A groundbreaking advancement in the application of artificial intelligence to thyroid cancer diagnosis has been unveiled by an interdisciplinary team of researchers from the University of Hong Kong’s LKS Faculty of Medicine (HKUMed), the InnoHK Laboratory of Data Discovery for Health (InnoHK D24H), and the London School of Hygiene & Tropical Medicine (LSHTM). This pioneering AI model distinguishes itself as the world’s first capable of accurately classifying both the cancer stage and risk category of thyroid cancer with an accuracy exceeding 90%. The system combines cutting-edge natural language processing technology with extensive clinical data analysis to redefine how clinicians approach this complex disease, ultimately promising to enhance diagnostic precision and profoundly reduce the time required for pre-consultation preparation.
Thyroid cancer, a prominent malignancy globally and within Hong Kong, is traditionally managed through a dual-system approach that relies heavily on manual integration of clinical information. The widely accepted American Joint Committee on Cancer (AJCC) Tumour-Node-Metastasis (TNM) system stratifies cancer by its pathological stage, while the American Thyroid Association (ATA) provides a risk classification framework crucial for prognostic evaluations and treatment planning. Despite their importance, these systems demand meticulous review and interpretation of multifaceted medical records, often resulting in a time-intensive process for healthcare professionals and leaving room for human error.
The innovation presented by the HKUMed-led team harnesses the power of large language models (LLMs), sophisticated AI frameworks capable of interpreting human language with remarkable nuance and contextual understanding. By adapting models such as ChatGPT and the newly introduced DeepSeek, the research team developed an AI assistant designed to parse complex clinical documents including pathology reports, operation records, and clinical notes. This AI leverages deep learning techniques to extract critical entities and information, bridging the gap between unstructured textual data and actionable clinical insights.
Central to the model’s development was the integration of four open-source LLMs—Mistral AI’s Mistral, Meta’s Llama, Google’s Gemma, and Alibaba’s Qwen. Unlike proprietary online models, these offline LLMs allow for local deployment, an essential factor in maintaining patient data privacy and complying with stringent health data regulations. Training occurred using pathology reports from 50 thyroid cancer patients sourced from The Cancer Genome Atlas Programme (TCGA), a well-regarded open-access database, followed by rigorous validation against an extended cohort of 289 TCGA cases alongside 35 meticulously crafted pseudo cases generated by experienced endocrine surgeons, ensuring robustness and clinical relevance.
Remarkably, the AI assistant’s fusion of outputs from all four language models elevated its performance to notable levels, achieving accuracy rates between 88.5% to 100% in ATA risk classification and between 92.9% to 98.1% for AJCC cancer staging. These figures compare favorably to manual chart reviews and highlight the system’s potential as a transformative clinical tool. Beyond accuracy, one of the most impactful outcomes of this technology is its capability to reduce clinicians’ preparatory workload by almost half, streamlining clinical workflows and enabling more focused patient interactions.
Professor Joseph T Wu, Sir Robert Kotewall Professor in Public Health and Managing Director of InnoHK D24H, emphasized the AI model’s dual advantage: high precision combined with offline operation. By enabling local analysis of sensitive clinical data, the AI solution prioritizes patient confidentiality without sacrificing technological sophistication—a critical balance in today’s healthcare landscape. This offline capability ensures that hospitals and clinics can adopt the system without concern for data breaches or regulatory hurdles associated with cloud-based solutions.
Further comparative analyses highlight the AI assistant’s competitive edge. Tests employing a “zero-shot approach” compared the model against recent versions of DeepSeek (R1 and V3) and GPT-4o, both leading online language models renowned for their vast training datasets and computational power. Impressively, the HKUMed AI model matched these high-caliber systems in performance, an achievement that underscores its engineering excellence and adaptability within resource-constrained environments.
Dr Matrix Fung Man-him, Clinical Assistant Professor and Chief of Endocrine Surgery at HKUMed, underscored the tangible clinical benefits rendered by the AI platform. The model not only excels in parsing intricately detailed pathological and surgical documentation but also condenses the interpretive burden on surgeons and endocrinologists. By delivering concurrent results for cancer stage and risk stratification based on internationally recognized frameworks, it provides a comprehensive clinical picture faster and with greater accuracy.
The versatility of the AI system hints at its broad applicability. Both public institutions and private healthcare providers, locally and internationally, stand to benefit from deploying this technology, which seamlessly integrates into existing clinical infrastructures. Dr Fung expressed optimism that the model’s real-world implementation will translate directly into enhanced efficiency for clinicians, improved quality of care for patients, and increased opportunities for physicians to focus on patient counseling and treatment planning rather than administrative burden.
Aligned with the Hong Kong Government’s commitment to leveraging AI in healthcare, as exemplified by recent developments like the LLM-based medical report writing system introduced by the Hospital Authority, the research team is preparing for subsequent phases. These involve large-scale validation using expansive, real-world patient data sets to ensure robustness and generalizability. Upon successful testing, rapid deployment into hospital systems and clinical workflows is anticipated, heralding a new era of AI-assisted medicine that could redefine operational and therapeutic efficiency.
The research team responsible for this breakthrough reflects a confluence of expertise spanning public health, clinical medicine, and family medicine research. Led by Professor Joseph Wu Tsz-kei, Dr Matrix Fung Man-him, and Dr Carlos Wong King-ho, the collaboration also includes first authors Dr Eric Tang Ho-man and Dr Tingting Wu. Such multi-disciplinary cooperation, under the auspices of HKUMed and supported by initiatives like the Hong Kong Jockey Club Global Health Institute and the Innovation and Technology Commission’s InnoHK program, exemplifies the integrative approach necessary for modern medical innovation.
The InnoHK Laboratory of Data Discovery for Health (InnoHK D²4H), spearheading the project, embodies a bold vision for precision medicine. They aspire to harness unparalleled data resources and apply frontier analytics to safeguard global health while advancing individualized medical care. By fostering collaborations across scientific disciplines and sectors, InnoHK D²4H positions itself at the forefront of transforming healthcare technology in Hong Kong and beyond, striving toward ambitious goals with wide-reaching implications for worldwide disease management.
With an article slated for publication in the prestigious journal npj Digital Medicine, this research heralds a promising intersection of artificial intelligence and cancer diagnostics. As thyroid cancer remains a critical public health challenge, innovations like this AI model offer a beacon of hope for more efficient, accurate, and privacy-conscious clinical practices that could set new standards for patient care around the world.
Subject of Research: Not applicable
Article Title: Developing a named entity framework for thyroid cancer staging and risk level classification using large language models
News Publication Date: 1-Mar-2025
Web References:
https://www.nature.com/articles/s41746-025-01528-y
http://dx.doi.org/10.1038/s41746-025-01528-y
References:
Wu, J. T., Fung, M. M-h., Wong, C. K-h., Tang, E. H-m., Wu, T., et al. Developing a named entity framework for thyroid cancer staging and risk level classification using large language models. npj Digital Medicine (2025). DOI: 10.1038/s41746-025-01528-y.
Image Credits: The University of Hong Kong
Keywords:
Thyroid cancer, Public health, Clinical research
Tags: advancements in cancer risk assessmentAI model for thyroid cancer diagnosisAmerican Joint Committee on Cancer TNM systemAmerican Thyroid Association guidelinescancer stage classification AIHKUMed research advancementsInnoHK Laboratory innovationsinterdisciplinary research in medicinenatural language processing in healthcarepre-consultation preparation efficiencyreducing diagnostic time in cancerthyroid cancer accuracy over 90%
