Research Seminar with Clinical Data Science

Schedule

LECTURE 1: 12:00 - 12:30

Speaker: Sander Puts 

Subject: Enhancing ICD-10 Coding with GPT-4, RoBERTa, and RAG: An AI-Powered Approach

Abstract:
This presentation explores the integration of GPT-4, RoBERTa, and Retrieval-Augmented Generation (RAG) to improve ICD-10 medical coding. Using data from the CodiEsp-X challenge, we developed a method that combines RoBERTa for term extraction with GPT-4 and RAG for assigning ICD-10 codes based on code descriptions. The approach shows potential in increasing both accuracy and efficiency. Key methods, results, and future steps for AI-assisted coding will be discussed.

LECTURE 2: 12:30 - 13:00

Speaker: Dennis Soemers (DACS)

Subject: A Kaggle Data Science Competition about dozens of algorithms playing over a thousand games.

Abstract: 
Monte-Carlo Tree Search (MCTS) is a popular search algorithm for autonomous game-playing agents. The literature on artificial intelligence (AI) for games includes dozens if not hundreds of different strategies that researchers have proposed for various subroutines of the algorithm. Experiments in publications generally show such newly proposed strategies to outperform baselines when averaging results over small sets of different games. 

However, we have a poor understanding of which MCTS variants work well or poorly in which types of games. We have created a large dataset of outcomes from a total of 72 different MCTS agents playing over a thousand different board games, including a wide array of features to characterise the games themselves. 
In collaboration with Kaggle, we are running a data science competition that challenges participants to predict the outcomes from pairs of different agents playing games, featuring a prize pool of $50,000. We hope that well-performing models will help to uncover insights into the relations between features of games, configurations of agents, and performance levels of agents, fundamentally improving our understanding of these algorithms and their strengths and weaknesses.