Yooseok Lim

Hi! I am currently working as a researcher at Seoul National University Hospital.

I am focused on advancing healthcare and medical AI. Recently, I have been working on solving sequential decision-making problems and developing AI models that are robust to diverse sensor signals. I believe these methods are powerful tools for developing advanced medical solutions that can meaningfully contribute to improving society.

I received an MS in Industrial and Information Systems Engineering from Soongsil University in South Korea in 2024. Fortunately, I was advised by Prof. Sujee Lee (currently at SKKU). I am currently working with Prof. Hyun-Lim Yang, and Prof. Byoungjun Jeon.

Email  /  GitHub  /  Google Scholar  /  CV

News

Publications

Large Language Model-augmented Offline Reinforcement Learning Framework for Sepsis Management in Critical Care

MORE-CLEAR is a multimodal offline RL framework that integrates clinical notes and structured data using LLMs, gated fusion, and cross-modal attention, achieving better survival estimates and policy performance than single-modal methods.

A Reinforcement Learning Framework for Personalized Anticoagulation Dosing in Critical Care: Integrating Batch-Constrained Policy Optimization and Off-Policy Evaluation

This study presents an RL-based decision-support framework for heparin dosing in the ICU, employing Batch-Constrained Q-Learning (BCQ) with rigorous off-policy evaluation. Trained and validated on MIMIC-III and MIMIC-IV, the learned policies outperform clinician strategies and align with therapeutic aPTT targets, establishing a foundation for reliable RL-based systems in critical care.

OMG-RL: Offline Model-based Guided Reward Learning for Heparin Treatment

Applying RL to medication dosing is challenging due to the reliance on narrowly defined reward functions and the diversity of clinical drugs. we propose Offline Model-based Guided Reward Learning (OMG-RL), an Inverse RL (IRL) framework that infers clinicians’ therapeutic intentions and derives an optimal heparin dosing policy.

Mapping Interconnectivity of Digital Twin Healthcare Research Themes through Structural Topic Modeling

These days, the integration of digital twins and helathcare is a particularly interesting ares. This study aims to identify key research trends related to application of digital twins to healthcare using structural topic modeling.

Multimodal sensor fusion models for real-time exercise repetition counting with IMU sensors and respiration data

A deep learning model is proposed to count exercise repetitions using multimodal data from smart earbuds, integrating IMU sensors and respiratory audio. Trained on 30 exercise types with data augmentation and interpolation, it enables accurate tracking for personalized fitness plans and improved health outcomes.

Machine learning-derived model for predicting poor post-treatment quality of life in Korean cancer survivors

Many cancer survivors experience poor quality of life (QOL) even after completing treatment. In this study, we developed predictive models using machine learning (ML) to identify poor QOL in post-treatment cancer survivors in South Korea.

In Progress

Assessing Senior Driver Safety through Trajectory-Level Inverse Reinforcement Learning: A Multidimensional Evaluation for Conditional Licensing

To address the rise in traffic accidents caused by the growing older driver population, this study propose an Inverse Reinforcement Learning-based framework for multidimensional and quantitative assessment of driving competence.

Trustworthy ICU Care: Enhancing Clinical Safety via Out-of-Distribution Detection-Augmented Offline Reinforcement Learning

An uncertainty-aware reinforcement learning framework is introduced to quantify policy uncertainty and mitigate unsafe dosing behaviors that arise in OOD patient states. Our method detect high-risk recommendations with strong accuracy, enabling clinician overrides and enhancing the safety and interpretability of ICU treatment policies.

Generalizable Repetition Counting via Metric-Based Few-Shot Learning on Wearable Motion Signals

We propose a universal exercise repetition counting method that analyzes IMU signals through a deep metric-based few-shot learning framework. The model generalizes to unseen exercises with high accuracy, demonstrating robustness for real-time applications in fitness and healthcare.

Education

Industry Projects

Development of an AI Model and Service Concept Based on Canine Activity Data

Development of a Course Recommendation System Based on Natural Language Data

Development of a Counting Algorithm Based on Human Activity Data

Personal Projects

3D Medical Segmentation

This project implements 3D medical segmentation algorithms.

Traffic Sign Detection

This project tackles traffic sign detection using multiple models across two datasets. It emphasizes cross-dataset evaluation to assess how well a model generalizes to different data distributions, addressing the challenge of distributional shifts.

Autonomous Driving Temperature Measurement Robot

Development of an autonomous driving robot using computer vision, reinforcement learning, and edge computing technologies

Design and source code from this site