Adverse drug events (ADEs) are a critical aspect of patient safety and pharmacovigilance, with significant implications for patient outcomes and public health monitoring. The increasing availability of electronic health records, social media, and patient forums provides valuable yet challenging unstructured data sources for ADE surveillance. To address these challenges, we introduce CONORM, a framework integrating named entity recognition (NER) and normalization (EN) for ADEs across diverse textual domains. CONORM comprises CONORM-NER and CONORM-EN, featuring a dual-encoder with dynamic context refining (DCR). DCR adaptively combines isolated entity embeddings with contextual representations. Our analyses demonstrate that CONORM adapts to text formality, enhances precision on out-of-distribution concepts, and reduces errors compared to context-agnostic baselines. CONORM was evaluated on tweets, forum posts, and structured product labels, covering 30,000 annotated ADE mentions across the SMM4H, CADEC, and TAC datasets. It achieved end-to-end F1-scores of 63.86%, 72.45%, and 84.99%, surpassing the strongest baseline by an average margin of 20%. These results highlight CONORM's robust adaptability across domains. CONORM offers a scalable solution for pharmacovigilance, with pre-computed target embeddings enhancing inference efficiency. Its generalization establishes it as a robust tool for ADE surveillance. The source code is available at https://github.com/ds4dh/CONORM .
Federated learning (FL) enables privacy-preserving multi-site clinical research, but orchestrating, monitoring, and interpreting federated analyses remains out of reach for the biostatisticians and clinicians who most need them. We present Starfish-FL, a multi-tier agent harness for cross-silo healthcare analytics built around the safety invariant that the LLM is advisory and cannot independently trigger consequential actions. Agents operate at three tiers. An Autonomous Experiment Agent orchestrates end-to-end experiments via CLI tools, an Embedded Router Agent handles adaptive aggregation, early stopping, and failure triage, and Controller Agent Hooks produce per-site summaries, outlier flags, and convergence signals. Three hybrid safety policies, SafeStop, SafeAggregate, and SafeTriage, gate every site exclusion, early stop, and recovery action behind deterministic preconditions, so LLM errors cannot cause unsafe state transitions, conditional on guard correctness. All agent features are opt-in and degrade gracefully to standard FL, and the platform provides 21 Python and R task implementations with built-in statistical diagnostics. Evaluated across four LLM providers with repeated-run confidence intervals, task selection reaches \({\sim}97\%\) accuracy with hosted models, the online SafeStop policy saves 50–70% of rounds with no accuracy loss, and safety-policy validation shows a 100% safe-action rate under injected failures, confirming that deterministic guards, not the LLM, preserve safety.
Recent advances in control methodologies have significantly improved prosthetic functionality. However, current systems are constrained by a limited number of controllable degrees of freedom and face challenges in effectively handling unexpected external events. Here, we propose a prosthesis interface system that combines ultrasound recording and intention decoding with a human-inspired somatosensory control, allowing the prosthesis to autonomously respond to thermal stimuli and slippage events. Ten participants without amputation and one person with amputation participated in the validation, which included an offline evaluation, three real-time functional tasks with increasing levels of complexity to assess system efficacy and user experience, and one sensory task to evaluate the system's capability to manage slip events and thermal stimuli. The offline analysis yielded an average gesture classification accuracy of 72% and 92% for participants with and without amputation, respectively. Real-time task analysis showed consistent performance among complexity levels. The sensory task showed an instantaneous response of the prosthesis to temperature stimuli ( \(0.53\pm 0.16\) s) and slip events ( \(0.43\pm 0.13\) s) that closely matched natural responses. Building upon previously validated studies, the integration of a human-inspired somatosensory control strategy with an ultrasound-based gesture decoding classifier paves the way for advancing prosthetic technology and improving user experience.
Frontotemporal Dementia (FTD) presents significant challenges in disease progression forecasting due to its complex temporal dynamics and the logistical difficulties of acquiring longitudinal imaging data. Convolutional Neural Networks (CNNs) excel in image-based tasks but struggle with tabular data lacking spatial relationships. To address these challenges, we propose a Tensorized Image Generator (TIG) algorithm that converts longitudinal tabular data into structured image representations. TIG systematically maps features to specific pixel positions while preserving the spatial proximity of related features, enabling the use of CNNs for effective spatial feature extraction. We utilize longitudinal neuropsychiatric data from the FTLDNI archive and develop a hybrid deep learning framework integrating CNNs with Recurrent Neural Networks (RNNs), including Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), to model temporal dynamics. This approach forecasts FTD progression up to four years into the future. Benchmarking results reveal that CNN-LSTM and CNN-GRU models trained on TIG-generated images achieve significantly lower mean squared and mean absolute errors compared to models trained on original tabular data and images from existing conversion methods (e.g., IGTD). Additionally, TIG-generated images outperform spatio-temporal ConvLSTM2D models by preserving feature neighborhood structure more effectively, offering a novel and efficient solution for FTD progression forecasting.
Legitimacy is a key concern when introducing data about work in healthcare, especially as data-driven systems move closer to real-world application and begin shaping how care work is represented, interpreted, and shared across stakeholders. This paper reports healthcare workers’ (HCWs’) perspectives on data sharing in a design project that uses workflow data to support care, including accounting for the needs of patients and caregivers. Guided by the principle of “minimum viable datafication” [62]- limiting data production and sharing to what is meaningful and contextually appropriate - we engage Dourish’s [26] notion of the “legitimacy trap,” in which participatory processes appear consultative, but ultimately reinforce designer or managerial assumptions. Aiming to pro-actively achieve legitimacy, rather than falsely assume it, we contribute three lessons for the Human-Computer Interaction (HCI) and Health community: 1) data representations should attune to workers’ lived experiences; 2) HCWs should be enabled to care both for their own data-related needs and the impact of data sharing on patients and caregivers; and 3) space should be created for contestation and negotiation, recognizing that data sharing must remain responsive to how stakeholders experience the representation of work.
