<?xml version="1.0" encoding="UTF-8"?><?xml-stylesheet href="/in/rss-styles.xsl" type="text/xsl"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Hospital AI Lab</title><description>Columns on the regulation of medical AI in Singapore: HSA’s software medical device guidelines, the Ministry of Health’s AI guidelines and the PDPC’s data breach guide.</description><link>https://hospital-ai-lab.com/in/</link><language>en</language><atom:link href="https://hospital-ai-lab.com/in/rss.xml" rel="self" type="application/rss+xml"/><lastBuildDate>Mon, 16 Nov 2026 00:00:00 GMT</lastBuildDate><item><title>AIHGle 2.0 after go-live: training, monitoring, off switch</title><link>https://hospital-ai-lab.com/in/blog/2026-11-16-aihgle-2-after-go-live-training-monitoring-off-switch/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-16-aihgle-2-after-go-live-training-monitoring-off-switch/</guid><description>Singapore&apos;s AI in Healthcare Guidelines 2.0 ask hospitals to train staff, monitor AI and plan a way to switch it off. Who reports what to whom.</description><pubDate>Mon, 16 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;An AI tool is switched on, the vendor leaves, and the ward carries on. Weeks later a nurse notices the alerts look odd. Who does she tell, and what happens to the patient in the meantime? Singapore’s Ministry of Health and the Health Sciences Authority (HSA) address exactly this stage in version 2.0 of the Artificial Intelligence in Healthcare Guidelines (AIHGle 2.0), published in March 2026. This post covers only what the guidelines say about the period after go-live, in plain words. They are guidelines, not a statute, and we read only the guidelines document itself.&lt;/p&gt;
&lt;h2 id=&quot;training-that-goes-beyond-how-to-click&quot;&gt;Training that goes beyond “how to click”&lt;/h2&gt;
&lt;p&gt;Section 6.5.2 says staff training should comprehensively cover a long list of topics. It includes roles and responsibilities, relevant professional standards and regulatory requirements, workflow integration, contingency plans, specific use cases, the limits of the AI solution, how to interpret its output, patient communication, data input restrictions (including data security and classification), spotting abnormal behaviour and incident reporting. It also says organisations should re-train staff when there are significant updates to the AI solution or to clinical workflows. It suggests several ways to train: self-paced online courses, hands-on demonstrations and knowledge-sharing sessions led by experienced colleagues, and structured seminars or programmes.&lt;/p&gt;
&lt;h2 id=&quot;monitoring-that-someone-owns&quot;&gt;Monitoring that someone owns&lt;/h2&gt;
&lt;p&gt;Section 6.6.1 asks organisations to set up a process to monitor the AI solution’s performance, both ad hoc and regularly, to keep it safe, effective and robust while it is in use. Section 6.6.2 says periodic evaluation should at minimum show that the tool is as good as or better than current practice for patient safety or welfare, and should check the accuracy and reproducibility of its decisions. User feedback channels should be set up and communicated, and feedback should reach organisational leadership, relevant authorities and the developers or technical teams. The guidelines list features of a feedback loop: roles and responsibilities, reporting timelines, indicators such as model drift and a response matrix.&lt;/p&gt;
&lt;h2 id=&quot;when-something-goes-wrong&quot;&gt;When something goes wrong&lt;/h2&gt;
&lt;p&gt;Section 6.7.1 asks organisations to set up a process to detect, respond to, investigate and report adverse events or other device issues from using the AI solution. Three points stand out:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Contingency plans&lt;/strong&gt; should include shutting down the AI solution and switching to a contingency protocol, for example one where the AI is not involved, so the impact on the patient is minimised, while working with developers on interim and permanent fixes.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Reporting:&lt;/strong&gt; organisations should set up internal reporting and abide by existing frameworks. Data-related incidents go to the Personal Data Protection Commission (PDPC), adverse events from medical devices to HSA, and critical cybersecurity incidents to the Cyber Security Agency of Singapore (CSA). A footnote adds that other AI-related cybersecurity incidents may be reported to SingCERT.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Investigation:&lt;/strong&gt; because the root cause may not be obvious, events should be documented and investigated, and the AI developer may be involved where appropriate.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;what-a-ward-or-department-can-prepare&quot;&gt;What a ward or department can prepare&lt;/h2&gt;
&lt;p&gt;These steps are our suggestions drawn from the guidelines, not requirements stated in them.&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Write a one-page “when the AI looks wrong” card.&lt;/strong&gt; It should say who to call, how to switch to the non-AI process and how to log the event.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Rehearse the off switch.&lt;/strong&gt; Check that the non-AI protocol still works on a normal shift, not only on paper.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Name the owner of monitoring.&lt;/strong&gt; Decide who reviews performance indicators and where user feedback goes.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Schedule re-training&lt;/strong&gt; whenever the vendor ships a significant update.&lt;/li&gt;
&lt;/ol&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;AIHGle 2.0 asks for broad staff training, re-training after significant updates, and periodic monitoring against the performance of current practice.&lt;/li&gt;
&lt;li&gt;It asks hospitals to plan for adverse events, including shutting down the AI and switching to a protocol without it, and to report through PDPC, HSA and CSA as applicable.&lt;/li&gt;
&lt;li&gt;Turn the guidelines into a one-page card and a rehearsed fallback. This is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health and HSA, AIHGle 2.0 (March 2026)&lt;/a&gt;, sections 6.5.2, 6.6.1, 6.6.2 and 6.7.1 with footnotes 24 to 26. Read on 16 November 2026.&lt;/em&gt;&lt;/p&gt;
&lt;hr&gt;
&lt;p&gt;&lt;em&gt;This site’s content is for general information, research, and education. It does not replace medical diagnosis or advice for any specific patient, and it is not legal advice.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>Column</category></item><item><title>AI model updates in Singapore: what HSA expects, what to ask the vendor</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-ai-model-updates-hsa-change-notification-singapore/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-ai-model-updates-hsa-change-notification-singapore/</guid><description>When a registered AI model is updated, HSA’s rules on change notification and its optional Change Management Program apply. What administrators can ask before an update reaches the wards.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;An AI tool that was tested last year may not be the same tool next quarter. Vendors retrain models, change inputs and push updates. For a hospital, the practical question is who checks those changes, and when you hear about them. HSA’s guidelines give part of the answer for registered devices.&lt;/p&gt;
&lt;h2 id=&quot;every-change-to-a-registered-ai-device-is-notified&quot;&gt;Every change to a registered AI device is notified&lt;/h2&gt;
&lt;p&gt;HSA’s GL-04 (revision 4, effective 31 December 2025) says a change notification is required for any change made to a registered machine-learning-enabled medical device. The category of the change, technical, review or notification, follows HSA’s separate guidance on change notification. For software in general the guidelines list technical review for Class C and D devices, review for Class B, and notification for Class B, C and D, depending on the change.&lt;/p&gt;
&lt;p&gt;GL-04 also describes devices that keep learning after deployment. The manufacturer must define the learning process and put controls in place. The items it lists include a built-in safety mechanism that can detect anomalies and roll back to a previous version, controls on the data used for learning, software version control, and traceability between real-world data, the software version and the output seen in clinical use.&lt;/p&gt;
&lt;h2 id=&quot;an-optional-fast-lane-the-change-management-program&quot;&gt;An optional fast lane: the Change Management Program&lt;/h2&gt;
&lt;p&gt;HSA’s GN-37 (revision 1, December 2024) introduces an optional pathway for SaMD, including machine-learning-enabled SaMD. A manufacturer can have anticipated “pre-specified changes” approved in advance and then implement them under its quality management system. The guidance says conformity with ISO 13485 and IEC 62304 is essential. Its scope note says it does not include continuous learning and generative AI.&lt;/p&gt;
&lt;p&gt;For a hospital, this means a vendor may say “our updates are pre-approved”. That can be true only for the changes that were actually pre-specified and approved, for a registered device that is in the program.&lt;/p&gt;
&lt;h2 id=&quot;what-hospitals-can-ask&quot;&gt;What hospitals can ask&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;Is the product registered, and does it use machine learning? (GL-04 says users must be told that the device contains ML.)&lt;/li&gt;
&lt;li&gt;How will we hear about an update, and can we delay it?&lt;/li&gt;
&lt;li&gt;Is any update covered by pre-specified changes under the Change Management Program, and which ones?&lt;/li&gt;
&lt;li&gt;Does the model keep learning after deployment? If so, how do you detect problems and roll back?&lt;/li&gt;
&lt;li&gt;How do you monitor real-world performance after deployment, and how does our feedback reach you? (GL-04 says companies must monitor the real-world performance of registered ML-enabled devices.)&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;The Ministry of Health’s AI guidelines add the hospital’s side: monitoring should check that performance is at least equivalent to current practice, with indicators such as model drift.&lt;/p&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Any change to a registered ML-enabled medical device needs a change notification to HSA.&lt;/li&gt;
&lt;li&gt;Continuous-learning devices must have defined learning, rollback and traceability controls.&lt;/li&gt;
&lt;li&gt;The Change Management Program is optional and covers only pre-specified changes; it excludes continuous learning and generative AI.&lt;/li&gt;
&lt;li&gt;Put the update process in the contract. This is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/409/26808a93-6a7a-4551-8c66-13046c6124c5/gl-04-r4-regulatory-guidelines-for-software-medical-devices---a-life-cycle-approach-(2025-dec)-pub.pdf&quot;&gt;HSA, GL-04-R4&lt;/a&gt;, &lt;a href=&quot;https://isomer-user-content.by.gov.sg/409/4e8620d9-267f-41f5-adc3-cc35456d3731/gn-37-r1-guidance-on-change-management-program-(cmp)-for-samd-including-machine-learning-enabled-samd.pdf&quot;&gt;HSA, GN-37-R1 Change Management Program&lt;/a&gt;, &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health, AIHGle 2.0&lt;/a&gt;. Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item><item><title>AI vendor data breach in Singapore: the PDPC’s 30-day and 3-day clocks</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-ai-vendor-data-breach-pdpc-30-day-3-day-clocks/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-ai-vendor-data-breach-pdpc-30-day-3-day-clocks/</guid><description>If an AI vendor has a data incident, when does your hospital’s clock start, and who must be told? What the PDPC’s breach guide and the Ministry of Health’s AI guidelines say.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Handing patient data to an AI vendor adds another place where a breach can start. Hospital administrators and IT leads should know the timing before the first incident, not during it. The Personal Data Protection Commission (PDPC) publishes a Guide on Managing and Notifying Data Breaches under the PDPA. The version read for this article was revised on 15 March 2021; check the PDPC website for later revisions.&lt;/p&gt;
&lt;h2 id=&quot;the-two-clocks&quot;&gt;The two clocks&lt;/h2&gt;
&lt;p&gt;According to the guide, once an organisation has credible grounds to believe a data breach has occurred, whether through self-discovery, an alert from the public or a notification from its data intermediary, it must take reasonable and expeditious steps to assess whether the breach is notifiable. That assessment should be completed within 30 calendar days, and every step must be documented. Once the organisation determines that the breach is notifiable, it must notify the PDPC within 3 calendar days.&lt;/p&gt;
&lt;h2 id=&quot;what-makes-a-breach-notifiable&quot;&gt;What makes a breach notifiable&lt;/h2&gt;
&lt;p&gt;The guide describes two grounds. A breach is notifiable if it is likely to result in significant harm to the affected individuals, or if it is of significant scale. Significant scale means the personal data of 500 or more individuals. In that case the organisation must notify the PDPC even if no prescribed personal data is involved. If the exact number is unknown, the organisation should notify when it has reason to believe that at least 500 individuals are affected.&lt;/p&gt;
&lt;h2 id=&quot;who-else-hears-about-it&quot;&gt;Who else hears about it&lt;/h2&gt;
&lt;p&gt;The Ministry of Health’s AI guidelines (AIHGle 2.0) tell hospitals to set up internal reporting and to follow existing frameworks: data-related incidents go to the PDPC, adverse events from the use of medical devices to HSA, and critical cybersecurity incidents to the Cyber Security Agency of Singapore. They also say contingency plans should include shutting the AI solution down and switching to a protocol that does not involve it.&lt;/p&gt;
&lt;h2 id=&quot;what-to-put-in-the-vendor-contract&quot;&gt;What to put in the vendor contract&lt;/h2&gt;
&lt;p&gt;These are the author’s suggestions, not requirements:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;A duty to tell the hospital quickly, with a named contact who answers at night, so the hospital’s own assessment can start.&lt;/li&gt;
&lt;li&gt;A list of sub-processors, such as cloud and model providers, and where the data is stored.&lt;/li&gt;
&lt;li&gt;Return or erasure of data at exit, including logs and backups.&lt;/li&gt;
&lt;li&gt;Cooperation with the hospital’s investigation.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;The PDPC guide (revised 15 March 2021) sets a 30-calendar-day assessment and a 3-calendar-day notification once a breach is determined notifiable.&lt;/li&gt;
&lt;li&gt;Breaches likely to cause significant harm, or involving 500 or more individuals, are notifiable.&lt;/li&gt;
&lt;li&gt;AIHGle 2.0 names the PDPC, HSA and the Cyber Security Agency of Singapore as reporting routes for different incidents.&lt;/li&gt;
&lt;li&gt;This article did not read the Personal Data Protection Act itself. It is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://www.pdpc.gov.sg/-/media/files/pdpc/pdf-files/other-guides/guide-on-managing-and-notifying-data-breaches-under-the-pdpa-15-mar-2021.pdf&quot;&gt;PDPC, Guide on Managing and Notifying Data Breaches under the PDPA (revised 15 March 2021)&lt;/a&gt;, &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health and HSA, AIHGle 2.0&lt;/a&gt; (section 6.7.1). Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item><item><title>Generative AI in Singapore hospitals: four risks AIHGle 2.0 names</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-generative-ai-hospital-singapore-aihgle-risks-redlines/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-generative-ai-hospital-singapore-aihgle-risks-redlines/</guid><description>Hallucination, harmful content, data disclosure and adversarial prompts: what the Ministry of Health’s guidelines say a hospital should do about generative AI tools.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;Staff already try chatbots to draft letters, summarise notes and translate. The Ministry of Health’s AI in Healthcare Guidelines 2.0 (March 2026) have a chapter on generative AI that tells hospitals what to expect and what to put in place. For administrators and nurse leaders, it is a ready outline for a staff policy.&lt;/p&gt;
&lt;h2 id=&quot;four-amplified-risks&quot;&gt;Four amplified risks&lt;/h2&gt;
&lt;p&gt;The guidelines name four risks that generative AI amplifies:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Hallucination.&lt;/strong&gt; Models may generate and present incomplete or fictitious content as facts, and users may not notice.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Undesirable content.&lt;/strong&gt; Models may produce harmful content that could compromise clinical outcomes or patient safety, for example a mental health chatbot that responds dismissively to distress.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Data disclosure.&lt;/strong&gt; Models could leak sensitive data when users submit it, or when the model is prompted in specific ways.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Vulnerability to adversarial prompts.&lt;/strong&gt; Deliberate attempts to manipulate the model’s output or behaviour can lead to misinformation and security breaches.&lt;/li&gt;
&lt;/ol&gt;
&lt;h2 id=&quot;what-deployers-should-do&quot;&gt;What deployers should do&lt;/h2&gt;
&lt;p&gt;For hospitals, the guidelines say deployers should:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;design workflows that support fact-checking of outputs;&lt;/li&gt;
&lt;li&gt;train healthcare professionals on responsible use of generative AI, including “redlines”, the things never to do;&lt;/li&gt;
&lt;li&gt;support patient communication to address concerns about generative AI;&lt;/li&gt;
&lt;li&gt;where possible, show clear front-end labels where content is AI-generated;&lt;/li&gt;
&lt;li&gt;use testing strategies such as red teaming to improve safety and performance.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Users, meaning healthcare professionals, should be trained to evaluate outputs critically and keep clinical judgment as the primary decision-making tool.&lt;/p&gt;
&lt;h2 id=&quot;turning-it-into-a-staff-policy&quot;&gt;Turning it into a staff policy&lt;/h2&gt;
&lt;p&gt;A one-page policy can follow the same headings. Which tools are approved? Which data types may be typed into each? Who checks drafts before they are used or entered in a record? Where is AI content labelled? Who gets the report if patient data went into an unapproved tool? The templates on this site give a starting draft.&lt;/p&gt;
&lt;p&gt;The guidelines also say, more generally, that hospitals should specify the permissible data types for each AI solution, taking into account their data security classification.&lt;/p&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Generative AI amplifies hallucination, undesirable content, data disclosure and adversarial prompts.&lt;/li&gt;
&lt;li&gt;Hospitals should build fact-checking into workflows, train staff, support patient communication and label AI content.&lt;/li&gt;
&lt;li&gt;Clinical judgment remains the primary decision-making tool.&lt;/li&gt;
&lt;li&gt;This is general information, not legal advice. The guidelines are guidance, not a statute.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health and HSA, AIHGle 2.0&lt;/a&gt; (sections 6.1.1(d) and 8.2). Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item><item><title>Hospital AI governance in Singapore: what AIHGle 2.0 asks before go-live</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-aihgle-2-hospital-ai-governance-before-go-live/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-aihgle-2-hospital-ai-governance-before-go-live/</guid><description>The Ministry of Health’s AI in Healthcare Guidelines 2.0 give hospitals a list of things to set up before AI reaches patients. Five of them, in plain language.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;The Ministry of Health and HSA published version 2.0 of the Artificial Intelligence in Healthcare Guidelines (AIHGle) in March 2026. It splits responsibilities between developers, deployers and users. A hospital is a deployer, and the chapter written for deployers reads like a project checklist for administrators and nurse leaders. It is guidance, not a statute.&lt;/p&gt;
&lt;h2 id=&quot;1-a-governance-platform-and-a-registry&quot;&gt;1. A governance platform and a registry&lt;/h2&gt;
&lt;p&gt;The guidelines say organisations should set up internal governance platforms to oversee all healthcare AI solutions through their life cycle. The platform assesses and approves solutions (including that they are fit for purpose and perform at least as well as current standard processes), keeps a registry of every deployed solution with its risk classification, gives guidance on safe deployment, and ensures deployments meet cybersecurity and data protection requirements. It recommends three distinct functions: a governance platform, a unit for day-to-day implementation and monitoring, and a unit that audits compliance.&lt;/p&gt;
&lt;h2 id=&quot;2-a-risk-assessment-tied-to-human-oversight&quot;&gt;2. A risk assessment tied to human oversight&lt;/h2&gt;
&lt;p&gt;Hospitals should assess the potential harm to a patient and how probable it is. The guidelines give an illustrative table separating human-in-the-loop, human-over-the-loop and human-out-of-the-loop, for clinical and clinical-operations uses. For clinical use, they say deployers should not deploy AI that makes clinical decisions independently without a human making the final decision.&lt;/p&gt;
&lt;h2 id=&quot;3-testing-in-your-own-workflows&quot;&gt;3. Testing in your own workflows&lt;/h2&gt;
&lt;p&gt;Testing and validation is usually done by the implementing organisation, in its own clinical context. For higher-risk solutions the guidelines suggest a pilot phase with a smaller patient population. They mention the AI Verify testing framework, and say hospitals should also measure clinical outcomes, at minimum ensuring that professionals and patients benefit from or maintain current care standards.&lt;/p&gt;
&lt;h2 id=&quot;4-training-and-change-management&quot;&gt;4. Training and change management&lt;/h2&gt;
&lt;p&gt;Training should cover roles, workflow integration, contingency plans, limitations, output interpretation, patient communication, data input restrictions and incident reporting, and staff should be re-trained after significant updates. The guidelines also suggest user champions, as a case study from one public healthcare group shows.&lt;/p&gt;
&lt;h2 id=&quot;5-monitoring-and-a-way-out&quot;&gt;5. Monitoring and a way out&lt;/h2&gt;
&lt;p&gt;Monitoring should check that performance is at least equivalent to current practice and track indicators such as model drift, with feedback channels and reporting timelines. For adverse events, the guidelines call for detection, response, investigation and reporting, including contingency plans that can shut the AI down and switch to a protocol without it. Before changing provider or decommissioning, an exit management plan should be agreed with the developer.&lt;/p&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Set up a governance platform and a registry of AI solutions with risk classification.&lt;/li&gt;
&lt;li&gt;Tie the deployment model to human oversight; keep a human making final clinical decisions.&lt;/li&gt;
&lt;li&gt;Test in your own workflows, pilot higher-risk tools, and plan training, monitoring and exit.&lt;/li&gt;
&lt;li&gt;Read the full guidelines and the deployers’ toolkit on the Ministry of Health’s page. This is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health and HSA, AIHGle 2.0&lt;/a&gt; (sections 6.1 to 6.8), &lt;a href=&quot;https://www.moh.gov.sg/others/health-regulation/emerging-regulatory-policy-issues/&quot;&gt;Ministry of Health, AIHGle 2.0 page&lt;/a&gt;. Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item><item><title>Is hospital AI software a medical device in Singapore? What HSA says</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-is-hospital-ai-a-medical-device-singapore-hsa/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-is-hospital-ai-a-medical-device-singapore-hsa/</guid><description>How HSA’s software guidelines decide what counts as a medical device and which class it falls in. A plain reading for administrators and nurse leaders before they sign a proposal.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;When a vendor says its AI tool is “registered with HSA”, or that it needs no registration at all, an administrator has no easy way to judge. The Health Sciences Authority (HSA) publishes two documents that answer most of the first questions: GL-04, its life-cycle guidelines for software medical devices, and GL-07, which explains how software is classified by risk.&lt;/p&gt;
&lt;h2 id=&quot;what-counts-as-a-software-medical-device&quot;&gt;What counts as a software medical device&lt;/h2&gt;
&lt;p&gt;GL-04 (revision 4, effective 31 December 2025) applies to software whose intended use meets the definition of a medical device under the Health Products Act 2007. That includes software meant for medical purposes such as investigating, detecting, diagnosing, monitoring, treating or managing a medical condition, disease, anatomy or physiological process. It covers software embedded in a device and software that runs on general computing platforms, including standalone mobile applications, which are often called software as a medical device (SaMD).&lt;/p&gt;
&lt;p&gt;The key word is “intended use”. GL-07 quotes the regulations: it is the objective purpose reflected in the specifications, instructions and information the product owner provides. Two products with the same model can differ in status because of what the manufacturer says they are for.&lt;/p&gt;
&lt;h2 id=&quot;how-the-class-is-decided&quot;&gt;How the class is decided&lt;/h2&gt;
&lt;p&gt;GL-07 (revision 2, July 2025) follows an international framework. It asks two questions. How significant is the software’s information for care: does it treat or diagnose, drive clinical management, or only inform it? And how serious is the patient’s situation: critical, serious or non-serious?&lt;/p&gt;
&lt;p&gt;In its table, software that treats or diagnoses in a critical or serious situation is Class C. Software that only informs clinical management in a serious or non-serious situation is Class A. A footnote moves some software up to Class B: software meant to analyse or monitor a vital physiological process, or anatomical structure images, to drive clinical management. GL-07 gives an example of Class A: software that lets healthcare professionals predict the risk of migraine from patient inputs such as sleep and stress.&lt;/p&gt;
&lt;p&gt;GL-04 adds that Class B, C and D software medical devices require product registration. Software of every class, including Class A, must be safe and perform as intended throughout its life cycle.&lt;/p&gt;
&lt;h2 id=&quot;decision-support-is-not-automatically-exempt&quot;&gt;Decision support is not automatically exempt&lt;/h2&gt;
&lt;p&gt;GL-07 also has a section on clinical decision support software. It sets out how such software qualifies as a regulated medical device or not. A vendor’s claim that its tool is “only decision support” is therefore something to test against that section, not a reason to stop asking.&lt;/p&gt;
&lt;h2 id=&quot;three-questions-for-the-vendor&quot;&gt;Three questions for the vendor&lt;/h2&gt;
&lt;ol&gt;
&lt;li&gt;Do you treat this product as a medical device in Singapore, and on what intended use is that based?&lt;/li&gt;
&lt;li&gt;If so, which class is it, and what are its HSA registration details?&lt;/li&gt;
&lt;li&gt;Does our planned use match the intended use you state?&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;Ask for the answers in writing and attach them to the contract. Registration does not replace the hospital’s own testing: the Ministry of Health’s AI guidelines say the hospital usually validates a tool in its own workflows.&lt;/p&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Software whose intended use meets the medical device definition falls under HSA’s guidelines, whether embedded or standalone.&lt;/li&gt;
&lt;li&gt;The class follows the significance of the information and the seriousness of the situation; Class B, C and D software needs product registration.&lt;/li&gt;
&lt;li&gt;Decision support software is not automatically outside the rules.&lt;/li&gt;
&lt;li&gt;Ask for status, class and registration details in writing. This is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/409/26808a93-6a7a-4551-8c66-13046c6124c5/gl-04-r4-regulatory-guidelines-for-software-medical-devices---a-life-cycle-approach-(2025-dec)-pub.pdf&quot;&gt;HSA, GL-04-R4 Regulatory Guidelines for Software Medical Devices&lt;/a&gt;, &lt;a href=&quot;https://isomer-user-content.by.gov.sg/409/ae814059-a3ff-4ce3-8192-4534a43ddc6f/gl-07-r2-guidelines-risk-classification-samd-cdss-(2025-jul)-pub.pdf&quot;&gt;HSA, GL-07-R2 Risk classification of SaMD and qualification of CDSS&lt;/a&gt;, &lt;a href=&quot;https://isomer-user-content.by.gov.sg/3/6636a140-7986-43f6-9861-096333138d59/AIHGle_2.0.pdf&quot;&gt;Ministry of Health, AIHGle 2.0&lt;/a&gt;. Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item><item><title>Machine-learning medical devices in Singapore: what HSA expects from vendors</title><link>https://hospital-ai-lab.com/in/blog/2026-11-14-machine-learning-medical-device-what-hsa-expects-vendor/</link><guid isPermaLink="true">https://hospital-ai-lab.com/in/blog/2026-11-14-machine-learning-medical-device-what-hsa-expects-vendor/</guid><description>HSA’s software guidelines set extra expectations for machine-learning devices: a quality system, model documentation and instructions for use. A buyer’s reading for administrators.</description><pubDate>Sat, 14 Nov 2026 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;A vendor brochure usually says what a model can do. HSA’s guidelines say what a vendor of a machine-learning-enabled medical device has to put in writing. Reading that list helps a hospital decide what to ask for in a proposal. The document is GL-04 (revision 4, effective 31 December 2025), section 9.&lt;/p&gt;
&lt;h2 id=&quot;a-quality-system-around-the-whole-model-life&quot;&gt;A quality system around the whole model life&lt;/h2&gt;
&lt;p&gt;HSA says that all activities for designing, developing, training, validating, retraining and deploying a machine-learning-enabled medical device must be managed under a quality management system based on ISO 13485. This covers retraining too, so a vendor that retrains its model should be able to show how that step is controlled.&lt;/p&gt;
&lt;h2 id=&quot;what-the-registration-file-describes&quot;&gt;What the registration file describes&lt;/h2&gt;
&lt;p&gt;For pre-market registration, the guidelines list additional information. It includes the intended use and indication for use, a description of the ML feature and its clinical association with the condition, the model’s inputs and outputs, the clinical workflow during deployment, and how the model was trained. A vendor should describe the model used, show that it suits the intended purpose, and explain its limitations and how shortcomings are mitigated.&lt;/p&gt;
&lt;h2 id=&quot;what-the-instructions-for-use-must-tell-you&quot;&gt;What the instructions for use must tell you&lt;/h2&gt;
&lt;p&gt;In addition to ordinary labelling, the guidelines say users must be told:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;that the device contains machine learning, and what the ML function does, including its inputs and outputs;&lt;/li&gt;
&lt;li&gt;how to install, use, calibrate, validate locally and monitor performance on an ongoing basis;&lt;/li&gt;
&lt;li&gt;which hardware and software versions it works with;&lt;/li&gt;
&lt;li&gt;the clinical workflow of the device, including any human intervention;&lt;/li&gt;
&lt;li&gt;the risks and limitations of the ML function, and any exclusion criteria in the training and validation data.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Those items map neatly onto questions a hospital can ask: where was it validated, and on whom; what does our local validation need to look like; and where must a human step in?&lt;/p&gt;
&lt;h2 id=&quot;after-go-live&quot;&gt;After go-live&lt;/h2&gt;
&lt;p&gt;The guidelines say companies must monitor the real-world performance of all registered machine-learning-enabled devices after deployment, and that developers and distributors should set up processes with implementers and users to ensure traceability. For devices that continue to learn, the guidelines add controls on the learning process, rollback and traceability.&lt;/p&gt;
&lt;h2 id=&quot;key-takeaways&quot;&gt;Key takeaways&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Retraining and deployment of an ML-enabled device sit inside an ISO 13485-based quality system.&lt;/li&gt;
&lt;li&gt;Registration and the instructions for use should describe the model, its inputs and outputs, the workflow, its limits and the training data exclusions.&lt;/li&gt;
&lt;li&gt;Vendors must monitor real-world performance after deployment; hospitals are part of that loop.&lt;/li&gt;
&lt;li&gt;Ask for these in writing. This is general information, not legal advice.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;&lt;em&gt;Sources: &lt;a href=&quot;https://isomer-user-content.by.gov.sg/409/26808a93-6a7a-4551-8c66-13046c6124c5/gl-04-r4-regulatory-guidelines-for-software-medical-devices---a-life-cycle-approach-(2025-dec)-pub.pdf&quot;&gt;HSA, GL-04-R4 Regulatory Guidelines for Software Medical Devices&lt;/a&gt; (sections 9.1 to 9.3). Read on 14 November 2026.&lt;/em&gt;&lt;/p&gt;</content:encoded><category>News</category></item></channel></rss>