Design
GPS data collection occurred within a seven-day-Event Sampling Methodology (ESM) phase within a longitudinal, controlled clinical effectiveness trial [13].
Participants
Participants (n = 106; Inpatients: n = 69; outpatients: n = 37) were recruited from two specialized units within a psychiatric hospital with an open-door policy in Switzerland, from ongoing intake procedures. Inclusion criteria were: ≥ 18 years, ability to speak German. Exclusion criteria were: acute suicidal intent, acute substance dependency, active mania, and inability to read or complete assessments. Otherwise, all diagnoses were eligible (e.g., Affective Disorders, Anxiety Disorders, Somatoform Disorders, Mood Disorders, Anxiety-stress related Disorders, Somatic Disorders, Obsessive-Compulsive Disorder, Impulse Control Disorders, and Personality Disorders). Participants completed informed consent and were explicitly informed about and consented to GPS data collection.
Procedure
Participants carried a study-issued smartphone (i.e., not their personal phone), which was set to automatically collect GPS data as soon as the smartphone was turned on to avoid data loss if the phone was shut down during the assessment week. Participants were instructed to carry the phone with them during the study week. Participants explicitly gave permission to activate GPS on the study phones. They were informed that the studyphones (with disabled wi-fi and no SIM-card) would not be trackable and location data would be saved locally on the phone. Patients were free to leave the ward any time. Further, all patients were highly encouraged to partake in individual engagement exercises, which involved engaging in activities that kept them in contact with the important aspects of their life. Most patients went home weekends and some went home nights.
Assessments
GPS
GPS data was automatically logged every five seconds as a balance between high frequency data collection and battery life. GPS data were subsequently converted for analysis (latitude, longitude, date, time of day). GPS data for one patient on one particular day were only included if ≥ 1000 signals were available, which corresponded to GPS data for at least 1.4 h per day. The theoretically maximum number of GPS points was 17,280 for 24 h with recordings every 5 s. Subsequently, the ST-SBCAN algorithm – a state-of-the-art density-based clustering algorithm [14] – was applied individually for each patient and day and the obtained spatiotemporal clusters were merged with the GPS coordinates of the hospital (in the case of inpatient) and home (in case of outpatient) location of each patient. Coordinates of the hospital and home were defined in decimal degrees, and all destinations with centroid coordinates within a radius of 200 m of the hospital or home coordinates were given the respective label. All data points included in any estimated cluster (see below) were labeled by “not in-transit”. Finally, all data points that were not included in any of the estimated clusters were grouped in one cluster labelled “in-transit” to indicate that patient were moving between two clusters.
Questionnaires
Symptoms were assessed using the Brief Symptom Checklist (BSCL) [15], a 53-item self-report inventory measuring levels of psychopathology on a scale from 0 (not at all) to 4 (extremely). The nine subscales show sufficient to good internal consistency (Cronbach’s α = .75 to .90). Wellbeing was assessed using the Mental Health Continuum – Short Form (MHC-SF) [7], a self-report inventory consisting of 14 items that show high internal consistency (Cronbach’s α > .80). Each item assesses how often a statement was true during the past month, ranging from 0 (never) to 5 (almost every day). Psychological flexibility was measured using the Psyflex scale ([16]; Gloster AT, Block VJ, Klotsche J, Villanueva J, Rinner MTB, Benoy C, et al: Psy-Flex: A Contextually Sensitive Measure of Psychological Flexibility, In review) which measures core skills of psychological flexibility on a scale from 1 (very rarely) to 5 (very often). The Psyflex shows very high internal consistency (Raykov’s r = 0.91) and produces a single score with higher scores indicating better psychological flexibility qualities.
Data processing
Patients with more than 1000 GPS coordinates per day were included into the analysis. The ST-DBSCAN algorithm was used to identify unique destinations for each patient within a day [17]. The ST-DBSCAN algorithm contains three parameters that had to be assigned before estimating unique destinations: spatial distance, temporal distance, and number of points needed to form a cluster (see Additional file 2 for details). Thus, a destination was comprised of at least 10 data points which were all within a Euclidean distance of 200 m with a temporal proximity of 20 min [12]. The following measures were calculated from the GPS data points for each patient and day after completing the ST-DBSCAN algorithm (see Fig. 1 for an example).
-
a)
The individual hull area, i.e. the area (in km2) under the minimum convex polygon that includes all GPS data points of a unique cluster. It is the shortest possible line that surrounds all GPS points of a unique cluster with an outward curvature. The resulting area does not have any indentation.
-
b)
The total individual hull area (in km2) is the sum over all individual hull areas across one day.
-
c)
The total hull area (total activity area; in km2) is the area under the minimum convex polygon (see explanation above for the individual hull area) that includes all GPS data points across one day (disregarding unique clusters). The total hull area is always greater or equal to the sum over all individual hull areas.
-
d)
The distance travelled within destinations is the cumulative distance (in km) among all GPS data points within a cluster.
-
e)
The distance travelled across the entire day is the cumulative distance (in km) across all GPS data points, disregarding unique clusters.
-
f)
The time spent within a cluster (in minutes).
-
g)
The variability measures entropy and normalized entropy [11]. It is a measure to describe the variability of the time within a destination for a patient at a given day. For more details, see Saeb et al. [11]. A high entropy means that the patient has spent his/her time more evenly distributed over the clusters per day. In contrast, low entropy values mean that the time spent within a cluster varies within a patient at a given day.
-
h)
The location variance is a summary measure of the statistical variances of latitude and longitude [11] to determine the variability in patient’s destination. “In transit” destinations were not included in the calculation of location variance. Location variance is estimated by the logarithm of the sum of variances of latitude and longitude per patient and day.
The ST-DBSCAN function and all other functions necessary to calculate the measures mentioned above were taken from the statistical software R [18].
Statistics
Descriptive statistics included mean and standard deviation and median and interquartile range (25th percentile and 75th percentile). Due to the hierarchical nature of the data (clusters within patients per day), we used weighted statistics, i.e., patients with more clusters gave more weight to the calculated statistics. Results are reported for the entire sample, inpatients and outpatients, as well as for different types of destinations. The home address was not known for some patients (n = 17, covering 52 days) hence the cluster status could not be unequivocally allocated. P-values are based on Mann-Whitney test (comparison between in- and outpatients) and Spearman rank correlation coefficients (associations with continuous variables), whereby spatiotemporal data were first averaged across all clusters and days to obtain one value per patient.