Presentation is loading. Please wait.

Presentation is loading. Please wait.

Detecting Myocardial Infarctions (Heart Attack) using Neural Network

Similar presentations


Presentation on theme: "Detecting Myocardial Infarctions (Heart Attack) using Neural Network"— Presentation transcript:

1 Detecting Myocardial Infarctions (Heart Attack) using Neural Network
Meital Lian Rannon, Fall 2018

2 Executive Summary Terminology
MI - Myocardial Infarctions (Heart Attack) ECG – Electrocardiogram Facts Every year ~735K Americans experience MI. One major mean of MI diagnosis is the ECG.

3 Executive Summary The project’s tasks: Get 12 channels of the ECG
Handle the Data Train and test 12 NN Use committee classifier Combine the results The network was compared to a CNN The NN met the expectations when using channel i Did not perform as well when using all 12 channels

4 Approaches: 01 Get ECG files from PhysioNet Database: PTB diagnostic ECG database from PhysioNet. Network: 2 layers, 10 neurons at the hidden layer, 1 at the output (binary). Platform: Laptop Get ECG 02 Cut ECG to frames- two methods: 1. Arbitrarily 2. Peak-to-peak . Framing 03 06 Confusion matrix Downsample Downsample to 192 samples Database contains 207 ECG files: 80 files of healthy patients and 127 filed of MI patients. Combine the confusion matrices of all 12 networks Committee classifier Train and test NN Create final labels using majority voting NN was created by Matlab’s NN train tool 05 04

5 Youden’s J-statistic =
Results: Statistics: Sensitivity = 𝑇𝑃 𝑇𝑃+𝐹𝑁 Specificity = 𝑇𝑁 𝑇𝑁+𝐹𝑃 Youden’s J-statistic = 𝑆𝑒𝑛𝑠+𝑆𝑝𝑒𝑐−1 12 ECG channels: Channel I only: J-stat. Sens. Spec. Fully connected shallow NN 1st framing method 0.401 0.733 0.669 Fully connected shallow NN 2nd framing method 0.661 0.847 0.814 Fully convolutional architecture 0.827 0.933 0.897 J-stat. Sens. Spec. Fully connected shallow NN 2nd framing method 0.715 0.767 0.947 Fully convolutional architecture 0.703 0.875 0.828

6 Discussion: Main conclusions:
The Peak-to-Peak framing method is preferred. There are differences between the performance of the NN for each channel. Main conclusions: For shallow NN - framing method that doesn’t change the ECG signal position over the feature vector. Only channels with high performance for classification. More complex NN to achieve better results.

7 Thank you!


Download ppt "Detecting Myocardial Infarctions (Heart Attack) using Neural Network"

Similar presentations


Ads by Google