This work focuses on improving blood glucose prediction by incorporating glycemic-aware training strategies that better capture hypo- and hyperglycemic events.

We propose Artificial Intelligence for Modeling and Explaining Neonatal Health (AIMEN), a deep learning framework that predicts adverse labor outcomes from risk factors and provides the model’s reasoning.

A hybrid attention framework designed for accurate and efficient blood glucose forecasting using multimodal data using feature decomposition and knowledge distillation

An innovative self-supervised learning framework developed for real-time detection of Freezing of Gait (FoG) in Parkinson’s Disease (PD) patients, using a single triaxial accelerometer

We present a structured overview of methods, challenges, and solutions, aiming to support researchers and practitioners in developing reliable and explainable AI solutions for digital health. This paper is further enriched with detailed discussions of the contributions toward robustness and explainability in digital health, the development of trustworthy AI systems in the era of LLMs, and various evaluation metrics for measuring trust and related parameters such as validity, fidelity, and diversity.

The presentation outlines the later stages of conducting a PRISMA-style scoping review, focusing on the transition from initial screening to final synthesis and manuscript preparation. It begins with a recap of completed steps, including search strategy development, deduplication, and initial screening, and then details the progression to final screening, paper chaining, data extraction, data synthesis, and paper writing. The presentation emphasizes the importance of clearly defined Population, Concept, and Context (PCC) criteria, structured screening rules, and exact exclusion reasons, highlighting their role in ensuring consistency, transparency, and defensibility in the review process. It discusses practical strategies for final screening, including resolving ambiguous “maybe” cases through stricter rule interpretation and verifying alignment between wastewater catchments and population-level outcomes. Data extraction is presented as a structured and integral step, often performed alongside screening using standardized templates to ensure consistency across studies. The presentation then frames data synthesis as an analytical process that goes beyond summarization to identify patterns, inconsistencies, and gaps, with preliminary observations noting variability in benchmarking, temporal clustering of studies, and inconsistent reporting practices. Finally, it describes the writing phase, where prior steps are translated into detailed methods, flow diagrams, and synthesized findings, and concludes with reflections on the high effort required but overall value of the scoping review process.

GainLoRA is a continual learning method for large language models that builds on LoRA by adding a small input-dependent gating mechanism for each task-specific LoRA branch, so instead of always combining all learned branches equally, the model learns how much each branch should contribute for a given input. The key goal is to reduce forgetting in the realistic setting where tasks arrive sequentially, task identity is unknown at test time, and no old data are replayed. Its main insight is that the newest LoRA branch should have little or no effect on inputs from older tasks, so the paper designs the gate to stay near zero on old-task inputs through special initialization and update constraints, while still allowing strong activation for the current task. The authors evaluate the method on continual instruction-learning and long-sequence benchmarks using models like T5-Large, T5-XL, Llama-2, and Llama-3, and report consistently better average performance and lower forgetting than prior LoRA-based continual learning methods such as O-LoRA and InfLoRA. In short, the paper’s contribution is a smarter way to combine task-specific LoRA adapters, making LoRA-based continual learning more selective, more stable, and substantially less prone to interference across tasks.

PaperBanana is an agent-based framework that automates the generation of publication-ready academic illustrations using VLMs and image models. It uses multiple specialized agents and iterative refinement, and outperforms existing methods across several quality metrics.