Hospital AI Lab Practical information on medical AI for Singapore hospitals

Updated:

Templates: generative AI policy and 21 vendor questions

“We don’t know where to start writing” is a common problem for hospitals. Here are four drafts: press the button, copy, and replace “__” with your own details.

Checked: November 14, 2026 · For: hospital administration and IT

Before you use these
These are the author’s drafts, not legal advice and not a version that is ready to copy as is. Adapt them to the size, systems and contracts of your hospital, and show them to counsel or your compliance office before they become documents.

Contents

Related pages: Roadmap, Patient data checklist, Checklist: records and AI vendors.

1. Generative AI use policy

A page you can hand to staff. Replace "__" with your hospital’s details.

GENERATIVE AI USE POLICY, "__" HOSPITAL (DRAFT)

1. Purpose
   Sets rules for staff of "__" when they use generative AI tools (chatbots, summarising tools, translation tools).

2. What must NOT be entered
   - Patient names, NRIC or FIN numbers, medical record numbers, dates of birth, addresses, phone numbers, or the content of patient records.
   - Images or recordings that could identify a patient.
   - Contracts, or restricted financial and personnel data.
   Exception: a tool that "__" has approved in writing for that data type (see clause 3).

3. Approved tools and permitted data
   Only tools on the approved list kept by "__" (owner: __). For each approved tool the list states which data types may be entered, following our data security classification. Using an unapproved tool for work is not allowed.

4. Review of output
   A person is responsible for every AI-generated text used in work. Generative AI can state fictitious content as fact, so check facts, numbers and names against the source before use.
   AI output is never entered in a patient record until the responsible clinician has reviewed and approved it. Where AI-generated content is shown to others, label it as AI-generated.

5. Reporting
   If patient information was entered into an unapproved tool, report it the same day to __ (name, phone __). Do not try to delete it yourself first.

6. Training
   Staff who use AI tools are trained on this policy, including the data input restrictions and how to report an incident, and are re-trained after significant updates. Owner: __. Review date: __.

Basis of the template

  • The data input rules follow AIHGle 2.0: organisations specify permissible data types per AI solution (section 6.1.1(d)), train staff on data input restrictions (section 6.5.2), and manage generative AI risks such as hallucination and data disclosure, fact-checking workflows and AI labels (sections 8.2.1 and 8.2.2(d)).
  • The rest is the author’s proposal.

2. Twenty-one questions for a vendor

Print it and take it to the vendor presentation. Write the answers down.

TWENTY-ONE QUESTIONS FOR AN AI VENDOR (DRAFT)
Write the answers down and attach them to the contract.

Device status
1. Do you treat this product as a medical device in Singapore? If not, why not, based on its intended use?
2. If yes, which risk class, and is it registered with HSA? May we see the registration details?
3. What is the intended use statement, and does our planned use match it?
4. Is the software a stand-alone product or part of another device? If part of a device, which registration covers it?
5. Does the product contain machine learning, and where do the instructions for use say so, with the model’s inputs and outputs?
6. How will we be told before an update reaches our wards, and can we delay it?
7. How is an update rolled back, and who tells our staff?

Performance and monitoring
8. On which patients was the tool tested, and how does that compare with ours, including age groups, ethnic groups and languages?
9. Which real-world performance figures do you monitor after go-live, how often, and will you share them with us?
10. Does the model keep learning after deployment? If so, what stops its performance from drifting, and can we revert to an earlier version?

Patient data
11. For what purposes will you process our patient data, and nothing else? Will you use it to improve your own product?
12. Where is our data stored and processed, and which sub-processors (cloud, model providers) touch it?
13. Can you trace the data used for training or learning to a software version and to the output we see?
14. What security safeguards do you apply, and how do they fit our data security classification?
15. How and how fast do you tell us of a personal data breach, and who answers at night?
16. How is data erased when the contract ends, including in logs, backups and by your sub-processors?
17. How does your tool let us label AI-generated content for staff and patients?

Operations
18. Who at our hospital gets more work with this tool, and who reviews its output?
19. What does the exit plan look like, including return or erasure of data?
20. Which decisions affecting patients will the tool’s output feed into, and does a clinician make the final decision?
21. What is your support commitment and escalation path for safety issues?

Basis of the template

  • Questions 1–7: HSA’s GL-04-R4 (sections 1.3, 3, 9.1, 9.4) and GL-07-R2 (intended use, risk classification).
  • Questions 8–10 and 13: GL-04-R4, sections 9.1 to 9.3 (model description, continuous learning, post-market monitoring, traceability).
  • Questions 11, 12 and 14–16: AIHGle 2.0, sections 6.1.1(d) and 6.7.1, and the PDPC breach guide (revised 15 March 2021).
  • The other questions are the author’s proposals, not legal requirements.

3. Text for patients

A short example of how to tell a patient that the hospital uses an AI tool. Match it to your actual practice.

TEXT FOR PATIENTS: OUR USE OF AN AI TOOL (EXAMPLE)

"__" hospital uses an AI tool for __ (for example: drafting discharge summaries, which a clinician then checks).
The tool does not make the final decision about your treatment. A member of our staff reviews what it produces.
We tell you which of your information the tool uses and why. You can ask us about it and about alternatives.
Questions: __ (name or office, phone, email).

Basis of the template

  • AIHGle 2.0, section 6.9, says patients should receive accurate and sufficient information to make informed decisions about their care, and that not all AI use is relevant to a patient’s medical management (for example billing or scheduling).
  • The wording is the author’s example, not required language.

4. Incident record

Turn the procedure into one fillable page and keep it where people will find it at night and on weekends.

PERSONAL DATA INCIDENT RECORD (DRAFT)

Date and time discovered: __   Discovered by: __   Reported to: __
What happened (one sentence, no patient names): __
Systems and data involved (type of data, approximate number of individuals): __
Credible grounds to believe a data breach occurred from (date): __   Assessment must be completed within 30 calendar days.
Who assesses whether the breach is notifiable: __   Steps taken in the assessment (document all): __
Likely significant harm to individuals? __   500 or more individuals affected (or believed to be)? __

Containment done: __   By: __   Time: __
Determined notifiable on (date and time): __   PDPC notified (within 3 calendar days): who files: __   Filed on: __
Affected individuals notified: who sends: __   Method: __   Sent on: __
Other authorities, if relevant: HSA (adverse event from a medical device) __   CSA (critical cybersecurity incident) __
Vendor involved: __   (contact: __, 24-hour line: __)

Recorded: __ (who contacted whom, and when).

Basis of the template

  • The 30-day assessment, the 3-day notification and the 500-individual threshold come from the PDPC’s Guide on Managing and Notifying Data Breaches (revised 15 March 2021). The reporting routes to HSA and CSA come from AIHGle 2.0, section 6.7.1. Check the PDPC website for any later revision.

Sources and date checked

Sources checked November 14, 2026. Rules and procedures change, so rely on the current text on the official source page.

The material on this site is general information for education. It does not replace diagnosis or medical advice for a specific patient, and it is not legal advice or a purchasing recommendation. For rules and policy, rely on the primary official sources.