Introduction

Summary

In this lesson, we are going to build an activity classifier using data from an accelerometer from a wrist wearable. Activity classifiers can be useful directly in that people like to keep track of the activities they are doing over the day. But they can also be used in more clinical contexts. For example, if a company is doing a drug trial and wants to know if their drug makes study subjects more or less active, they can look at the activity classifier output and see if subjects are spending more time walking around or if they are mostly idle.

To build this activity classifier, you will learn how to featurize a high-rate time-series signal. How do you take a few seconds of a 200Hz signal -- so that’s 1000s of data points -- and reduce it to a handful of features that traditional machine learning models know how to deal with. We’ll then use a random forest model to train our activity classifier and use leave-one-subject out cross-validation to evaluate its performance. We’ll then talk about our model’s hyperparameters and do hyperparameter optimization. To be successful in any modeling task, however, we need to dive into our data and become familiar with its intricacies, so we’re going to begin with some data exploration.

Outline

Understanding Your Data

• Wrist PPG Dataset
• Data Exploration and Visualization Understanding The Literature
  • Feature Engineering and Extraction
    Modeling
    Performance Evaluation
    Hyperparameter Optimization

Concepts

We will follow an algorithm development process, as outlined below.

Before jumping into a new domain, we first need to understand the data and the literature. For this lesson, we will be building an algorithm that first featurizes the raw high-rate IMU time series into a manageable number of data points. We can then put these features into a classical machine learning model, optimize the hyperparameters, and evaluate our performance.