Many people, especially those with higher risk of cerebrovascular/cardiovascular diseases, are bothered by the risk of having an acute condition like stroke. Inspired by the paper (, also by the enlarging market on personal health equipment, we decided to try to approach this by training a model that help predicts the risk given the HRV data.

What it does

Our prototype included a web based application and a mobile application that takes data from various sources and make predictions using models that is trained with federated training. Federated training is considered because health data are considered private and we may not like to collect them in any ways.

How we built it

We read many papers and extracted few candidates from many of the parameters of a HRV. After testing them on the actual data, we narrowed it to nine parameters that generated the most accurate model. We then train a machine learning model whose inputs are those parameters and whose output is a number between 0 and 1.

Challenges we ran into

The greatest challenge that we ran into was insufficient time. We have spent too much time doing research and modeling, resulting in insufficient time to implement the actual applications. Also another great challenge that we faced is that we have insufficient data. Federated learning once implemented should be able to help with our data and should further increase the accuracy of our model (current accuracy=0.8921).

Accomplishments that we're proud of

We are proud how we managed to extract the 9 most relevant parameters from about 40 parameters that we started with. The decreased number of parameters would make it easier to generate these parameters and would increase the performance of the model since it's going to run on user's end.

What we learned

Our most important lesson is the importance of time management, as well as the importance to divide tasks into subtasks.

What's next for Risk Indices Based on Heart Rate Variability

The next step would be to implement an application on the user's end, and to implement federated learning to further increase the performance of our model.

Built With

Share this project: