Peripheral artery disease is a cardiovascular condition affecting over 10M people in the US. It is the formation of plaque/blockage of arteries that supply blood to the legs and can cause a lot of leg pain.

One of the simple tests to diagnose PAD is ABI — Ankle Brachial index. This is the ratio of blood pressure in the ankles to the blood pressure in the arms. Normal is >= 0.9, Mild 0.9-0.8, Moderate 0.8-0.5, Severe <0.5.

Often clinically functional exercise capacity is tested, specifically using the 6 minute walk test. To perform this test patients walk up and down a 100 foot hallway course at their own pace for 6 minutes. The total distance walked is what is classically measured.

Using smart phone data, such as accelerometers, we’re trying to see if we can detect peripheral artery disease early -- potentially helping diagnose earlier than clinical visits, just through the devices in our pockets.

What it does

Given accelerometer data every 0.01 seconds for about 2.5 minutes, try to predict a patient's ABI (between 0 - 1.4), if they're using a walker, and if they have PAD.

How I built it

Trained many, many deep learning models on Intel's super power Nervana AI cluster. Nervana allowed me to train lots of large models that would have been practically infeasible on my computer, and many models at the same time to do many experiments.

Models include:

  • LSTMs
  • Bidirectional LSTMs
  • Temporal Convolutions with LSTM
  • Temporal Convolutions with Bidirectional LSTM

Also tried training multilayer LSTMs.

Challenges I ran into

  • There were on the order of 100 patients
  • Modeling time series data can be complex

Looking Forward:

Would try to get a free app to collect accelerometer data from as many people as possible, recommending if they have PAD. With more users, the models will keep getting better!

Built With

Share this project: