NETWORK ANALYTICS AND MACHINE LEARNING: APPLICATIONS TO ONLINE SOCIAL NETWORKS AND HEALTHCARE SYSTEMS
Open Access
- Author:
- Sung, Yi Shan
- Graduate Program:
- Industrial Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 15, 2017
- Committee Members:
- Soundar Rajan Tirupatikumara, Dissertation Advisor/Co-Advisor
Soundar Rajan Tirupatikumara, Committee Chair/Co-Chair
Guodong Pang, Committee Member
Tao Yao, Committee Member
Reka Z Albert, Outside Member
Kamesh Madduri, Outside Member
LiYing Cui, Special Member - Keywords:
- network analytics
machine learning
healthcare analytics
online social networks
dominance attribute
community detection
provider order recommendation system
length of stay
remaining length of stay
classification
prediction - Abstract:
- Data proliferation, referring to the explosion in data generation by organizations, leads us to spend considerable efforts on managing the data and retrieving the useful information from such huge amounts of data. This drives researchers to explore efficient approaches for analyzing the data and deriving insights to make better decisions. In this thesis, we develop four different decision support models based on network analytics and machine learning methods to solve four different research problems in online social networks and healthcare systems. In most of the real-world networks such as Facebook networks, each node usually contains multiple attributes representing the node’s characteristics. It is difficult to identify the dominant attributes, which have determining effects on community formation. In this thesis, we disclose the association of node attributes to the community topology by defining dominance ratio and applying correlation metrics. The method is tested on Facebook data of 100 universities for uncovering how the offline lives infer online friendship construction. Healthcare systems are known to have huge amounts of data due to recording patient clinical history. In this thesis, we develop three different recommendation and prediction models using network-based methods and machine learning to support healthcare providers in making daily operational decisions. The first model, SuperOrder, is an automated recommendation system for outpatient clinics to predict what medical orders the providers would like to place for the upcoming appointments. The implementation of SuperOrder will increase order effectiveness and ease order documentations. The second model is to predict hospital length of stay (LOS) for patients who are admitted through emergency departments (ED). The LOS model can help ED physicians make better decisions on hospitalization at the time of admission, as well as reduce ED boarding time and solve ED overcrowding problem. The third model is to predict the remaining length of stay (RLOS) for general adult inpatients to determine which patients can be discharged from hospitals today or tomorrow. Unlike the LOS model which helps in admission decisions and demand forecasting, the RLOS model supports in discharge planning and bed availability forecasting. All of these three healthcare prediction models are expected to be implemented in Geisinger Health System and integrated into their daily workflows to support providers’ daily decisions.