1 AI education and interdisciplinary talent development
Help physicians, other healthcare professionals, researchers, IT engineers, healthcare management professionals, and students develop a shared language. Plans include AI literacy courses, hands-on workshops, research methods courses, and governance courses to promote mutual understanding between healthcare and technical professionals.
The existing clinical pathway plans for Level 1 AI literacy → Level 2 Clinical AI builder → Level 3 Clinical AI leader; this is one part of Medical AI education. Expanded curricula, certificates of completion, certification policies, and validity periods are still subject to review and approval by the relevant committees.
Corresponding project: P-20 Society operations platform
Course schedule and registration — in preparation
2 Medical AI research and trustworthy evidence
Promote research in Medical AI and medical data science, and develop shared specifications, evaluation metrics, and collaboration across institutions. Different uses require different evidence: research reproducibility, educational outcomes, improvements in management processes, and clinical safety cannot be assessed solely using a single model score.
Models from a single hospital are difficult to generalize to other settings, and multicenter validation is an important method for trustworthy clinical AI. P-12 and P-13 currently use synthetic hospital-site data for federated validation and a shadow-mode drill, exploring an architecture in which execution takes place locally at each hospital and only aggregate statistics are returned; this does not yet constitute real multicenter retrospective validation, a silent prospective trial, or regulatory-grade evidence.
Corresponding projects: P-11 CQL rule library, P-12 federated validation, P-13 prospective silent trials, and P-14 Model Card registry
3 Healthcare applications, quality, and process improvement
Starting with practical needs and human–machine collaboration, the directions for work encompass clinical care, health promotion, healthcare quality, patient safety, and healthcare management. First clarify the intended use, beneficiaries, workflow burden, and evaluation approach, then determine the appropriate AI methods.
“Workflow” before “model”; rules provide the skeleton, and the LLM adds the flesh. The first batch of v0.1 prototypes focuses on reducing documentation burden and on cardiovascular CDSS; draft rules, mock, golden sets, and synthetic data are currently used to validate engineering processes, and neither clinical committee review and approval nor real clinical deployment has been completed. The design boundaries remain limited to clinical reminders and summaries, with no automatic prescribing and no automatic entry of medical orders.
Corresponding projects: P-06 pre-visit SOAP, P-07 clinic summaries, P-08 heart failure GDMT, P-09 AF anticoagulation, P-10 CAD/ACS secondary prevention, and P-11 CQL rule library
4 Governance, policy, and translation of innovation into practice
Support research and industry–academia collaboration through data governance, ethics, privacy, and academic independence, and promote policy exchange and responsible dissemination of results. Not all Medical AI falls within SaMD; applications involving medical devices must undergo appropriate evaluation based on their intended use. For details, see Regulatory and research responsibility boundaries (Mandarin only).
P-17 and P-18 in the existing SaMD pathway currently consist of Docs-as-Code document scaffolds and a synthetic drift drill; they are not submission-ready documents or formal monitoring services. They do not indicate that review and approval by regulatory consultants, submission to or approval by Taiwan’s Food and Drug Administration (TFDA), or deployment of post-market surveillance has been completed. Collaboration must also comply with D-03 Role separation and conflict of interest management (Mandarin only).
Corresponding projects: P-17 SaMD document generator and P-18 post-market surveillance
5 Secure digital infrastructure and Agent/MCP
Reusable digital tools can support research, education, knowledge management, and Society administration, and may also be explored for clinical use under more stringent approval conditions. Least privilege, traceable provenance, human oversight, and data protection are shared design requirements; when patient data are involved, the existing boundary that data must remain within the hospital must still be maintained.
Agent and MCP require “forward-looking exploration under strict controls.” P-15 and P-16 currently use an internal checklist, static analysis, signing tests, and synthetic red-team attacks to validate v0.1 prototypes; internal review does not amount to external security certification, and neither a clinical MCP repository nor a production agent runtime has been deployed. Security review, an allowlist of signed tools, tiered permissions, auditing, and physician confirmation are required elements of the target architecture for subsequent clinical adoption.
Corresponding projects: P-01 PHI scrubbing gateway, P-02 FHIR minimization, P-03 audit ledger, P-15 MCP/skill repository, and P-16 clinical agent execution framework