Thyroid cancer AI assistant developed in Hong Kong

A large language model-powered AI assistant developed in Hong Kong has demonstrated high accuracy in thyroid cancer staging and risk classification.

A team of researchers from the Li Ka Shing Faculty of Medicine of the University of Hong Kong (HKUMed), the InnoHK Laboratory of Data Discovery for Health, and the London School of Hygiene and Tropical Medicine conducted the study which built what could be the world’s first AI assistant for classifying thyroid cancer stage and risk categories.  

FINDINGS

The AI model leverages four open-source LLMs, namely Mistral by French startup Mistral AI, Meta AI’s Llama model, Google’s Gemma, and Qwen by China-based Alibaba Cloud, to analyse free-text clinical documents, including clinical notes, pathology reports, and operation records. 

It provides cancer staging and risk classification based on the widely used 8th edition of the American Joint Committee on Cancer’s (AJCC) TNM cancer staging system and the American Thyroid Association (ATA) classification system.

The model was trained with and validated against open-access pathology reports from The Cancer Genome Atlas Programme. It was also validated against some 35 pseudo-cases created by endocrine surgeons.  

Based on findings published in npj Digital Medicine, the AI assistant achieved overall accuracy of 92.9%-98.1% in the AJCC cancer staging and 88.5%-100% in the ATA risk classification. 

“We conducted further comparative tests with a ‘zero-shot approach’ against the latest versions of DeepSeek – R1 and V3, as well as ChatGPT-4o. We were pleased to find that our model performed on par with these powerful online LLMs,” added the study’s lead, HKUMed professor Joseph Wu Tsz-kei.

WHY IT MATTERS

Cancer staging and risk classification are done to guide treatment decisions and predict patient survival. Usually done manually, this task can take much time, the research team said, and so they started developing the AI assistant.

Considering its high accuracy, researchers suggest that the AI tool could help cut the time clinicians spend on pre-consultation preparation by half. 

Prof Wu also shares that they integrated offline capability into their AI assistant to allow its deployment without the need for sharing or uploading sensitive patient information.

“The AI model is versatile and could be readily integrated into various settings in the public and private sectors, as well as local and international healthcare and research institutes,” added Dr Matrix Fung Man-him of HKUMed, who also led the study. 

The research team now plans to further validate their AI assistant with a larger real-world dataset before it can be deployed in hospitals and other clinical settings.

THE LARGER TREND

There have been innovations in Hong Kong recently that have also leveraged large language models and generative AI to enhance the efficiency of disease diagnosis and management. 

Early this year, HKU engineers introduced their genAI-based system for label-free tumour imaging, which they proposed as a cost-effective way to do single-cell analysis. 

Over at the Chinese University of Hong Kong, engineers have integrated DeepSeek into a blood pressure management system, which could scale its rollout, especially in rural and remote areas, as it does not require costly equipment.