PCA (Preventing Cardiac Arrest)

Inspiration

Currently, about 9 in 10 people who have cardiac arrest outside the hospital die. Also, sudden cardiac arrest accounts for nearly 50% of the deaths in the United States. However, it has been established that if CPR is performed within the first few minutes of cardiac arrest, it can exponentially increase the person’s chance of survival.

What it does

PCA uses transfer learning to classify a segment of ECG signal into three classes: normal, irregular, and dangerous. Ultimately, we plan to implement an alarm system where a wearable device (e.g. Apple Watch) collects ECG data from the user on a continuous basis and monitors any dangerous signal patterns that lead to the onset of sudden cardiac arrest. Whenever a dangerous signal is detected and there is a 'sudden fall' signal from the device, a loud alarm will turn on and an emergency message will be sent to a designated group of people.

How we built it

To implement our model, we initially tested the efficiency and accuracy of a pretrained model that makes use of PTB-XL ECG heart dataset. This dataset includes 21,000 fragments of ECG recordings labeled by their diagnostic statements. These ranged from normal sinus rhythm to some fatal ones such as atrial fibrillation. We benchmarked 3 different pretrained neural networks including LSTM, ResNet, Inception and determined that ResNet displays the best runtime and accuracy on the dataset.

In the next step, we found another ECG dataset with 11 diagnostic labels and manually ranked each data sample cohort as normal, irregular, and dangerous. Then, we transferred the existing weights of ResNet and trained the model with the newly annotated dataset. Through the method of transfer learning, we intended to transform the predictor of cardiac rhythm into the predictor of clinically severe condition. After training for about 5 number of epochs, the model displayed True Positive Rate of 0.941.

Challenges we ran into

• Lack of ventricular fibrillation dataset
• Preprocessing signal data

Accomplishments that we're proud of

• Achieving TPR of 0.940 through transfer learning using ResNet
• Building a clinical predictor of patients' heart conditions

What we learned

We learned that putting an idea into reality requires a collaboration from various disciplines of industry, academia, and healthcare.

What's next for PCA (Preventing Cardiac Arrest)

• A comprehensive dataset of ECG signals including the most fatal ones (e.g. ventricular fibrillation)
• Using recurrent neural networks to enhance model performance

Built With

• fastai
• python
• pytorch