are not eligible for heart transplantation. Thus, a possibility is to receive a circulatory device known as a

ventricular assist device (VAD). This kind of device shows some problems, like thrombogenesis. The thrombus

formation in a VAD can cause patient death, and a previous, non-invasive diagnostic is quite complex. The

objective of this work is to develop an algorithm to reproduce time signals that indicate the presence and

absence of a thrombus, use these signals to train an artificial neural network to classify them, and use these

algorithms in a predictive algorithm for early thrombus detection. The results show that it was possible to detect

the thrombus formation in its early stages, but the noise level interferes with the accuracy of the ANN,

especially when signals in the time domain are used.

Key-Words: - Ventricular assist device, thrombus, artificial neural network, signal analysis, predictive

Affecting millions of people around the world,

cardiovascular diseases (CD) are responsible for the

first cause of death and hospitalization, making

them a serious public health problem, [1]. Many

patients suffering from Congestive Heart Failure

(CHF) need heart transplantation, although many of

them will die until they receive the donated heart,

[2]. To minimize this death rate, some of them

receive a circulatory device known as a ventricular

exposed to any surface not fully covered by

endothelium, blood flow inside the pump, and

others, [6]. One of the most common therapies to

avoid thrombus formation in VAD is the

prescription of anticoagulants that will dissolve the

thrombus, [7]. However, this therapy will be

effective only if the patient starts it in the first stages

Fig. 2: Thrombus formation in a removed VAD.

Adapted from, [9]

disability or even cause the patient's death, [3],

[10]. Immediate action is necessary, but in the

current scenario is indecisive about which treatment

needs to be carried out, [11]. The thrombus

formation and its consequent release into the

Based on data from an experimental study that

performed tests with artificial thrombus in a pump

prototype, the objective of this work is to develop a

Python language algorithm to reproduce simulated

signals obtained by that study that indicate the

disturbances caused by the presence of artificial

thrombus adhered in three different regions of the

for their study, and Figure 4 shows the areas where

the thrombosis was simulated. The material used to

simulate the thrombus had a density of 0.97 g/cm2,

and the liquid used to simulate the circulatory

smoother the curve representing the state change.

The value used for ti was 1000. In plotting the

results, the argument cmap=jet was used,

representing a color palette with smooth transitions

from blue to red. Thus, the higher the amplitude, the

redder the coloration presented.

by a fully connected Feed Forward ANN

architecture. Four classes and four labels for data

classification (Table 1) were chosen. To train and

test the ANN, the dataset was divided as follows:

The number of epochs was 4000 and the optimizer

used to reduce overall loss was, [19]. The metric

used to check performance was accuracy. To assess

the noise sensitivity of the ANN, it was tested and

trained with signals from both the time and

frequency domains, with the background noise

being gradually increased. The best-performing

ANNs in terms of accuracy sensitivity will be used

This model utilizes the best ANNs with the dataset

generated by the SRA to predict the probability of

comprehensive analysis of the temporal variations in

the probabilities of thrombus presence or absence

and was designed to better simulate real-life

applications.

point of the region, indicating the absence of a

thrombus, adjacent to the region indicating

thrombus presence. In each iteration, the thrombus

probability is calculated for this evolving data range,

and the calculated probabilities are stored. At the

algorithms. To facilitate comprehension, these

results have been divided into the following four

subsections.

The spectrum pattern representing the absence of

thrombi is characterized by peaks at the

fundamental frequency of 30 Hz and its respective

harmonics (60, 90, and 120 Hz). The presence of a

thrombus at the base is identified by a peak near the

progressive increase in the amplitude of this

frequency, directly related to the appearance of the

thrombus at the rotor's base. At a certain moment,

the areas corresponding to the frequencies of 90 and

120 Hz began to become less visible. The

the rotor’s vanes and the rotor’s spiral are presented

in Figure 8 (b) and (c), respectively. In both cases,

we observe an intense peak at 90Hz growing with

time. For the thrombus in the rotor’s vane, Figure 8

To demonstrate the capability of the ANN to predict

thrombus formation with a small frequency

signature, the output probabilities are presented in

Figure 10 for three thrombus positions. In all

scenarios, illustrated throughout the iterations,

where the iterations represent the degree of mixing

of signals with the absence and presence of

thrombus, the probabilities of thrombus absence and

presence vary. As expected, the probabilities always

to thrombus absence, and the remaining 70% are

signals of thrombus presence.

pump, where more effective monitoring and control

systems can be incorporated into VADs, making

them more efficient, as well as suggesting the

feasibility of implementing prescriptive

maintenance strategies. By analyzing the data, the

Technology of São Paulo for their support

throughout the research process.

References:

[1] G. Renugadevi, G. Asha Priya, B. Dhivyaa

[3] J. I. Glitza, F. Müller-von Aschwege, M.

[14] T. D. Rossing, Shock and Vibration

[17] G. M. Khan, ‘Artificial neural network

[18] S. Biswal and G. R. Sabareesh, ‘Design and