Skip to main content
eScholarship
Open Access Publications from the University of California

Multimodal Analytics for Healthcare

  • Author(s): Torres, Carlos
  • Advisor(s): Manjunath, Bangalore S
  • et al.
Abstract

The ailing healthcare system demands effective autonomous solutions to improve service and provide individualize care, while reducing the burden on the scarce healthcare workforce. Most of these solutions require a multidisciplinary approach that combines healthcare with computational abilities. The work presented in this thesis introduces a multimodal multiview network along with methods and solutions that leverage inexpensive visual sensors and computers to monitor healthcare. One of the most prominent outcomes of this work includes enabling the medical analysis of ICU conditions such as sleep disorders, decubitus ulcerations, and hospital acquired infections, which are preventable and negatively affect patients' health population. The problems tackled include patient pose classification, pose motion analysis and summarization, role representation and identification, and activity and event logging in natural hospital settings. These problems are addressed via a non-intrusive non-disruptive multimodal multiview sensor network (Medical Internet-of-Things). The multimodal data is combined with coupled-optimization to estimate source weights and accurately classify patient poses. Pose patterns such as pose transitions are represented using deep convolutional features and pose duration is modelled via segments. The proposed techniques serve to differentiate between poses and pseudo-poses (transitory poses) and create effective motion summaries. The role representation is tackled using novel appearance and semantic interaction maps to assign generic labels to individuals (doctor, nurse, visitor, etc) without using identifiable information (e.g., facetracking or badges), which is prohibited in healthcare applications. Finally, activity and event analysis is tackled using a new contextual aspect frames where aspect bases and weights are learned and then used to reconstruct activities. The objective of this thesis is to enable the development, evaluation, and optimization of individualized therapies, standards-of-care, room infrastructural designs, and clinical workflows and procedures.

Main Content
Current View