Task

Our task was to develop highly efficient machine learning models running on a wearable medical device (embedded ML or edge processing). The algorithms ranged from gait analysis and step counting to temperature monitoring, vomiting and coughing detection, and sleep analysis. They utilized audio and accelerometer time series data captured by sensors on the device. We also developed a cloud-based joint audio-accelerometer classifier to monitor coughing and other symptoms.

Challanges

• Algorithms must be highly efficient to maximize battery life as much as possible.
• ML models must have a very low level of false positives and perform well for a wide range of patients in varied environments.
• Project deliverables needed to be completed within a rapid timeline due to the startup structure of the product.

How we helped

• Carried out core R&D activities and developed multiple solutions related to embedded ML and backend AI.
• Assisted the client in creating annotated datasets used to train the models.
• Created reference designs with reference tests for the developed solutions.
• Managed the project and coordinated a small team of five engineers.

Results

We created and trained a machine learning algorithm that met the client's requirements for true positives and false positives while remaining efficient enough for implementation on a microcontroller (ARM Cortex). Additionally, we developed a cloud-based audio and accelerometer classifier (Convolutional Neural Network (CNN) in TensorFlow) that exceeded the client's precision and recall requirements. Many other algorithms and custom solutions were also created.