Predictive models for pressure ulcers from intensive care unit electronic health records using Bayesian networks

Area of Science:

• Clinical Informatics
• Data Science in Healthcare
• Predictive Modeling

Background:

• Pressure ulcers (PUs) pose significant risks to intensive care unit (ICU) patients.
• Accurate identification of PU risk factors is crucial for effective prevention strategies.
• Electronic health records (EHRs) contain valuable data for understanding PU risk.

Purpose of the Study:

• To develop predictive models for understanding and exploring patient clinical data and risk factors for pressure ulcers in ICU patients.
• To construct Bayesian networks to predict PU incidence and elucidate the structure of related risk factors.
• To simplify Bayesian network nodes and edges using statistical network techniques.

Main Methods:

• A three-stage framework for predictive analysis of patient clinical data was implemented.
• Feature extraction from EHRs was developed with clinician assistance, followed by feature simplification.
• Bayesian network predictive models were constructed and evaluated using various algorithms, scoring functions, and feature sets.

Main Results:

• Bayesian network models were constructed from 86 medication, diagnosis, and Braden scale features using EHR data from 7,717 ICU patients.
• The model identified known and suspected high PU risk factors and increased prediction sensitivity nearly threefold compared to logistic regression models, maintaining overall accuracy.
• Clinician collaborators identified strong relationships between risk factors through model visualization.

Conclusions:

• The Bayesian network model offers a novel framework for significantly improving PU prediction sensitivity.
• Early and accurate prediction enables caregivers to respond proactively to conditions associated with PU incidence.
• This approach can reduce the adverse impact and high treatment costs associated with pressure ulcers.