Creating Clinically Useful Outputs from Wearable Sensor Technology

Wearable sensor technology holds significant potential to enhance healthcare by allowing for remote monitoring which will reduce patient visits to hospital, allow for earlier disease detection and ensure that our health professionals are prioritizing their time on patients that need help the most.

Getting a new solution to provide a clinically meaningful measures from sensor technology is not an easy task. Generally, these four steps are involved:

Planning

Without clear requirements on what should be measured and on who, the project is bound to fail. This stage should be completed with a clinical / scientific / patient advisory board to get insight from all relevant stakeholders. Saying you want to measure mobility or cardiac function is not enough, you need to pick a specific measurement you want to replicate, such as gait speed or heart rate variability.

Data Collection

The quality of the data will define the quality of the algorithm. In terms of quantity, machine learning algorithms typically require at minimum 50 data points. Deep learning algorithms typically require over 1,000 data points. In the healthcare setting, independent samples are very important, so each data point must come from a different subject. As you can imagine, this makes the data collection portion of the project costly and time intensive. The data-set should reflect the patient population of interest. Collection of gold standard data to train and test an algorithm on is a vital part of this process.

Algorithm Development

A lot of data scientists spend all their time at this point, without any consideration of the big picture. Playing with advanced machine learning techniques is fun, but if it’s not done with clear planning and an appropriate data-set, it isn’t going to get you anywhere. Three main types of algorithms to use are:

1. Heuristic algorithm – manually created features and thresholds are used to estimate a clinical measure.

2. Machine learning – in this case, a supervised statistical learning method is used to estimate a clinical measure based on extracted features.

3. Deep learning – Advantageous because there is no underlying assumption about the model fit, this method is very adaptable to what the data tells it. The main disadvantage is the amount of data required to create models, generally over 1,000 distinct data points required, which is a big ask for clinical data.

Validation

Proving that your algorithm works well on data it has not seen before is very important. This provides a true test of the algorithm and tells you how it is likely to perform in the real world. This is a very important step for health board regulatory approval as well as for proving to potential clients that your solution is worth while.

I have experienced the good and the bad at every stage of this process. I can help your group get the most out of your solution development no matter what stage you are at. Send me an email and let’s start the conversation, patterson.m@biomed-data.com.