BPNet

Project Idea

The goal of this project is to make a model to maximize single channel ECG to continuously predict blood pressure. Heart disease is the leading cause of death in the US and monitoring cardiac activity is one of the key reactive measures taken to combat it. Reducing the amount of hardware and time required to acquire cardiac data by not relying on a blood pressure cuff to get blood pressure data would be a significant advancement and is possible through the research this project is on. There are also plenty of other applications for single and multi channel ECG analysis in research and everyday healthcare. The papers we are looking at replicating use an LSTM with extra residual features. Databases are relatively small in their population size and diversity which can be seen in the MIMIC III (Multi-parameter Intelligent Monitoring in Intensive Care III) database used by most researchers on this topic. Another component of this project could be comparing MIMIC III with other databases as a way of validating this method. The minimum goal is to reimplement a single channel ECG to blood pressure LSTM with accuracy similar to what other researchers are getting. Further goals are comparing different datasets and architectures, and expanding to multi channel ECGs and other cardiac data. We will likely create our own codebase since there is no public implementation available for this specific paper. The dataset we will likely be using is PulseDB and possibly MIMIC III if we can get access to it.

Key Limitations

Already it has been difficult to gain access to quality datasets that have continuous ECG and blood pressure data. These datasets are also limited in their population and diversity which will likely lead to limited generalizability in the model we create. Another large issue with single channel ECG data is the amount of noise and the signal ceiling created by the smaller amount of data compared to a multi channel ECG. Getting the most out of the little data available in the single channel will be very important and difficult. Since there is no codebase for the paper we are reimplementing, it will be up to us to determine the best way to reduce noise. Testing different cleaning algorithms and possibly using a VAE will be important for overcoming this limitation. Choosing hyperparameters might also be difficult without a codebase to reference. With the variety of implementations we want to try, time and computational power may also become a limitation.

Built With

• python