PlaySafe is a digital health wearable that predicts whether someone is about to have a sudden cardiac arrest.
PlaySafe is focused towards three target groups:
- Athletes (According to the NIH, sudden cardiac death is the most common cause of death in young athletes)
- The elderly (According to Nature Cardiology, sudden cardiac death is a major cause of mortality in the elderly)
- ICU patients (Critically ill hospital patients are constantly at risk of dying from sudden cardiac arrest)
Also, these are two “metrics” I’m using to evaluate the overall PlaySafe system:
- Raw accuracy (what is the false positive rate?, how accurately does it identify cardiac arrest, etc;)
- Time to event (How much prior warning can PlaySafe give while maintaining its accuracy?)
Week 1 Updates:
I accomplished four things this week.
1) Background Knowledge
I spent this week reading lots of literature about ECG signal processing, digital collection of heart data, and how to apply recurrent neural networks (such as LSTMs) to ECG data. I took notes on the following papers/tutorials, which I found interesting and relevant to the problem:
2) Algorithm Design
Based on the knowledge I gained from Step 1, I identified a problem with previous approaches to cardiac arrest prediction: current algorithms only use a single modality to generate features and make predictions; either ECG signals or heart rate signals, but never both.
My hypothesis was that an ensemble model that leverages both these modalities in a multimodal framework could significantly improve not just raw accuracy but also prediction of time to event (how much prior warning a person gets).
Based on this hypothesis, I designed a novel architecture for this new PlaySafe ensemble model, one that uses both ECG and heart-rate to generate features and make predictions.
3) Software Prototyping
I used my design to write a prototype version of the prediction model, using an Amazon Web Services EC2 instance to train the model and training data from the PhysioNet Holter Sudden Cardiac Death database.
For now, I’m just looking at raw accuracy. Given a heart signal window, my model has to classify it into two classes:
-> Normal Rhythm
-> Cardiac arrest/abnormal
I didn’t have time to make any pretty visualizations, but preliminary results indicate that the model has a ~73% classification accuracy.
4) Hardware CAD Design
The final thing I did this week was make a CAD design of the hardware component of PlaySafe. In terms of sensors, I plan to use the following materials to assemble the initial prototype:
Here is an angled view of the CAD design:
- Raspberry Pi 3
- SparkFun Open-Source Pulse Sensor (https://www.sparkfun.com/products/11574)
- HeartyPatch: single-lead ECG patch (https://github.com/Protocentral/protocentral_heartypatch)
- Wires, etc;
Ultimately, I want to miniaturize the entire system by replacing the Raspberry Pi with a specialized IC, but for now, I’m using it because it’s easy to integrate with the backend software through the AWS IoT platform.
Goals for Week 2: (updated daily)
1) Acquire parts and make v1 of the hardware using the CAD model I created
2) Make v2 of the software by adding a front-end UI
3) Improve prediction accuracy to 80-85%
4) (Hopefully) use data collected from the hardware to make predictions on a real person’s heart data (mine)