Prediction of Postoperative Complications in Cardiac Surgery

1.

Prediction of Postoperative
Complications in Cardiac Surgery
Master Thesis
Dina Zverinski
Supervisors: Prof. Thomas Hofmann, Dr. Carsten Eickhoff, Dr. Alexander Meyer

2.

Outline
Problem Definition
Data Set
Methods
Results & Discussion
Conclusion
2

3.

Problem Definition
3

4.

Motivation
Huge amounts of data collected at the
intensive care unit (ICU)
High workload for the ICU staff
Harder to recognize postsurgical
complications
Early recognition can lower the risk of
late complications
No clinical real-time decision support
system
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
4

5.

Postoperative Bleeding
Coagulation Problems:
Surgical Bleeding:
Bleeding due to non-clotting
Unstaunched bleeding
Treatment: transfusion (blood products)
Treatment: transfusion at first, if no
improvement, surgical re-exploration
Early recognition can be crucial
Hard to distinguish at the beginning!
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
5

6.

Problem Statement
Predicting the need for surgical re-exploration due to
postoperative bleeding in real-time.
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
6

7.

Related Work
Electronic Health Records (EHRs) for prediction
Mortality prediction in real-time at the ICU
Methods: e.g. logistic regression, deep learning
Risk factor analysis of surgical bleeding
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
7

8.

Data Set
8

9.

Patients
Bleeding patient: surgical re-exploration within 25 hours after initial surgery
Control group: no surgical re-exploration after initial surgery
All initial surgeries are open heart surgeries
Adult patients only (18+)
3650 patients in total (50% bleeding patients)
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
9

10.

Features
Continuous or categorical
Static features: e.g. age, gender, initial surgery type, …
Dynamic features: e.g. bleeding rate, blood pressure, laboratory results, …
72 features in total
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
10

11.

Time Slices
Time window: end of initial surgery until start of surgical re-exploration
Time slice: feature vector (one per half an hour) labelled with its patient’s class
69996 time slices in total
Missing values imputed with:
last measured value
default value
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
11

12.

Representation
a
a
Time Slice Representation
Sequence Representation
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
12

13.

Methods
13

14.

Clinical Baseline
Decision in favor of a surgical re-exploration, if the bleeding rate is
> 400 mL/h for 1 hour
> 300 mL/h for 3 hours
> 200 mL/h for 4 hours
Otherwise, no surgical re-exploration needed
from: Robert M. Bojar. Manual of Perioperative Care in Adult Cardiac Surgery. John Wiley & Sons, 2005.
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
14

15.

Machine Learning Approaches
Naive Bayes
AdaBoost (Decision Trees)
Logistic Regression
Support Vector Machines (SVM)
K-Nearest Neighbors (KNN)
Feedforward Neural Network (FNN)
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
15

16.

Recurrent Neural Network (RNN)
x: input
s: hidden state
o: output
U, V, W: weight matrices
Figure from Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015.
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
16

17.

Results & Discussion
17

18.

Evaluation Metrics
Accuracy:
N: number of negative time slices
TP + TN
P+N
ROC AUC: area under the true positive vs.
false positive rate curve
TP: number of true positive time slices
Precision:
TN: number of true negative time slices
Recall:
FP: number of false positive time slices
F1 score:
P: number of positive time slices
TP
TP + FN
TP
P
FN: number of false negative time slices
2*
precision * recall
precision + recall
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
18

19.

Results
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
19

20.

Accuracy
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
20

21.

Different Feature Sets
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
21

22.

Possible Time Savings
Given actual time s until reexploration and the first time f RNN
predicts re-exploration, the relative
saved time d is defined as:
<
Per-Patient-Specificity:
number of true negative patients
number of negative patients
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
22

23.

Problem Complexity and Limitations
Ground truth unknown
Real-time prediction
Missing or incorrect data
Coarse temporal resolution
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
23

24.

Conclusion
24

25.

Conclusion
All approaches perform significantly better than the clinical baseline
RNN performs with
accuracy of 0.818
ROC AUC of 0.889
F1 score of 0.802
RNN could help decrease the time until re-exploration by up to 65%
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
25

26.

Thank you!
26

27.

Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
27

28.

ROC Curve
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
28

29.

Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
29

30.

Distribution of Patients
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
30

31.

Likelihood
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
31

32.

Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
32

33.

RNN Classification Options
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
33

34.

Likelihood
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
34

35.

Feedforward Neural Network (FNN)
Final model:
Hidden layers: 1
Hidden nodes: 20
Activation function: sigmoid
Regularization: L2-norm
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
35

36.

Recurrent Neural Network (RNN)
Final model:
Hidden layers: 1 (GRU)
Hidden nodes: 40
Activation function: sigmoid
Problem Definition - Data Set - Methods - Results & Discussion - Conclusion
36