AI in Healthcare

AI IN
HEALTHCARE
Deep Learning Electrocardiographic
Analysis for Detection of Left-Sided
Valvular Heart Disease
01.
INTRODUCTION
INTRODUCTION
● Valvular heart disease (VHD) is a critical condition impacting
cardiovascular health.
● Its prevalence is rising and late-stage diagnosis leads to poorer
outcomes.
● Early and accurate diagnostic tools are crucial to improve
prognosis.
● Deep learning analysis of ECG shows potential for precise VHD
detection.
02.
HYPOTHESIS &
PURPOSE
Can the deep analysis of the ECG be the basis for
the development of a valvular heart disease (mild
to moderate AS, AR or MR) screening program?
Deep learning ECG-model is introduced in this study with the following
goals:
1- Accurate identification of valvular heart disease e.g AS, AR and MR

individually using the deep learning analysis of ECGs.
2- Detection of any of the 3 pathologies from a single unified prediction.
3- Perform independent algorithm validation.
4- Identify potential uses of an ECG deep learning- based screening

program for VHD based on ValveNet algorithms.
03.
Appraisal of the
evidence base
● Gold standard for VHD detection is screening with physical
examination and echocardiography.
● However, due to limited expense and expertise it is not

widespreadly implemented.
● When only using Echo to assess valvular disease it was

discovered to be subject to interobserver variability.
04.
Methodology
Inclusion Exclusion
Patients 18 years and Patients’ ECG showing
above who had a evidence of pacemakers,
nonpaced, analyzable and or significant artifacts.
12-lead ECG before an Echocardiogram
echocardiogram from identifying repaired or
January 2000 to April 2021 replaced Mitral or Aortic
valves.
Randomization and Selection
ECG data were accessed through the MUSE data
management system to obtain demographic and
ECG specific information.
Echocardiographic data was accessed from the
Syngo Dynamics (Siemens) to include the final
echocardiographer diagnosis.
The program used was python version 3.7
Model development and testing
TOTAL OF 77,163 PATIENT RANDOMLY ASSIGNED TO:
Training set Validation set Test set

(n=43,165) (n= 12,950) (n= 21,048)
Multiple ECGs per Most recent ECG per Most recent ECG per
patients were used to patient was included patient was included
maximize the model
learning
Description of the Deep Learning Model
● Valve-net deep learning model is a new
convolutional neural network
architecture developed to detect
moderate to severe left sided valvular
diseases.
● The model inputs : raw ECG waveforms

(12 leads) , demographics (ex.age) , and
ECG data (ex.PR interval)
● The model output determines whether

the patient has moderate or severe
valvular heart disease with a number
that ranges from 0-1.
05.
RESULTS &
DISCUSSION
SENSITIVITY THRESHOLDS
● Optimum sensitivity thresholds for correct disease classification were
selected using the Youden's J-statistic and were as follows:
○ AS: 0.429;
○ AR: 0.398;
○ MR: 0.459;
○ any of AS, AR, Or MR: 0.45.
PERFORMANCE ASSESSMENT TOOLS:
● Performance assessment tools (AU-ROC, ODDS RATIO , PRECISION
RECALL CURVES) :
AU-ROC: area under the receiver operating characteristic shows how well the
deep learning algorithm distinguish between positive and negative cases:
The model accuracy was as follows:
AS AU-ROC: 0.88 (95% CI: 0.87-0.90);
AR AU-ROC: 0.77 (95% CI: 0.72-0.81);
MR AU-ROC: 0.83 (95% CI: 0.81-0.85);
Composite of any of AS, AR, or MR AU-ROG:0.84 (95% CI: 0.82-0.85)
DISCUSSION
WHY AORTIC STENOSIS DETECTION IS SUPERIOR TO OTHER LESIONS :
AS is a chronic, progressively worsening syndrome that results in left

ventricular hypertrophy and recognizable ECG anomalies. In contrast, MR
and AR are more influenced by variations in hemodynamics between
studies.
PERFORMANCE ASSESSMENT TOOLS:
Precision-Recall Curves: the area

under this curve is used as a
summary measure of model
accuracy.
SALIENCY MAPPING :
Purpose: to highlight the specific ECG waveform patterns that were
influential in the deep learning algorithm detection of AS, AR & MR.
This information can aid in the development of diagnostic criteria and

improve the understanding of disease mechanisms.
+BONUS POINT : it provides a visual interpretation of the deep learning model's

decision-making process. This can enhance the trust and acceptance of the model's
predictions by clinicians.
RESULTS OF SALIENCY MAPPING :
The QRS complex was the most important feature for
determining model predictions with lesser effects seen for P
waves and R-R interval variance.
09.
APPLYING TO THE
GENERAL POPULATION
Screening Program
To determine whether the model may operate in a population screening
application, the researches modeled the precision-recall curves based on
various disease prevalence estimates using the test data.
For these models, prevalence was manipulated by including ECGs that

were positive for VHD along with randomly selected disease-negative
ECGs and then accuracy was measured based on precision-recall curves
for each conditions at each disease prevalence.
10.
CONCLUSION
Benefits in Clinical Use:
● Non-invasive and cost-effective method for detecting left-sided
VHD.
● Improves diagnostic accuracy
● Facilitates early intervention.
● Offers a more efficient approach to identifying at-risk patients in
screening programs : ~5 echocardiograms would need to be
performed in ValveNet ECG-positive patients to identify a single
case of moderate or severe VHD leading to a reduced need for
echocardiogram expense.
Weak Points and Limitations:
● Lack of assessment in real-time or prospective clinical scenarios

(retrospective study).
● More accurate in younger patients
● Low accuracy in wide QRS and LBBB

CONCLUSION
● Deep learning analysis of ECG accurately detects left-sided VHD.
● Provides a valuable tool for physicians in diagnosing AS, AR, and

MR.
● Enhances diagnostic accuracy, facilitates early intervention, and

supports the development of VHD screening programs.
● Further research and validation are necessary to establish its

clinical utility and overcome potential limitations.
Menna El Zohairy 201800105
Nadine Hesham 201800545
Nour Ghazy 201800660
Mohy eldin Mohamed 201801464
Reference
https://pubmed.ncbi.nlm.nih.gov/35926935/
THANK YOU!
CREDITS: This presentation template was created by
Slidesgo, including icons by Flaticon, and infographics &
images by Freepik and illustrations

AI in Healthcare

Uploaded by

Copyright:

Available Formats

AI in Healthcare

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AI in Healthcare

Uploaded by

Copyright:

Available Formats

AI IN

1- Accurate identiﬁcation of valvular heart disease e.g AS, AR and MR

2- Detection of any of the 3 pathologies from a single uniﬁed prediction.

3- Perform independent algorithm validation.

4- Identify potential uses of an ECG deep learning- based screening

● However, due to limited expense and expertise it is not

● When only using Echo to assess valvular disease it was

Training set Validation set Test set

● The model inputs : raw ECG waveforms

● The model output determines whether

AS is a chronic, progressively worsening syndrome that results in left

Precision-Recall Curves: the area

This information can aid in the development of diagnostic criteria and

+BONUS POINT : it provides a visual interpretation of the deep learning model's

For these models, prevalence was manipulated by including ECGs that

● Lack of assessment in real-time or prospective clinical scenarios

● More accurate in younger patients

● Low accuracy in wide QRS and LBBB

● Provides a valuable tool for physicians in diagnosing AS, AR, and

● Enhances diagnostic accuracy, facilitates early intervention, and

● Further research and validation are necessary to establish its

You might also like