AIGolem

AIGolem

Video
Project / product name: AIGolem
Team leader: Stanislav Jirák
Challenge: 8. MediDetect AI: Czech Diagnosis Challenge
Problem: Detect diagnosis, detect other features and analyse from written unstructured patient records. Do not use model that stores prompt or other data outside hospital used to analyze.
Solution: We use paid GPT-4 turbo model via openai API, which allows programmatically fine tune queries and do not store prompt for other learning (which ChatGPT do store prompts for future learning). (1) We build web UI/python backend that prompts to openai directly for short prompts with 10-20 records max. (2) However, we also prepared scripts and trained model that analyse 50000 records in advance based on expected usecase to classify the records - the result can be used to train offline AI models.
Impact: Do clever analysis. Use AI assistant for unexpected or hard research, which was almost impossible before. Unstructured data can be classified or interpretted in non-english language. Potential assist and detect other diagnoses and risks for future monitoring. Model can standardize form of records.
Feasibility & financials: Current GPT4-turbo API can be used to train the 50000 or more records - in terms of max 10 USD + GCP 1-10USD, It also can add prompt accesibility for each query cost 0.01 USD. However it took 2 hours for trained person to classify 20 records. Thus spare 2000 hours man time. Our solution is enterprise ready.
What is new about your solution?: Use GCP for scripting and control GPT4-Turbo to classify/train on large data(50000 recods) as an input for smaller model for offline data model. Use chat prompt GPT4-Turbo with data security ensured to do general queries on smaller chunk of records (up to 30).
What you have built at the hackathon - text explanation + code (e.g. GitHub link): We have learned backend frontend to control openai gpt4-turbo model and GCP to do notebook script for preparing large data and clasify them.
What you had before the hackathon, please mention open source as well: Only general knowledge of some tools and api. Knowledge of GCP, ...
What comes next and what you wish to achieve: Develop a platform where doctors can chat to a database of patient records, and invidual medical records. Implement early health risk warning, and suggest possible medical conditions that has not been diagnoses by a doctor. Monitoring and observability. Solid backend and user-frindly UI. Open-source LLM model fine-tuned for medical data.