6.3 Signals of the Cardiovascular System

|

279

(a) Control group (CG): Vascularly healthy subjects. No stenoses in the aorta or the

pelvic arteries, diameter of the aorta in the entire thoracic and abdominal course

< 40mm (except aortic sinus).

(b) Aneurysm (A): detection of a thoracic or abdominal aortic aneurysm with a dia-

meter of ≥45mm.

Six photoplethysmography probes of the measuring device were attached to the pa-

tient who is in supine position, two on the temples, two on the thumbs, and two on

the toes (see Figure 6.47).

The position of the sensors was optimized using the single signal of the respect-

ive channel. ECG electrodes were placed. Measurement was started as soon as clear

signals were visible in all channels. The raw signals were recorded with a sampling

rate of 2048 Hz. The patient was measured in quiescent condition (no physical strain

15 min before the measurement, no meal or smoking 1h before the measurement) and

the room temperature was fixed to 23 °C to ensure that there is normal blood flow in the

peripheral parts of the body. The room was darkened to avoid perturbation by other

light sources.

A cut-out of the raw signals are shown in Figure 6.47, the raw signals undergo the

following pre-procssing:

1.

the spikes at the start and the end of the measurement due to the removal of the

sensors have been removed by cutting the signal between 5 s and 55 s, leaving us

with 50 s of data

2.

the baseline was corrected by subtracting a signal filtered by a sharp lowpass filter

at 0.5 Hz

3.

the signal was filtered with a sharp filter at 30 Hz (digital filter that does not shift

phase). This was done due to the removal of low and high frequency noise.

4.

the signal was scaled such that the amplitude of the first harmonic is 1 because

the absolute amplitude is assumed not to be relevant due to large differences in

positioning the sensor and skin thickness etc.

The resulting band limited signal of frequency range 0.5 Hz - 30 Hz and a signal length

of 50 s serves as basis for further computation. Based on a these pre-processed PPG

signals from different locations, the above diagnostic problem of aneurysms is de-

scribed as a classification problem. For all feature extraction methods described be-

low, we split the signal in five 10 s intervals and perform an average on the resulting

features. This is beneficial if e.g. a single extrasystole would be present, which could

throw off the transfer function computation. Subsequently, the characteristic Fourier-

coefficients were calculated from these pre-processed signals and specific harmonics

of the transfer-functions between two measuring locations were used as features for

a subsequent classification of the aneurysm (A) and control group (CG).

These classification problems are a special case of machine learning which is in

turn a part of artificial intelligence. With the help of machine learning, IT systems are