Medical education in Asia may be headed to a future where institutions are not only training students but also training AI models.
In the HIMSS24 APAC session, “Implementing AI in Healthcare,” Dr Hyung-Chul Lee, deputy CIO of Seoul National University Hospital (SNUH), shared how their hospital develops, validates, and implements AI technologies.
“For more than 100 years, our hospital has been educating medical students and residents. But now we also have to train LLMs (large language models).”
When LLMs eventually made their way to healthcare recently, researchers at SNUH quickly validated their potential application. While LLMs scored high in one of the world’s toughest medical examinations and outperformed existing predictive AI models in such use cases as predicting in-hospital mortality and real-time admissions, their ability remains limited.
“Our research shows that no LLMs have yet scored higher than 60 [in the Korean Medical License Exam], especially in medical law.”
“Also, these LLMs have limited ability [in performing] multimodal tasks. For example, if I uploaded images related to ventricular tachycardia – a condition requiring emergency resuscitation – from VitalDB (a multi-parameter vital signs database), even the [latest] ChatGPT-4o [wrongly] answers that it’s a normal ECG and no resuscitation is required.”
Acknowledging the need for LLMs to get fed with more diverse data, Dr Lee and his team built a vector database and a model for fine-tuning search. “We are currently working on vector embedding of all data in our hospital and loading it into the VectorDB.”
This initiative led to the development of SNUH’s in-house LLM platform to safely develop AI models for their clinicians. The LLM service, Dr Lee said, required 40 H100 GPUs and six petabytes of storage.
“VectorDB will be our textbook and fine-tuning it will be [in] our curriculum.”
“I look forward to seeing AI models trained at our hospital and [how these] will change our practice and improve patient outcomes.”