This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mhealth and uhealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.
Is someone at home, at their friend’s place, at a restaurant, or enjoying the outdoors? Knowing the semantic location of an individual matters for delivering medical interventions, recommendations, and other context-aware services. This knowledge is particularly useful in mental health care for monitoring relevant behavioral indicators to improve treatment delivery. Local search-and-discovery services such as Foursquare can be used to detect semantic locations based on the global positioning system (GPS) coordinates, but GPS alone is often inaccurate. Mobile phones can also sense other signals (such as movement, light, and sound), and the use of these signals promises to lead to a better estimation of an individual’s semantic location.
We aimed to examine the ability of mobile phone sensors to estimate semantic locations, and to evaluate the relationship between semantic location visit patterns and depression and anxiety.
A total of 208 participants across the United States were asked to log the type of locations they visited daily, using their mobile phones for a period of 6 weeks, while their phone sensor data was recorded. Using the sensor data and Foursquare queries based on GPS coordinates, we trained models to predict these logged locations, and evaluated their prediction accuracy on participants that models had not seen during training. We also evaluated the relationship between the amount of time spent in each semantic location and depression and anxiety assessed at baseline, in the middle, and at the end of the study.
While Foursquare queries detected true semantic locations with an average area under the curve (AUC) of 0.62, using phone sensor data alone increased the AUC to 0.84. When we used Foursquare and sensor data together, the AUC further increased to 0.88. We found some significant relationships between the time spent in certain locations and depression and anxiety, although these relationships were not consistent.
The accuracy of location services such as Foursquare can significantly benefit from using phone sensor data. However, our results suggest that the nature of the places people visit explains only a small part of the variation in their anxiety and depression symptoms.
Passive and unobtrusive detection of the physical location of individuals has been made possible over the years by embedding global positioning system (GPS) systems into commonly used devices, such as mobile phones. Physical location alone is usually not very useful for understanding human activity, or the motivations that underlie that activity. In contrast to physical location, semantic location carries additional information about the meaning of the location [
A growing number of papers have shown that a variety of location features, measured by GPS, can detect mental health problems such as depression [
Local search-and-discovery services, such as Foursquare, can estimate semantic locations based on GPS coordinates and the data they have globally collected from populated areas in the world. When these services are embedded in a mobile app using an application programming interface (API), they can passively provide location-specific information for the locations that users visit. Since its launch in 2009, Foursquare has been used in research applications to accomplish diverse tasks, ranging from the analysis of individuals’ food and drink habits across cultures [
However, asking search-and-discovery services such as Foursquare about semantic locations, based on a given GPS coordinates, has limitations. First, GPS can be inaccurate, and particularly in denser urban environments, variability in GPS may lead to the detection of false locations. For example, one might be at a restaurant within a shopping mall, and the search-and-discovery service may classify the person as at a shop rather than a restaurant. Second, although these services can detect “residential” locations, they cannot distinguish a person’s home from another home they are visiting. These limitations prevent such services from being a reliable source of information, especially for behavioral sensing and intervention, where it is crucial to know exactly when a person is at home, work, a friend’s home, or other locations.
In addition to GPS, mobile phones can sense many more variables in the environment, such as light, sound, and Wi-Fi signals. Using a mobile phone, we can also determine what type of physical activity an individual is performing, how much time they spend in a location, and how they interact with their phones. Semantic locations may have distinct signatures, such as the length of time a person spends at a location, time of the day and day of the week that they visit, type of activities that they perform, and the sound and light conditions in the environment. These features may help us to determine if the place is home, a grocery store, place of worship, or a library. As an obvious example, a place that a person spends time over night is most likely home, and a bright place visited during the day, with intermittent walks and stops, is likely a store. Therefore, detection of semantic locations using mobile phone sensors seems feasible.
The aim of this paper was first to develop methods for improving mobile phone-based detection of semantic locations by incorporating sensors beyond the simple GPS. We developed methods for detecting semantic locations, and compared their accuracy to that of Foursquare. While improving semantic location detection is worthwhile and could further serve clinical and consumer-driven purposes, our second aim was to explore the relationship between semantic location detection and depression and anxiety. We specifically investigated the relationship between semantic location visits and the severity of depression and anxiety symptoms, as well as the differences between individuals with and without those symptoms.
We recruited participants between October 28, 2015 and February 12, 2016. The recruitment was done in collaboration with Focus Pointe Global (FPG), a company that specializes in market and scientific research strategies and participant recruitment and retention [
Interested individuals from the general population of the United States contacted FPG and were screened for eligibility using a brief questionnaire. Individuals were eligible for our study if they were at least 18 years old, able to read and understand English, owned a mobile phone with Android 4.4 through 5.1, and had access to Wi-Fi for at least one 3-hour period per day. We excluded individuals who indicated on self-report that they were diagnosed with any psychotic disorders, were unable to walk more than half a mile (4 city blocks), or had positive screens for alcohol abuse (Alcohol Use Disorders Identification Test [
Depressive symptoms were measured using the Patient Health Questionnaire, 9-item (PHQ-9) [
For anxiety assessment, we used the Generalized Anxiety Disorder, 7-item (GAD-7) [
We wanted to have a wide range of depression and anxiety symptoms in our sample, and therefore we selected roughly equal numbers of participants in four groups, based on their screening assessments: depressed and anxious, depressed and nonanxious, nondepressed and anxious, and nondepressed and nonanxious. In addition to assessment at baseline, we also assessed each participant’s depression and anxiety at week 3 and week 6.
Eligible participants were consented using procedures approved by the Northwestern University Institutional Review Board. Consenting was done using a website: participants were directed to a webpage that contained information about the study procedures, benefits, and potential risks. Specifically, participants were informed about the sensor data that were going to be collected from their mobile phones, the types of questions that would be asked throughout the study, and the procedures undertaken to protect their private information. After digitally signing the consent form, participants were enrolled in our study.
Each participant was enrolled for a period of 6 weeks. First, a study identification (ID) number was assigned to the participant by FPG. Participants were then asked to complete an online questionnaire regarding their demographic information, which consisted of their age, gender, race, and ethnicity, along with their US state of residence, and information about various aspect of their lives that could impact movement patterns (eg, health difficulties, number of jobs, and job locations). Finally, participants downloaded two apps:
After participants were enrolled, we started collecting two categories of data from their mobile phones: (1) sensor data, which contained data from the physical sensors as well as software services such as phone and short message service (SMS) communications; and (2) ecological momentary assessment (EMA) data, which consisted of daily questions that showed up on participants’ phones asking them about the locations they visited throughout the day.
The phone sensor data were captured using the
For the collection of EMA data, we used a second Android app,
A list of likely location names was provided to the user to choose from. This list was obtained from the Foursquare location API. The participant could also enter their own location name if it was not provided.
This list was adapted from Foursquare venue categories, and included Arts & Entertainment, Food, Nightlife Spot, Outdoors & Recreation, Professional or Medical Office, Spiritual, Shop or Store, Travel or Transport, and Home. In addition, we added Work, Another's Home, and Other. If the participant answered Other, they were asked to enter the location type. The EMA app saved the cluster center corresponding to each detected location, the visit times, and the participant’s answers to the questions regarding that location.
Location category labels reported by our study participants (left) and their corresponding high-level Foursquare location categories.
EMA app Location Category | Foursquare Location Category |
Nightlife Spot (Bar, Club) | Nightlife Spot |
Outdoors & Recreation | Outdoors & Recreation |
Arts & Entertainment (Theater, Music Venue, Etc.) | Arts & Entertainment |
Professional or Medical Office | Professional & Other Places |
Food (Restaurant, Cafe) | Food |
Home | Residence |
Shop or Store | Shop & Service |
Travel or Transport (Airport, Bus Stop, Train Station, Etc.) | Travel & Transport |
Work | - |
Another’s Home | - |
Purple Robot and EMA app anonymized any sensitive information before storage and transmission. Specifically, the apps used an MD5 hashing algorithm [
We wanted to assess how well Foursquare could predict the type of locations that users reported daily. To do so, we used the Foursquare wrapper library [
Foursquare’s response to our queries was in JavaScript object notation (JSON) format [
The location category returned by the Foursquare website was too specific, being as detailed as “Cambodian Restaurant” or “College Math Building”. Since we did not need this level of detail in our study, we used Foursquare’s Category Hierarchy [
After querying the Foursquare category for each location cluster, we compared it to the category reported by the participant, and calculated the accuracy (see section: Classifier Evaluation). We skipped locations reported as Work, Another’s Home, or Spiritual for this comparison, since these did not exist in Foursquare categories. The calculated accuracy gave us the performance of Foursquare in predicting semantic locations.
To classify semantic locations from phone sensor data, we first calculated their features. These features were extracted from all sensor data that were gathered during a visit to a location. In this way, for every location visit, we obtained one feature vector. This vector consisted of 45 features, which will be described in the following sections.
Light features were calculated from light intensity, in
Sound features captured different aspects of the sound in the environment. Specifically, we sampled the audio using the phone’s microphone every 5 minutes, each time for 15 seconds. From each 15 second audio recording, we extracted the power and the dominant frequency. Power was calculated as described in
To calculate the dominant frequency, we obtained the amplitude of the fast Fourier transform of the audio signal, and found the frequency that maximized the amplitude.
We used screen activity to measure the amount of participants’ interaction with their phones. We calculated the number of times the screen state transitioned from
We used the physical activity states provided by the Android Activity Recognition API. We sampled this API every 10 seconds. The Physical Activity API uses the accelerometer sensor to detect the following physical activities:
Communication features consisted of the total number of incoming, outgoing, and missed phone calls. In addition, we derived the number of incoming and outgoing SMS text messages.
These features were calculated from the latitude and longitude values provided by the GPS sensor, sampled every 5 minutes. GPS features included average latitude, average longitude, and
In addition to these features, by filtering out the data points that were outside the 50-meter radius of a location’s average latitude and longitude during a visit, we approximated the
We sampled the current access point’s media access control address and the number of available Wi-Fi networks every 5 minutes. We only used the number of Wi-Fi networks as a feature.
We calculated the visit duration, the
We obtained the weather conditions at the location and time of visits. For this data, we used the Weather Underground service [
We wanted to see how successfully we could detect semantic locations, reported by the participants, using the sensor features that were passively collected from their mobile phones. For this classification problem, we used ensembles of decision trees with the gradient boosting optimization method [
A decision tree, shown in
Each decision tree in the ensemble is assigned to one class, and provides a
The final class probabilities are calculated as a softmax function of the predictions scores using the equation shown in
Therefore, for each given feature vector, the ensemble provides a probability distribution over the classes.
Sound power calculation; where S(n) is the sound amplitude (dB) at sample n, and N is the total number of samples.
Location variance feature; calculated as the logarithm of the sum of variances in latitude and longitude values.
An example of a single decision tree in the ensemble of decision trees. Each circle is a tree node, where a decision is made by comparing a feature value fxx to a threshold. For a given feature vector, depending on which path is taken, a single prediction score is generated, shown in the boxes. Note that one of the outgoing branches of each node is also dedicated to the situation where the data is missing.
Aggregation of prediction scores made by individual trees; where gk,m represents the decision tree k in class m, that maps a feature vector x to a prediction score gk,m(x), and ym is the ensemble prediction score for class m. K is the total number of trees for each class.
The goal of training is to push the class probabilities
While the logistic loss term in
Class probability calculation; where pm represents the probability for class m, ym is the ensemble prediction score for class m, and M is the number of classes.
General form of the cost function; where l(yi,yi*) is the logarithmic loss [
Cost function for training a new tree added at iteration t; where gt(.) is the prediction score provided by the new tree. See
In the gradient boosting method, trees are added to the ensemble one by one. The ensemble starts with one tree, which is fit to the training data using the cost function in
The parameters of the new tree are chosen such that
We tuned the hyperparameters of the XGBoost classifier by grid search, and used data from 10% of participants. Within this subset of data, we performed a 10-fold cross-validation to estimate the area under the curve (AUC; see Classifier Evaluation). We chose the set of parameters on the grid that maximized this AUC.
The parameters included in hyperparameter tuning were
Our goal was to create algorithms that could determine the semantic locations for unseen individuals, so we trained and evaluated the classifiers using a subject-wise cross-validation scheme. Specifically, we randomly selected 70% of the subjects to train the classifier, and used the remaining 30% to evaluate its prediction accuracy. We repeated this procedure 100 times. The distribution of prediction errors on held-out participants used as
To calculate the prediction error in each round of cross-validation, we estimated the receiver operating characteristic curve, and calculated the AUC. The AUC ranges between 0 and 1, with 0.5 indicating chance level performance. The advantage of using AUC is that it is robust to the imbalance in the number of samples in the classes. Therefore, by iterating over all participants as
We evaluated the relationship between the amount of time participants spent at each semantic location and their level of depressive and anxious symptoms, measured by PHQ-9 and GAD-7, respectively. We performed two analyses. First, we calculated Pearson’s correlation between the scores and the time spent in each location, across all participants. For the second analysis, we divided participants into depressed and nondepressed, as well as anxious and nonanxious, based on their scores. For depression, we defined the two groups by considering participants who consistently had PHQ-9 <10 (termed nondepressed) or PHQ-9 >10 (termed depressed) across all three assessment time points. Likewise, for anxiety, we defined the two groups by considering participants who consistently had GAD-7 <10 (termed nonanxious) or GAD-7 >10 (termed anxious). Therefore, in both analyses, we excluded the participants who crossed the PHQ-9=10 or GAD-7=10 thresholds. The main reason was that these participants could not be reliably classified. Furthermore, if we had included them, it would have added two additional categories (those who improved and those who got worse), which would have reduced power. It is also unclear how we would interpret any relationships with participants transitioning from one clinical state to another. After dividing subjects into these groups, we compared the duration of time that participants spent at each semantic location between the groups, using two-sample
A total of 208 individuals passed the eligibility criteria for participating in our study, and were recruited. One participant did not install the software on their phone, and another had invalid GPS data. These two participants were removed from all analyses. Of the remaining 206 participants, 22 (10.7%) stopped providing data before the end of the 6-week period. However, many continued to send data after the end of 6 weeks, with 27 (13.1%) providing more than 60 days of data.
The 206 participants included in the analyses were 170 females (82.5%) and 36 males (17.5%). Participants’ ages ranged between 18 and 66 years, with a mean of 39.3 (SD 10.3). The participants’ locations were diverse, covering most of the populated states and major cities in the United States. Most of these locations (86.8%, 178/206) were in “mostly urban” areas, as defined by the United States Census Bureau [
In response to a question on employment status, 61.2% (126/206) indicated that they were employed, 20.9% (43/206) were unemployed, 8.3% (17/206) had a disability which prevented them from working, and 1.9% (4/206) were retired. Sixteen participants (7.8%, 16/206) did not specify their employment status. Of the 126 employed participants, 98 (77.8%) had one, 23 (18.3%) had two, 4 (3.2%) had three, and one (0.8%) had four jobs. In addition, of these 126 participants, 36 (28.6%) worked in more than one location.
The semantic locations reported by the participants were diverse. While most participants reported the predefined locations in Purple Robot, as the example in
The optimized hyperparameters for the XGBoost classifier were the following: for sensor-only classification, we set the number of trees to 200, the fraction of samples seen by each tree to 0.2, and the fraction of features to 0.5. For classification based on both sensor and Foursquare features, these three parameters were set to 300, 0.25, and 0.2, respectively. In both scenarios, we set
We first measured how accurately Foursquare could detect the semantic locations reported by participants. To obtain the locations detected by Foursquare, we used the GPS coordinates of that location, and queried Foursquare about its closest match to that location. We then compared the results to the locations reported by participants, and calculated the AUC for each category. The results are shown in the left column of
Mean area under the curve (AUC) in detecting each location category, using Foursquare only, mobile phone sensor features only, and both. Note that we could not use Foursquare to detect Work, Another’s Home, or Spiritual locations; hence there are no results.
Semantic Location | Foursquare | Sensor | Sensor+Foursquare |
Travel or Transport, mean (CI) | 0.54 (0.49-0.60) | 0.79 (0.72-0.86) | 0.84 (0.78-0.91) |
Nightlife Spot, mean (CI) | 0.61 (0.53-0.72) | 0.87 (0.78-0.94) | 0.89 (0.79-0.95) |
Spiritual, mean (CI) | N/A | 0.82 (0.75-0.88) | 0.87 (0.80-0.92) |
Outdoors & Recreation, mean (CI) | 0.59 (0.53-0.64) | 0.81 (0.71-0.88) | 0.86 (0.75-0.92) |
Arts & Entertainment, mean (CI) | 0.67 (0.61-0.73) | 0.88 (0.85-0.91) | 0.92 (0.88-0.95) |
Work, mean (CI) | N/A | 0.86 (0.82-0.90) | 0.87 (0.83-0.91) |
Professional or Medical Office, mean (CI) | 0.65 (0.58-0.73) | 0.85 (0.80-0.91) | 0.88 (0.83-0.93) |
Another's Home, mean (CI) | N/A | 0.77 (0.69-0.82) | 0.83 (0.75-0.89) |
Food, mean (CI) | 0.64 (0.59-0.68) | 0.79 (0.74-0.83) | 0.83 (0.78-0.87) |
Home, mean (CI) | 0.53 (0.51-0.56) | 0.96 (0.95-0.97) | 0.96 (0.95-0.97) |
Shop or Store, mean (CI) | 0.76 (0.73-0.79) | 0.86 (0.82-0.90) | 0.89 (0.85-0.92) |
Mean AUC | 0.62 | 0.84 | 0.88 |
(A) Location report data from one example participant, collected between 11/07/2015 and 11/28/2015. Each rectangle shows the period of time the participant has been in a specific location. The sensor data during that time period is used to create a feature vector, which is then used to detect that semantic location. (B) Top locations visited by all participants, sorted by how many participants visited them. As the total number of unique reported locations was 370, we only included the ones that had been visited by at least two participants.
We wanted to determine whether mobile phone sensors alone could detect the semantic location of participants. We used 45 features that were extracted from a variety of sensors during the time that the participant was visiting a location (see section: Sensor Features). We trained the XGBoost classifiers to map these features to semantic locations, and tested these classifiers on participants that they had not seen during training. Compared to Foursquare, the AUC of detecting certain locations was considerably higher (
Next, we used both Foursquare and phone sensor data to see if this approach could further increase the accuracy of our classifiers. To this end, we added two extra features to the 45 features that we previously used for training the classifiers: the Foursquare location type, which was represented by a binary vector with 9 elements (each corresponding to one category); and the distance to the nearest Foursquare location. Therefore, the total number of features increased to 55. Using this new feature set further increased the average AUC to 0.88 (
Finally, we asked which features contributed the most to detecting semantic locations by estimating their
We evaluated the relationship between the time spent at different semantic locations and the level of depression and anxiety symptoms, measured by PHQ-9 and GAD-7, respectively. First, we evaluated the linear correlation between these two groups of variables (
Linear correlation coefficients (
PHQ-9 Week 0 | PHQ-9 Week 3 | PHQ-9 Week 6 | GAD-7 Week 0 | GAD-7 Week 3 | GAD-7 Week 6 | |
0.057 | 0.073 | 0.089 | 0.083 | 0.101 | 0.097 | |
Shop or Store | -0.010 | 0.0183 | -0.020 | 0.001 | -0.030 | -0.038 |
Work | -0.084 | -0.139 | -0.140 | -0.083 | -0.085 | |
Food | -0.088 | -0.093 | -0.089 | -0.086 | -0.115 | |
Another's Home | 0.046 | -0.065 | -0.064 | -0.016 | -0.003 | 0.000 |
Professional or Medical Office | 0.029 | 0.049 | -0.069 | 0.019 | 0.051 | |
Outdoors & Recreation | 0.016 | -0.123 | -0.101 | -0.065 | -0.131 | -0.109 |
Arts & Entertainment | -0.092 | -0.090 | -0.044 | -0.055 | -0.057 | |
Travel or Transport | -0.070 | -0.037 | 0.082 | 0.012 | ||
Spiritual | -0.041 | -0.078 | -0.094 | |||
Nightlife Spot | -0.045 | 0.041 | -0.063 | -0.045 |
We also performed a group difference analysis, by dividing the participants into two groups (once based on their depression scores, and another time based on their anxiety scores). We compared the duration of time participants spent at each semantic location between these groups. For depression, the nondepressed group consisted of 51 participants and the depressed group consisted of 68 participants. The remaining 88 participants crossed the PHQ-9=10 threshold between the assessments, and were excluded from this analysis because they could not be clearly classified. For anxiety, the nonanxious group consisted of 51 participants while the anxious group consisted of 61 individuals. The remaining 96 participants crossed the GAD-7=10 threshold and were excluded.
The results for depression are shown in
Mobile phone sensor feature importance in detecting semantic locations. Features are sorted based on their importance, from top to bottom. The importance of each feature is calculated by computing the decrease in the cross-validated area under the curve when that feature is removed from the feature set. Negative values indicate an increase in performance. Each value is the mean feature importance across cross-validation folds, and error bars show the standard error of the mean.
The relationship between semantic location visit duration, depression, and anxiety. Each bar shows the average time spent at each location by (A) depressed versus nondepressed and (B) anxious versus non-anxious groups, relative to the total time spent by all participants in that location. Error bars show 95% CIs. Both mean and CIs are obtained by bootstrapping over 1000 iterations. Stars indicate significant difference between the means, obtained using a 2-sample t-test at the P<.05 level. However, adjusting for multiple comparisons, these differences are all nonsignificant.
In this paper, we were able to detect the type of locations that individuals visited, using data passively collected from their mobile phones. The phone sensor data were especially crucial in detecting these semantic locations. Sensor features alone produced accuracies that were more than 20% greater than those reported by Foursquare, and combining the sensor features with Foursquare produced even greater accuracy. This result is not surprising since detecting semantic location based on GPS alone is not necessarily accurate, especially in urban areas [
The performance of the classifiers considerably varied across the location types. While Home could be detected with an AUC of above 0.95, the classification AUC for Another’s Home and Food was 0.83. This variability may have multiple causes. First, visits to certain locations, such as home or work, are more regular in time, which makes them easier to detect based on the time of visit. Another cause might be that some semantic locations such as Travel or Transport were less represented in the data, since participants visited those locations less often. This factor has likely made it difficult for classifiers to find the feature patterns that are distinct indicators of those locations. Finally, while some locations (eh, Home) have a clear definition, participants may have been confused about which location type to report for some other locations. For example, a participant might have had food in a store, and have reported that location as either “Food” or “Shop or Store”. Overall, although classification performance varied across different semantic locations, it significantly benefited from incorporating mobile phone sensor data.
While we could detect the types of locations, we found only few significant relationships between the amount of time spent in those locations and self-reported symptoms of depression and anxiety. Furthermore, these few relationships were weak and inconsistent. This failure may have multiple explanations. First, our categorization of semantic locations was based largely on Foursquare categories, which was not developed with mental health or wellness in mind and may not be accurate, useful, or relevant to mental health. These categories were also often imprecise (eg, “Professional or Medical Office”). For mental health research, we may need to create location categories that are mostly relevant to the factors that influence mental health.
Second, the lack of a consistent relationship between semantic location and depression or anxiety may reflect larger problems in the literature. Past research has examined smaller, discrete samples of participants, such as university students [
It is possible that this finding is accurate: that the kinds of places we go is
There are a number of limitations that need to be mentioned. First, when detecting semantic locations we did not consider the transitions between locations. Knowing the transition probabilities can be useful; for example, it may be more likely to visit Home after Shop or Store. One reason for not considering transitions was that we only considered the top 11 most-visited locations for the classification problem, and therefore the sequence of semantic locations in the training data were not necessarily consecutive in time. Another reason was the existence of gaps in the data, which caused further separation between consecutive visits. Incorporating transition probabilities in detecting semantic locations, when possible, will likely increase the classification accuracy of the resulting algorithms.
Second, semantic locations may have signatures that we failed to capture through our phone sensors. For example, the type of phone apps people use, or individuals who they contact, can be a good distinguishing feature between locations. Using such sources of information as features in future studies may improve the performance of semantic location detection.
Third, our study participants differed from the general population in a few aspects. Approximately 83% of the participants were women, significantly different from 50.8% in the general population of the United States [
Fourth, the assessment of depression and anxiety in this study was based on self-report, and therefore may not generalize to assessments based on diagnostic interview. A clinical diagnosis usually involves an in-depth interview and consideration of confounding factors, based on the criteria in the Diagnostic and Statistical Manual of Mental Disorders [
Fifth, data collection took place from late October to early February, and thus most participants were providing data during the winter holiday season. While the geographic diversity of the sample allows us to account for variations in weather (eg, participants from Florida experienced a much different climate than those in Minnesota), we recognize that holiday-related travel, such as spending time at other family members’ homes, and holiday-related time away from work presents a departure from an individual’s typical behavior. The holiday season may have served as a confounder, as participants may have been engaged in activities not representative of how they would behave during other times of the year. Furthermore, the 6-week study period may not have been long enough to detect changes or meaningful relationships between behavioral patterns and mood. Ultimately, we aim to develop models to ascertain the relative components of these factors. However, as this is a relatively new field of inquiry, the timing and length of this study protocol may have interfered with our ability to detect true signals.
In conclusion, mobile phone sensors promise considerably more accurate estimations of individuals’ daily life behaviors. In this study, we have shown that semantic location (the type of locations that people visit) can be detected using a combination of phone sensors and a mapping service such as Foursquare. We performed this study in a sample that was diverse in terms of geographic location, climate, education, employment, and lifestyle. However, there were no consistent relationships between the time spent at different locations and depression or anxiety. Future research should focus on those semantic locations that are more likely to be relevant to depression or anxiety. In addition, longer studies that extend across seasons, and larger studies that are more adequately powered to manage the level of dimensionality in human subject data, will be better positioned to investigate the relationships between semantic locations and mental health. The advancement of mobile phone technology will facilitate the design of these future studies.
application programming interface
area under the curve
ecological momentary assessment
Focus Pointe Global
Generalized Anxiety Disorder, 7-item
global positioning system
identification
just-in-time adaptive intervention
JavaScript object notation
Patient Health Questionnaire, 9-item
standard deviation
short message service
extreme gradient boost
This study was supported by the following National Institute of Health grants: 5R01NS063399, P20MH090318, and R01MH100482. The authors would like to thank Weather Underground for providing access to weather history data.
None declared.