Can AI patients improve medical training?

A partnership between The Chinese University of Hong Kong (CUHK) and technology company Votee AI offers a glimpse of what scientific collaborations between universities and industry might look like in an AI age.

Votee AI and the Nethersole School of Nursing, part of the Faculty of Medicine at the CUHK, have worked together to build an AI-driven virtual elderly client for training social workers in the field of acceptance and commitment (ACT) therapy. 

The technology, which allows students to practise scenarios with a virtual client in real time, took home a gold medal at the 51st International Exhibition of Inventions in Geneva earlier this year.

CUHK co-developed the AI layer of the tool, training the virtual client system using academic literature, approved clinical simulation frameworks, anonymised research datasets and synthetically generated dialogues developed under expert medical supervision.

The way it works is that the virtual client presents scenarios and the system examines whether the responses of the students (learning to counsel) align with ACT principles.

Connie Chong, associate professor at the Nethersole School of Nursing, said that AI-simulated clients can save medical schools the cost, in time and money, of employing human actors, and also help to save time.

In higher education globally, a scarcity of placements, supervisors and time to practise continues to be a challenge for ACT training, argued Chong.

Using the virtual client in supervisions means that “clinical sites can be reserved for what only a real clinic can teach”, Chong continued.

AI simulation in health and social care education is not entirely novel in UK settings: Coventry University announced the development of a virtual patient with AI firm PCS last year. But the difference is the language medium: Votee AI, which is a for-profit company entirely independent of the university, has built the world’s first Cantonese LLM, a language spoken by 85 million people. 

Jack Ng, head of corporate communications at Votee AI and former assistant director at Hong Kong University (HKU), argued that this makes it different from other frontier models like ChatGPT or Gemini.

“You cannot train a Cantonese-speaking mental-health worker on an English-speaking virtual patient,” Ng explained.

This year, increasing attention has been concentrated on the linguistic inequalities of AI model development and infrastructure, and various UN agencies have cautioned about possible repercussions. 

In July, a report by the UN’s Independent International Scientific Panel on AI warned that: “Artificial intelligence leaves most languages behind. Although more than 7,000 languages are spoken worldwide, current AI models are trained and optimized for only a small fraction of them.”

Cantonese, in particular, is a language that presents unique challenge for AI technologies, because there is a lot of “code switching”, according to Ng.

Code switching is when speakers mix languages within a single sentence; in this case, combining Cantonese grammar with English professional terms. 

LLMs can struggle to capture those “nuances”, Ng said, citing slang, colloquialisms and idioms as areas where an LLM might struggle.

Votee AI has began to collaborate with other regional academic institutions in East Asia. The company signed a memorandum of understanding (MOU) with HKU’s Faculty of Medicine on student internships, and is working on an MOU with the Singapore government’s national AI programme to build LLMs for other local languages including Indonesian, Vietnamese and Thai.

Leo Ma, chief scientist (responsible for Asia-Pacific) at Votee AI and former visiting research fellow at University of Oxford, said that patient data and privacy are huge considerations for medical students using a virtual assistant, with the immediate concern for a university being data protection and ensuring that student voice recordings, assessment scores and internal clinical scenarios are not sent to commercial cloud providers.

Meanwhile, on the side of patient privacy, any underlying clinical case data used to inform training scenarios should be secure since some of the clinical simulation scenarios are constructed using consensual real case studies or medical transcripts.

Ma said that Votee AI’s LLM runs entirely on CUHK’s own infrastructure, which means the solution is deployed on the premises of the institution.

“There are a few good things about this,” he continued. “One is the security so that there’s no data leakage to outside.”

Deploying on premises also means clinical training materials remain in the university’s own infrastructure.

Nethersole’s Chong agreed that an on-premises design which “keeps that analysis, and the trainee’s performance data, inside the university rather than depending on an external language model”, is critical for any virtual assistants used in medical school settings. 

rosalind.skillen@timeshighereducation.com