Health AI Hub

ONCOTIMIA is a novel GenAI-powered clinical tool designed to automate the completion of lung cancer tumour board forms, addressing the significant documentation burden in oncology. Leveraging a multi-layer data lake, RAG, and adaptive forms, the system transformed unstructured clinical data into standardized records. The evaluation demonstrated high performance, achieving 80% correct field completion with clinically acceptable latency, proving its technical feasibility and operational viability for multidisciplinary lung cancer workflows.

This paper introduces RHSIA, a Graph Transformer deep learning model that enables real-time prediction of temporal Wall Shear Stress (WSS) patterns for intracranial aneurysms (IAs) directly from surface meshes. This addresses the current clinical translation barrier of Computational Fluid Dynamics (CFD) due to its high cost and time. The model achieves high accuracy and effectively captures the pulsatile nature of hemodynamics, further enhanced by a novel data augmentation strategy using low-cost steady-state CFD data.

This paper introduces the concept of "hyperdisorder" to describe a distinct architectural regime in tumor growth, characterized by spatially nonstationary, multiscale disorder that persists or amplifies with increasing observation scale. Analyzing H&E whole-slide images of gastric cancer, the study demonstrates that architectural disruption is unevenly distributed and partially decoupled across scales within individual tumors, suggesting growth-driven multiscale incoherence.

This paper introduces an interpretable recommendation model specifically designed for psychometric data, addressing key challenges faced by recommender systems in healthcare, such as data scarcity, interpretability, and risk assessment. Applied to gerontological primary care, the model provides visual explanations to assist care professionals in creating personalized care plans. Evaluations demonstrate its promising performance and the interpretability of its explanations, suggesting an advancement in applying recommender systems within this critical healthcare niche.

This paper introduces an interactive evaluation framework for medical diagnostic agents, addressing the oversight of evidence-gathering processes in existing evaluations. It proposes Information Coverage Rate (ICR) and the EviMed benchmark to quantify and systematically study evidence elicitation. The core finding is that strong diagnostic reasoning alone is insufficient for effective information collection, leading to a performance bottleneck, which is mitigated by the proposed REFINE strategy.

This paper introduces a novel backpropagation-free multi-task Green Learning (MTGL) framework designed for simultaneous Left Ventricle (LV) segmentation and Left Ventricular Ejection Fraction (LVEF) classification from echocardiograms. The MTGL model achieves state-of-the-art performance on the EchoNet-Dynamic dataset with superior computational efficiency and interpretability, addressing key limitations of existing Deep Learning approaches in medical image analysis.

Health-SCORE introduces a novel framework to significantly reduce the cost and effort of creating domain-specific rubrics for evaluating and training Large Language Models (LLMs) in safety-critical healthcare. It achieves evaluation quality comparable to human-created rubrics while simultaneously serving as a structured reward signal for safety-aware reinforcement learning and enhancing response quality via in-context learning. This approach aims to make rubric-based evaluation and training of Health-LLMs substantially more scalable.

This paper advances infectious disease modeling by introducing chemotaxis-inspired Partial Differential Equation (PDE) models, which more realistically capture airborne infection spread towards susceptible populations compared to traditional reaction-diffusion models. It provides an epidemiologically interpretable framework, including a novel spatially-aware basic reproduction number ($R_0$), and demonstrates improved accuracy in simulating spatiotemporal disease dynamics in real-world scenarios like the COVID-19 outbreak in Lombardy.

This paper investigates the effectiveness of temporal embedding models trained on longitudinal electronic health records (EHRs) for chronic kidney disease (CKD) patients. It demonstrates that a time-aware LSTM (T-LSTM) learns more structured and clinically meaningful representations, and that learning embeddings as an intermediate step significantly improves predictive performance for tasks like in-ICU mortality compared to direct end-to-end models.

This paper introduces a diagnostic tool to assess the adequacy of global clinical prediction models by identifying patient subgroups for whom personalized, localized regression models are more appropriate. It addresses the challenge of high-dimensional data by leveraging an autoencoder to learn a latent representation, facilitating robust localized regression. The tool ultimately aims to characterize patients whose outcome associations deviate from a global model, providing insight into specific clinical needs.

Advances in single-cell and spatial transcriptomic technologies have transformed tumor ecosystem profiling at cellular resolution. However, large scale studies on patient cohorts continue to rely on bulk transcriptomic data, where variation in tumor purity obscures tumor-intrinsic transcriptional si...