6.3 Signals of the Cardiovascular System

|

283

reliable way to implement it on top of an existing classification method. For the cross-

validation itself, it is always ensured that two measurements of the same patient are

not split between test and training set.

Due to the existing measurement devices and other limitations, the mean of the

accuracies for the classifications on correct labels is around 60%, whereas it is 50%

with randomly permuted labels, see Figure 6.49.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

accuracy

0

50

100

150

Normalised probability density

KNN, accuracy with correct label

KNN, mean of accuracy with

randomly permuted label

Fig. 6.49: Resulting accuracies with frequency response coefficients as features: The blue histogram

shows the distribution of accuracies achieved with 100 different partitions on correctly labeled data

as described in Figure 6.48. In comparison, the red histogram shows the distribution of the accuracy

of a classifier trained and evaluated on randomly permuted labels 100 times, where each time the

accuracy for 100 different partitions was calculated and averaged. (This figure is quoted from [25,

Figure 7].)

This proves the influence of aneurysms on the collected data, but the resulting ac-

curacies are still not suitable for technical purposes. The method can therefore detect

an aneurysm with 60% accuracy, actual results using four instead of two PPG signals

lead to 64%; this gives a first indication that can be further controlled by image pro-

cessing methods.

Listing 6.3.4: Matlab function for K-next neighbour classification.

function KNN(obj,previousMethod,data,labels,previousLog,previousModel)

% Tests whether the right parameters are given

...

% Use data all together

X = reshape([data.data],length(data(1).data),[])';

Y = [labels.data];