical overlap (F1: 49-52%), and high retrieval performance (Recall@10: 94-96%), despite the challenges posed by free-form medical questions. The system was developed and deployed entirely locally making it suitable for clinical contexts where patient data privacy is essential. In future work, we plan to integrate HemaRAG into an empathetic conversational agent designed to support patients and clinicians in the field of hematologic oncology.
Gene regulatory networks (GRNs) orchestrate cell fate decisions, yet conventional transcriptomic analyses often overlook subtle but critical structural disruptions in bioinformatics. We present a spectral framework that reveals local GRN collapse after GATA1 knockout, a key transcription factor in erythroid and eosinophil differentiation. Using Laplacian-based spectral descriptors, we detect a marked collapse in a granulocyte subpopulation, despite minimal global transcriptomic change. This collapse is characterized by low-frequency eigenvalue accumulation, reduced connectivity, and high localized instability. Our findings suggest that GATA1 maintains hidden regulatory attractors in hematopoietic GRNs, and their loss causes lineage-specific structural failure. This is the first application of graph spectral theory to capture cell-type-specific GRN fragility in single-cell perturbation data, offering a theoretical framework for evaluating transcription factor function and cell identity resilience.
Function is increasingly recognized as an important indicator of whole-person health, although it receives little attention in clinical natural language processing research. We introduce the first public annotated dataset specifically on the Mobility domain of the International Classification of Functioning, Disability and Health (ICF), aiming to facilitate automatic extraction and analysis of functioning information from free-text clinical notes. We utilize the National NLP Clinical Challenges (n2c2) research dataset to construct a pool of candidate sentences using keyword expansion. Our active learning approach, using query-by-committee sampling weighted by density representativeness, selects informative sentences for human annotation. We train BERT and CRF models, and use predictions from these models to guide the selection of new sentences for subsequent annotation iterations. Our final dataset consists of 4,265 sentences with a total of 11,784 entities. The inter-annotator agreement (IAA), averaged over all entity types, is 0.72 for exact matching and 0.91 for partial matching. We train and evaluate common BERT models and state-of-the-art Nested NER models. The best F1 scores are 0.83 for Action, 0.69 for Mobility, 0.60 for Assistance, and 0.67 for Quantification. Empirical results demonstrate promising potential of NER models to accurately extract mobility functioning information from clinical text. The public availability of our annotated dataset will facilitate further research to comprehensively capture functioning information in electronic health records (EHRs).
Pulse oximeters are essential in neonatal care for monitoring blood oxygen saturation, however their accuracy can be affected by skin pigmentation. The discrepancy between arterial oxygen saturation (SaO2) and saturation measured by pulse oximeters (SpO2) is more pronounced for darker skin tones, increasing the risk of occult hypoxemia. This study introduces a personalized machine learning approach aimed at reducing measurement bias by integrating objective, non-invasive skin pigmentation metrics alongside individual physiological parameters. Using the OpenOximetry Repository, several feature sets were constructed to compare the performance of various machine learning models. XGBoost achieved the lowest root mean square error and was selected for further analysis. The model demonstrated improved SpO2 accuracy, resulting in corrected values which are more closely aligned with actual SaO2 values across a range of skin pigmentation levels. These results support the potential of personalized models to improve measurement accuracy and reduce disparities in clinical monitoring.
Obstructive sleep apnea and hypopnea syndrome (OSAHS) is a significantly underdiagnosed condition that can lead to dangerous and sometimes life-threatening complications such as heart failure, stroke, and sudden cardiac death. Traditional diagnostic methods for OSAHS, such as polysomnography, are resource-intensive and not readily accessible for large-scale screening. In this study, we compared the efficacy of machine learning (ML) algorithms using non-invasive physiological data-pulse oximetry and heart rate variability, which can be recorded using wearable sensors, to detect OSAHS in a large dataset consisting of 6399 recordings (53% women and mean age 62 +/- 13 years). The ML algorithms were trained and tuned using nested cross-validation on a subset of the dataset (training set, 80% of the dataset) and separately validated on the independent test set (20% of the dataset) to showcase the generalizability of our model performance. Furthermore, we investigated the performance of ML algorithms with respect to the sampling frequency, available data length, and presence of noise in physiological signals to understand the impact of real-world constraints on OSAHS detection. We also explored the model explainability with SHapley Additive exPlanations (SHAP) and an ablation study to enhance the clinical interpretation of the results. Our comparative analysis of ML algorithms (Random Forest, Support Vector Machine, eXtreme Gradient Boosting, Multi-Layer Perceptron, etc.) demonstrated the best performance for eXtreme Gradient Boosting algorithms with an F1-score of 0.896 +/- 0.012 and 0.897 on the cross-validated training set and independently validated test set, respectively. The algorithm's performance deteriorated with reduced data availability in the independent test set, with an F1-score of 0.897, 0.89, 0.887, 0.885, and 0.879 using physiological data with eight (full-night), four, two, one-hour, and 30-minute recording lengths, respectively. Algorithm performance was highest in models using pulse oximetry data with a 0.5 Hz sampling rate compared to 1 and 0.25 Hz sampling rates. The findings highlight the potential of various ML-driven analyses of unobtrusive physiological signals for scalable OSAHS screening and consideration of real-world constraints on the ML algorithm performance.
Despite significant advancements in brain-computer interface (BCI) technology, systems capable of leveraging physiological signals to detect and recognize human intentions in real-time are still underdeveloped. To achieve a new level of human-machine interaction, it is essential to integrate motor activity correlates with state-of-the-art artificial intelligence (AI) architectures. In this study, we present the first demonstration of handwriting decoding - a complex motor task - using a novel myographic method called Optomyography (OMG). Unlike previous electromyography (EMG)-based approaches that treat handwriting decoding as a classification problem, we frame it as a continuous trajectory reconstruction challenge. We evaluated GRUScribe (GRU-based decoder) and TransScribe (transformer-based decoder), successfully decoding 10 numerical digits and 33 Russian letters from 20 able-bodied and 4 amputee participants, without requiring elaborate preprocessing. Our results demonstrate the remarkable potential of OMG for recognizing complex motor activity. We believe that our work sets a new benchmark in non-invasive muscle activity decoding, offering direct applications in advanced prosthetic control and human-machine interfaces.
Sleep plays a crucial role in human well-being, while insufficient sleep affects cognitive function, decision-making, and overall health. Sleep assessment via polysomnography (PSG) is time-consuming, resource-intensive, and limited to in-laboratory sleep testing. To address the challenges of PSG, wearable sleep screening devices have been widely used, especially to detect wakefulness and sleep stages. This study proposes deep models for the detection of wakefulness versus different stages of sleep using heart rate and wrist actigraphy extracted from the multi-ethnic study of atherosclerosis (MESA) sleep dataset. First, two sets of features were extracted from heart rate and actigraphy, which were separately fed into two separate branches of convolution neural network (CNN), then merged and fed to a deep classifier. The model detected wakefulness versus sleep and different sleep stages with the accuracies of 88.19% and 79.6% respectively. This work showed that combining heart rate, actigraphy signals, and demographic data in a deep framework could improve sleep stage-staging performance. This study offers a subject-specific approach for sleep assessment based on convenient wearables.
