AI Case Study
Fraunhofer Heinrich Hertz Institute and University Medical Center Schleswig-Holstein, a German University Hospital, has developed an algorithm to detect the probability of getting heart attack from ECG and EEG, matching accuracies achieved by cardiologists
Using convolutional neural networks, two German researchers have developed a model to predict risk of heart attacks from ECG and EGG. The algorithm analyses electrical signals and identifies patterns such as voltage and time intervals along with patient demographic data to calculate the risk of developing heart attack and to identify arrhythmia. The algorithm has achieved results matching accuracies of cardiologists.
Industry
Healthcare
Healthcare Providers And Services
Project Overview
"Electrocardiograph records electrical signal from 12 different leads and patterns in the electrical behavior such as time intervals and voltage values, presence of ST elevation etc for each beat to diagnose the risk of developing heart attack. Currently there are pre-defined rules limited to the input of a few signals. Using machine learning, hidden patterns can be identified to build better models in predicting the risk of heart attack.
Two German researchers have developed a CNN based model which has shown promising results."
Reported Results
Stage 1 testing resulted in 93.3% sensitivity and 89.7% specificity, evaluated with 10-fold cross validation, which is the performance level of human cardiologists.
Technology
"Recurrent neural networks have been successfully applied to time series classification problems"
"Convolutional networks for time series classification and applied to the UCR Time Series Classification Archive datasets. Also used a sliding window approach similar to the one applied in this work and feed differently downsampled series into a multi- scale convolutional neural network also reaching state-of-the- art results on UCR datasets"
"Models were implemented in TensorFlow"
Function
R And D
Product Development
Background
Existing rules-based automated heart monitoring systems are less reliable than trained cardiologists.
Benefits
Data
"Trained using UCR Time Series Classification Archive"
Testing on 549 records from 290 subjects