Project 45: AI-Driven Time Series Prediction in Healthcare
Contact Information
Prof. Chenglin Li
Email: LCL1985@sjtu.edu.cn
Project Description and Objectives
Electronic Health Records (EHRs) have amassed vast Irregularly Sampled Medical Time Series (ISMTS) data, comprising multiple medical variables with distinct attributes and varied sampling patterns. Techniques like Ordinary Differential Equations (ODEs) address irregular sampling but often neglect variable-specific temporal patterns and dynamic inter-variable correlations, which are crucial for accurate predictions. Additionally, current models rely heavily on data-driven approaches, potentially overlooking valuable expert knowledge and limiting their effectiveness. Enhancing prediction performance also requires addressing multiple related clinical tasks simultaneously. Multi-Task Learning (MTL) facilitates knowledge sharing across tasks, boosting efficiency and accuracy in prediction. However, existing MTL frameworks depend on manual task grouping and model design, leading to inefficiencies and task interference.
This project aims to develop an intelligent medical time series prediction system by integrating advanced time series processing, knowledge-driven modeling, and automated multi-task learning to enhance clinical prediction accuracy and efficiency. Specific objectives include developing methods to handle varying temporal patterns and sampling intervals, incorporating medical expertise to understand relationships between variables for better prediction outcomes, and implementing an automated MTL framework to optimize training and reduce task interference. Achieving these objectives will improve the analysis and prediction of medical time series data, supporting precise clinical decision-making.
Eligibility Requirements
Basic knowledge of machine learning and deep learning, such as Transformer and Graph Neural Networks.
Proficiency in at least one programming language, such as Python, and familiarity with relevant machine learning frameworks (e.g., TensorFlow, PyTorch).
Main Tasks
Develop techniques to handle irregular and heterogeneous time series data, capturing unique characteristics of medical variables in ISMTS.
Incorporate medical expertise by extracting insights from medical literature and clinical texts to improve model interpretability.
Implement a multi-task learning framework to automate task grouping and optimize the model training process.
Website
School: http://english.seiee.sjtu.edu.cn/