Automatic Learning Filters to Improve the Accuracy of Sensing Algorithms

The technology combines novel filters and a convolutional neural network model to automatically learn what kind of data will make an algorithm produce inaccurate results. While the idea of rejecting certain data is not new, existing filters are manually-developed and rely on assumptions about what causes the algorithm to be unreliable.

The approach applies machine learning methods to automatically learn when sensor processing algorithms will be reliable or unreliable and discards unreliable data, making the overall system more accurate.

In-the-wild continuous sensing on mobile devices has the potential to revolutionize fields such as personalized health care. However, a key problem with current methods is the diverse nature and noise associated with incoming sensor data. Running sensor data processing algorithms on this diverse data can lead to unexpected and poor results because it is difficult for algorithms to anticipate the variety of data that can occur. This tunable system is able to achieve error rates significantly lower than existing approaches, especially for:

  • At‑home monitoring
  • Wearable tech
  • Clinical studies
  • Diagnostics

STATUS

Patent application has been filed and a proof of concept has been created and tested in a clinical setting with patients who have COPD by implementing a respiratory rate monitor on a smartwatch. Existing smartwatch-based respiratory rate monitors rely on the assumption that motion will cause the respiratory rate monitor to be unreliable. This filter automatically learns when respiratory rate detection from a smartwatch will be reliable and yields significantly better results than current methods.

ID:

2114

Keywords:

Artificial Intelligence (AI) , Machine Learning , Data Analytics , Neural Networks , Wearable Tech , Software

VPRI Contact

Kurtis Scissons

Director, UTEST & Innovations & Entrepreneurship Manager
Innovations & Partnerships Office (IPO)
(416) 978-3557

Related Resources