Patient-Reported Flares Were Correctly Predicted By an Algorithm Using [...]

Gossec, L., Guyard, F., Leroy, D., Lafargue, T., Seiler, M., Jacquemin, C., Molto, A., Sellam, J., Foltz, V., Gandjbakhch, F., Hudry, C., Mitrovic, S., Fautrel, B., Servy, H. (September 2017).


Voir le poster


Background/Purpose: The natural history of RA and axSpA comprises periods of low disease activity and flares. However, there are few data linking patient-reported flares to quantifiable outcomes. We previously indicated in the ActConnect study that flares were related to a moderate decrease in physical activity (1). Objective: to predict patient-reported flares based on activity-tracker-provided continuous flows of steps per minute.


Methods: This prospective multi-center observational study (ActConnect) included patients with definite RA (ACR/EULAR criteria) or axSpA (ASAS criteria), owning a smartphone. Over 3 months, physical activity was sampled continuously (each minute) using an activity tracker, and flares were self-assessed weekly using a specific flare question. In this reanalysis of the dataset, Machine Learning statistical methods were used. Physical activity data were first normalized at patient level using each patientÕs mean and standard deviation of steps for a similar timeframe without flares. Then the data were analysed by multiclass Bayesian methods with a Machine Learning software belonging to Orange (2). The software was instructed to find the best predictive model of patient-reported flares. Sensitivities and specificities were calculated. Several sensitivity analyses were performed using different physical activity timeframes, different definitions of flares.


Results: In all, 170/178 patients (91 RA and 79 axSpA patients; 1228 weekly flare assessments and 24,972 1-hour physical activity assessment timeframes) were analyzed: mean age 45.5±12.4 years, mean disease duration 10.3±8.7 years; 60 (35.3%) were males and 90 (52.9%) received biologics. Disease was well-controlled (mean DAS28: 2.3±1.2; mean BASDAI: 3.3±2.1) but flares were frequent: reported in 24% of all the questionnaires. The Khiops generated model detected correctly both flares and absence of flare (Table) with a sensitivity of 96% and a specificity of 97%. The corresponding positive and negative predictive values were respectively 89% and 99%. Sensitivity analyses were confirmatory.


Conclusion: Machine Learning methods are useful to deal with repeated data in big datasets. The results confirm objectively the functional impact of patient-reported flares. Furthermore, the correct detection of flares by the activity tracker and adapted statistics opens the way for future studies of flares using connected devices with great precision and minimal patient burden.


(publication) Gossec, L., Guyard, F., Leroy, D., Lafargue, T., Seiler, M., Jacquemin, C., Molto, A., Sellam, J., Foltz, V., Gandjbakhch, F., Hudry, C., Mitrovic, S., Fautrel, B., Servy, H. (September 2017). Patient-Reported Flares Were Correctly Predicted By an Algorithm Using Machine-Learning Statistics on Activity Tracker Data on Steps, in a Longitudinal 3-Month Study of 170 Patients with Rheumatoid Arthritis (RA) or Axial Spondyloarthritis (axSpA) [abstract]. Arthritis Rheumatol, 69(10).

S.A.S.U fondée en 2010
capital social de 8 000 Euros

R.C.S. Marseille 528233190
188, av 2nd Division Blindée
13420 Gémenos – FRANCE

tel : +33 4 84 83 00 60

  • LinkedIn - Black Circle

Echangeons sur votre besoin !

* obligatoire

SANOIA_Real_World_Digital_CRO-logo_edite
noun_Upload_1358833.png