Use of wearable activity trackers to detect disease flares in rheumatic diseases
Flares are fluctuations of disease activity levels, which have consequences on long-term outcomes in inflammatory diseases such as rheumatoid arthritis (RA) or axial spondyloarthritis (axSpA). Regular monitoring of patients to assess these flares is a key aspect of a treat-to-target approach. But how can this monitoring be performed? In this article, we will explain how connected devices such as wearable activity trackers can help, when coupled with artificial intelligence.
Why are flares important in inflammatory articular diseases?
The evolution of chronic rheumatic diseases is marked by alternated periods of flares and stable disease activity. Flares are of interest to patients and clinicians since they reflect disease activity fluctuations, which should be avoided according to a treat-to-target approach.
These fluctuations in disease activity have deleterious consequences, because the poor control of inflammation may lead to irreversible joint damage and functional disability, which have an impact on patients’ quality of life as illustrated below.
Image: Concept of disease flares.
How can we assess flares?
Flares can be assessed by questions or questionnaires, through increases in disease activity levels (increased inflammation as measured; on composite scores), or through modifications in the patients’ lives.
Methods to assess flares
Disease activity levels
Increase in DAS28 ≥1.2 or ≥0.6 if DAS28 ≥3.2
Increase in ASDAS ≥0.9
There is no consensus on the best way to assess flares: questionnaires and changes in disease activity status are the usual ways to detect flares, but do not provide the same information. Remote monitoring of patient-reported outcomes may facilitate a treat-to-target approach and help to measure flares, as it enables regular assessment of disease status. However, questionnaires completed remotely online necessitate patient engagement in care, which may lessen over time.
Physical activity, and in particular walking, is an essential part of daily life and may be influenced by flares. Physical activity can be objectively and precisely measured using activity trackers. Therefore, we have previously hypothesized that activity trackers could be used in the assessment of disease flares.
Detecting flares by activity trackers: The ActConnect Study
We performed a three-month longitudinal observational study (ActConnect) to explore the link between flares and physical activity measured by activity trackers. Patients with either RA or axSpA were included consecutively within several centers in France. Patient-reported flares were assessed weekly through an app on the patient’s smartphone, which asked a dedicated question: “Has your disease flared up during the past week?”. Physical activity was collected continuously using a connected activity tracker (Activité Pop watch; Withings®, Issy-les-Moulineaux, France) over the 3 months.
During the study we found that:
1. Flares were frequent
Most of the 170 patients had long-standing disease and around half of them were receiving a biologic. Although the disease appeared well controlled, patients reported having experienced a flare in 28% of the weekly assessments, on average.
2. Physical activity was moderate overall
The mean number of steps over three months was 7067 (± 2770) steps per day, and 24–30% of patients fulfilled the World Health Organisation recommendations for physical activity.
3. Flares were related to a moderate decrease in physical activity
At the group level, during weeks with flares, there was a relative decrease in physical activity of 12–21%, corresponding to an absolute decrease of 836–1462 steps per day. However, we were unable to find a precise cutoff value that allowed detection of flares based on steps.
Using AI to investigate the link between flares and activity
In the second phase of the ActConnect study, we analyzed the link between patient-reported flares and activity-tracker-provided steps per minute using artificial intelligence or, more specifically, machine learning through selective (multiclass) naive Bayesian statistical methods.
Artificial intelligence allows analyses of huge amounts of data with minimal aggregation. To explain: with machine learning, the patient’s walking profile and their corresponding response to the flare question over the first weeks are used to ‘calibrate’ the machine, at the patient level. Then, variations in steps and in their patterns are analyzed, and the machine informs you if the patient is flaring, or not.
Such techniques necessitate significant computing power and dedicated programs. In this case, we worked with a telecommunications company, Orange Healthcare (Paris, France). It is remarkable that the machine learning model correctly detected both patient-reported flares and absence of flares with a sensitivity of 96% and a specificity of 97%. The corresponding positive and negative predictive values were respectively 91% and 99% (See table below).
Table: Prediction of flares by steps per hour using a machine learning approach via pooled analyses (4030 weeks)
Flare according to the patient (N=920 weeks)
No patient reported flare (N=3110 weeks)
Flare according to the computer program
No flare according to the computer program
Some healthcare professionals do not ‘believe’ patient-reported flares are true flares. However, the ActConnect study objectively confirms the reality of the impact of patient-reported flares, using an objective measure of daily life and functioning.Patient-reported flares are real
Activity trackers and disease activity
The ActConnect study results indicate that connected activity trackers may give indirect information on disease activity. Thus, activity trackers may be useful for more than measuring steps and motivating patients to move more.
We found that patient-reported flares were strongly linked to physical activity and that patterns of physical activity could be used to predict flares with great accuracy. Automatic monitoring of steps may lead to improved disease control through potential early identification of disease flares, with high convenience for patients since the data collection is passive.
Potential practical applications in the clinic of these findings would include a three-step process.
- The patient would wear an activity tracker linked to the healthcare center. In cases of decreased physical activity, an alert appears for the healthcare professional.
- The center contacts the patient to check if indeed they are flaring or if the decrease in physical activity is not due to a flare (eg, a beach holiday).
- If a flare is confirmed by the patient, a visit is organized so that the patient can be assessed fully and the treatment changed, if necessary.
Of course, this theoretical framework has caveats, including data confidentiality, anonymity, and privacy concerns, as well as responsibilities taken on by the healthcare professionals in relation to alerts (eg, what happens if the physician does not contact a patient after an alert, and the patient is having a heart infarct?)
The ActConnect study is one of the first to demonstrate the usefulness of machine learning applied to large rheumatology datasets. Machine learning applied to large datasets should yield interesting results, particularly in the -omics field, but also when applied to clinical datasets. We believe that an explosion in such analyses will be seen in the years to come.
About the authors