Responsible implementation of generative artificial intelligence in healthcare:Together

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Artificial intelligence (AI) has long promised to transform healthcare, yet most AI models fail to progress beyond the research setting, a phenomenon known as the implementation gap. With the arrival of large language models (LLMs) such as ChatGPT, closing that gap has become both more urgent and more complex: this new generation of AI extends well beyond the narrow predictive algorithms that previously dominated the field, encompassing documentation, communication, and clinical decision support. The inherent flexibility of these models carries a compounding expansion of risk, and the field has not kept pace with the practical and governance challenges this creates, particularly in high-stakes settings such as the intensive care unit (ICU).<br/><br/>This thesis advances the responsible use of LLMs in healthcare across the full trajectory from development to clinical implementation. Through a systematic review of AI maturity in critical care, governance and regulatory frameworks, clinical benchmarking, and empirical evaluation studies, the work establishes the prerequisites for safe LLM deployment and provides practical tools to achieve it. To guide healthcare institutions through the specific risks of LLM deployment, from safeguarding patient privacy to ensuring human oversight, the thesis develops a step-by-step governance framework at the application level. Complementing this, it offers original analysis of when LLM-based tools qualify as medical devices under European law, providing clarity for developers and institutions navigating the MDR and AI Act. This thesis furthermore demonstrates that leading LLMs possess competent medical knowledge, and that AI-generated clinical documentation, including patient summaries and ICU family meeting notes, is of comparable quality to physician-written equivalents while substantially reducing administrative burden.<br/><br/>This thesis converges on a four-pillar model of responsible implementation integrating clinical validation, technical governance, regulatory compliance, and organizational readiness. Successful AI integration in healthcare depends as much on institutional preparedness and a structured, phased approach to adoption as it does on technological capability, beginning with lower-risk administrative applications before advancing to clinical decision support. This thesis offers concrete guidance for navigating that path, demonstrating that the AI implementation gap in healthcare is neither inevitable nor insurmountable.<br/>

Links and sources

Comments