Minisymposium

Evidence from randomised controlled trials is often reported as aggregate effects, while individual patients may differ from the trial population in many ways. Determining the most applicable evidence for a given patient’s unique characteristics and prior treatments can be challenging. Large Language Models (LLMs) enable large-scale processing of textual information, offering a step change in how we extract and apply trial evidence. This presentation will cover recent advancements in leveraging LLMs to accelerate the translation of evidence into personalised treatments.

Integrating computing-based solutions into translational biomedicine presents significant challenges, particularly in healthcare institutions underlying technical infrastructure and the adaptation of these solutions. Many facilities, especially in remote areas, lack infrastructure such as computing rooms and robust datacenters to support high-performance computing (HPC) systems required to handle the massive volumes of diverse generated biomedical data due to budget restraints preventing investment in cutting-edge technology. Secured AI tools must be tailored to complex clinical workflows, requiring resources for customization, testing, and regulatory compliance. Network reliability and cybersecurity issues complicate integration efforts for deploying AI-driven solutions. Data exchange across systems is hindered by fragmented sources, interoperability challenges, and varying data formats. The mix of structured and unstructured data, such as clinical records or genomic data, complicates integration. Data governance is a concern, especially regarding patient privacy, regulatory compliance like HIPAA, and ensuring ethical use of AI tools. For remote ICUs, limited bandwidth and the need for real-time decision-making add pressure, where delays can have critical consequences. Overcoming these physical, technical, and financial barriers, along with addressing the complexities in infrastructure, dataflow, and governance complexities, is crucial for successful integration of AI in healthcare and realizing its potential in improving patient care and outcomes.

The rapid growth and complexity of medical data necessitate innovative initiatives in medical AI, particularly in integrating diverse data types for holistic analysis and robust decision-making. This talk explores how NVIDIA, through collaborations with research institutions and healthcare organizations, is advancing a scalable multimodal medical AI ecosystem. Central to this effort is MONAI Multimodal, an open-source framework that integrates agentic AI architectures to bridge healthcare data silos. By leveraging autonomous agents for multistep reasoning across modalities, MONAI Multimodal enables seamless diagnostic analysis, combining medical imaging such as CT and MRI with surgical recordings, electronic health records (EHRs), clinical documentation, and other data streams. Key components include specialized large language models (LLMs) and vision-language models (VLMs) tailored for medical applications. The ecosystem supports research innovation across radiology, surgery, and pathology, enhancing clinical collaboration and accelerating the translation of AI research into clinical practice. This presentation will delve into how NVIDIA's technologies, including MONAI Multimodal, are empowering clinicians and researchers to develop scalable and reliable AI systems in healthcare.

Advanced AI systems like LLMs and LMMs are transforming medical applications, but face adoption barriers in clinical settings. This panel will discuss the challenges and solutions to develop reliable, multitask clinical AI agents and integrating them into hospital infrastructure.

Discussion Points:+ Current challenges: task-specific limitations, knowledge integration gaps, single-modality constraints, and reasoning deficiencies+ Capabilities of foundation models for multitask medical applications (classification, segmentation, reporting)+ Challenges to clinical implementation: 1. Knowledge integration: Embedding biomedical expertise into AI systems. 2. Multimodal integration: Combining imaging, text, and diverse data streams. 3. Evidence grounding: Ensuring outputs are verifiable and trustworthy. 4. Technical feasibility: Optimizing models to run efficiently on standard hospital infrastructure. 5. Workflow integration: Bridging research innovations to practical clinical use.