202410021400 - After lunch - AI implementation to healthcare
[[2 Oct 2024 - HIMSS APAC Day 1]]
Speaker: Dr Lee Hyung-Chun Lee. Deputy Head, Office of Hospital Information, Seoul National University Hospital. Head, Department of Data Science Research, Innovative Medical Technology Research Institute. Head, Center for Data Innovation, National Strategic Technology Research Institute
![[Pasted image 20241012163555.png]]
Content
1. Development and validation
2. Application
3. LLM and LLM Models
## Developing AI software at SNUH.
![[Pasted image 20241012163622.png]]
Patient data are Pseudonymized. Allow use of data without consent from the original patients for research and scientific purpose.
![[Pasted image 20241012163642.png]]
Can decide approval for commercial / non-commercial project based on guidelines (Do we have this in SIngapore?)
![[Pasted image 20241012163703.png]]
**Clinical Data Warehouse**
Using this to build models. (Do we have such a big warehouse for patients in Singapore?)
![[Pasted image 20241012163718.png]]
For labeling clinical event, without accessing raw data in EMR.
(Something about researcher can access patient's EMR but without going into the actual system)
![[Pasted image 20241012163745.png]]
![[Pasted image 20241012163802.png]]
![[Pasted image 20241012163809.png]]
Access of data of Pseudonymized data - for research. Permission need to be granted.
![[Pasted image 20241012163819.png]]
SHINE network ? what it’s that?
Federation Network, sharing data
![[Pasted image 20241012163844.png]]
VitalDB Open Dataset (2022) -?Why do they release data to share with others?
Biosignal set.
![[Pasted image 20241012163855.png]]
Do we have such dataset for schizophrenia, mental illness patient data set?
![[Pasted image 20241012163930.png]]
Using all these data to develop AI tools. Such as there, Rings, and Smart Gadget. To measure blood pressure.
![[Pasted image 20241012163942.png]]
K-MIMIC Dataset.
![[Pasted image 20241012163950.png]]
![[Pasted image 20241012164000.png]]
![[Pasted image 20241012164010.png]]
Smart Apps @ SNUH Demo apps.
![[Pasted image 20241012164019.png]]
⁃ Developed internal apps - ICU - patient in ICU.
![[Pasted image 20241012164030.png]]
⁃ another platform, real data developed in JSON format.. or web monitoring.
![[Pasted image 20241012164040.png]]
⁃ Calculating with AI
Can be access by their researcher, .. many hospital use the vitalDB to get research on the platform.
![[Pasted image 20241012164056.png]]
ChatGPT already better than medical student knowledge.
![[Pasted image 20241012164111.png]]
## Limitations of Generalist AI models
⁃ no LLM scan higher then 60
![[Pasted image 20241012164120.png]]
⁃ LLM have limited to multi-model test,
That’s why they need to build a more narrow LLM.. Because of the training set..
![[Pasted image 20241012164129.png]]
Building a more specialist LLM?
![[Pasted image 20241012164144.png]]
Building their own in house LLM.
Will need GPU and national data Center.
![[Pasted image 20241012164154.png]]
⁃ how to train LLM? Now need to train not only medical student but LLM.
⁃ Vector DB is like their text book.
![[Pasted image 20241012164204.png]]
Training LLM model as if they are training medical student, sending LLM across for cross-organisational training. That’s interesting.
——-