The Digital Tracking of Arthritis Longitudinally (DIGITAL) study was an ancillary study of the ArthritisPower registry (Advarra Institutional Review Board [IRB] protocol number 00026788). A cohort of 470 eligible patients with RA was recruited by researchers at the Global Healthy Living Foundation and the University of Alabama at Birmingham, and 278 (59.1%) of them qualified for participation in the main study after successfully meeting adherence thresholds during an initial 2-week lead-in period. The main study followed the 278 qualifying patients for 12 weeks (84 days), and each patient was given a Fitbit Versa 2 to record physical metrics such as heart rate, physical activity, calories burned, and sleep progression. All collected data, including various PRO surveys that patients were asked to fill out on a weekly basis, were inputted by patients into a study-specific app (the ArthritisPower registry, with a custom workflow unique to this study). Data were analyzed and stored in the Health Insurance Portability and Accountability Act–compliant Amazon Web Services cloud and analyzed at the University of Alabama at Birmingham and the University of California, Los Angeles (UCLA). Data were deidentified prior to being made available at UCLA. Analyses were done to summarize trends between and within participants over time, and descriptive statistics quantified adherence or compliance rates for the various measures and technology as well as any differences in the enrolled populations and those compliant with the study protocol. The mean age of the patient cohort was 50.2 (SD 11.1) years, the mean number of years since RA diagnosis was 9.4 (SD 10.1), and 91.7% (n=255) of subjects identified as female. Patient features included demographics, race or ethnicity, gender, age, years since RA diagnosis, comorbidities, and medications, and some of these data were also used for training our machine learning models. Based on adherence criteria we describe later in the Data Imputation and Preprocessing section, we excluded 59 of these 278 patients, resulting in 219 patients in the final analytic data set.

The Fitbit Versa 2 is a commercially available activity tracker whose algorithms have been shown to accurately track metrics relating to physical activity, sleep, and energy expenditure [24-27]. Whenever the Fitbit Versa 2 was worn during the clinical phase of the study, it passively collected patient metrics and was synced to the ArthritisPower app continuously, but at a minimum, occurred every 5 days. Overall, there were 15 distinct physical features that were used to train our machine learning models, and the respective feature means and SDs among patients using all available data are depicted in Table 1. To account for noise and redundancy, each activity tracker–derived feature was aggregated on a daily level (ie, a 24-hour interval, from 8 PM on one day to 8 PM on the next day) based on the data collected with second- or minute-level precision.

PROMIS questionnaires are a library of instruments developed and validated to measure domains of physical and mental health [16]. In this study, patients were asked to actively input information regarding 6 PROMIS scores on a weekly basis through several questionnaires on the study-specific ArthritisPower app [21]. Within the ArthritisPower app, PROMIS computer adaptive testing instruments were used to record scores relating to pain, physical function, fatigue, sleep disturbance, and satisfaction with participation in discretionary social activities (“social activity” for concision). The t score metric was used to standardize each of these scores to a mean of 50 and an SD of 10, with a range between 0 and 100 [15]. Most scores fall between scores of 20 and 80. In addition to these 5 PROMIS scores, a score related to exercise frequency and intensity was assessed through the in-app Godin Leisure-Time Physical Activity Questionnaire [28,29]. For the Godin score, a patient scoring ≥24 was considered active, and a patient scoring <24 was considered inactive as a generalization [30]. PROMIS computer adaptive testings have previously shown efficacy in previous studies, in addition to low barriers to form completion [14].

“Symptom” (pain interference, fatigue, and sleep disturbance) scores of 60 (1 SD above the average of 50) or higher were defined as moderate to severe symptom severity [31]. Similarly, “function” (physical function and social activity) scores of 40 (1 SD below the average of 50) or below were defined as moderate to severe symptom severity, meaning less functional ability than normal. In this study, multiclass classification techniques were used to classify PRO state or score transitions over time, and binary classification techniques were used to determine whether patients’ PRO scores were above or below the critical threshold for at least moderate symptom or functional severity.

Missing data are a concern in any study that involves consistent patient adherence to a study protocol over time. In this study, missing activity tracker data could have resulted from patients either forgetting to wear their devices or removing them for charging, and missing weekly PRO scores could have resulted from patients forgetting to fill out their PROMIS surveys on a given week. Patients may have also gotten tired of completing PRO surveys consistently as the study progressed.

At first, prior to any data preprocessing, each of the 15 smartwatch-derived feature columns consisted of daily aggregated measurements. In other words, the original data set had 12 weeks (84 days) of data for each of the 278 originally qualifying patients in total, with each row corresponding to a given day’s measurements of the activity tracker features. However, weekly PRO measurements only presented themselves once a week, and this meant that we needed to aggregate data weekly rather than daily. To address this, we preprocessed the data set such that each of the 15 physical feature columns was split into 7 columns corresponding to the 7 days in a week. For instance, the DailyTotalSleepMins column, which measured how many minutes the patient was asleep in a given day, was split into 7 “unmelted” columns (eg, DailyTotalSleepMins1, DailyTotalSleepMins2, …, DailyTotalSleepMins7) such that there would not be any missing PRO values for that given week if the patient filled out all of their PROMIS surveys. It was first decided that the criteria for dropping a given week of data prior to model training and evaluation would be if there were at least 3 days of missing data for any of the 15 Fitbit-derived features. We excluded patients from our data set who had more than 2 weeks of dropped rows at any time during the 12-week study period. Based on these specifications, 59 of the 278 originally qualifying patients were excluded, and thus, 219 patients ended up remaining in the data set. In addition, prior to training our machine learning models to generate predictions for each of the PRO scores, we imputed remaining missing feature data based on the corresponding feature means from the previous week.

To establish distinct states that correspond to PRO score ranges, 8 “bins” of scores were created within a range of 5 (0.5 SD) units around the score’s established threshold for a given PRO score. Specifically, we created 3 score groupings below the threshold and 3 score groupings above the threshold, with 2 “catch-all” bins that grouped together all scores below and above the 6 central score groupings. As an example, this grouping process resulted in 8 unique states for the sleep disturbance PRO score: 40 (grouped all scores that were below 40 together), 45, 50, 55, 60 (the threshold grouping), 65, 70, and 75 (grouped all scores that were above 75 together). We manually merged bins that had less than 40 observations with the adjacent bins to avoid sparsity. When it came to measuring the success of the machine learning models themselves, we trained and evaluated each model 10 times and outputted the average metrics of those 10 rounds since the random seed of each iteration was different.

A naïve approach to model creation would involve treating each week independently, as PROMIS survey scores were generated on a weekly basis. A 1-subject example of the independent model is illustrated in Figure 1, as there would be 12 weeks of evaluable data that are treated as independent observations. All the features for the 7 days of the week were appended into a single feature vector, which was then used to train the machine learning models for the task of PRO score classification. Ensemble methods such as RF, Adaboost, and gradient-boosted regression trees have previously been shown to be robust over unbalanced data sets and are capable of generating better classification accuracies compared to other independent machine learning models [32]. In this particular set of analyses, the RF model was used to establish a baseline performance that the HMM could potentially improve on. Through hyperparameter tuning, we determined that each RF instance should consist of 100 estimators (decision trees), use the Gini Index as criteria for splitting, allow a maximum depth of 25 to prevent overfitting, and require at least 10 training samples as the minimum threshold for splitting. In addition, through use of the RF model, we were able to assess the relative importance of the 15 different activity tracker–derived features in carrying out PRO score classification.

A significant shortcoming of models that treat each week independently (such as the RF) is that they do not take temporal factors into account, or how previous PRO score states influence the current PRO score state. The use of the HMM addresses this weakness and effectively incorporates temporal correlations of PRO score states across weeks. As illustrated in Figure 2, the HMM’s state at each week corresponded to the PRO score state for that given week, with the activity tracker–derived features treated as observations. Following the score grouping detailed in the Data Imputation and Preprocessing section, we indexed each of the unique scores or states starting from 1 and created transition matrices by counting the S state transitions from week to week. These transition matrices were then normalized based on the Gaussian standard distribution and were used to generate label predictions in the testing and evaluation phase from the prediction probabilities generated by the RF model.

The forward algorithm computed the probabilities across states at time t, with the maximum computed probability representing the state classification,

where the weekly PRO score was treated as state y_t with observations of features x_t [33]. The emission probability, P(x_t|y_t), computed the probability of the observed feature vector x_t given state y_t, computed from the RF classifier and P(y_t).

At the first step, the transition probability distribution is undefined, so the state probability was defined as:

S(y₁|x₁)∝P(x₁|y₁)P(y₁)

For the task of binary classification, states were binarized according to whether or not they were normal or mild (ie, within 10 units of the US population mean of 50) or moderate or severe, based on the thresholds described in the PRO Measures section. Because dichotomizing PRO score values loses some information and adds noise, multiclass classification of PRO score states was conducted as well without the binarization process. For the binary classification task (distinguishing between a “normal” and “severe” PRO score), we chose to use the receiving operating characteristic area under the curve (ROC-AUC) metric, as it would give us a comprehensive understanding of the relationship between the true and false positive rates as we vary the classification threshold. For the multiclass classification task (PRO state prediction), we chose to use the Pearson correlation coefficient metric to evaluate the statistical relationship between the models’ predicted state labels and the true state labels, as well as the quadratic weighted Cohen κ coefficient to measure agreement or consistency between predicted state labels and true state labels. For each PRO score, the RF and HMM were trained and evaluated 10 times each to establish robustness in the observed metrics.

The funders of the study had no role in the data analysis, data interpretation, or writing of this report.

For all patient-level data used in this study, we obtained patients’ written informed consent. Patients’ consent statements have been kept on file. Our research activities related to this study were governed by the University of Alabama at Birmingham IRB (160708003).

In this analysis of passive data collected by a health activity tracker device (Fitbit Versa 2), we found generally high correlations with weekly PROMIS, other PROs related to RA, and self-reported exercise. Correlations between the passive data and the actively contributed PROs were highest for weekly physical function score, weekly fatigue score, and the exercise score from the Godin Leisure Time Exercise Instrument. This might be expected given that the health activity tracker provides the most information about actigraphy. Performance metrics for the multiclass and binary tasks of classifying PRO scores were significantly higher for the HMM compared to the RF. This general trend makes sense, as the HMM essentially builds from the independent RF model by factoring in sequential state changes when it comes to predicting a PRO class for a given week. A similar study, focused on classifying PRO scores in patients with stable ischemic heart disease, generally observed similar results, but in that study, the independent RF model outperformed the HMM for some of the PRO scores [23]. The results of our study could indicate that the HMM’s usage could vary from disease to disease and that it is effective in using activity tracker data to predict changes in RA symptom progression or improvement in particular. Although our results show promise for the use of temporal models for clinical applications such as passive remote monitoring, there is always room for future improvement. Even though the HMMs consistently outperformed the independent RF models for the multiclass classification task, the average Pearson correlation coefficient of HMM predictions and true labels across all PRO scores was 0.357, and the average quadratic weighted Cohen κ coefficient was 0.345. Although these values should not be regarded as “poor,” they suggest that physical activity trackers may not capture all the relevant information needed to make accurate predictions about changes in different aspects of a patient’s disease state. As we saw, our models perform relatively better for measures related to actigraphy since that type of data was directly captured by the activity tracker. Data about mental or emotional states were not incorporated into our models, which might be possible by examining heart rate variability or electrodermal activity and could potentially help with PRO score classification for emotional states. Another objective for future studies would be to further increase patient adherence rates and reduce the percentage of missing data, as this is generally one of the largest obstacles for accurate mHealth telemonitoring, including our study [9]. With more precise data about a patient’s state at a given point in time, statistical models could be better tuned to classify PRO score states over time and thus improve overall accuracy metrics and usefulness in a clinical setting. In the future, similar studies could also be carried out using deep learning methods such as long-short-term-memory networks (a class of recurrent neural networks), which have both shown prowess when dealing with sequential clinical data and missing data [34,35]. These kinds of approaches effectively model varying length sequences and capture long-range dependencies and could serve as another approach to classifying PRO scores over time [34]. A caveat is that these types of deep learning methods require a large amount of training data (ie, more patients in the clinical study) in order to perform robustly, which was not available for this study. As the efficacy of collecting clinical data from patients continues to improve over the next few years, such complex approaches will become more and more attainable and practical.

PROMIS instruments themselves could also be a source of potential improvement for model development. Even though recorded PROs have face validity and provide useful information about the patient experience, some doubts revolve around how some PRO scores may be partially redundant with each other and may not serve as the most effective way to inform treatment decisions [36]. In addition, patients may experience response fatigue when taking PROMIS surveys due to their question-by-question format, which could lead to measurement error and misclassification [37,38]. Improving the patient experience while completing a more limited set of PROMIS surveys could be a way to address the issues of low completion rates and reduce response fatigue that may take place in clinical studies such as this one. In a clinical setting, we envision a hybrid approach where PRO data collection is complemented by passive actigraphy measures, with sophisticated algorithms applied to both, such as those that we used. As advances in technology such as more widespread availability of the internet and mobile devices become more prevalent, patient accessibility increases, and thus, patients’ ability to participate in studies related to PRO score classification and related clinical intervention applications can improve [4].

A related use case for these kinds of models would be to detect worsening from a patient’s baseline state over time. In a clinical setting, it would serve greater usage to notify physicians when a patient’s state has significantly worsened from an initially measured state (ie, after the start of medication administration or a previous physician checkup). In our analysis, patients’ initial state (time 0) was a random, unimportant time. However, if “time 0” was instead marking the start of a new drug treatment, then there would be an underlying slope of PRO score trends and an expected trajectory of score improvements. In a clinical study such as this, our models would be trained and evaluated to detect the first week of significant decrease in PRO scores from baseline or initial scores over the study period.

We acknowledge several study features that are useful to contextualize this work. First, based on the demographics of the ArthritisPower registry, participants are predominantly White and have higher amounts of education (ie, at least some college), so these results may or may not generalize to other demographic groups, and they may be influenced by other social determinants of health. Ongoing work in the registry is explicitly recruiting a more diverse community of patients with respect to these factors. Our sample size of only 219 patients was admittedly modest to train a machine learning model, but the number of patients recruited was constrained by project resources. With larger sample sizes or the ability to pool across several similar data sources (ie, from other ongoing studies being conducted in the ArthritisPower registry), additional features may emerge as important. As described earlier, we also recognize that deep learning models may provide performance improvements in some settings, which also would be facilitated by a larger sample size.

Machine learning methods can be used to classify self-reported PRO scores over time in patients with RA, and temporal methods such as HMM have been shown to consistently outperform independent models such as the RF for binary and multiclass classification tasks. Our study indicates that passively generated data from activity trackers can be useful in a machine learning environment to classify health status over time and that additional research should be conducted to further validate the use of such frameworks for clinical applications such as remote health monitoring and early intervention for patients with chronic illnesses. Encouragingly, we note that clinical use of monitored mobile apps and passive biosensor devices are now reimbursable by insurance in real-world settings. Since 2019, Medicare and other commercial insurance companies now compensate providers for care provided through remote physiologic monitoring and remote therapeutic monitoring programs [39]. These potentially transformative opportunities offer the potential to accelerate the adoption of approaches like ours and move remote monitoring out of research settings and into routine, clinical settings where they can positively impact care.