Geolocation as a Digital Phenotyping Measure of Negative Symptoms and Functional Outcome

Abstract

Objective

Negative symptoms and functional outcome have traditionally been assessed using clinical rating scales, which rely on retrospective self-reports and have several inherent limitations that impact validity. These issues may be addressed with more objective digital phenotyping measures. In the current study, we evaluated the psychometric properties of a novel “passive” digital phenotyping method: geolocation.

Method

Participants included outpatients with schizophrenia or schizoaffective disorder (SZ: n = 44), outpatients with bipolar disorder (BD: n =19), and demographically matched healthy controls (CN: n = 42) who completed 6 days of “active” digital phenotyping assessments (eg, surveys) while geolocation was recorded.

Results

Results indicated that SZ patients show less activity than CN and BD, particularly, in their travel from home. Geolocation variables demonstrated convergent validity by small to medium correlations with negative symptoms and functional outcome measured via clinical rating scales, as well as active digital phenotyping behavioral indices of avolition, asociality, and anhedonia. Discriminant validity was supported by low correlations with positive symptoms, depression, and anxiety. Reliability was supported by good internal consistency and moderate stability across days.

Conclusions

These findings provide preliminary support for the reliability and validity of geolocation as an objective measure of negative symptoms and functional outcome. Geolocation offers enhanced precision and the ability to take a “big data” approach that facilitates sophisticated computational models. Near-continuous recordings and large numbers of samples may make geolocation a novel outcome measure for clinical trials due to enhanced power to detect treatment effects.

Introduction

Rapid developments in mobile technology now make it possible to measure negative symptoms and functional outcome in the context of everyday life in ways that account for limitations inherent to clinical rating scales.¹ Clinical rating scales are designed to assess symptoms using retrospective interviews covering epochs ranging from 1 week to 1 month; however, state-level fluctuations are seen even in individuals with the most severe and persistent negative symptoms.^1,2 This suggests that state-level fluctuations (ie, waxing and waning within and across days, locations, or activity) are important to assess and provide information about temporal dynamics that clinical rating scales are not sensitive to.²

The broad umbrella term for measurement in the context of everyday life is “Digital Phenotyping,” ^3,4 which involves using mobile devices to trigger the collection of data in real life. Data collection can be either active or passive. Active data collection entails collecting self-report or task behavior (eg, cognitive tests or a video recording) from participants in situ. As the name implies, active data collection only occurs when the participant actively chooses to engage and complete the survey/task. Passive data collection is unobtrusive and typically done through the internal sensors of a device (eg, smartphone or band) or by examining phone usage data (eg, call logs, text logs, and screen time).³ Several forms of passive data (eg, accelerometry, speech segregated from background noise, and ambulatory psychophysiology) have been evaluated in psychiatric disorders.^5–8 However, they have yet to be extensively validated in psychotic and mood disorder populations, where they may be valuable objective measures of negative symptoms and functional outcome.

One passively collected variable that holds promise for examining negative symptoms and functional outcome is geolocation. This method involves collecting GPS coordinates at predetermined intervals or every time the participant moves a certain radius in space. From this data, a number of variables can be calculated that provide information about social and motivated behavior, such as entropy, location variance, total amount of change in distance from time t to t + 1, total distance traveled, percentage of time spent at home, and amount of deviation from home location. Three published studies with small sample sizes have demonstrated the utility of geolocation for use with individuals diagnosed with mood disorders, demonstrating mean group differences from healthy controls (CN).^8–10 Ben-Zeev et al have used geolocation in their Cross-Check platform, reporting that basic geolocation variables are useful for prediction and relapse monitoring.^11–14 To our knowledge, the validity of geolocation as a measure of negative symptoms and functioning has been examined in only 1 published study.¹⁵ Results provided preliminary support for the validity of geolocation as indicated by: (1) high convergence between geolocation measures and ecological momentary assessment (EMA) reports of location; (2) differences between schizophrenia or schizoaffective disorder (SZ) and CN; (3) associations between clinically rated negative symptoms and geolocation; and (4) minimal associations with positive symptoms, depression, and neurocognition.

Although preliminary results are promising,¹⁵ important gaps in knowledge must be addressed before geolocation can be implemented as an outcome measure in clinical trials. For example, many sophisticated geolocation indices have been developed but not yet evaluated in SZ or mood disorders. Psychometric properties have also not been comprehensively evaluated. For example, reliability has not been established, convergent validity in relation to the most updated conceptualization of negative symptoms as 5 domains has not been achieved, and convergent and discriminant validity has not been examined in relation to temporally proximal EMA measures.^16–22 Comparisons between SZ and a clinical control group have not been undertaken. However, this will be necessary to control for effects common to serious mental illness (eg, lower education, unemployment, and social isolation) that confound interpretations of validity.

In the current study, we conducted a comprehensive psychometric evaluation of geolocation to address these gaps in the literature. Our sample consisted of outpatients with SZ, a clinical comparison sample of outpatients with bipolar disorder (BD) who are also known to display negative symptoms and poor functional outcome,^21,23 and demographically matched CN. To assess convergent and discriminant validity, passive GPS data were collected concurrently with active self-report data regarding subjective and behavioral components of negative symptoms (avolition, anhedonia, and asociality), as well as standard interview-based measures of symptoms and functional outcome. The following hypotheses were evaluated: (1) compared to CN, SZ and BD would have reduced behavioral output on geolocation measures of: entropy (ie, less variability in time spent in different locations), location variance, time transitioning between stationary locations, time at home, and distance traveled from home; (2) geolocation measures would demonstrate convergent validity with concurrently collected active digital phenotyping self-reports of negative symptoms and clinical interview-based measures of negative symptoms and community-based functional outcome; (3) discriminant validity would be supported by low correlations with measures of positive symptoms, disorganization, depression, and anxiety; and (4) reliability would be supported by good internal consistency and stability of scores across days.

Methods

Participants

Data were collected from 3 participant groups: (1) 51 individuals with Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) diagnoses of SZ; (2) 20 individuals with DSM-5 diagnoses of BD; and (3) 55 psychologically and neurologically CN. Of this total sample, 10 SZ, 1 BD, and 9 CN did not have valid geolocation data. This resulted in a final sample of: SZ, n = 44; BD, n = 19; and CN, n = 46. The final groups did not significantly differ on age, ethnicity, sex, or parental education. As anticipated, SZ had lower personal education than CN (see table 1).

Table 1.

Group demographic characteristics and digital phenotyping adherence and accuracy levels

Variable	CN (n = 46)	SZ (n = 41)	BD (n = 19)	Test statistic	P
Age; M (SD)	38.33 (10.53)	39.59 (12.64)	37.32 (12.82)	F = 0.27	.766
Male; n (%)	16 (34.8)	13 (31.7)	3 (15.8)	χ 2 = 2.37	.305
Personal education; M (SD)	15.37 (3.03)	13.2 (2.04)	14.68 (2.36)	F = 7.91	<.001
Parental education: M (SD)	13.72 (2.84)	13.78 (3.11)	13.84 (2.01)	F = 0.01	.987
Race; n (%)				χ 2 = 16.36	.090
African American	15 (32.6)	13 (31.7)	3 (15.8)
Asian American	3 (6.5)	0	0
Biracial	2 (4.3)	3 (7.3)	0
Caucasian	19 (41.3)	23 (56.1)	16 (84.2)
Hispanic/Latino	6 (13)	2 (4.9)	0
Other	1 (2.2)	0	0
Survey adherence; M (SD)	65.62% (26.26)	55.88% (27.98)	54.71% (27.17)	F = 1.81	.168
Survey adherence with paired geolocation; M (SD)	59.46% (28.29)	50.21% (27.62)	53.29% (27.84)	F = 1.21	.303
Home categorization accuracy percentage; M (SD)	89.90% (12.58)	90.22% (11.70)	89.53% (12.36)	F = 0.02	.979
GPS accuracy in meters; M (SD)	7.69 (3.99)	7.03 (4.27)	7.34 (5.30)	F = 0.25	.778

Note. Adherence is the percentage of momentary surveys (out of 8 per day).

SZ, schizophrenia or schizoaffective disorder; BD, outpatients with bipolar disorder; CN, demographically matched healthy controls.

Individuals with psychiatric diagnoses (SZ and BD) were recruited from local community outpatient mental health centers and advertisements. SZ and BD diagnoses were determined via the Structured Clinical Interview for DSM-5 (SCID) ^24. CN participants were recruited from the local community using posted flyers and electronic advertisements. CN had no current major psychiatric diagnoses (eg, former axis I) and no current cluster A personality disorder diagnoses as determined by the SCID-5 and Structured Clinical Interview for DSM-5 Personality Disorders (SCID-5-PD).²⁵ Participants in both groups were free of lifetime neurological disorder and did not meet criteria for a substance use disorder within the last 6 months. All participants received monetary compensation for their participation—$20 per hour of laboratory assessment, $1 per survey completed, and $80 when returning the phone—and provided written informed consent for a protocol approved by the University of Georgia Institutional Review Board.

Procedures

Study procedures were completed in 3 phases that occurred over 1 week.

Phase I: Initial Laboratory Visit

On the initial visit, participants provided informed consent, completed EMA training procedures, and the following diagnostic/symptom interviews: SCID²⁴ (SZ, BD, and CN) and SCID-5-PD²⁵ (CN only), Brief Negative Symptom Scale (BNSS)²⁶ (SZ and BD), Positive and Negative Syndrome Scale (PANSS)²⁷ (SZ and BD), and Level of Functioning Scale²⁸ (SZ and BD). EMA training consisted of instructions on how to use the phone, how to initiate and complete surveys, and basic troubleshooting.

Participants were provided with a Blu Vivo 5R smartphone and written instructions for key questions and encouraged to contact the researchers if there were any technical problems. All participants received a follow-up call on the first day of phase 2 to ensure that surveys were delivering properly. Smartphones were provided to ensure the consistency of phone sensors and to allow participation regardless of whether participants owned a smartphone.

Phase II: Digital Phenotyping

Surveys were preprogrammed and delivered using the mEMA application from Ilumivu (https://ilumivu.com/). Participants were asked to complete 8 quasi-random surveys scheduled within 90-min epochs between 9 am and 9 pm. Each momentary survey was available for a 25-min window; surveys became available 10 min before their scheduled delivery and were available for 15 min after delivery. Participants were notified to complete surveys by a tone and vibration of the phone at the scheduled time and again 5 min and, then, 10 min later if the survey was not completed. Any momentary surveys not completed within 15 min of the scheduled time became unavailable.

Surveys were designed to capture different information based on the context in which it was to be completed, including current location, activity, and social interaction. Emulating procedures from the BNSS, EMA negative symptom items were constructed that had both subjective and behavioral components that were designed to measure avolition, asociality, and anhedonia (see figure 1). Confirmatory factor analysis (CFA) results supported these score calculations (see supplementary material).

Fig. 1.

Derivation of Ecological Momentary Assessment negative symptom items. Squares on the right represent individual items where pills and circles represent construct composites. All behavior and internal experience composites are scored such that higher values indicate greater negative symptom severity. Other social context responses: Coworkers/Classmates, Doctor/Therapist, Strangers, and No one/Alone. Other activity context responses: Eating/Drinking, Internet/Computer use, Resting, TV/Music, Smoking, Bathing/Hygiene, Pacing restlessly, Nothing, and Coworkers.

Phone sensors were programmed to collect geolocation every 10 min or when participants moved more than 10 m. This combination of interval- and movement-based sampling did not require us to perform data estimation procedures for missing data, which was necessary in some previous studies that used periodic sampling without movement-triggered sampling.²⁹ Data was stored as GPS coordinates and change in meters from the previous sample. Each completed survey instance was also coded with GPS coordinates at the time of survey completion. All data were encrypted and stored using unique identification codes on the Ilumivu servers, separate from identifying information, until downloaded by the research team. Once downloaded, all data with GPS coordinates were password protected and kept on an approved, secure, university server.

Phase III: Final Laboratory Visit

After the 6-day period ended, participants completed a final laboratory visit where they returned their study phone, received payment, completed a poststudy debriefing interview (eg, tolerability of procedures and paranoia over phone monitoring).

Geolocation Data Filtering

As numerous variables affect geolocation coordinate accuracy (eg, number of available satellites, weather, and interference), every instance of passively collected geolocation was paired with an accuracy rating of the sample. This rating is best interpreted as a range in meters of possible coordinates from which a single coordinate pair is selected. As inaccurate values would introduce error to results, only samples with an accuracy range of less than 35 m (~115 ft) were retained (95% of all samples). To further filter erroneous coordinate samples, any values with a change in coordinates >200 km/h (~125 mph) were excluded (similar to procedures used by Palmius et al; as some participants indicated abnormal travel [eg, travel for a conference], samples indicating a distance from home greater than 400 000 m [~250 miles] had variables related to the distance from home deleted [3% of all samples]. Other variables, including distance changed between samples and number of stationary locations, were retained).⁹

Geolocation Variable Creation

Our approach to selecting geolocation metrics of interest was to be broadly inclusive of variables used in prior geolocation studies to determine which might be most relevant for measuring negative symptoms and functional outcome (for text descriptions of variables, see table 2; see figure 2 for visual description of key variables). Variable calculations were based on prior studies^8,9,29

Table 2.

Geolocation variable definitions and formulas

Variable	Abbreviated names	Definition
Home time	Home	0 or 1 per sample, 1 if sample within 200 m of home. Home location determined as mean coordinates when home location was endorsed in surveys. See supplementary materials for analyses supporting 200-m threshold.
Distance change	Δd MPD Range	Distance traveled in meters from previous sample calculated by Haversine formula where Δlat and Δlong are latitude and longitude at time t (a) minus latitude and longitude at time t + 1 (b), respectively, and R = 6 371 000. Total meters traveled per day and range of distance per day were also calculated.29
Distance from home	Δmh maxmh	Meters from home for each sample calculated by Haversine formula, where Δlat and Δlong are latitude and longitude for the sample (a) minus latitude and longitude of home coordinates (b), respectively, and R = 6 371 000. Maximum distance from home was also calculated per day.29
Stationary location clusters8,9	nc	To determine a stationary location for sample at time t, total change in distance within t ± 10 min was calculated. If the total distance change was less than 500 m, a sample was identified as stationary. The 20-min window is based on sampling frequency in the event of no detected movement (every 10 min) and the 500-m area is based on previous research8,9 and has been found to reflect meaningful locations.45 Stationary locations were identified through k means clustering. Seven clustering solutions were identified for each participant, see supplementary materials for descriptions. Agreement between the metrics was satisfactory (Cronbach’s α = .71). The median of cluster solutions was taken as the final number of clusters.
Location variance8,9	lv	Variance within a specified timeframe: $lv=\log (({\sigma }_{\mathrm{l}\mathrm{a}\mathrm{t}}^2+1)+({\sigma }_{\mathrm{l}\mathrm{o}\mathrm{n}\mathrm{g}}^2+1))$
Entropy and normalized entropy8,9	ent nent	Entropy reflects equity of time spent in different locations; higher values indicate a more uniform distribution of time. Four entropy related variables were calculated for each participant by first calculating ent and nent for each day where p is the percentage of samples taken from a given cluster i on day d and N is the total number of clusters: $\mathrm{e}\mathrm{n}{\mathrm{t}}_d=-\underset{i=1}{\sum ^n}{p}_{di}\log (p_{di})$ $\mathrm{n}\mathrm{e}\mathrm{n}{\mathrm{t}}_d=\frac{\mathrm{e}\mathrm{n}{\mathrm{t}}_d}{\log (N_d)}$ From each of these, mean (ent-m and nent-m) and standard deviation (ent-sd and nent-sd) were taken for each individual.
Transition time8,9	tt	Percentage of samples taken in transit (ie, those samples that did not meet criteria to be labeled stationary).
Flights29	f.dur f.dist f.num	Flights—discrete trips—were identified as consecutive samples in transit. At least 2 samples within 10 min of each other were required to identify a flight. Flight variables (flight duration, flight distance, and number of flights) were calculated per day, then averaged over the 6-day study period to reduce the influence of outlier flights.

Note. Haversine formula⁴⁶:

$${\displaystyle \begin{array}{l}{\displaystyle a=\sin {\left(\frac{\mathrm{\Delta }\mathrm{l}\mathrm{a}\mathrm{t}}{2}\right)}^2+\cos \left(\mathrm{l}\mathrm{a}{\mathrm{t}}_a\right)\times \cos \left(\mathrm{l}\mathrm{a}{\mathrm{t}}_b\right)\times \sin {\left(\frac{\mathrm{\Delta }\mathrm{l}\mathrm{o}\mathrm{n}\mathrm{g}}{2}\right)}^2}\end{array}}$$

$${\displaystyle \begin{array}{l}{\displaystyle c=2\times {\sin }^{-1}\sqrt{a}}\end{array}}$$

$\mathrm{\Delta }d=c\times R$,

where ∆lat = lat_a− lat_b, ∆long = long_a− long_b, and R = 6 371 000.⁴⁷

Suffixes -m, -sd, and -ac represent mean, standard deviation, and autocorrelation, respectively.

Fig. 2.

Visual demonstration of geolocation variables. Pilot data from a staff member in the Strauss lab is presented to illustrate the nature of the method and score calculations. Each dot represents a single sample of GPS coordinates, each circle represents an identified cluster, and each diamond represents the reported home location of that individual. Each sample provides values for distance changed since previous sample, distance from home, and if that sample was taken at home. Clusters allow for aggregated data to examine entropy, location variance, transition time, etc.

Data Analysis

All variables fall into one of 3 levels of analysis: (1) passively observed; (2) active, or EMA, samples; and (3) summary level. Passive-level data are collected continuously without active effort by the participant. Active-level data consists of the signal-contingent EMA self-report survey data and aggregated passively collected data from within 30 min of each survey sample. Summary-level data consists of means and SDs of passively and actively recorded variables with a single value for each individual for the entire sample period. Aggregated data in the active and summary levels noted with suffixes -m and -sd, are the mean and SD of the aggregated variable, respectively. Geolocation variables not already representing variability (eg, time at home, meters from home, and number of clusters) were used to generate autocorrelation scores (suffix -ac) to represent stability across time. Composites were made separately at the summary level by standardizing and adding negative symptom items within each domain.

Individuals with any geolocation data were used for analyses and cases with missing variables for a given analysis were removed on a pairwise basis. Group differences were analyzed using ANOVAs at the summary level and linear mixed models (LMMs) at the active and passive levels. Geolocation variables demonstrating extreme skew (|skew| >2) were transformed prior to LMMs. Geolocation variables may follow nonlinear relationships with symptoms.¹⁵ As such, relationships between variables were tested for the best-fitting regression (linear, quadratic, or logarithmic). Active-level regressions were multilevel. All regressions were conducted with untransformed data. See table 2 for variables used and their definitions and figure 1 for a visual explanation of variables. Analyses were conducted with and without personal education and employment status as covariates. As these do not change the results, they are reported only in the supplementary material.

Data transformation, cleaning, and analyses were conducted using R³⁰ and the following packages: cluster,³¹ effsize,³² ggmap,³³ ggpubr,³⁴ gmodels,³⁵ Hmisc,³⁶ lubridate,³⁷ nbclust,³⁸ lme4,³⁹ lmerTest,⁴⁰ emmeans,⁴¹ pROC,⁴² psych,⁴³ and tidyverse.⁴⁴

Results

Criterion Validity

Criterion validity was assessed by examining differences in geolocation variables based on group, day of the week, and context. While many summary-level group results were nonsignificant (see supplementary material for results where P >.1), distance from home, number of clusters autocorrelation, flight duration, and transition time variables showed significant group effects (see table 3). Follow-up pairwise least significant difference contrasts on these variables showed, generally, more activity in the BD and CN groups relative to SZ.

Table 3.

Significant group effects

					F
Variable	Level	CN	SZ	BD	Group	Weekday	Group × Weekday	Contrasts
entsd	Summary	0.25 (0.13)	0.19 (0.12)	0.18 (0.09)	2.39a	—	—	CN > SZ, BD
f.durm	Summary	14.47 (18.48)	9.46 (14.09)	20.78 (21.97)	3.89*	—	—	BD, CN > SZ
f.dursd	Summary	20.18 (30.15)	14.14 (29.21)	32.43 (37.24)	3.79*	—	—	BD, CN > SZ
Home	Summary	0.50 (0.25)	0.62 (0.28)	0.50 (0.25)	2.40	—	—	CN > SZ
Δmhm	Summary	11 522.80 (15 318.33)	5871.52 (11 547.89)	8860.72 (9785.15)	4.28*	—	—	CN, BD > SZ
Δmhsd	Summary	14 491.69 (18 693.52)	7522.04 (11 536.43)	14 062.68 (18 343.72)	4.04*	—	—	CN, BD > SZ
Δmhac	Summary	0.94 (0.15)	0.88 (0.20)	0.96 (0.06)	2.58	—	—	BD, CN > SZ
Rangem	Summary	27 618.13 (66 152.01)	12 115.04 (12 481.15)	22 864.51 (28 444.11)	2.58	—	—	CN > SZ
Rangesd	Summary	33 753.70 (100 687.09)	14 627.78 (18 127.02)	33 594.50 (64 645.06)	2.37a	—	—	CN, BD > SZ
ncac	Summary	−0.23 (0.41)	0.07 (0.41)	−0.24 (0.37)	5.34**	—	—	CN, BD > SZ
ttm	Summary	0.21 (0.13)	0.17 (0.14)	0.28 (0.15)	4.15*	—	—	BD > SZ, CN
ttsd	Summary	0.19 (0.09)	0.17 (0.11)	0.25 (0.11)	4.29*	—	—	BD > SZ, CN
Home	Active	0.49 (0.49)	0.71 (0.44)	0.57 (0.48)	2.48a	4.71***	1.22	SZ > CN
tt	Active	0.14 (0.33)	0.10 (0.29)	0.17 (0.37)	4.02*	3.85***	0.89	BD, CN > SZ
Δd	Passive	71.90 (2301.81)	54.53 (985.66)	73.31 (459.67)	5.08**	2.03	0.39	BD > SZ, CN
tt	Passive	0.26 (0.44)	0.25 (0.43)	0.38 (0.49)	4.74*	3.03**	1.09	BD > SZ, CN

Note. Descriptives provided are for raw data, test statistics are for transformed data (as needed, see Data Analysis). Abbreviations are explained in the first footnote to table 1. Also, please refer to table 2 for the geolocation variable abbreviations.

^aThe addition of covariates into the analysis made this result nonsignificant.

*P < .05, **P < .01, ***P < .001.

Similar results were observed at the active and passive levels (table 4). Active- and passive-level LMMs also indicated the effects of day of the week such that weekends show less activity and more time at home than the week (see supplementary material).

Table 4.

Summary-level regressions

Convergent
Clinical variable	GPS variable	FitSZ	β  SZ	F  SZ	FitBD	β  BD	F  BD
BNSS anhedonia	MPDsd	quad	−0.30 + 0.76	0.31	quad	−0.52 + −2.18	3.28a
BNSS anhedonia	Rangem	quad	−2.17 + 0.84	2.89	quad	−0.90 + 0.12	0.38
BNSS anhedonia	Maxmhm	quad	−1.16 + 2.00	2.86	quad	−0.86 + −0.75	0.60
BNSS anhedonia	f.numac	quad	−3.19 + 1.38	9.63***	quad	0.23 + 0.59	0.17
BNSS asociality	Δdm	quad	−0.32 + −0.25	0.08	quad	−2.26 + 0.55	3.44a
BNSS asociality	Home	log	0.69	2.85a	quad	0.90 + 1.56	1.75
BNSS asociality	Δmhac	quad	−0.22 + −1.46	1.09	quad	−1.76 + −1.65	3.86* ,a
BNSS asociality	Rangeac	quad	−0.02 + −0.40	0.07	quad	−1.59 + 1.59	3.05a
BNSS asociality	nentm	quad	−1.24 + 1.30	1.66	quad	0.17 + 2.29	3.12a
BNSS asociality	nentsd	quad	−0.92 + −1.02	0.94	quad	−2.23 + 1.26	4.47* ,a
BNSS avolition	Δmhac	quad	1.59 + −1.00	1.83	quad	−0.06 + −2.41	3.79* ,a
BNSS avolition	Rangeac	quad	0.56 + −1.21	0.88	log	−1.00	3.41a
BNSS avolition	Maxmhsd	quad	−0.57 + −0.12	0.16	quad	−1.84 + 1.40	3.42a
BNSS avolition	Maxmhac	quad	0.07 + −1.49	1.12	quad	−0.67 + −2.19	3.51a
BNSS avolition	f.distm	quad	−0.18 + 0.37	0.08	quad	−2.07 + −1.16	3.63a
BNSS avolition	f.numac	quad	−1.91 + 1.47	3.32	quad	0.10 + 0.45	0.09
BNSS avolition	lvsd	quad	−0.34 + 0.89	0.43	quad	−0.51 + 2.11	2.79a
BNSS avolition	ttsd	quad	1.48 + −1.94	3.29*	quad	−0.24 + −0.31	0.07
LOF social	Δmhac	log	0.94	3.11a	quad	0.02 + −0.04	0
LOF social	f.numac	quad	0.77 + −0.47	0.38	quad	1.38 + −2.25	4.55* ,a
LOF social	nentm	quad	0.40 + 0.29	0.08	quad	−0.65 + −2.95	8.55**
LOF social	nentsd	quad	−0.61 + 0.27	0.15	quad	2.63 + −1.12	6.74**
LOF social	Ent	quad	−0.55 + 1.70	1.57	log	1.10	3.31a
LOF social	Nent	quad	−0.72 + 1.79	1.91	quad	2.40 + −2.45	11.72**
LOF social	ttsd	quad	−0.36 + 1.03	0.56	quad	2.27 + 0.64	3.59a
LOF work	Δdm	quad	0.49 + −0.13	0.12	quad	2.17 + −1.19	3.50
LOF work	MPDm	quad	0.17 + −0.60	0.18	quad	−5.57 + −6.39	7.56**
LOF work	f.durm	quad	4.55 + 5.22	1.94	quad	2.11 + −1.05	2.99
LOF work	f.dursd	quad	29.90 + 12.93	1.68	quad	2.16 + −1.00	3.14
LOF work	Nc	quad	−1.23 + −0.42	0.81	quad	−1.30 + −2.86	4.88*
LOF work	ttsd	quad	−2.87 + 1.10	5.89**	quad	2.25 + −0.14	3.08
LOF work	ttm	quad	−1.74 + 1.80	3.47*	quad	2.41 + −1.10	4.75*
EMA anhedonia	Homeac	quad	0.71 + −2.44	3.76*	quad	−0.25 + −0.24	0.05
EMA anhedonia	f.numsd	quad	−2.39 + 0.96	3.77*	quad	1.26 + −1.25	1.70
EMA anhedonia	f.numac	quad	−1.09 + 1.84	2.54a	log	−0.52	0.93
EMA anhedonia	f.durac	quad	−1.44 + 1.75	2.91a	log	−0.92	2.55
EMA anhedonia	Nc	log	−0.64	3.67a	quad	0.47 + 0.83	0.43
EMA anhedonia	ncac	quad	−1.49 + −1.58	2.63a	quad	−2.57 + −0.84	6.17*
EMA anhedonia	nentm	quad	−0.10 + −0.31	0.05	quad	0.35 + 3.05	8.75**
EMA anhedonia	nentsd	quad	0.35 + −0.16	0.07	quad	−2.11 + 2.17	8.25**
EMA asociality	MPDm	quad	−0.81 + 2.07	2.58	quad	0.71 + −1.35	1.19
EMA asociality	Δmhac	quad	−1.90 + 1.80	3.96*	quad	−1.16 + 0.50	0.78
EMA asociality	Rangeac	quad	−1.60 + 1.15	2.01	quad	−1.81 + 1.55	3.77* ,a
EMA asociality	Homeac	quad	0.50 + −2.35	3.25	quad	−0.92 + 0.87	0.79
EMA asociality	f.numsd	quad	−1.98 + 1.36	3.15	quad	0.97 + −0.34	0.50
EMA asociality	ncac	quad	−0.89 + −1.28	1.24	quad	−2.43 + −1.38	7.07**
EMA asociality	nentm	quad	−0.22 + −0.61	0.20	quad	−0.67 + 3.35	14.71***
EMA asociality	nentsd	quad	0.50 + 0.30	0.16	quad	−3.09 + 0.51	9.60**
EMA asociality	Nent	quad	0.32 + −0.37	0.11	quad	−1.73 + 1.68	3.82*
EMA avolition	Home	quad	1.96 + 1.54	3.51*	log	0.67	1.26
EMA avolition	MPDm	quad	−1.12 + 2.12	3.15a	quad	−0.49 + −1.79	1.90
EMA avolition	MPDsd	quad	−0.78 + 2.15	2.96a	quad	−0.85 + −0.77	0.63
EMA avolition	Rangem	quad	−1.08 + 2.24	3.50*	quad	0.05 + 0.78	0.28
EMA avolition	MPDac	quad	0.03 + 0.90	0.39	quad	−2.25 + 0.37	3.36a
EMA avolition	Homeac	quad	0.47 + −2.27	3.01a	log	−0.58	0.86
EMA avolition	f.numm	quad	−2.19 + 1.24	3.50*	quad	−0.11 + −1.01	0.49
EMA avolition	f.numac	quad	−0.74 + 0.82	0.59	log	−1.00	4.18
EMA avolition	f.distac	quad	0.38 + 0.46	0.17	log	−1.27	6.66*
EMA avolition	f.durac	quad	−0.66 + 0.87	0.58	log	−1.47	8.81**
EMA avolition	nc	log	−0.65	3.85a	quad	−0.07 + 1.34	0.89
EMA avolition	ncac	quad	−1.81 + −0.83	2.15	quad	−2.49 + −1.26	7.02**
EMA avolition	nentm	quad	−1.04 + −0.06	0.53	quad	−1.43 + 2.78	9.51**
Discriminant
Clinical variable	GPS variable	FitSZ	β  SZ	F  SZ	FitBD	β  BD	F  BD
PANSS, depression/anxiety	f.durac	quad	−0.40 + 2.82	5.34* ,a	quad	0.51 + 1.58	1.20
PANSS, depression/anxiety	entsd	log	0.75	3.08a	quad	0.64 + 0.55	0.24
PANSS, depression/anxiety	nentsd	quad	0.05 + −2.53	3.17a	quad	−0.27 + −0.23	0.05
PANSS, excitement/hostility	Δdsd	quad	−1.25 + −0.58	0.94	quad	−4.95 + −5.75	3.59a
PANSS, excitement/hostility	Rangem	quad	0.19 + −2.24	2.62a	quad	−0.25 + −0.53	0.13
PANSS, excitement/hostility	MPDac	quad	−2.18 + 1.05	3.42* ,a	log	0.48	0.51
PANSS, excitement/hostility	Rangeac	quad	−2.04 + 1.00	2.81	quad	1.50 + 1.55	2.01
PANSS, excitement/hostility	Maxmhsd	quad	−0.99 + −2.23	2.71a	quad	0.85 + 0.55	0.46
PANSS, excitement/hostility	Maxmhac	quad	−2.04 + 0.88	2.67a	quad	1.87 + 1.53	1.89
PANSS, excitement/hostility	f.numm	quad	2.92 + 1.16	6.23**	quad	0.27 + −1.95	1.26
PANSS, excitement/hostility	f.durm	quad	4.20 + −0.07	4.42* ,a	quad	1.97 + −1.02	2.17
PANSS, excitement/hostility	f.numsd	quad	2.75 + 2.26	8.94***	quad	1.00 + −1.65	1.10
PANSS, excitement/hostility	f.dursd	quad	47.60 + 15.92	2.51a	log	1.14	5.59* ,a
PANSS, excitement/hostility	ent	quad	−0.14 + 2.97	5.12*	quad	−0.05 + 1.43	0.93
PANSS, excitement/hostility	nent	quad	−0.69 + 3.36	8.27**	quad	0.02 + 0.28	0.03
PANSS, excitement/hostility	ttm	quad	2.52 + 0.72	3.93* ,a	quad	1.67 + −1.45	2.25
PANSS positive	MPDsd	quad	−1.25 + −0.27	0.79	quad	−1.70 + 2.60	4.85*
PANSS positive	Maxmhac	quad	−2.11 + 1.00	3.04a	quad	1.06 + 0.98	0.25
PANSS positive	f.numm	quad	2.00 + −1.50	3.46*	quad	0.02 + −1.65	0.77
PANSS positive	f.durm	quad	−1.32 + −4.22	3.97*	log	−0.60	1.05
PANSS positive	f.dursd	quad	32.06 + 9.76	2.77a	quad	−1.01 + 0.91	0.59
PANSS positive	entm	quad	0.67 + 0.30	0.25	quad	2.84 + 2.34	7.85**
PANSS positive	ttm	quad	2.11 + −1.01	2.99a	quad	−0.50 + −1.47	0.95

Note. BNSS, Brief Negative Symptom Scale; EMA, ecological momentary assessment; LOF, Level of Functioning Scale; PANSS, Positive and Negative Syndrome Scale. Abbreviations are explained in the first footnote to table 1. Also, please refer to table 2 for the geolocation variable abbreviations.

^aThe addition of covariates into the analysis made this result nonsignificant.

*P < .05, **P < .01, ***P < .001.

We examined active-level reports of location (home, public, or work) and activity (commuting) to evaluate criterion validity of the geolocation measures. These results generally suggest more mobility during contexts where greater activity is required (see supplementary material).

Convergent Validity

Convergent constructs included negative symptoms and functional outcome. The majority of models fit best using a quadratic regression line. Generally, results indicated that greater activity was associated with lower negative symptoms and greater functioning (table 4).

Regressions at the active level indicated that linear and logarithmic models generally provide the best fit. In general, results showed that greater activity was associated with lower negative symptoms. Results were strongest with avolition measures (see supplementary material).

Discriminant Validity

PANSS excitement/hostility was associated with geolocation variables in the SZ group but the direction of relationships was inconsistent. PANSS positive symptoms were associated with variability in meters per day and mean entropy in the BD group but number of flights and mean flight duration in the SZ group (see table 4). Relationships between PANSS depression/anxiety and geolocation variables were nonsignificant when controlling for education and employment.

Active-level regressions show some relationships between anxiety and sadness measured via EMA and geolocation variables in the CN and BD groups (supplementary material). However, the only significant relationship observed in the SZ group was a positive logarithmic relationship between sadness and time at home. Delusions were not significantly associated with geolocation variables in any group. Additionally, the magnitude of the associations between EMA-measured negative symptoms and geolocation was higher than between geolocation with EMA delusions, sadness, and anxiety.

Incremental Validity

Incremental validity was explored using hierarchical regressions based on the best-fitting regression lines. Level 1 predicted a BNSS domain score (eg, anhedonia) from the active-level composite of that same domain score. Level 2 examined if the addition of a geolocation variable accounted for significant additional variance (based on the change in R²). By and large, geolocation variables did not account for additional variance above active-level composite scores, although there were a few exceptions where certain geolocation variables combined with EMA self-reports added incremental value for predicting avolition, asociality, and anhedonia (see supplementary material).

Temporal Stability

Intraclass correlation coefficients (ICC) were calculated within each day for each group to evaluate 2 sets of variables reflecting distance from home (meters from home and time spent at home) and variability in location (SD of meters from home, change in meters between samples, time spent in transit, and location variance). Resulting ICC values were high for home distance variables (ICC_CN = .98; ICC_BD = .98; ICC_SZ = .98) and moderate for location variability (ICC_CN = .82; ICC_BD = .87; ICC_SZ = .76). These ICC values suggest that home distance variables are more consistent than measures of variability.

The temporal stability of each measure was calculated by creating averages for each day and examining the ICC across days. Results indicate generally moderate to high temporal stability for most measures: change in meters (ICC_CN = .78; ICC_BD = .11; ICC_SZ = .55), distance from home (ICC_CN = .84; ICC_BD = .76; ICC_SZ = .87), time at home (ICC_CN = .81; ICC_BD = .85; ICC_SZ = .88), and time in transit (ICC_CN = .70; ICC_BD = .46; ICC_SZ = .57). ICC for change in meters is lower in the BD group, suggesting greater variability in transit time between days for that group compared to the others. Location variance generally showed low temporal stability (ICC_CN = 0.0; ICC_BD = .10; ICC_SZ = .53) but was more stable in the SZ group, who were more stationary.

Discussion

The current study evaluated the psychometric properties of geolocation as a passive digital phenotyping measure of negative symptoms and functional outcome in outpatients with SZ and BD. Our results provide a significant advance in knowledge in several ways.

First, this represents the first comprehensive psychometric validation of one of the most promising digital phenotyping methods for measuring negative symptoms and functioning. Results supported the psychometric properties of geolocation by indicating: (1) criterion validity with lower activity observed in SZ for several geolocation variables, (2) convergent validity with temporally proximal EMA measures of negative symptoms, (3) discriminant validity with EMA and clinically rated positive symptoms and mood/anxiety, and (4) good internal consistency and moderate temporal stability. Collectively, these findings provide proof-of-concept that geolocation is feasible for use as a measure of behavioral components of avolition and asociality.

Second, this is the first study to extensively examine a range of computationally sophisticated geolocation variables (38 in total) that hold promise for measuring negative symptoms and functional outcome. The lone prior study in this area by Depp et al¹⁵ examined 3 basic geolocation metrics. Our study identified a number of highly innovative and computationally intensive variables that hold promise as outcome measures of negative symptoms and functional outcome whose validity had not previously been explored in this population. We also extended prior work by demonstrating the relevance of these geolocation variables to specific negative symptom subdomains (ie, avolition and asociality), accounting for confounds inherent to serious mental illness via a clinical comparison group and demonstrating the relevance of collecting active (ie, EMA surveys) and passive (ie, geolocation) digital phenotyping variables concurrently to measure negative symptoms and functioning with higher resolution.

Third, these results are timely as digital phenotyping appears to be the next generation of symptom measurement, and there is considerable interest in incorporating these methods in clinical trials. To date, there has been limited guidance on how to achieve this goal. We recommend the following: (1) use a recording system that can collect both interval- and movement-based GPS data. The combination of both approaches allows for more computationally sophisticated calculation of variables and avoids issues related to imputing large amounts of data.²⁹ However, not all smartphones or operating systems have this functionality, and care should be taken in selecting devices and apps to record data. (2) Based on study designs, levels of geolocation variable resolution can be calculated at the level of survey instance, day, week, month, etc. Selecting the appropriate level of granularity for score calculations will be critical for observing treatment effects for drugs with more rapid or gradual effects. (3) The combination of active and passive methods is more advantageous than recording passive data alone. Our incremental validity analyses suggested that EMA surveys provide good coverage of negative symptom and functional outcome constructs and that the inclusion of geolocation with EMA surveys has added benefit. However, measuring geolocation devoid of context that is provided by EMA surveys may be problematic as it becomes less clear when the geolocation data is most relevant or valid. (4) Geolocation measures were more highly correlated with behavioral than inner-experience components of avolition, asociality, and anhedonia. Geolocation may, therefore, measure the objective rather than subjective component of negative symptoms. Correlations with clinical measures of social and work functioning suggest that geolocation may also hold relevance for assessing functional outcome. However, fuzzy conceptual distinctions between functional outcome and negative symptoms⁴⁸ make it difficult to determine which clinical construct geolocation is most relevant for. Clinical trials may, therefore, benefit from using geolocation as a measure of the objective components of negative symptoms and functioning but might want to supplement data collection with EMA surveys to capture the subjective component. (5) Data processing requirements are intensive. Pharmaceutical companies will benefit from partnering with academic partners with backgrounds in data science who have expertise in working with this data.

Certain limitations should be considered when interpreting these findings. First, although geolocation provides an impressive number of samples and most analyses were highly powered, the total sample size for each group was modest and replication in larger samples is warranted. Second, participants used study phones provided by the researchers. It is unclear whether results might change with the use of personal phones. Third, participants were paid for completing surveys. It is unclear how results might change when payment is not offered, which might be necessary for study designs using much longer data collection timeframes. Fourth, the timeframe of data recording was 1 week. Longer timeframes should be tested in future studies, which will be relevant for future applications to clinical trials. Finally, convergent validity with clinical ratings of negative symptoms and functional outcome was modest. This likely reflects differences in scaling and precision among these variable types. Associations between more temporally proximal EMA measures and geolocation were more robust than the associations with rating scales, suggesting that this approach is more ideal for determining convergent validity.

There are several considerations for adopting geolocation as a negative symptom or functional outcome assessment. It is clear that geolocation has potential benefits in terms of improving efficiency, being more time and cost effective than rating scales, and allowing scalable resolution to view symptom dynamics over user-defined time points and situations. However, much work is needed before digital phenotyping measures like geolocation can be incorporated into routine clinical care. The current study took an important first step in this direction. However, other critical processes must take place before geolocation can be used practically, such as creating a back-end system that supports the collection, storage, data processing, and analysis. Given the sheer volume of data, this will not be a straightforward process. It would also be beneficial to develop demographically stratified norms on large and representative populations of healthy and patient samples, similar to processes used for standardizing and norming neuropsychological tests. Second, there are issues related to ensuring trust. By its nature, geolocation calls into question issues of privacy and agency. This is particularly important in a population where some of the symptoms can revolve around surveillance and threats from being monitored. Although no patients in the current study indicated that geolocation data collection made them feel paranoid in our poststudy debriefing, our sample was clinically stable with mild positive symptoms. It is possible that geolocation could exacerbate delusions in more symptomatic samples and there is, thus, a need for further research to inform the limitations of the method. There are currently no standards for ensuring trust. Guidelines should be created in a collaborative effort with input from researchers, consumers, and industry. Finally, an important next step is to use geolocation as an outcome measure in clinical trials. Prior to doing so, it will be necessary to take the steps necessary for ensuring HIPAA and FDA safety and security compliance. Given that the rate of smartphone ownership is increasing in the world, including in people with schizophrenia, use of digital phenotyping methods is becoming increasingly more feasible.⁴⁹ However, there may be some subgroups of people with schizophrenia (eg, lower socioeconomic status, greater cognitive impairment, and older) for whom smartphone ownership is a challenge and these methods may be less practical.^50,51 With appropriate trial entry criteria and study designs, geolocation could be a valuable tool for pharmaceutical companies wishing to engage in siteless trials.

Author contribution: Initial manuscript draft was written by G.P.S. and I.M.R. with further contributions and edits from all authors. G.P.S. designed the study. G.P.S. and I.M.R. planned and performed statistical analyses. C.M..G., I.M.R., and H.C.C. collected data. I.M.R. conducted data management and processing. I.M.R. and S.H.J. interpreted and wrote statistical results.

Conflict of interest: G.P.S. and B.K. are original developers of the Brief Negative Symptom Scale (BNSS) and receive royalties and consultation fees from ProPhase LLC in connection with commercial use of the BNSS and other professional activities; these fees are donated to the Brain and Behavior Research Foundation. B.K. has received honoraria and travel support from ProPhase LLC for training pharmaceutical company raters on the BNSS, consulting fees and travel support from Genentech/Roche, Minerva Neurosciences, Lundbeck, and ProPhase LLC, consulting fees from anonymized pharmaceutical companies through Decision Resources, Inc, and from an investment capital company through Guideposts, and from Wockhardt Bio AG for consulting on a legal issue. G.P.S. has consulted for Minerva Neurosciences, Acadia, and Lundbeck. A.S.C., C.M.G., I.M.R., and H.C.C. have no conflicts to report.

Funding

Research was supported by R21-MH112925 to G.P.S. from the National Institute of Mental Health. Study sponsors had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.