FORCE: To Save lives by Early Detection

Jan 28, 2019 | Siddharth Biswal, RCE AI

Critical heart events such as myocardial infarction(MI) or heart attack occur when blood flow to the heart is blocked, causing damage to heart muscle. Treatment usually involves intervention to widen the artery, allowing blood flow back to the heart muscles. In fact, early detection of MI events make a huge difference in treatment and prognosis of patients.

Automated ways of accurately and reliably spotting these events is really critical. At RCE, we have developed artificial intelligence based models capable of detecting and alerting physicians of early morphological progression in critical heart events. While some of these events can escape the scrutiny of even the trained eye, well trained deep learning models are able to detect these events with high accuracy.

In the past few years, deep learning models have been used to solve some of the really challenging problems. We have built a model using deep learning techniques to tackle the challenge of early detection of heart attacks. The following diagram shows the schematics of RCE smart wearable collecting 12 channel ECG data and the data being processed by RCE AI model to detect heart attacks

We describe our artificial intelligence model building process in the following diagram. This shows the pipeline of using ECG datasets from PTB dataset, preprocessing the data into train, test split. We then use a convolutional deep learning model to train a model which is evaluated on the unseen test dataset to accurately evaluate model performance.

agent viz scII
RCE ML Pipeline.

As the diagram shows, we use PTB dataset as the source of ECG dataset for representative cases of MI and healthy controls. PTB dataset contains data from 290 subjects with 590 records. This dataset contains a mix of diverse cases of MI and some subtypes. We preprocess the dataset to only consider subjects with MI. We also apply steps of downsampling and normalization to the dataset before creating train and test sets.

We are able to achieve 94.21% accuracy after training on the PTB dataset, producing state-of-art results using our deep learning models. Our models achieve 94.68% sensitivity and 92.23% specificity which indicates that we are able to detect with high true positive rate and low false positive rate.

agent viz scII
Confusion Matrix indicating the results

Why Deep Learning?

While we were thinking of different approaches to solve the task of detecting and classifying ECG signals into different classes such as myocardial infarction and healthy sample, we did search for existing literature for methods for classifying ECG signals. While in the past, different methods have been proposed which used hand crafted features such as RR rate, ST elevation, heart rate etc. to build classifiers for MI classification, the generalizability of these models is a difficult challenge. Convolutional Neural networks on the other hand are capable of learning from raw data. Natural data such as image, text, audio signal contain a lot of information and convolutional neural networks are capable of learning the inherent hierarchical structure of information. Various non-linear features can be learned through convolutions and non-linearities(ReLU blocks). Ultimately, various high level features are learned as compositions of lower level features.

As seen in the following diagram, we use a stack of convolutional layers which are combined with pooling layers and activation functions to classify ECG samples as healthy or myocardial infarction events. We pose this as a binary classification task and learn using binary cross entropy loss function.

Convolutional Neural Network used in our deep learning pipeline

Our system uses multiple machine learning models as a combination to detect false positives in MIs, such as Left Bundle Branch Blocks (LBBB), Left Ventricular Hypertrophy (LVH), Pericarditis etc.. We also use a combination of machine learning models to detect if the patient is walking, running or in various physical states at rest. We additionally leverage the data from the accelerometer on the RCE smart wearable towards a multi modal approach to our deep learning. Thereafter, we apply combination of dynamic filters which are used to clean the ECG before it is passed to RCE artificial intelligence detection system.

The true power of deep learning systems is that the accuracy of the model increases as the amount of training data increases. So we intend to use online Interactive training procedures to train our models, which get better over time. We also intend to do a clinical evaluation of on real world test subjects.