Separating The Signal From The Noise – Finding Clinical Relevance In Consumer Health Data

In it’s unadulterated form, consumer health data is 99+% noise.  If we present it to health providers in this form, we’re going to very quickly turn them off the idea of gathering data from patients and caregivers.  Any who could blame them?  If Digital Health is going to add value to busy health providers, we need to separate the signal from the noise for them.

 

In the rush to gather vast amounts of data from consumer health devices and wearable technology, there is a significant risk that we will make health providers’ lives worse, not better.  Wearable technology is the sexy part of Digital Health right now, but frankly, we haven’t put enough thought into how we integrate these large data streams into models of care that actually make a difference.

 

The Noise

I’ve been wearing a Fitbit every day for over 3 years now, and have almost continuous heart rate data for just under 2 years.  As a consumer, looking at that data is interesting.  During the 3 years I’ve covered a bit over 11,500 kms (on foot), climbed just short of 50,000 floors and taken a little less than 16 million steps.  My heart rate has ranged between 36 and 182 and typical resting heart rate is around 50 beats per minute.  I’ve also gathered data on weight and blood pressure (albeit less regularly) over that period, and can quickly pull out charts and data points to show my health provider.

But what do you think would happen if I pulled out all of this data and tried to show it to my GP, or other health providers.  Quite rightly they would tell me that a) they don’t have time to go through it all and b) most of it is not (or at least, doesn’t seem) that clinically significant anyway.  And I’d be in complete agreement on both points.

Firstly, outside of evidence that I’m getting the right amount of exercise, physical activity data has relatively little clinical utility.  It might, therefore, be interesting in aggregate form, showing the total time I’ve spent exercising in a given period at different intensities, and my resting heart rate as a basic proxy for fitness.  Second, the sheer volume of data we’re talking about here is largely noise.  It shows a number of repeating patterns (that I can see with the human eye), like what happens to my heart rate over the course of a night’s sleep, or what my typical patterns of movement are on a given day.

It’s also worth pointing out, as I’ve discussed previously here, it’s my view that Patient Reported Outcome (PRO), also known as Patient Reported Measures, offer a far more interesting (and clinically significant) source of patient data than what is generated from wearables or consumer health devices.  But, regardless, both types will generate large quantities of data that, unfiltered, could easily overwhelm health providers.

In it’s unadulterated form, consumer health data is 99+% noise.  If we present it to health providers in this form, we’re going to very quickly turn them off the idea of gathering data from patients and caregivers.  Any who could blame them?  If Digital Health is going to add value to busy health providers, we need to separate the signal from the noise for them, not present them with vast quantities of data and pretend it doesn’t make life worse.

 

The Signal

So what signals can we find amongst the noise?  Well, I think over time we’re going to learn how to decode all sorts of interesting signals from streams of consumer health data, some in real time.  But for now, here are a few ideas:

  • Aggregated data – This is simple, but critical.  As with the exercise data example I gave earlier, we need to aggregate large quantities of data into more meaningful forms, supporting health providers in interpreting the clinical significance of what is shown.  For example, current physical activity trackers still don’t do this.  Much more work is required on visually accessible dashboards making the health implications of physical activity easier to determine.
  • Exceptions / changes – We need to support the filtering of streams of consumer health data to capture exceptions to the norm.  Whether it’s Patient Reported Outcome data or information from wearables, what is of greatest potential clinical interest is what is different to normal values.  If I’m self-reporting a “Pain” PRO and its typically in the 2-4 (out of 10) range, and then it jumps to a 7 and stays there for a few readings, that’s of interest.  It could be significant.  If my resting heart rate is typically between 50-55 and rises over a period of time to 70, that could be significant.  We need to assist health providers in using thresholds that allow the filtering of signal from noise.
  • Complex conditional logic – Once we can capture exceptions in data, we need to be able to chain them together using conditional logic.  For example, as a health provider I might not care about a slight rise in blood pressure.  But I might be interested in a slight rise in blood pressure when combined with weight gain and an increase in resting heart rate.  By combining changes and exceptions in data using conditional logic, I can build more subtle ways to tease out trends and changes in data that are potentially clinically significant.
  • Patterns – There are a range of techniques, from simple through to very complex, for finding patterns in a stream of data.  At the most simple, the ability to detect simple trends, i.e. rising or falling weight, blood pressure, etc.  At the more complex end, there is significant potential for the use of Artificial Intelligence to detect patterns in data streams that are too complex for the human eye.  For example, what if machine learning could teach us what early stage patient deterioration looked like in conditions like heart disease, COPD, etc.?  What if learning from large quantities of data could allow us to intervene days or weeks before a patient becomes sick enough to go to hospital, simply based on consumer health data from PROs and wearables?

 

All this potential exists with consumer health data, but for the time being, why don’t we just start with the simple stuff.  Like recognising the need to separate the signal from the noise…

 

Leave a Reply

Your email address will not be published. Required fields are marked *

(Required)