51 patients and 50 non-depressed controls (matched for sex, age, and body mass index) participated in this study. Patients were recruited from outpatient clinics and local psychotherapists in Berlin, Germany. Controls were recruited from advertisings in local newspapers. The data of 14 control participants and 16 patients were lost due to hardware failure either in the palm handhelds, the accelerometer, or in the time-based assignment of accelerometer and subjective mood data. Thus, the final sample consisted of 36 controls (n = 15 male, n = 21 female), and 35 depressed patients (n = 13 male, n = 22 female), with age ranging from 18–60 (M = 39.31, SD = 11.12) years, and body mass index (kg/m2) ranging from 18–39 (M = 24.60, SD = 4.14). Controls and patients did not differ with respect to sex, age, and body mass index (see Table 1). Patients and controls were diagnosed with a standardized diagnostic interview for DSM-IV disorders (Schneider and Margraf 2011) by trained and certified psychotherapists. All patients fulfilled DSM-IV criteria for current diagnosis of major depression (n = 25 recurrent major depression, n = 6 single episode of major depression). In total, n = 22 patients had additional diagnoses (n = 10 one additional diagnosis, n = 9 two, n = 2 three, n = 1 five additional diagnoses). Among these comorbid diagnoses, anxiety disorders were the most frequent (n = 13). Non-depressed controls had no current and no history of mental disorders. Patients had significantly higher scores on the Beck Depression Inventory (BDI) than controls (see Table 1). A total of n = 18 patients took psychotropic medication (SSNRI n = 3, SSRI n = 7, tricyclic antidepressant n = 3, antiepileptic medication n = 4, neuroleptics n = 4, dopaminergic antagonists n = 1, MAO inhibitors n = 2, amphetamines n = 3). Most patients received a single substance (n = 10), seven patients two types of medication simultaneously, and one patient four types of medication.
Assessment of Gait Parameters and Body Motility
Body activity was assessed with two accelerometers, in accordance with published recommendations (Foerster and Fahrenberg 2000). One accelerometer (three-channel: sagittal, longitudinal, transversal axis) was placed over the participant's trunk and another accelerometer (one-channel: sagittal axis) placed on the outer part of the upper right leg (Foerster and Fahrenberg 2000; Reichert et al. 2015) (both accelerometers provided by Vitamove, NL). Acceleration data (range: ± 6 g) were sampled continuously with a sampling rate of 128 Hz. From the raw data, periods of active (i.e., activities in upright position, for example walking, running, bicycling) and sedentary behavior (i.e., lying, sitting) were identified automatically offline according to published movement pattern identification algorithms (Bussmann et al. 2001; overview in Bussmann et al. 2009) using commercial software (Vitascore, Temec, NL). The participant's nonspecific amount of overall movement intensity (i.e., body motility) was calculated in accordance with previous reports (von Haaren et al. 2016) using customized scripts (Matlab). For periods of walking, four specific gait parameters were additionally extracted, corresponding to the laboratory study of Michalak et al. (2009).
In brief, body posture was calculated as the arctangent of the low pass filtered (0.5 Hz) raw signal from the sagittal axis of the trunk sensor (range ± 180°). Positive values of body posture signify forward bowing of the participants. Lateral swaying movements and vertical up-and down movements were estimated from the transversal and longitudinal axis of the trunk sensor, respectively, and calculated using customized scripts.
Both signals were bandpass-filtered within the frequency range of walking (i.e., 0.5–3 Hz), rectified and smoothed using a moving average function (width 1 s). Walking speed was calculated with commercial software (for details see Vitascore, Temec, NL).
Assessment of current mood. Self-report measures were assessed with Palm Tungsten T3 handhelds running the freely available Experience Sampling Program software (ESP, version 4.0). Handhelds were configured to request the participants' input regularly in 1-h intervals by releasing a series of single beeps (i.e., time-based assessment, see Shiffman et al. 2008). Upon participants' response, a number of questions assessing the participants' mood and physical activity were released. Questions and assessment of positive and negative affect was assessed based on the procedure described earlier (Mata et al. 2011). In brief, participants were asked to indicate on a 4-point scale from “not at all” to “completely” how anxious, sad, disgusted, angry, guilty, ashamed, happy, frustrated, excited, alert and active they were at that moment.1
Beck Depression Inventory (BDI). We used the BDI (German version by Hautzinger et al. 2006) to assess depressive symptoms by self-report. The BDI is a widely used 21-item measure covering affective, cognitive, motivational, behavioral, and biological symptoms of depression with good psychometric properties (Beck et al. 1988).
To increase reliability, data were assessed on two consecutive days, preferably a weekday and a weekend day (see Buchowski et al. 2004). Distribution of weekdays and weekend days did not differ between patients and non-depressed controls, p = 0.318. On the day prior to the first day of data assessment, participants were introduced to set up the two accelerometers and to the use of the Palm handhelds, and were asked to fill in the BDI questionnaire. Participants were informed that the aim of the study was to assess the association between physiological parameters and mood. They were instructed to wear the accelerometers during the entire day until going to sleep and to put on the accelerometer again in the morning of the next day as early as possible. They were then told that they should respond to every prompt of the Palm handheld, whenever possible, but that they should refrain from responding while driving a car. Participants were informed that the Palm would prompt them every hour from 8:00 am to 10:00 pm on both days (Total response to prompts did not differ between non-depressed controls, 89,4% and MDD patients, 87,9%, p = 0.666). At the end of the second day, participants returned the equipment and were thanked for participating.
Data Reduction and Statistical Analysis
First, those segments were extracted during which the participant was walking (as identified by movement pattern identification algorithm). Data, where the participants were running were excluded from analyses (Bussmann et al. 2001). Figure 1 exemplifies the chronology of movement detection within the individual data sets of each participant. These segments were subsequently used for further analyses. To evaluate group differences between patient and control participants, we computed the mean of each gait parameter per participant during these walking phases over the entire two days of data assessment (i.e., mean overall movement intensity, body posture, lateral swaying movements, vertical up-and-down movements and walking speed). Simple t-tests were performed to evaluate differences between patients and non-depressed controls. In addition, mean positive and negative affect was calculated by averaging all responses over the entire 2 days of data assessment. Simple t-tests were calculated to assess differences between patients and controls in mean overall positive and negative affect.
To evaluate the predictive value of the four gait patterns on self-reported mood, again those segments were used during which the participant was walking. Because affect self-report was requested every 60 min, and five minutes time had to be allotted to the respective subsequent self-reporting, data were analyzed for the remaining 55 min of each 60-min interval. For each of the identified walking segments within the 55 min prior to a respective self-report, data of each of the four gait parameters (i.e., 'body posture', 'vertical movement', 'lateral movement', and 'walking speed') were averaged. Additionally, overall movement intensity was calculated by aggregating overall movement intensity data over the entire 55 min of each interval.
The study duration of two days provided a number of self-reports and related gait parameters for each participant. This hierarchical dataset with repeated measures was modeled using multilevel regression, with self-reported positive and negative affect as the two dependent variables, the four gait parameters as the fixed effects, and participant as the random effect. The intercepts as well as the slopes of the fixed effects were allowed to vary freely, and therefore 'participant intercept' and 'participant slope' were entered as the random effects into the models.
We computed a total of 10 multilevel models considering the four gait parameters (body posture, vertical up-and-down movements, lateral swaying movements, and walking speed) predicting both affect variables (positive and negative affect). Gait parameters as observed during the 55 min preceding the respective self-report assessments were used as predictors. Within each model, affect-ratings from the preceding self-report assessment (i.e., positive or negative affect one hour ago) were entered as a covariate. We additionally added group membership (i.e., patients vs. controls) and the interaction of group membership with the respective gait parameter to test whether the gait parameters differentially predicted mood within depressed patients or non-depressed controls. Finally, overall movement intensity was used as a covariate in each model to test whether the predictive power of gait parameters would exceed the mere overall motor activity level. Taken together, this approach enables, under control of participants’ overall movement intensity, if a change in mood from one hour to another is significantly predicted by a previous change in the four gait parameters. Since previous research has shown that MDD patients and non-depressed controls most likely differ in overall positive and negative affect (e.g., Mata et al. 2011), adding group membership also enabled us to control for these overall differences in mood in our multilevel analyses. Thus, the models in our study aimed at analyzing the association between measures of mood (i.e., positive and negative affect) and gait patterns as recorded over time.
Finally, a set of control analyses was performed to account for the possibility of a bi-directional association of affect and gait. To this end, we computed a total of 10 multilevel models considering positive and negative affect predicting the four gait parameters (body posture, vertical up-and-down movements, lateral swaying movements, and walking speed) and overall movement intensity as observed during the 55 min after the respective self-report assessments. Within each model, affect ratings of the preceding self-report assessment (i.e., positive or negative affect one hour ago) were entered as a covariate. Again, we included group membership (i.e., MDD patient vs. controls) and its interactions with the respective predictors (in these models, negative and positive affect) as factors to our analyses. All statistical analyses were performed using JMP Pro 14 statistical software (SAS Institute Inc.).