CSCD-SOCR Demonstration of Visualization of High-dimensional Diabetes Data using TensorBoard

This CSCD Resource demonstrates:

1. t-distributed stochastic neighbor embedding (t-SNE) statistical method for manifold dimension reduction,
2. The TensorBoard machine learning platform, and
3. Hands-on Big Data Analytics using the SOCR Diabetes Case-Study,
4. Interactive Visual Analytics of user-provided data.

Load the data as a TSV file of vectors, e.g., 4 cases in 5D:
0.1 0.2 0.3 0.4 0.5
0.2 0.3 0.4 0.5 0.6
0.2 0.3 0.4 0.5 0.6
0.2 0.3 0.4 0.5 0.6

See these examples of the CSCD Diabetes Case-Study Data and the SOCR Country Ranking Data



Optionally, load a TSV file of labels/metadata, e.g., 4 cases in 2D (metadata). If there is more than one column, the first row must include column names.
Class Phenotype
A Patient
A Control
B Control
C Patient

See these examples of the CSCD Diabetes Readmission Outcome Labels and the SOCR Top-30 Country Rank Indicator Labels. More details about these Country Ranking data are available here.

System Description

• System Objective: To provide a scalable, distributed, highly-functional, collaborative, and portable webapp for interrogating complex biomedical and healthcare data.
• Design/Approach/Technology: This architecture utilizes modern machine learning and artificial intelligence techniques (e.g., DSPA), powerful algorithms (e.g., t-SNE), and robust architecture (e.g., TensorFlow/TensorBoard) to deliver a user-friendly interface to advanced back-end computational resources.
• System Status: Fully functional system that provides an embedded data analytic case-study (diabetes) and allows the user to import their own high-dimensional data using the specified data-import format (TSV).

Study

• Purpose/Objective: Do patient phenotypes (e.g., race, gender, and age), clinical settings (e.g., admission type, time in hospital, medical specialty of admitting physician), and treatment regiments (e.g., number of lab test performed, HbA1c test result, diagnosis, number of medication, diabetes medications, number of outpatient, inpatient, and emergency visits in the year before the hospitalization) affect diabetes treatment outcomes? Examples of specific driving healthcare challenges that can be addressed include:
  • Develop risk predictive models using available risk factors associated with 30-day readmission rate of diabetes patients.
  • Is polypharmacy (number of medications received) importance as a risk factor for hospital readmission?
  • What are the combination of health conditions (comorbidity) that greater impact the hospital readmission rate?
  • Is there a relationship between comorbidity and diabetes severity as measured by HbA1C level.
• Case-Study: This diabetes case-study includes a 10-year span (1999-2008) of clinical care observations of 130 US hospitals and integrated delivery networks. There are over 50 features representing patient phenotypes and hospital outcomes related to diabetes care. The original data includes over 75,000 encounters with over 50 biomedical and health measures per encounter. Following a preprocessing step, the data complexity is reduced to 24,084 encounters with 45 features. Data attributed, references, acknowledgments, inclusion/exclusion criteria, data dictionary, and additional information is available here.
• Methods: We first employed data science and predictive analytics (DSPA) data wrangling methods to preprocess the data and generate a computable data object. Then, we use linear (PCA) and non-linear (t-SNE) dimensionality reduction methods to project the high-dimensional data into 2D or 3D space. Finally, we employ visual and exploratory data analytics to interrogate the low-dimensional projection, identify clusters of patients, and explore the intrinsic lower-dimensional structure of the data.
• Results & Conclusion: We find strong associations between derived computed phenotypes (clusters) in the low-dimensional projections of the original data and the actual diabetes (30-day) readmission outcomes. This open-science project provides a low-cost solution for interactive visual analytics, hypothesis generation, and pattern identification for complex biomedical and healthcare case-studies. It does not require any special software or licensing and supports evidence-based discovery science and provides semi-automated clinical decision support for health practitioners.

Acknowledgments