Trustworthiness in clinical LLM systems depends not on fluency, but on grounding models in authoritative knowledge and distilling that signal into reliable behaviours. In this talk, Dr Yuan-Fang Li, Chief AI Scientist at Oracle Health & AI, presents two case studies demonstrating this principle in practice.
For automated ICD-10 coding, integrating coding standards and structured validation reduces hallucinated predictions and improves robustness as guidelines evolve. For radiology report generation, verifiable, fact-level feedback and preference-based supervision enhance accuracy while enabling more data- and compute-efficient reinforcement fine-tuning.
Together, these examples outline a practical blueprint for trustworthy clinical AI: combine grounding with verification, then distil the resulting signal so models not only sound correct, but behave reliably.
Dr Li leads a team of over 100 applied scientists at Oracle Health & AI in Australia, developing advanced AI solutions for electronic health record systems. His research spans large language models, knowledge graphs, multimodal learning, and neuro-symbolic AI.
Dr Yuan-Fang Li is the Chief AI Scientist at Oracle Health & AI in Australia, where he works with a team of over 100 applied scientists to develop cutting-edge AI solutions for electronic health record (EHR) systems, aimed at transforming healthcare. In this role, Yuan-Fang provides strategic scientific leadership to ensure the delivery of impactful, high-quality, and innovative AI-driven products. His research interests include large language models, knowledge graphs, multimodal learning, and neuro-symbolic AI.
