Seizures pose an imperative risk due to their frequency rising after the COVID-19 pandemic, with between 25-30% of patients going misdiagnosed or undiagnosed for seizure related conditions such as epilepsy. Electroencephalogram data can determine the location and timing of seizures to prevent future ones from occurring, but needs to be filtered to reduce interference and noise from biosignals. Current hospital research focuses on implementing common filters such as the bandpass filter to filter EEG data, resulting in highly distorted signals that create inaccurate diagnosis. The current research gap is the lack of an effective filter in hospitals to identify seizures and the absence of coding for a filter to be implemented in wearable seizure devices such as IoT devices. The objective of this research is to evaluate the optimal digital filtering technique for isolating brain wave frequencies while minimizing signal distortion and construct code utilizing the most optimal filter to build a usable EEG system for IoT seizure detection devices and hospitals. It was hypothesized that the Butterworth Bandpass filter will be the most effective at filtering signals because it is designed to have a flat response and doesn’t distort frequencies that are needed while having a steep roll-off to remove unnecessary frequencies that can cause misinterpretation of the data. This research was conducted on an anonymous pediatric EEG dataset from the Children’s Hospital in Boston, with the dataset containing 22 subjects, each contributing 9-42 files, with 23 EEG signals per file recorded during a seizure. All signals were sampled at 256 samples per second with 16-bit resolution to capture the detailed data. Specific signals were identified and tested in R using the commonly used Butterworth bandpass filer, the moving average filter, and the zero-phase filter. The Butterworth bandpass filter was tested in three orders (2, 4, and 6) to determine the effects of different orders on highlighting specific seizure episodes. The moving average filter was compared with the Butterworth bandpass filter and calculated with a window size of 5 data samples. The zero-phase filter was used in comparison with conventional filters utilized in hospitals to determine its impacts on time preservation relative to current usage. The results were quantified using noise removal comparison via power spectral density and signal-to-noise ratio. Power spectral density was calculated to translate the data from the time-domain to the frequency-domain. Signal power was defined to be between 1 and 40 Hz and noise power was considered to be over 40 Hz. Order 4 Butterworth bandpass filters were found to have a SNR of 25.5 dB, moving average had an SNR of 27.37 dB, and zero-phase filtering had an SNR of 79.06 dB, over 3 times greater than the SNR for the Butterworth bandpass filter. The research concludes by providing novel data that the most commonly used Butterworth filter is ineffective in seizure patients and finds that Moving Average and Zero-Phase filters should be used. The research also constructs code that can directly be implemented in wearable IoT devices using the Moving Average filter to utilize a low-energy filter that can analyze data on command. Moreover, the zero-phase filter can be used in diagnosis of seizures and to analyze episodes in patients, and the research constructs an accessible website utilizing the filter that can be rolled out in hospitals. Doctors can upload the data, adjust which frequencies they consider to be noise, and generate a resulting graph to analyze results in patients.
Electronic cigarettes, i.e., vaping, remain a major public-health concern, and Reddit posts provide naturalistic information about vaping cessation. In this study, we evaluated whether a set of refined prompting strategies and a multi-agent consensus workflow improve automated behavioral stage classification of vaping cessation posts on Reddit. First, clinicians annotated sample Reddit posts using the original five Transtheoretical Model (TTM) stages including (1) precontemplation, (2) contemplation, (3) preparation, (4) action, and (5) maintenance, plus what they added a new stage, namely "no stage" for this study. Ultimately, the first three stages were deterministically collapsed into pre-Action while Action, Maintenance, and No Stage were retained. We then evaluated six large language models (LLMs) including GPT-5, GPT-5.1, OpenAI o1, DeepSeek R1, DeepSeek V3, and LLaMA-4 Maverick, across five prompting strategies, consisting of a basic prompt, a basic prompt with TTM definitions, a basic prompt with a clinical annotator role, a structured prompt with role, definitions, and few-shot examples, and a chain-of-thought prompt, as well as a four-agent workflow, i.e., extractor, classifier, critic, and judge. Using 1,184 vaping-related Reddit posts, the best configuration is DeepSeek R1 with agentic adjudication, achieving macro-F1 = 0.862 and weighted F1 = 0.875. Structured prompts and agentic adjudication outperformed minimal prompts, showing that simplified TTM staging and clearer decision rules can support large-scale annotation of vaping cessation discourse.
The adoption of artificial intelligence (AI) in hospitals requires systematic assessment of institutional readiness. This study presents a domain-specific, readiness-oriented AI maturity model developed using a Design Science Research methodology. The model comprises six conceptual dimensions, namely Technology, Data, Strategy, People, Organization and Regulations. These are linked to 36 critical success factors and operationalized through 107 measurable items. All components are embedded in a web-based artefact that enables structured self-assessment, quantitative scoring and rule-based generation of improvement recommendations. The tool produces dimension-specific maturity profiles and identifies targeted areas for advancement based on Likert-scale responses. By translating abstract constructs into a practical evaluation instrument, this research offers a transparent and context-sensitive framework for guiding strategic AI development in hospitals. The artefact supports organizational reflection, benchmarking and capability building, contributing to the structured advancement of AI readiness in healthcare environments.
Drug development is notoriously costly and slow, with most candidates failing before reaching patients. Drug repurposing is the discovery of new therapeutic uses for existing compounds. It offers a faster and lower risk alternative by leveraging known safety profiles. Computational approaches, particularly machine learning, are increasingly applied to drug repurposing to accelerate candidate discovery. However, many machine learning approaches still lack transparency, limiting their adoption in translational research.We present RePurposeX, a mixed-initiative decision support system that integrates linked evidence, expert feedback mechanisms, and intuitive visual modules to support usability and trust in repurposing workflows. In a think-aloud study with 20 participants, RePurposeX achieved strong System Usability Scale (SUS = 84.3) and enabled diverse expert strategies, including visual-first, model validation, and clinical-priority approaches. Our findings yield design implications for health informatics, demonstrating how visual-first evidence, override mechanisms, and linked knowledge graphs improve usability and trust in AI predictions.
Recent advances in artificial intelligence have substantially improved histopathological image analysis. However, reliance on a single deep learning paradigm-such as convolutional neural networks (CNNs) or vision transformers (ViTs)-often constrains model robustness, amplifies correlated errors, and results in poorly calibrated confidence estimates. CNNs effectively capture fine-grained cellular morphology, while transformers encode global tissue context, yet neither alone adequately addresses the heterogeneity and uncertainty inherent in clinical pathology data. To address these limitations, we propose a reinforcement learning (RL)–guided adaptive fusion framework that integrates heterogeneous experts, including ConvNeXt, ResNet-50, ViT, and a hybrid classical–quantum model based on low-qubit parametrized quantum circuits (PQCs). Fusion is formulated as a confidence-aware decision-making process in which a policy network observes batch-level uncertainty statistics from each expert and learns adaptive fusion weights via policy-gradient optimization. This design enables dynamic weighting of expert predictions while directly optimizing a recall-sensitive clinical objective, thereby reducing false negatives in tumor detection. The PQC component is incorporated as a peer expert producing probabilistic predictions, contributing complementary quantum-derived decision patterns that enhance ensemble diversity rather than acting as a standalone classifier. Fusion is performed deterministically in probability space during inference to ensure stability and interpretability, while stochastic policy updates are confined to training. We evaluate the proposed framework on the BreaKHis breast cancer histopathology dataset (7,909 images from 82 patients across 40×, 100×, 200×, and 400× magnifications) using patient-level stratified splits. The fusion model achieves 99.4% accuracy, with macro-averaged precision, recall, and F1-score of 0.99, and malignant recall of 1.00 on the held-out test set. Ablation studies demonstrate that RL–based fusion performed better than the strongest individual backbone, with quantum-derived features providing higher performance.
Individuals with intellectual and developmental disabilities face a public health crisis that personal safety technology has struggled to address. Prior studies including findings cited by the Commonwealth of Massachusetts report that over 90% of individuals with Intellectual and Developmental Disabilities (I/DD) experience sexual abuse in their lifetime, with 49% experiencing ten or more abusive incidents [1]. Despite this, existing solutions were built around five neurotypical cognitive assumptions: that users can consciously recognize threats, make deliberate decisions to act, execute multi-step interface navigation, maintain fine motor control, and tolerate standard alert stimuli under acute stress. These assumptions may not hold true for I/DD users as interoceptive dysregulation produces delayed or absent conscious threat recognition. Further, Alexithymia, present in approximately 50% of autistic individuals within the I/DD spectrum [2], impairs the emotional self assessment required to initiate a safety response. Associated polyvagal dorsal vagal shutdown eliminates the executive function and motor control that app navigation requires. This paper terms this convergence the design-threat mismatch: a structural incompatibility between how existing safety applications are designed and how I/DD users actually respond to threats, which causes these applications to fail at exactly the moment and for the population that needs them most. Building on this analysis, the paper introduces SafetyCompanion a personal safety application currently in its Minimum Viable Product (MVP) stage, which works as a proposed architecture to directly bridge the design-threat mismatch found in existing safety SafetyCompanion’s core design inverts the standard activation model through a finite state machine escalation engine, incremental encrypted streaming to Firebase Cloud Storage, dead man’s switch logic that escalates first to trusted contacts via SMS and subsequently to emergency services if no check-in is received, pre-written communication scripts, and sensory-safe interface design. This work makes two original contributions to health informatics and assistive technology by introducing a design-threat mismatch and evaluates a novel evidence-informed architectural response to the documented and urgent public health crisis, grounded in the neurobiological realities of I/DD users under acute threat.
Objective Structured Clinical Examinations (OSCEs) are widely used to evaluate clinical and communication skills in healthcare trainees but require significant faculty time and institutional resources. This study investigates whether large language models (LLMs) can assist in evaluating OSCE encounters. Ten practice OSCE encounters were analyzed using diarized transcripts generated from extracted audio. Multiple LLMs were prompted with a structured rubric to evaluate checklist items and provide supporting evidence. Model outputs were compared against a consensus ground truth established by three blinded faculty reviewers. Results demonstrate strong agreement between several models and faculty evaluations, suggesting that LLM-based systems can provide scalable decision support for OSCE grading while reducing faculty workload.
Intensive care unit (ICU) vital signs and laboratory tests form irregularly sampled multivariate time series with severe missingness, where the observation process itself can be informative. We propose TRACE-D, a concept-wise representation learning framework that partitions temporal variables into physiologically motivated concepts and encodes each concept sequence using missingness-aware LSTM-D that explicitly leverages missingness masks and time-gap information. Using the first 48 hours of data from the PhysioNet/Computing in Cardiology Challenge 2012 dataset to predict prolonged ICU stay (length of stay > 3 days), TRACE-D achieves 82.55% AUROC with extreme gradient boosting and 98.20% AUPRC with random forest, outperforming TRACE and competitive baselines. Ablation studies show that injecting missingness masks into gate computations is a major contributor to the performance gains. Concept drop and Shapley Additive Explanations analyses identify clinically plausible concept-level associations (e.g., kidney function and electrolytes) and show that the relative contributions of observed values and missingness indicators vary over time. Sample-efficiency experiments further show that TRACE-D retains its advantage in low-data regimes. Overall, these results suggest that combining physiologically grounded concept decomposition with explicit modeling of missingness and observation intervals yields robust and interpretable prediction from highly sparse ICU time series.
Asthma self-management depends on timely recognition of environmental triggers, yet existing tools rely on subjective recall, regional air-quality indices, and siloed symptom diaries that capture neither personal indoor exposures nor their physiological context. This paper presents AllerAlert, a smartphone-based mHealth system whose contribution is architectural: a unified, privacy-preserving pipeline that synchronizes continuous indoor air-quality sensing (PM, VOCs, temperature, humidity, pressure) over BLE, outdoor meteorology, wearable physiology (HRV, HR, respiratory rate, SpO2, sleep, activity) from HealthKit and Health Connect, validated questionnaires (ACT, PHQ-9), medication logs, and voice-based symptom capture into a single high-resolution timeline. We describe the engineering decisions that make continuous multimodal sensing tractable on consumer devices, including foreground-service BLE promotion, incremental health-data backfill, Firestore write batching that reduces cloud cost by ≈98%, and a two-stage NER pipeline combining a sub-100ms on-device lexicon fallback with an asynchronous transformer pass. On top of this timeline we implement a DBSCAN-based exposure-clustering stage with symptom-to-cluster alignment as an exemplar analytics layer. Two controlled engineering bring-up protocols verify end-to-end pipeline function with 98.1% data completeness; these are not a clinical validation but a foundation for the free-living, multi-participant studies needed to evaluate automated, personalized trigger discovery.
Clinical Decision Support Systems (CDSSs) based on evidence-based clinical guidelines (GLs) enable real-time, consistent, and cost-effective medical decision-making. However, developing or modifying GLs remains a complex, expert-driven process. As a result, guidelines are often static, require local adaptation, and are not calibrated to institutional data or temporally complex patient trajectories, limiting their ability to generalize and evolve. We introduce TP-DRL, a framework that combines knowledge-based temporal abstraction, frequent temporal pattern mining, and conservative offline deep reinforcement learning to derive adaptive treatment policies directly from retrospective data. Applied to hypokalemia management using the MIMIC-IV ICU dataset, the pipeline transforms raw laboratory measurements, vital signs, and intervention sequences into temporally aware discrete dynamic states. A multi-horizon clinical reward function guides learning by jointly optimizing short-term biochemical correction, medium-term complication prevention, and long-term survival and discharge outcomes. The reward incorporates trajectory-efficiency penalties and treatment-balance constraints to discourage unnecessary interventions and looping behavior while ensuring safety. In patients clinically identified for potassium chloride treatment, use of TP-DRL policies would have potentially reduced mortality by 2%, increased discharge rates by 9.96%, and reduced the session duration by 28.1% while avoiding over-treatment. These results demonstrate that temporal-pattern-aware offline reinforcement learning can serve as a foundation for data-adaptive "living guidelines" that align with institutional practice, improve patient outcomes, and generalize to other complex clinical pathways.
When clinicians encounter patients with rare, complex, or difficult to diagnosis conditions, Artificial Intelligence (AI), specifically Large Language Models (LLMs) can be ‘consultant-like’ assistants at the point of care. LLMs can synthesize information from medical text which can assist clinicians in arriving at evidence-based final diagnoses. The Department of Biomedical Informatics at the University at Buffalo, Jacobs School of Medicine and Biomedical Sciences has built a domain-specific LLM upon an open-source 70B parameter Llama 3.2 model enhanced by retrieval augmented generation (RAG) to include formally represented semantic clinical knowledge into the LLM, named Semantic Clinical Artificial Intelligence (SCAI, pronounced, sky). This doctoral consortium extended abstract discusses the programmatic research project benchmarking the LLM for clinical use; beginning with usability testing to explore how clinicians interact with the interface and identify areas for improvement. Followed by comparative effectiveness analysis of diagnostic accuracy between the Native LLM (70B Llama model without RAG) and SCAI (70B Llama model with RAG). Future work will incorporate analysis of the next diagnostic steps and treatment recommendations output by the LLMs. Notably, bias testing will be conducted to determine how well the LLM can handle information related to the Social Determinants of Health (SDoH) and how this information impacts outputs across domains.
AI-based voice analysis shows promise as a noninvasive tool for disease diagnostics, yet existing systems often fail to generalize due to sex-related acoustic differences and scarce data for rare pathologies. We propose a clinically motivated two-stage framework that first captures sex-specific vocal representations using ResNet-50 on Mel spectrograms, followed by sex-conditioned disease classification. We address class imbalance through multi-scale resampling and time-warping augmentation. Evaluated on a merged dataset from four public repositories, our framework achieves state-of-the-art performance, attaining 97.2% accuracy and 94.5% MCC on average, consistently outperforming a single-stage baseline with an average MCC improvement of approximately 3%. Fairness evaluation shows reduced sex-based performance gaps, while feature representations reveal sex-specific patterns that provide clinical evidence supporting a hierarchical, sex-aware design. Together, these results demonstrate that sex-aware hierarchical modeling improves both classification accuracy and fairness in voice pathology detection, with potential applications in telemedicine and clinical decision support.
Biomedical informaticians have made extensive contributions for knowledge representation of health information and data. Of these contributions are biomedical ontologies which spawned a variety of tools and methods to engineer these artifacts. Evaluation of biomedical ontologies to verify and validate them remains to be an unrealized area. Extending our previous endeavors to address this area, we provide a downloadable software application to provide an accessible option to evaluate and analyze newly developed ontologies. Ontokeeper-Desktop utilizes a semiotic theoretical framework to measure the intrinsic aspects of an ontology. The tool was developed with open-sourced Java libraries and currently available on GitHub. Future direction will lead us toward a usability and user experience study and using the tool’s features for comprehensive studies on biomedical ontology archives.
This study benchmarks the performance of four large language models (Llama3.1-8B, Llama3.1-70B, Gemma3-27B, and MedGemma-27B) for the extraction of social determinants of health (SDoH) from electronic health records. We evaluated model performance on clinical notes from non-small cell lung cancer patients treated at Emory University’s Winship Cancer Institute, including a triple-annotated subset to enable comparison against individual human annotators. MedGemma-27B achieved the highest micro-F1 score for the extraction of any SDoH mention (0.84 in a triple-annotated dataset, 0.79 in our full annotated Emory corpus) and approached human-level performance across multiple SDoH categories when compared to individual human annotators. Performance varied by SDoH variable, with rarer variables such as parental status and education level proving the most challenging. The models were better able to identify mentions of SDoH than to determine whether or not they were adverse, but their high recall relative to human annotators suggests their potential utility in human-in-the-loop systems for identifying and adjudicating SDoH. These findings demonstrate that domain-adapted LLMs can perform scalable, near-human-level SDoH annotation on clinical text without task-specific fine-tuning, supporting downstream clinical modeling and health equity research. Future work will explore ensemble and lexicon-augmented approaches to improve performance for low-prevalence and semantically complex SDoH categories.
Non-invasive hemoglobin (Hb) prediction from smartphone-captured fingertip photoplethysmography (PPG) signals possesses transformative promise for point-of-care anemia screening. However, the fairness and generalizability of regression models trained for the task can be severely limited by small, class-imbalanced clinical datasets and demographic representational bias. Our previous research indicated that vanilla GANs amplified bias across most regression models, whereas conditional GANs (cGAN) improved class balance but ignored demographic factors and privacy risks. We introduce FedHemo, a federated CTAB-GAN+ framework that tackles these three challenges in a single pipeline. Jensen-Shannon Divergence (JSD) is a widely used statistical measure for quantifying the disparity between probability distributions. The lower JSD values indicate that the synthetic model generates more representative samples. Pearson Pairwise Correlation (PPC) was utilized to assess feature dependency. Subsequently, our proposed system’s higher PPC values reflect stronger preservation of feature relationships, ensuring that FedHemo accurately captures the complex physiological inter-dependencies inherent in PPG signals. Furthermore, FedHemo consistently surpassed vanilla GAN and cGAN across every fidelity metric. It most notably reached a mean JSD of 0.267, which is more than doubling the distributional accuracy of the cGAN (0.682), while maintaining a superior PPC of 0.654. Downstream regression results show FedHemo-augmented training achieves a best MAE of 0.912 g/dL with SVR (RBF), an improvement over the 0.92 g/dL baseline from our prior work on the same model. The privacy of the training data is further confirmed by a Distance to Closest Record (DCR) of 0.193. We also provide a three-panel SHapley Additive exPlanations (SHAP)-driven visualization and analytics dashboard to facilitate health data analysis.
Discharge summaries are lengthy medical documents that summarize a hospital in-patient visit. Automatically generating them can reduce documentation burden and return clinician time to patient care. Whereas Large Language Model (LLMs) could be used for this task, their Achilles heel is hallucinations, which can have drastic consequences for clinical documentation. We present an evidence-driven alignment framework for discharge summarization at the clinical encounter level, that treats provenance as a first-class constraint, using semantic graphs and deep learning models. Each summary sentence is selected and organized via cross-document semantic alignment and is accompanied by explicit evidence links to its source spans. We show our results on two corpora: a publicly available corpus (MIMIC-III) and clinical notes written by physicians at the University of Illinois Hospital (UIC Health). Additionally, we make source code and trained models available.
Gait is a sensitive marker of mobility impairment and fall risk, motivating unobtrusive sensing for continuous monitoring beyond specialized laboratories. This pilot feasibility study presents a wearable, privacy-preserving 60 GHz mmWave radar system with an embedded Raspberry Pi workflow for remote gait measurement. Fixed corridor and shoe-mounted configurations were evaluated in healthy participants using simulated Parkinsonian-like gait patterns as a feasibility step. Step events were extracted from radar-derived velocity-time signatures to estimate step count, cadence, step time, and variability. The shoe-mounted configuration improved step-count agreement under irregular gait conditions while enabling untethered acquisition, real-time operator feedback, and clinician-ready data packaging.
This descriptive feasibility study applied Work Domain Analysis (WDA) to examine ICU medication administration. We integrated medication-safety policy documents with interview data from three ICU nurses to develop an abstraction hierarchy. The analysis represents medication administration as a sociotechnical work domain and illustrates how adaptive strategies may emerge under system constraints and competing demands. These preliminary findings suggest that WDA is useful for contextualizing adaptive nursing behaviors as responses to domain constraints. The approach may support future work on risk-relevant adaptations and system-level intervention targets.
Hospital discharge quality directly affects post-discharge quality of life through missed follow-up, medication confusion, unsafe recommendations, symptom burden, and avoidable readmissions. We studied whether a six-step compliance-aware discharge agent can make those failure modes more visible during intensive care unit (ICU) discharge planning by emitting auditable self-checks alongside generated action items. Using structured records from the eICU Collaborative Research Database Demo, we identified 1,183 eligible ICU stays, located 50 saved paired outputs, and evaluated the 14 cases in which both the agent and a single-prompt baseline produced parseable, non-empty discharge plans. On this successful-run pilot subset, the baseline was numerically better on completeness (mean F1 0.404 vs. 0.362) and missing follow-up rate (0.560 vs. 0.648), while calibration error was similar (0.428 vs. 0.418). The agent’s distinct contribution was explicit self-audit output: it surfaced 29 unsupported, 4 contradictory, and 10 potentially unsafe recommendations, together with 31 missing follow-up items, none of which were represented explicitly by the baseline. The multi-step agent is therefore preferable only in audit-focused, human-in-the-loop discharge review settings where those additional self-audit outputs justify higher inference cost despite similar-or-worse completeness. These results support the feasibility of auditable, QoL-relevant discharge generation, but they do not establish overall performance superiority. Larger evaluations with clinician adjudication and direct simpler-baseline comparisons are needed.
Accurate grading of renal cell carcinoma (RCC) is critical for prognosis estimation and treatment planning. Current grading relies on post-biopsy histopathological manual assessment of whole-slide images (WSIs), which is time-consuming and subject to inter-observer variability. Automated multiclass RCC grading remains challenging due to subtle inter-grade morphological and textural differences that cannot be captured just by histomics features or deep learning features individually. This study proposes a decision level-fusion based novel framework using histomics, machine learning and deep learning methodologies to derive multiclass RCC grading, which gives significantly higher accuracy. The approach integrates convolutional network with reduced information loss using residual links, channel and spatial attention to derive prominent deep learning features, Minimum Redundancy–Maximum Relevance (mRMR) guided prominent histomics features selection, augmented with classical machine learning classifiers. Convolutional networks capture localized nuclear and textural patterns. The model was evaluated on a five-class grading task (non-cancerous, Grades 1–4 of clear cell RCC) using a publicly available standard kidney histopathology dataset comprising 4,003 WSIs. The experiment exhibits that histomics contribute significantly in deriving maximum accuracy. Yet deep features significantly enhance accuracy beyond histomics analysis. Experimental results exhibit a high validation accuracy of 96.0% and a test accuracy of 96.1%, which can assist clinical decisions.
