Machine Learning Approaches to Classify Self-Reported Rheumatoid Arthritis Health Scores Using Activity Tracker Data: Longitudinal Observational Study

BACKGROUND: The increasing use of activity trackers in mobile health studies to passively collect physical data has shown promise in lessening participation burden to provide actively contributed patient-reported outcome (PRO) information. OBJECTIVE: The aim of this study was to develop machine learning models to classify and predict PRO scores using Fitbit data from a cohort of patients with rheumatoid arthritis. METHODS: Two different models were built to classify PRO scores: a random forest classifier model that treated each week of observations independently when making weekly predictions of PRO scores, and a hidden Markov model that additionally took correlations between successive weeks into account. Analyses compared model evaluation metrics for (1) a binary task of distinguishing a normal PRO score from a severe PRO score and (2) a multiclass task of classifying a PRO score state for a given week. RESULTS: For both the binary and multiclass tasks, the hidden Markov model significantly (P<.05) outperformed the random forest model for all PRO scores, and the highest area under the curve, Pearson correlation coefficient, and Cohen κ coefficient were 0.750, 0.479, and 0.471, respectively. CONCLUSIONS: While further validation of our results and evaluation in a real-world setting remains, this study demonstrates the ability of physical activity tracker data to classify health status over time in patients with rheumatoid arthritis and enables the possibility of scheduling preventive clinical interventions as needed. If patient outcomes can be monitored in real time, there is potential to improve clinical care for patients with other chronic conditions.