Page 123 - AIH-2-3

Artificial Intelligence in Health Opportunities for AI-based arrhythmia screening

Subsequently, subtracting the mean signal from the T U r (XXI)

signal data stream holds: x

w w ˘ 11 t (XVII) In this case, the amplitude of T (τ) at specific moments
,, ., 1
’
xφ
U r in relative time (τ) directly indicates the degree of cross-
Representing the alternating signal free of drift. correlation agreement between the template and the
measured signal. The amplitude maxima are located where
Applying these matrix multiplication procedures to the match is highest.
a healthy baseline signal will generate the template for I
analytical computations using the matched filter method. As the data compression ratio (1− J ) increases, the
If the healthy signal is unavailable for the patient being estimated matches of the cross-correlation function
investigated for pathological cardiac conditions, an become more susceptible to noise. When all these steps are
alternative approach is to compute an average from a broad implemented in the signal processing pipeline for the
dataset of healthy individuals within the same boundary temporal data acquisition of the 12-lead vectorial ECG
conditions as the patient. The more closely the boundary signal, the continuous cross-correlation process enables
conditions match, the more effective the analytical the identification and isolation of matches and miss-
template will be in the matched filter approach, helping matches within subsets of the input signal over time. These
to identify deviations within the repetitive full PQRST(U) subsets can then be analytically compared against matched
pulse trains. filters representing any of the over 30 pathological
In the matched filter approach, the correlation that conditions related to cardiac rhythm abnormalities

reveals the optimal estimates for the unknowns shift, ξ (Table 2). Following this, a root-cause analysis is conducted

with respect to ξ , and the unknown scaling factor, η with to determine the underlying electrochemical and
0
respect to η , is implemented by utilizing the probing electrophysiological factors contributing to the observed
0
sequence or template (Η ), defined by the dot product, or ECG deviations. These factors may include cell damage,
ξ
-k
scalar product, of the two respective vectors: malfunction, and congenital predisposing that disrupt
normal cardiac function.
  2
ξη⋅ = arg min ( x − η ⋅Η ) = In the AI and machine learning framework, the
ξη , ∑ k k ξ −k
arg max  ∑ Η v − η 2 Η ξ −  2 2 (XVIII) computational processing of the ECG data stream begins
ξη, k ξ ∑ k k −k k  with a comparison of the temporal pattern against a
baseline normal ECG template (Η ). Subsequently, the
ξ
-k
It then follows the removal of noise (with an average signal is compared against templates corresponding to the
value of 0), and the reversal of the sign of the equation, pathological conditions outlined in Table 2. This machine
converting the minimum into a maximum. This results learning-driven process aims to identify the highest
in a least-squares minimization within the domain of the likelihood of a pathological match, such as a diseased
data stream. By applying the Cauchy-Schwarz inequality, left ventricle, as illustrated in Figure 6. A probability
the two estimators reduce to, for the offset: distribution is computed, summarizing the degree of
 agreement between the acquired signal and various
i v i 2 Constant (XIX) templates. This distribution highlights both the interval
matches and specific pathological conditions, providing a
And for the scaling factor: 40-44 comprehensive diagnostic overview for physicians. 44-47 The
physician then evaluates these findings in the context of
v H the patient’s background and may consult colleagues for
 k k
k (XX) a second opinion. Based on this thorough evaluation, a
H 2
k k tailored treatment plan is developed.
An estimation is introduced into the correlation between
the compressed cardiac muscle depolarization signal and 4. Results and discussion
a compressed template (i.e., the average depolarization Long-term ECG recordings are often analyzed manually,
pattern complex in the compressed domain) using a relying on the expertise of a physician, typically an
matched filter template. The process employs the impulse electrophysiologist, to identify episodes of arrhythmogenic
response, which is equated to the temporal inversion of the cardiac depolarization or sequences of depolarization
root base signal description, φ(τ). The computed output pulses that indicate the development or progression of
can be represented by Equation XXI: a pathological cardiac condition. These findings guide

Volume 2 Issue 3 (2025) 117 doi: 10.36922/aih.8468