ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsQuick take3 min readDisconfirmAI Editorial

Chi Mei Hospital President Lin Hung-jung: Before AI Agents Enter Hospitals, a Secure Sandbox Environment Must Be Built

Lin Hung-jung presented Chi Mei Hospital's four-stage path from basic digitalization in 2005 to targeting Agentic AI by 2025 at the Taiwan Hospital CIO Conference. His core argument is that hospitals must first complete data governance and establish secure sandbox environments before allowing AI agents to read medical records or operate systems, thereby preventing data leaks or improper actions. He uses the metaphor of a 'lobster tank' to describe this prerequisite, noting that the currently deployed A+ Medication Safety Specialist utilizes Retrieval-Augmented Generation (RAG) combined with the Medi-Span pharmaceutical database to reduce hallucination risks, while the integration method between AI agents and existing Robotic Process Automation (RPA) remains under exploration.

2026-10-30 15:51 GMT+8

Lin Hung-jung divides Chi Mei Hospital's AI development into four steps: starting with data warehousing and business intelligence tools in 2005; establishing an AI Center in 2019 to extend biostatistical capabilities to machine learning; entering the generative AI assistant stage in 2023; and targeting Agentic AI starting in 2025. The logic driving each step is not technological availability, but whether the organization has already met the conditions for safely using the next layer of capability.

During the disease prediction phase, Chi Mei deliberately avoided medical imaging AI. Lin Hung-jung explained that imaging models require significant computing resources and manual annotation labor, whereas Chi Mei has accumulated years of clinical data and biostatistical experience. Focusing on disease risk prediction based on these existing capabilities is more pragmatic. This choice later yielded 32 online predictive models covering 13 departments, allowing physicians to view patient-related risks directly on the Hospital Information System (HIS) page.

Upon entering the generative AI stage, Chi Mei started with documentation tasks such as nursing records, converting vital signs and catheter logs into drafts, and then expanded to other professions like physicians and pharmacists. This collaborative mechanism operates independently of the existing HIS, avoiding the need to modify original programs and thus preventing version-change risks and cascading system crashes.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

The A+ Medication Safety Specialist is currently the most notable achievement. Lin Hung-jung cited an example where a pharmacist might be responsible for the medication safety of over a hundred patients in an entire ward. Previously, rule-based systems filtered patients according to preset conditions, but medical knowledge updates with literature and clinical guidelines, causing existing rules to become outdated. The new system allows AI to first analyze medical records and medication data, screening out 10 patients out of 100 who may have medication issues and compiling their histories for pharmacists to prioritize checking. Lin stated this system saves pharmacists approximately 150 minutes per day and won the National Healthcare Quality Award (NHQA). To reduce hallucination risks, Chi Mei combines AI with the Medi-Span pharmaceutical database, using RAG to introduce continuously updated pharmaceutical knowledge. Medi-Span first screens for drug interactions, contraindications, and dosage issues, which are then processed by AI before final professional judgment by the pharmacist.

However, Lin Hung-jung explicitly stated that Chi Mei has not yet officially launched AI agents because the hospital has not established a sufficiently robust secure execution environment. He uses the 'lobster tank' metaphor: you cannot let lobsters run wild, cause damage, or escape. For a hospital, if AI agents are to read medical records, operate systems, or execute tasks, it must first ensure they only operate within permitted scopes. Building such an environment presupposes effective data governance—Lin points out that this is a technical capability lacking in many hospital IT departments.

Another unresolved challenge is how AI agents will integrate with existing Robotic Process Automation (RPA). Chi Mei has used RPA to handle many repetitive tasks, but RPA executes predefined rules and steps. Theoretically, AI agents could understand tasks and plan steps, which RPA would then execute as fixed operations, forming agentic automation. Lin emphasizes that the specific integration method between the two remains an area of exploration.

Facing rapidly changing AI technology, Lin Hung-jung's strategy is not to chase the cutting edge, but to stay 'two steps behind the technological frontier,' maintaining caution while continuing to follow developments. This positioning means Chi Mei is willing to leave a buffer between capability verification and organizational readiness, rather than using technological availability as the sole criterion for deployment.

Sources

  • ithome.com.twiThome↗
Continue researching →