Studies have reported the efficacy of physical activity in reducing the risk of disease; however, physical inactivity is on rise in the United States [1]. Considering that physical inactivity was the fourth leading cause of mortality in 2016 [2], there is much emphasis on developing effective methods to maintain healthy levels of physical activity. One promising solution is wearable fitness trackers that enable individuals to monitor their activity levels and patterns to ensure a healthy level of physical activity [3].

It is reported that about 20% of the general health-tracking population use smart devices such as medical gadgets, mobile phone apps, or online tools to track their health data [4]. The use of technology to objectively monitor physical activity is associated with higher levels of activity [5]. However, the potential benefits derived from the use of physical activity trackers are challenged by the limited and transient adoption of devices that necessarily require sustained use to exert their intended effect. Continued engagement with fitness trackers is an issue that warrants further investigation [5] A previous study [6] found two factors associated with the adoption and sustained use of physical activity trackers: (1) the number of digital devices owned by the participants, and (2) the use of activity fitness trackers and other smart devices by the participants’ family members. The existence of these two factors bode well for the increased use of activity trackers; one study [1] demonstrated that motivational factors are associated with physical activity levels. Time constraints, fatigue, and aversion to exercise are some of the barriers to engaging in physical activity [1]. It has been reported that adjustments to activity targets are likely to enhance the users’ commitment to physical activity and engagement with fitness trackers [7].

With the rise of machine learning and availability of activity tracker data, it is possible to create a model that can learn from users’ behavior and adjust activity targets. Machine learning methods have broadly been applied to many health care areas such as cancer staging, risk assessment, and drug recommendation systems [8]. Researchers have studied the accuracy of activity trackers for energy expenditure assessment [9]. Having an automated personal trainer enabled by data mining techniques can be useful for an amateur athlete who cannot afford a personal trainer [10]. Similarly, effective feedback methods can be used for helping both athletes and coaches [11].

Although, there have been some attempts to study the benefit of activity trackers, there is room for studies on how to make use of activity trackers on a sustained basis. This study, to the best of our knowledge, is the first of its kind to develop a machine learning method to adjust the activity target for activity trackers. Machine learning techniques such as, but not limited to, lasso regression, ridge regression, Bayesian ridge regression, neural networks, random forest regression, and support vector regression have been used for prediction in the medical field [12].

Feature selection is an important step for improving model performance [13]. Prior to applying machine learning techniques, it is essential to study the data to find features that might negatively or positively affect the model [8]. In this work, we applied two feature selection techniques with a support vector machine [14]: principal component analysis [15] and recursive feature elimination.

We compared predictive models developed over (1) all features, (2) features generated by principal component analysis, (3) features selected by recursive feature elimination, and (4) features found from the authors’ previous study [6] that characterizes participants environments. Figure 1 shows the study flow diagram highlighting the key steps undertaken in this study.

The purpose of this study was to develop a predictive model to estimate achievable weekly step goals. By setting realistic targets that are neither too easy nor too difficult to achieve, activity tracker users can be encouraged to continue to meet their weekly activity goals, and at the same time, to find utility of the tracker. We chose individuals who were overweight as our first use case, because the benefits of sustaining or even increasing physical activity in this population are well known, while there have been a lack of sustained interventions addressing this issue [16-18].

The study was designed as a 9-week, nonrandomized pilot in which the data were analyzed retrospectively. For this purpose, adults (N=30) with a BMI of 25 kg/m² or greater were enrolled from a local Massachusetts General Hospital–affiliated clinic. After screening the participants and seeking their consent, the research team directed the study participants to visit the Wellocracy website [19] to read information regarding the study and review different types of activity trackers (and their features) available for their use during the 9-week study. The study staff assisted the participants with the device setup process as needed. For the study, 27 participants chose to use the FitBit Charge (Fitbit Inc) model, 2 chose the FitBit One (Fitbit Inc), and 1 chose the FitBit Zip (Fitbit Inc); 10 participants data were not used for analysis (7 participants either did not use the activity trackers at all or used less than 3 days a week, and an additional 3 participants demonstrated irregular use of their activity trackers on week by week basis). The data from the remaining 20 participants were used for model building.

The following surveys were collected from each participant both at enrollment and at the closeout stages: Behavioral Regulation in Exercise Questionnaire (BREQ-2) [20]; Barriers to Being Active (BBA) [21]; Patient Health Questionnaire (PHQ-8) [22]; Prochaska Stage of Change [23]; and Patient-Reported Outcomes Measurement Information System (PROMIS) Global-10 [24]. These surveys included questions about technology use and the ownership of electronic devices (BREQ-2), questions related to perceived barriers to exercise and activity (BREQ-2 and BBA), depression screening questions (PHQ-8), stages of change (Prochaska), and general health questions (PROMIS Global-10).

The BREQ-2 is designed to gauge the extent to which reasons for exercise are internalized and self-determined based on the following categories: motivation, external, introjected, identified, and intrinsic. In contrast, the BBA assesses whether participants gauge certain categories as reasons for inactivity and includes energy, willpower, time, and resources. A score of 5 or greater for a category indicates that it is a substantial barrier to a person’s ability to exercise.

Participants were instructed to continuously wear the activity tracker for the entire period of the study. The first week was treated as a run-in period. Participants were contacted minimally during the remaining 8-week period to facilitate observation of participants’ activity tracker habits without interference. At the end of the study, participants completed a closeout survey either online or in paper format and underwent a phone interview to gather information on their experiences during the study. All interviews were conducted and transcribed for analysis by a trained neuropsychologist.

We divided the data into disjoint training and testing sets. We chose 16 participants’ data as training data set and the remaining 4 participants’ data were used as the testing data set. We explored and fine-tuned hyperparameters using the training data set. Each participant was informed and had a fixed weekly activity goal for weeks 2 through 8; it was fixed for each participant at 110% of their week 1 average daily step count. Since we didn’t have a means to adjust the activity goal for each week for each participant, this value was used as an estimate of the personalized average daily step count goal for each week. We ignored the week if a participant used the tracker device for less than 3 days during the week.

We collected all data from questionnaires and the participants’ daily step count data from their activity trackers. The questionnaires generated 96 variables. The variables were screened to determine candidate predictors for building a machine learning model. In the first pass of the variable screening process, 11 variables whose variance was zero were eliminated. In the second pass, we examined pairwise correlations among all remaining variables to eliminate redundant ones. Since we found that all pairwise correlations were less than 0.6, we considered all variables to be nonredundant.

Figure 2 presents the models used to predict participants’ activity target. Selection of these models was based on their suitability and capabilities. Bayesian ridge regression estimates a probabilistic model of a ridge regression [25]. Lasso regression is robust to overfitting due to its regularization penalty [26]. Random forest models are versatile for numeric and categorical predictors and for classification and regression tasks. Random forest models can also be interpreted easily and are less susceptible to underfitting [27]. Neural network models consist of hundreds or thousands of neurons that perform mathematical operations to recognize patterns [28]. Support vector regression models are regression models whose optimization is unaffected by the dimensionality of the data [14].

We employed these models in four cases: (1) using all features, (2) using new features built by principal component analysis, (3) using important features found by recursive feature elimination, and (4) using a subset of features found from the authors’ previous study [6] that characterizes participants environments.

In this study, we used participants’ mental health, behavioral information, and weekly activity performance as inputs. For every 7-day moving window, the subsequent 7 days were considered as the forthcoming week. The questionnaires provided a highly redundant and low variance data set. These features coupled with 7 daily steps counts of the week and a normalized Shannon entropy (E_s) value of the weekly step count were considered candidate predictors for target step count for each individual participant for the forthcoming week [29]. The normalized Shannon entropy was computed as where i denotes the day, p_i denotes the portion of the total weekly steps completed on day i, and N denotes the number of days per week (ie, 7). The normalized Shannon entropy varies between 0 and 1. A value close to 0 indicates that the daily step counts of the participant throughout the week were irregular; in contrast, a value close 1 indicates that the step counts were consistent.

In total there were 93 candidate features for building the predictive models: 85 features from the questionnaires, 7 features from the daily steps of the week, and one weekly feature (ie, Shannon entropy). We used all 93 features for models developed in case 1.

For case 2, we developed a principal component analysis model. Principal component analysis is a dimension reduction technique that is widely used for extracting uncorrelated features components from correlated variables [14]. The principal component analysis used the 85 features from the questionnaires. We combined the principal components with the 7 daily and 1 weekly step count features.

In case 3, to identify the important features, we performed recursive feature elimination with support vector regression using the 85 questionnaire variables. We augmented these features with the 7 daily step count features and the normalized Shannon entropy value of the weekly step count.

Lastly, for case 4, we selected 2 features that were found to be important from the questionnaire in our previous work [6] studying the link between the participants’ environmental factors and the adherence to the use of activity trackers. These features were (1) the number of digital devices owned by the participants, and (2) the use of activity fitness trackers and other smart devices by the participants’ family members. We appended these two features with the daily step count features, and Shannon entropy of the weekly step count.

We trained the models in this study with 10-fold cross-validation using the training data set. We compared the performance of the models using mean absolute error (MAE) and adjusted R². We tested the final models (models A, B, C and D) of each case (cases 1, 2, 3, and 4) on an unseen test data set.

We used Python (version 3.5.0) and R (version 3.4.1) for model development and statistical tests. We performed a one-sided paired t test (P<.05 were deemed significant) for statistical comparison between the models. The null hypothesis was that the mean error of two models were equal, with the alternative hypothesis that the candidate model had a lower mean error than the mean error given by the comparison model.

We present the characteristics of the study participants in Table 1. Among them, 10 participants were eventually removed from the study since they stopped using the activity trackers.

Over the span of 8 weeks, not all participants met their week-by-week step objectives. By and large, under half of participants met their progression objective every week (see Table 2). We present the distribution step count for all participants for each of the 9 weeks in Figure 3. We also presented the distribution of steps for each day of the week during the 9-week study period (see Figure 4).

We performed a grid-search with 10-fold cross validation over the training data set. Lasso regression with α=0.01, Bayesian ridge regression with α=0.0001, and support vector regression with polynomial kernel of degree 2 with a regularization parameter of 0.00001 gave the best performance. Optimal parameters for random forests and neural networks depend on feature selection techniques. We found 30 variables from the questionnaires to be important using recursive feature elimination with support vector regression (see Figure 5). In the case of principal component analysis, the top 14 principal components explained 100% of the variance of the variables in questionnaires as shown in Figure 6.

We report the performance for all models of 10-fold cross-validation with the training data set in Table 3. We used mean absolute error (MAE) and adjusted R² for the model comparison. Among all models developed, the Bayesian ridge regression model that used 93 features (case 1) gave the best performance for the training set (see Figure 7; MAE 1672, 95% CI 1640-1704; adjusted R²=0.85). The Bayesian ridge regression model is referred to as model A in the rest of this paper.

Similarly, in case 2, we appended 14 principal components extracted from principal component analysis to the 7 daily and 1 weekly step count features resulting in 22 features. A random forest model (MAE 1700, 95% CI 1664-1737; adjusted R²=0.91) gave the best performance among all models (see Figure 8). This random forest model is referred to as model B for the rest of this paper.

In case 3, we appended variables found by recursive feature elimination to the 7 daily step counts and 1 weekly entropy feature which resulted in 39 features. A random forest model (MAE 1774, 95% CI 1739-1809; adjusted R²=0.81) offered the best performance among all the models developed with these features (see Figure 9). We refer to this random forest model as model C.

Finally, for case 4, we coupled two features found in the previous study [6] with the 7 daily step counts and 1 weekly entropy feature which resulted in 10 features. A neural network gave the best performance across all models (MAE 1622 steps, 95% CI 1589-1655; adjusted R²=0.89) (see Figure 10). We refer to this neural network model as model D.

Model D gave the best predictive performance among all the models. It had the lowest MAE across all models explored. We compared the predictive power of the model D (neural networks) with those of models A, B, and C using the testing data set. We found that model D gave a better predictive performance as shown in Figure 11. We performed comparisons using t tests between the errors generated by the model D and errors generated by models A, B, and C as shown in Table 4. We observed that model D’s lower errors in comparison to those of Bayesian ridge regression (model A: P=.01), random forest (model B: P<.001), and random forest (model C: P=.01) models were significant.

We evaluated the naïve rule: using participants’ average daily step count of the week as the prediction for the subsequent week’s activity goal [30]. This is a reasonably competitive approach because the weekly target exhibited strong autocorrelation. The naïve rule achieved an MAE of 1664 steps while model D achieved an MAE of 1545 steps (95% CI 1383-1706) for the 4 test participants over a period of 8 weeks.

Its reported that the Fitbit activity devices have margin of ±5% error in their step count readings [31,32]. We tested the performance of our best model (model D) on three noisy data sets generated from the original test data set by adding ±1%, ±3%, and ±5% noise. The first noisy data set was generated by adding random noise between –1% and +1% to the Fitbit readings. Similarly, the second and third noisy data sets were generated by adding ±3% and ±5% noise to the original data set. To conduct the sensitivity analysis, we generated 100 replications of each noisy data sets by adding different random noises with limit specified for each data set. These data sets were used to evaluate the performance of model D.

The model D achieved an MAE of 1606 steps (95% CI 1490-1755) for the noisy test data set generated using margin of ±1% error which is approximately 4% higher than the MAE achieved for the original noise-free test data set. Similarly, model D’s MAE was 1670 steps (95% CI 1571-1840) for the margin of ±3% error which was approximately 8% higher than the MAE achieved for the original noise-free test data set. Finally, for the data set generated using margin of ±5% error, model D’s MAE of 1710 steps (95% CI 1621-1908) was approximately 10% higher than the MAE achieved for the original noise-free test data set.

This led to the empirical observation that 2%, 6%, and 10% noise in Fitbit readings leads to approximately 4%, 8%, and 10% increases in prediction error, respectively. Therefore, we assume that the model error sublinearly increases with margin of error in activity tracker readings. However, one might need to validate this observation with further evaluation data.

We experimented with integrated gradients [33] in order to analyze the features of model D. This method provides a score that reflects the contribution of each variable to the response variable by calculating the integral of the gradient of the response variable with respect to that variable. We report the top features for model D in Figure 12.