Transcript BMI PowerPoint Presentation 48x36

Predicting Sepsis in the ICU using Dynamic Bayesian Networks
Anthony Wong,
1 Department
Introduction
1
MTech ,
Senthil K. Nachimuthu,
of Biomedical Informatics, University of Utah,
1
MD ,
2Intermountain
Peter J. Haug,
1,2
MD
Healthcare, Salt Lake City, Utah
Sepsis Temporal Model
Data
 Early diagnosis of sepsis is key for early
intervention and reducing mortality due to severe
sepsis or septic shock.
We analyzed retrospective data obtained from
Intermountain Healthcare’s ICUs.
 Clinicians often consider prior events and
temporal trends when they monitor their patients.
 Temporal relationships play an important role in
the decision making process but are often
overlooked when modeling the problem.
Data set: Clinical data from 2 ICUs
 January 2006 – February 2009
 100 randomly chosen adult patients from a total
of 3,336 (18 years old and above)
 6,469 temporal observations
 Dynamic Bayesian Networks (DBN) can provide
a generalized causal probabilistic framework that
can explicitly model temporal relationships
between clinical variables.
Discussion
Objectives
 DBN with K-Means L2 Norm discretization
performed slightly better. However, the area under
the ROC curves were not significantly different.

To design a temporal model using DBN for
predicting sepsis.
To analyze and evaluate the performance of
sepsis temporal models with different
discretization methods.
Methodology
Dynamic Bayesian Networks
We designed a Dynamic Bayesian Network
(HMM), to represent the causal and temporal
relationships found in sepsis.
Kevin Murphy’s Bayesian Network Toolkit (BNT)
and Projeny were the two main tools utilized in
this project.
Discretization with Clustering Technique
Clustering methods were used to discretize
continuous data in each observed field. 114,752
temporal observations were used for clustering.
We then compared two different parameters in Kmeans:
 Euclidean distance (L2 Norm)
 Manhattan (L1 Norm)
Inference
Parameters of the conditional probability table
(CPT) were obtained through Expectation –
Maximization (EM) method. Multiple time slices
were inferred using the Junction Tree Algorithm.
Results
 Discretization with K-Means L1 Norm seemed
to yield better sensitivity at higher cut-off.
ROC Curves
L1 Norm Discretization
Area under the ROC curve were
calculated using Trapezoidal rule:
L2 Norm Discretization
1
 DBN with L1 Norm: 0.52
Limitation
Due to the intensity of computation required for this
type of modeling, we could only perform the initial
test on a small set of patient data. We also
assumed that the diagnosis for sepsis is true for all
time slices in each patient.
0.8
 DBN with L2 Norm: 0.54
Confusion Matrix at 85% Cut-off (L1 Norm)
Positive Test
Negative Test
Sepsis
276
52
No Sepsis
61
30
Sensitivity

 Both sensitivity and specificity did not produce
satisfactory result.
0.6
0.4
Conclusion
0.2
Confusion Matrix at 85% Cut-off (L2 Norm)
Positive Test
Negative Test
Sepsis
289
39
No Sepsis
86
5
0
0
0.1
0.2
0.3
0.4
0.5
1-Specificity
0.6
0.7
0.8
0.9
1
Acknowledgements
Descriptive Statistics
Age (years)
Heart Rate (beats/min)
Respiratory Rate (breaths/min)
Systolic BP (mm/Hg)
Diastolic BP (mm/Hg)
Temperature (°C)
APCO2
White Blood Count
We have demonstrated that it is possible to
develop a temporal model using a DBN by
structuring the model using clinical knowledge.
Mean Standard Deviation Range Minimum Maximum
45.46
19.72
67
18
85
86.97
16.27
116
26
142
22.84
7.28
328
5
333
128.58
20.96
155
70
225
62.18
12.99
120
26
146
36.89
0.82
6.6
34.0
40.6
53.65
27.92
148.9
22.1
171.0
10.56
6.58
50.7
2.8
53.5
n
n (%)
100 100.00
6363 98.35
4510 69.71
4781 73.89
4781 73.89
2794 43.18
204
3.15
338
5.22
Data provided by Intermountain Healthcare.
Contact Information
Anthony Wong
[email protected]