Abstract
Data mining models for high-cost patient encounter prediction at the point-of-admission (HPEPP) in inpatient wards are scarce in literature. This is due to lack of availability of relevant features at such an early stage of treatment. In this study, we create a disease co-occurrence network (DCN) using a subset of the State Inpatient database of Arizona. We explore this network for community formation and structural properties to create new input features for HPEPP models. Tree-based data mining models are trained using input feature sets including these new network features and distinct disease communities in the DCN are identified. We propose community membership and high-cost propensity scores as two network based features for HPEPP modeling. We compare the performance of models with different input feature sets and find that the new features significantly improve the accuracy sensitivity and specificity of prediction models. This model has the potential to improve targeted care management and reduce health care expenditure.
Original language | English (US) |
---|---|
Journal | IEEE Journal of Biomedical and Health Informatics |
DOIs | |
State | Accepted/In press - Dec 12 2017 |
Keywords
- Data models
- Diabetes
- disease co-occurrence networks
- Diseases
- healthcare data analysis
- high cost patients
- Hypertension
- Informatics
- point of admission predictive modeling
- Predictive models
ASJC Scopus subject areas
- Biotechnology
- Computer Science Applications
- Electrical and Electronic Engineering
- Health Information Management