Download presentation
Presentation is loading. Please wait.
Published byLiani Kusuma Modified over 6 years ago
1
Zachi Attia, MSc (EE) Co-Director of AI in Cardiology Mayo Clinic
AI In Cardiology: Finding What's Hidden In Plain Sight and Peering Into The Future Zachi Attia, MSc (EE) Co-Director of AI in Cardiology Mayo Clinic
2
Disclosures Mr. Attia and Mayo Clinic have licensed potassium sensing technology and QT measurement technology to AliveCor., the LVD detection algorithm to Eko Devices and Arrhythmia detection to Myant . Mr. Attia and Mayo may benefit from commercialization of this technology, but not from its use at Mayo Clinic.
3
Natural History of Disease
I feel great! Sx + Signs Metabolic, physiologic derangements Diagnostic Testing Treatment Develop over years It may be too late: First event may be: CVA MI SCD Waiting for signs and symptoms before diagnosing has been done for 2,000 years. We can do better! CP Friedman, PA MN
4
Could you have heart disease and not know it?
Asymptomatic Left Ventricular Dysfunction (ALVD) – unaware Identification requires expensive, not readily available tests. Once identified, effective therapies reduce hospitalizations and mortality. ALVD: ~2% of the global population, >7 million Americans Up to 9% of people over 60 years of age Expensive and many American live more than 50 miles from a cardiologist Echocardiogram CT scan
5
ECG as Sensing Platform
Electrocardiography (ECG) is the recording of the heart electrical activity. Maybe it can tell us more ? Electrocardiography (ECG) is the recording of the heart electrical activity, the heart dipole is the sum of the different cells action potential. Sum of cell action potential!!! P Q R S T !!!!!!!!!!!!!!!!!!!!!!!11 Figures adapted from Jaakko Malmivuo & Robert Plonsey: Bioelectromagnetism
6
Is the Ejection Fraction Low? Weak heart pump?
ECG – Electrocardiogram. Painless, Inexpensive, Widely Available A clinician cannot assess EF from an ECG We hypothesized that a deep convolutional neural network (AI) could be trained to identify who has a low EF from an ECG Once the network identifies a low EF subject an echocardiogram can be arranged
7
Artificial Intelligence and Neural Networks
Artificial Neural Networks: Defining a task and a loss function Training from examples Convolutional neural networks were found very effective in vision tasks Requires a lot of data as the features are developed from the samples Feature 2 Feature 1 What is difference between AI and ML ? (funding , hireling ) Why did we advanced ? GPUs and amounts of Data Very Deep Convolutional Networks for Large-Scale Image Recognition - Simonyan 2014
8
Robust Digital Warehouse of Medical Information
625,326 patients with ECG - Echo paired data 461,434 ECG and Echo performed more than two weeks apart Conv Layer Pooling EF Fully connected Layers + dropout 100,000s 163,892 ECG-Echo pairs within 2-week interval, from 100,029 patients suitable for analysis 63,863 ECG-Echo pairs from patients with more than one pair used for follow up analysis 2,200 ECG-Echo pairs used to develop proof of concept First ECG-Echo pair of 97,829 patient was selected for analysis Figure 1 35,970 ECG-Echo pairs used to train the network 8,989 ECG-Echo pairs used to validate the network 52,870 ECG-Echo pairs used to test the network Robust Digital Warehouse of Medical Information
9
AUC of EF network = 0.93 (perfect test = 1.0)
Compares favorably with other medical tests: Mamography for breast cancer = 0.85 Cervical cytology (Pap smear) = 0.70 PSA (prostate cancer) = 0.92 Figure 2
10
Long Term Outcomes of Patients with Normal EF during screening
“False positive” or delayed true positive All patients had EF >=50% at baseline
11
Could the network perform better by feeding it age and sex?
No.
12
Age and Sex from ECG “Real” Age vs ECG Estimated Age
ROC -Reported Sex vs ECG Estimated Sex Can The model really not be affected from Age and Sex ? Blue Red For Sex- in par with Visual estimation Might be Heart Age
13
Progression of ECG Age in a Patient with Multiple ECGs
Aging Factor > 1 Real Age=Ecg Age Linear Regression of Estimated Age Real Age
14
Progression of ECG Age in a Patient with Multiple ECGs
Aging Factor < 1 Real Age=Ecg Age Linear Regression of Estimated Age Real Age
15
Progression of ECG Age in a Patient with Multiple ECGs
Heart Transplant “ECG Age “ Aging Factor < 0 ? Real Age=Ecg Age Linear Regression of Estimated Age Real Age
16
Can We Do It with a Single Lead ?
Lead I vs. 12 lead AUC = 0.9
17
What about Screening for Atrial Fibrillation to Prevent Stroke?
The presence of atrial fibrillation is critical for determining optimal therapy Paroxysmal AF is often asymptomatic and under detected – important for CS patients The longer we monitor, the higher the yield
18
Continuous Monitoring vs Mobile Form Factors
Will short episodes, detected with continuous or near continuous monitoring have the same prognostic significance?
19
What about Screening for Atrial Fibrillation to Prevent Stroke?
20
Scanning ECGs in NSR to Determine if AF was Present in the Past!
Conv Layer Pooling Hx of AF Fully connected Layers + dropout 1,000,000 Coded for HX AF Attia et al AHA 2018
21
Finding AF – During NSR! Area Under the Curve : 0.864
Sensitivity : 77.3 % Specificity : 80.0 % AHA 18 - Attia et al CP Friedman, PA MN
22
Conclusions The addition of AI to large digital data sets permits detection of “occult” disease (screening) It permits “prediction” of future disease (detection well before signs/symptoms) With Smartphone enable AI we can make these tools massively scalable Confidential
23
Thank you Questions ? Confidential
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.