At-Home Device Can Predict Crohn’s Disease Flares

A new at-home monitoring device developed at Massachusetts Institute of Technology (MIT) found that increases in breathing rate, more awakenings at night, and slower walking speed accurately predicted that a person’s Crohn’s disease activity was about to flare, according to a study presented at the annual Digestive Disease Week conference.

In this exclusive MedPage Today video, lead investigator Joshua Korzenik, MD, a gastroenterologist at Brigham and Women’s Hospital and assistant professor of medicine at Harvard Medical School in Boston, explains the technology and why it is exciting for both physicians and patients.

Following is a transcript of his remarks:

There’s a good deal of interest these days in remote monitoring. When we take care of people with Crohn’s, we are particularly eager to know whether someone is flaring or not, or someone’s in remission. And as we all know, people can have lots of symptoms that can be difficult to distinguish between a flare or not.

And so most of the types of approaches we have, such as whether it’s colonoscopy, whether it’s calprotectin, whether it’s CRP [C-reactive protein], or other modalities, all have their limitations. And the biggest limitation perhaps is that it’s, for the most part, a one-time snapshot. It’s fixed in time and you might have done a colonoscopy one month earlier, but now they’re having change in symptoms and so what’s going on.

And so there’s a good deal of interest in remote monitoring. And what’s particularly exciting about remote monitoring is it gives a more continuous assessment of disease activity. And so this study was looking at one particular device developed by MIT investigators who are collaborators in this study. That is a WiFi-type box, it uses WiFi-type waves, but a lower energy wave, and it sits in the background in your house or home or apartment. And it uses those waves, sort of like radar, to analyze the surrounding wireless signals to infer someone’s gait, their physical activity, their sleep quality, and other vital signs such as respiration.

The particular thing that’s so exciting about this technology is, one, its accuracy to be able to distinguish, and that’s part of what the study was aimed to do, but also that it’s touchless, so it doesn’t require the patient to wear anything. They don’t have to carry it with them. They don’t have to input any information. And so it’s completely passive.

And this study was using this device to determine whether we could detect activity in Crohn’s or whether we could determine whether someone was in remission.

In particular, what we were looking at was sleep stages and quality, breathing, and gait. And so we were then looking at each of those individually and then putting them together with machine learning techniques.

The second thing was looking at the individual biomarkers and saying, does each of these, or can each of these, help distinguish how a patient is doing and determine that compared to, say, patient-reported outcomes or as other measurements? We did calprotectin measurements, we did other things. And then we wanted to see, if we integrate them with machine learning techniques, does that do better?

We enrolled about 120 patients, about 105 were really analyzable. Some didn’t complete the data, some didn’t bring in calprotectin, various things like that. And they were monitored for up to a year. We collected blood for CRPs, we collected stools, and patient-reported outcomes, and then used all their clinical data to sort of characterize all of their time during monitoring using patient-reported outcomes, CT scans, MRs, or other data that we would have to say, can we characterize that 6-week segment as a flare or a non-flare, or we don’t know?

We really collected an enormous amount of data with this. We monitored people on average with Crohn’s disease about over 300 days. And so that meant we were collecting measurements for 25,000 nights. Breathing signals were measured for about 200,000 hours, and gait, we had something like 400,000-plus measurements for gait.

And then we looked at each of these and we could show that, say, with gait, when people went from remission to activity, their gait slows down slightly. When we looked at breathing as a function of Crohn’s disease, there’s a slight increase. And this is particularly measured during sleep, so that standardized it. So if you had just gone for a run or had some other issue, that would be sort of a bias there. But this is looking at breathing during sleep and that standardized it and there’s a slight increase, and that was a very sensitive marker.

And then we also looked at sleep and sleep stages. So there’s awakenings at night, which increased. There’s REM sleep, which goes down a little bit. And there’s also deep sleep, which actually increases, which is particularly interesting because there’s some data that the pro-inflammatory cytokines such as TNF [tumor necrosis factor] actually increase during that period. And so this is prolonging possibly that period when those cytokines are increased.

And then we took all of that data, trained a machine learning classifier to use that data, and compared it then to the fecal calprotectin as the gold standard of a flare. And the ROC curve was 0.84. So very strongly predictive, behaved almost as well as calprotectin both in terms of sensitivity and specificity.

And then we looked to see can the machine learning classifier characterize the flares in terms of severity? And fecal calprotectin has its limitations. It’s not exactly a linear scale, but we could detect a higher classifier score with increased fecal calprotectin.

And then we wanted to say, can the monitor detect a flare earlier? Can we actually take that calprotectin data, look at patient-reported outcomes before that point? And is the machine learning classifier actually able to pick up a flare before the individual can say, yes, I’m having an increase in symptoms? And indeed, it was able to. We had data to look back as much as 25 days before that time point to say that even at that point, the classifier’s area under the curve was 0.76. So still remarkably strong capacity to predict a flare before the patient is aware of it.