Page 233 - Artificial Intelligence for Computational Modeling of the Heart

206 Chapter 6 Additional clinical applications

6.3.2.2 Deep learning based personalization
Setting up a large medical dataset is often infeasible due to
the costs. As a first step, a database of 10 000 synthetically gen-
erated data samples has been created, reflecting the anatomical
and functional variations representative of healthy and patholog-
ical cases. For the systemic circulation, the following input pa-
rameters were randomly chosen: heart rate, systolic blood pres-
sure, diastolic blood pressure, left ventricular end-diastolic vol-
ume, stroke volume, and left ventricular ejection time. The cycle-
averaged blood pressure was computed from the systolic and
diastolic blood pressures and from the heart. The left ventricu-
lar end-systolic volume was computed from the left ventricular
end-diastolic volume and the stroke volume. A similar approach
was chosen for the right ventricle and the pulmonary circula-
tion. Moreover, the sampling procedure was constrained by a set
of well-defined consistency rules to achieve physiological plausi-
bility [455], e.g., left and right ventricular similar stroke volume,
physiological pulse pressure, etc. The WBC model was then em-
ployed for computing the personalized output measures of inter-
est, thus creating a dataset of training examples.
Two distinct neural networks were defined, to predict time-
independent and time-dependent measures of interest respec-
tively. Following the standard approaches, 8000 randomly selected
data samples were used for training while the remaining 2000 are
used for testing [454]. All input parameters were used as features
for the network, leading to an input feature vector of 9 floating-
point values. For more stable and faster training, features were
rescaled to have the properties of a standard normal distribution
with mean 0 and standard deviation of 1.
A schematic representation of the overall modeling method is
shown in Fig. 6.12: the deep neural networks are trained offline on
the synthetically generated database, and are then applied to pre-
dict measures of interest as a function of patient-specific clinical
data.

6.3.3 Results and discussion
To assess the model performance, the mean absolute percent-
age error, which is typically employed in regression tasks evalu-
ation, is computed (see Table 6.4). Pearson correlation between
predictions and ground-truth was also considered to assess the fi-
delity of the AI model. (See Fig. 6.13.)
Important information for the patient condition can be re-
trieved by assessing the evolution in time of diverse quantities
within a heart cycle. One relevant example is the non-invasive