Upload
ganesh-n-prasad
View
45
Download
0
Embed Size (px)
Citation preview
Information Visualization for Population Health Management via
Cognitively-Guided Disease Risk Assessment
Rema Padman, PhD1, Robert Monte, RPh, MBA2, Gayathri Hebbar2, Ganesh N. Prasad1 1Carnegie Mellon University, Pittsburgh; 2VA Pittsburgh Health System, Pittsburgh, PA
Introduction
Assessing and responding to many patients’ risk of diabetes related complications are complex, high-dimensional
information processing problems faced by time-constrained clinicians. Innovative algorithms and tools which
combine statistical machine learning, information visualization and electronic health data may reduce clinicians’
information processing load and improve their ability to assess risk of disease onset and related complications. A
critical element in visualization is the incorporation of flexibility in customizing assessments to the needs of unique
patient populations. This study presents preliminary results on evaluating computationally driven visualization
techniques for improving risk assessment using high dimensional data on 8,611 patients with diabetes.
Methods and Data
Statistical and machine learning methods commonly used for dimensionality reduction are applied to find
informative two-dimensional projections and classify patient data composed of arbitrary numbers of variables that
are relevant to diabetes-related risk assessment [1]. Included in this step are the identification of appropriate data
normalization procedures, disparate measurement of the data attributes, procedures for overlaying decision
boundaries that provide stratification into risk groups, attracting anchor points and specifications for plotting them
that contextualize predicted risk with important risk factors, and use of color and/or shape to highlight patient groups
or important risk factors as interpretable elements in the visual models. Early results with a simple data set show that
the framework may generate models which visually classify a patient population with accuracy comparable to
common statistical methods [1]. In this study, we apply this method to a data set with 8,611 patients and 35 variables
comprising demographic and clinical data related to diabetes management from a large, integrated health system to
obtain multiple insights at the population, individual and intervention levels that may facilitate point-of-care risk
prediction, stratification and exploration of optimal interventions.
Results
A brief descriptive analysis of the data indicates that
97% of the population is male, with average age of
68 years, high systolic (165) and diastolic (94) blood
pressure, but with A1C (6.9), LDL (89) and HDL
(43.5) near acceptable thresholds. Figure 1 shows a
separation of high risk patients (above the decision
boundary) from those at low risk, and anchored by
the relevant risk factors for the data set in [1]. The
size and location of the risk factors provide insights
into the critical drivers of risk, such as smoking and
BMI. This analysis has also been extended to explore
risk predictions for individual patients and impact of
a specific intervention on modifying their risk level.
Figure 1. Population-level, contextualized, binary stratification of risk of heart attack in patients with diabetes.
Conclusion
We demonstrate that an integrated risk assessment and visualization tool that displays contextualized risk for
diabetes related complications at multiple levels could be a powerful educational and disease management tool that
may benefit multiple stakeholders, including clinicians and patients.
References
1. Harle C, Neill D, Padman R. Development and evaluation of an information visualization system for chronic
disease risk assessment. IEEE Intelligent Sys. 2102:27(6):81-85.