Transcript Diagnosing Diabetes and Predicting Complications

Diagnosing Diabetes and
Predicting Complications
Electronic Health Records Data
From Duke Hospital (2007-2011)
16, 604 Diabetic Patients
13 Diabetes Outcomes & Dataset
Prevalence
Health Information
Medications
Laboratory Tests
Procedures/Diagnoses
Goals
Use exploratory clustering and visualization to
identify patient subsets with related
characteristics.
K-Means Clustering
TSNE High Dimension Visualization
Priya Sarkar
Lily Zerihun
Anqi Zhang
Mentor: Elizabeth Lorenzi
Advisor: Ricardo Henao, PhD
Demographic Information
Age (4% children)
Sex (59.86% female)
Race (56.29% POC)
Smoker (13.89%)
Cholesterol (total + HDL)
Blood pressure
Build cluster-specific
models to improve
performance in the
prediction of diabetes
complications.
Exploratory Clustering
Example of Clustering Method
Goal: Explore medications, labs, diagnoses, outcomes,
and demographics. Identify meaningful clusters of
similar patients. Explore the sources of similarities.
Exploring Top Medications
Clustering Lab Tests
TSNE Reduction
Amlodopine
(Hypertension)
Cluster 1
Kidney and
Heart Disease
Cluster 2
2
1
3
Patient Clustering: K-means
6
2
7
8
10
5
4
1
9
Cluster 3
Cluster 5
Blood Pressure in 10 Clusters
Clustering “Sickest” Pt
With > 100 Diagnoses
Cluster 1:
Cluster 2:
Lung Complications
Mycosis Fungoides
Pancreatic Cancer
Sprain and Strain
Lung Cancer
Prostate Cancer
Renal Failure
Joint/Shoulder Pain
2
1
Predictive Modelling
Comparing ROC Curves for 13 Complications
ROC Curve Plots Sensitivity and Specificity of
the outcome prediction. It plots the true
positive rate against the false positive rate.
Goal: Create and test the accuracy of a model to predict
diabetes complications based on medications, lab tests, and
diagnoses.
.
Comparing AUC values for outcomes predicted with
medications, laboratory tests, diagnoses, & combined models.
The AUC (area under curve) specifies the accuracy of our model in
predicting patients who will have a certain complication based on
their medications, laboratory tests, and/or diagnosis data.