Dr. Peter Kelly, Dr. Robert Ellis, Sarah Morgan
At Koneksa, we help our clients to gather real-world data from patients participating in clinical studies to help better inform decision making. Data collected in the real-world setting complements and enhances data collected in the clinic.
We focus on collecting data using digital devices. We help our clients to deploy wearable sensors to collect more informative data and develop new digital endpoints - for example, new, more sensitive measures of disease activity and treatment response. This type of work requires scientific rigor and careful understanding of how these technologies work.
There are two main components to wearable sensors: hardware and software. Hardware includes the circuits and sensors that make up the device. Software converts the raw signal from sensors into data that can be reviewed and interpreted.
But how is the data processed? The techniques that we use to choose, gather, analyze and interpret data can seem complicated. This article will hopefully help them become easier to understand.
Two of the most important facets of our work are signal processing and feature extraction – ways to reduce enormous quantities of information into a data set that is simple enough to work with, but that still accurately and reliably represents the whole. In this post, we will give a quick overview of these concepts.
Signal processing has been around for decades, working to separate the signal you do want from the signal you don’t want. These signals may represent sound, radio waves, images or biological measurements and usually come from a sensor, such as a microphone, camera, antenna, or accelerometer.
You may not realize it, but you use signal processing every day. Every time you use a mobile phone, a car's satellite navigation, listen to digital music, browse the Internet, or watch television, you are benefitting from signal processing. More specialized uses include radar, sonar, earthquake detection, communication with satellites and most forms of medical imaging, such as ultrasound or MRI scans.
Feature extraction usually works hand-in-hand with signal processing. Typically sensors will be producing hundreds or thousands of measurements each second. This is too much data for statistical analysis. Therefore all these measurements need to be reduced to a handful of features that are salient to the condition we wish to detect. The picture below shows the end to end process.
Take an example of the ECG technology. The electrodes attached to the skin measure small changes in the current when the heart muscle depolarizes. Processing of these raw electrical signals into secondary information such as heart rate, provides very valuable information about how well a person’s heart functions.
In Koneksa’s work, we apply signal processing techniques to sensor data collected from devices used in clinical studies - for example, an activity tracker collecting accelerometer data. Signal processing makes it possible to extract useful information, but to use that information to measure an aspect of a disease requires many careful decisions, steps, and tests.
For example, people with Parkinson’s disease usually have difficulties moving around and it can get worse as the disease progresses. Measuring a person’s gait, balance, or their stride variability is important to understand their response to treatment. These are all possible, but they would first require you to pull the right characteristics from all of the raw data available from an activity tracker.
Once you have a characteristic of interest, you have a feature – a derived attribute of the data that you believe to be indicative or characteristic of the condition in which you are interested.
To prove it’s useful, you must next go through a set of tests to find out:
Is the feature truly relevant to the disease: does it statistically differentiate people with the disease from people without it?
Does the feature help you to determine disease severity: can you use it to differentiate one person with the disease from another person with a more advanced form of the disease?
If the answers to both of those questions are “yes,” then you’ve identified a potential feature that has some significance in the study of the disease. It may have the ability to help you measure whether a medication is effective or a disease is progressing.
It’s still only potential, though. You haven’t proved any clinical utility yet. The next round of the work is a validation study that determines if a technology is good enough to measure the outcome of interest compared to well-accepted measures of the disease - by determining:
The degree to which you can establish a statistically significant relationship between your new measure and any existing “gold standard”
The degree to which you can explain the variability in your measurements
The responsiveness of your new measurement and relevance to patients
The degree to which the new measure agrees with the standard.
If you can surmount these hurdles… congratulations! You’ve come up with a new measure that can be used to track an aspect of the disease you’re studying.
We hope this article has given you a window into the kind of work we do for our clients, as we help them to discover how they can learn the most from their clinical trial data and make informed decisions.