Photo: Phil Boorman/Getty Images
University of Hong Kong researchers are developing an AI-powered chatbot for training healthcare professionals in facilitating end-of-life communications at scale, ahead of new legislation making advance medical directives legally binding across the city of Hong Kong.
The chatbot being developed by a research team at HKU Li Ka Shing Faculty of Medicine (HKUMed) will be based on the Advance Care Planning Communication Assessment Tool (ACP-CAT), a training and assessment framework that the team recently tested and validated in real-world clinical settings.
Research lead and HKUMed clinical assistant professor Dr Jacqueline Yuen Kwan-yuk told Mobihealth News that the chatbot is designed to train students and physicians in advance care planning (ACP) communication.
"This is a conversational AI simulator that enables students and clinicians to practice ACP discussions in realistic scenarios with real-time feedback. The chatbot simulates patient responses and evaluates communication quality using criteria derived from our validated ACP-CAT framework, assessing behaviours, including information sharing, eliciting patient preferences, responding to emotions, and shared decision-making."
"The chatbot can also be used for providing assessment and performance evaluation of ACP communication," she added.
It will be trained on a database of de-identified real-world ACP conversation transcripts and will require "many iterations and refinements" to create simulated conversations "on par with real conversations."
"Over the next 24-36 months – subject to securing necessary resources and partnerships – we will design realistic conversation scenarios, build and train the AI platform, and conduct testing and validation first with students, and later with clinicians," Dr Yuen shared.
WHY IT MATTERS
Hong Kong is facing a critical shortage in palliative care capacity as it prepares to implement the legislation making advance medical directives (AMDs) legally binding in May.
As of 2025, the city has roughly 40 palliative medicine specialists serving a population of 7.5 million — equivalent to about one specialist for every 150,000 to 180,000 people, well below international benchmarks. While global recommendations call for at least 1.5 specialists per 100,000 people, Hong Kong's current ratio stands at 0.6.
The shortage of trained ACP facilitators, Dr Yuen notes, is "even more acute", with systematic communication training remaining limited across both undergraduate and postgraduate medical education: few structured postgraduate programmes exist to build ACP communication competency among practising clinicians, while continuing medical education pathways for end-of-life communications remain constrained, she claimed.
This workforce gap is expected to create multiple system pressures as the legislation takes effect, according to Dr Yuen.
One immediate risk, she said, is an implementation bottleneck, with fewer trained clinicians to conduct high-quality ACP discussions for a growing patient population that requires such conversations.
Quality variability also remains a concern; in the HKUMed study's assessments of ACP consultations, for example, surrogate decision-makers were discussed in fewer than 5% of cases, non-medical priorities explored in only 30%, and treatment plans explicitly linked to patient values in under one-third of conversations.
Legal and ethical exposure is another emerging risk, Dr Yuen said, as ACP conversations with inadequately trained clinicians may result in AMDs that do not fully reflect patient preferences.
"Our AI-enabled training approach directly addresses these risks by rapidly building workforce capacity in a standardised, scalable, and evidence-based manner," Dr Yuen said.
She stressed, though, that this effort "requires strategic investment in research and development to ensure clinical validity and safety."
Digital enablement, particularly AI simulation, will be critical to scaling ACP training that effectively supports the implementation of the AMD legislation this year, according to Dr Yuen.
She said traditional apprenticeship models cannot reach the thousands of clinicians requiring upskilling within the policy's implementation timeline.
AI platforms, she explained, could provide 24/7 practice access without faculty supervision, standardised training across institutions, and personalised feedback at scale, while giving clinicians a psychologically safe environment to rehearse complex end-of-life conversations.
Dr Yuen added that digital delivery also addresses cost and workforce constraints tied to resource-intensive simulated patient programmes and faculty-led workshops.
However, she warned that failure to scale training alongside legislative rollout could create system-level risks, including "superficial compliance," where AMDs are legally valid but clinically hollow or unmeaningful, as well as moral distress among providers who are unprepared for emotionally complex conversations.
Poor-quality ACP discussions may also require multiple consultations, adding pressure on already stretched palliative care services, while procedural implementation could fuel public scepticism and ultimately undermine ACP uptake, she said.
THE LARGER CONTEXT
The ACP training platform is based on the recently validated ACP-CAT framework developed by the same HKUMed research team led by Dr Yuen.
It was used to assess 137 actual ACP conversations between doctors, patients and their families, collected from five hospitals and a community hospice. HKUMed noted from the study's findings, published in Palliative Medicine, that the framework "reliably measured the quality of clinician communication and corresponded with how patients and families perceived these interactions."
The research team is exploring other pathways to embed the ACP-CAT framework into routine clinical workflows besides education.
In the near term, Dr Yuen said healthcare institutions could deploy ACP-CAT as part of quality improvement and peer-review audits, sampling recorded ACP consultations to identify communication gaps, surface training needs, and track improvement over time. The framework could also be integrated into EHR systems through structured documentation templates, she added.
Dr Yuen also sees AI-assisted ACP-CAT scoring supporting objective measurement of communication quality in programmes where ACP quality is tied to reimbursement or accreditation. Additionally, she said, chatbots can help patients explore their care preferences, values, and questions before having ACP conversations.
Ultimately, according to Dr Yuen, digital integration must be balanced with ethical and governance safeguards, noting that any recording or AI-enabled analysis would require robust informed consent and full compliance with Hong Kong’s Personal Data (Privacy) Ordinance.
AI, she also emphasised, should augment, and not replace, clinical judgement and the deeply personal nature of end-of-life conversations.
Meanwhile, Singapore has been promoting ACP in recent years. In August, it launched a free online tool, called myACP, which guides users in documenting their preferences for end-of-life care. It is available to Singaporeans aged 21 and over, and with no existing serious conditions like cancer or dementia.
Besides AI, virtual reality has also been used to train professionals in ACP communication. In 2023, Tokyo-based Jolly Good co-developed a VR module in palliative care communication with Harvard-affiliated Brigham and Women’s Hospital.
