Genetic_Programming_Theory_and_Practice_XIII

84 B. Hodjat and H. Shahrzad

Ta b l e 1 Example of a
decision list (top)andadata
package (bottom), and the
evaluation of a rule on a data
package

Number Condition1 Condition2 Action

1 ifx 1 .T/<10and x 2 .T1/ > 20 thenlD 0

2 ifx 2 .T/<10and x 2 .T3/ > 20 thenlD 1

3 ifx 1 .T1/ < 10 thenlD 0

T x 1 .T/ x 2 .T/ l

0 10 20

1 10 32

2 9 30

3 8 20 0

The example shows how the current Time,TD 3 is applied to

each rule. The first rule evaluates the first clause withx 1 at the

current time to true. The second clause of the first rule,x 2 at

current time1,T 1 D 2 is also true. The first rule applies

labellD 0 as the action, which matches the label forTD 3.

The first clause in rule number 2 comparesx 2 atTD 3 ,which

is false. The second clause comparesx 2 atT 3 D 0 ,whichis

again false (this second clause does not have to evaluated given

the first one is false). Thus, the rule takes no action. The third

rule comparesx 1 at timeT 1 D 2 andistrueandtakesthe

actionlD 0 , which is correct. When there is more than one

prediction the current heuristic for choosing the action is to

take the first prediction, in the same manner as a decision list.

Thus, the action from rule number one,lD 0 , is predicted,

which is correct. If no rules are true then the action will be

‘Null’, which is always incorrect

CD

0

B

B

@

011

500 0 1

600 1 0

600 1 1

1

C

C

A

Each data sample includes a time-series of 100 labeled events. The quality of
predictions is aggregated and normalized by the candidate’s age to calculate the
candidate’s fitness at any given time.
We demonstrate the efficacy of the approach on roughly 4000 patients Arterial
Blood Pressure (ABP) waveforms from MIMIC II v3. In MIMIC Waveform records
available are sampled at 125 Hz (125 samples/s) (Goldberger et al. 2000 ) and ABP
is recorded invasively from one of the radial arteries. The raw data size was roughly
1 TB. The labels in the data are imbalanced, the total number of Low event are just
1.9 % of the total number of events. In total we had 45,693 EC-Star data packages
from 4414 patient records. Of these, we used 32,898 packages with 100 events each
as the training set. 12,795 samples are used for the unseen set.