Deep learning and tensor methods for medical time-series

Yang, Chaoqi

Permalink

https://hdl.handle.net/2142/124254 Copy

Description

Title

Deep learning and tensor methods for medical time-series

Author(s)

Yang, Chaoqi

Issue Date

2024-04-12

Director of Research (if dissertation) or Advisor (if thesis)

Sun, Jimeng

Doctoral Committee Chair(s)

Sun, Jimeng

Committee Member(s)

• Abdelzaher, Tarek F.
• Tong, Hanghang
• Westover, Brandon M.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Medical Time-Series
• PyHealth
• Healthcare
• Deep Learning
• Tensor Methods

Abstract

Time-series data are common in healthcare AI research, and they exhibit diverse modality and complex characteristics, usually leading to various problems in modeling. One large category of medical time-series is the Medical Sequence Data. For example, electronic health records (EHR) are documented on a daily basis, containing all the interactions between patients and clinicians. However, it can be in various formats across different clinical institutes. Medical claims data reflects the patient billing, prescription, vaccination, insurance activities, which could be important for understanding the population level disease effects, such as COVID. However, data processing and modeling could be time confusing on various types of claims data where missing values could exist to add more complexity. Another category of medical time-series is called the Medical Signal Data, like electroencephalogram (EEG). These physiological signal data are collected in high frequency by smart sensory devices that requires advanced signal processing techniques. Usually, sample format mismatching could create a gap between similar datasets and prevent jointly training for large-scale models. My PhD researches mainly focus on building new methodologies for medical time-series on four different problem paradigms: (i) Future Target Prediction that leverages the historical information to predict some future targets of interests, such as heart failure onset prediction; (ii) Underlying Class Prediction that predicts the current disease class or health stages based on the symptom progression or other temporal phenomena; (iii) Missing Imputation that estimate the missing elements based on the surrounding (location-wise) or nearby (time-wise) observed values; (iv) Unsupervised Feature Extraction that learns meaningful embeddings from large unlabeled time-series data. My proposed methods could benefits the health- care domain by providing better and safe prescriptions, making disease prediction model more generalizable to new patients, understanding the spatio-temporal distributions of pediatric respiratory syncytial virus (RSV) disease, accurately estimating the missing values in streaming multi-dimensional medical data, and more. Apart from building strong models, I have also devoted a great amount of time building the health AI system, named PyHealth (https://github.com/sunlabuiuc/PyHealth), which provides various supports for medical time-series (and other modalities), including flexible data processing, medical code mapping if any, easy AI model application and evaluation. The package has attracted increasing attention since 2022 and has gained 868 stars and 27k downloads up to now.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Chaoqi Yang

Owning Collections