digital phenotyping for psychiatric disorders: ready for
TRANSCRIPT
Digital Phenotyping for Psychiatric Disorders: Ready for Primetime?
Gregory P. Strauss, Ph.D.
Assistant Professor
Department of Psychology
University of Georgia
Email: [email protected]
ACKNOWLEDGMENTS & DISCLOSURESDisclosures
▪ 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.
▪ Last 12 Months: Speaking/consultation with Minerva, Lundbeck, Acadia
▪ Grant support from NIH and Brain & Behavior Research Foundation
LIMITATIONS OF CLINICAL RATING SCALES
PRACTICAL & COST EFFECTIVE: Data collection and analysis can be automated, reducing cost of rater training, time and processing.
ASSESSMENT IN REAL TIME: Avoids influence of retrospective memory limitations, halo effects, biases, culture and other potential “noise”.
HIGH RESOLUTION/OPTICS: Allows for scalable resolution to view symptom dynamics over user-defined time and situation.
SPECIFICITY: Allows high “spectral” resolution for parsing symptoms from each other, and from global cognitive and other impairments.
FACILITATES DATA MINING FOR FUTURE ANALYSIS: Raw data is available for data mining, to fuel preliminary analyses, “proof of concept” studies and to uncover subtle medication effects and explain null findings.
ALL rating scales have certain limitations▪Cognitive impairment▪Rater biases▪Social desirability▪Halo effects that limit precision▪Arbitrariness of item anchors▪Limited scope of anchor scores▪Practicality (training, time, expense)
ADVANTAGES OF DIGITAL PHENOTYPING
DIGITAL PHENOTYPING METHODS
Patient Interface
Active Digital Phenotyping:
Subjective
Passive Digital Phenotyping:
Objective
DATA COLLECTION DATA TYPESANALYSIS/FEATURE
EXTRACTIONSPECIAL DATA
PROCESSING NEEDS
Self-Report Data(automated & “offline”)
Geolocation(automated & “offline”)
Accelerometry(automated & “offline”)
Ambient Acoustics(automated & “offline”)
None
DATA DISSEMINATION
None
UserInterface
Active Digital Phenotyping:
Objective
Vocal Acoustics(Requires simple, shareware processing
tools)
Requires minor human processing (e.g.,
transcribing, optimizing audio, video signal).
All analyses can be batched and automated and conducted “off-line”
(hence, “off the shelf” software solutions can
be used).
Semantic “Coherence” Analysis(Requires proprietary software and
“corpus”)
Video Analysis (facial, head, eye analysis)
Requires Proprietary software .
Lexical Analysis(Requires simple text-search tool). Proprietary “Dictionary” needed
ACTIVE DIGITAL PHENOTYPING: SURVEYS
▪ 7 Days of EMA▪ Scheduled within 90 minute epochs 9AM-9PM▪ 15min window▪ Takes ~5min per survey to complete
SURVEY ADHERENCE AND ACCEPTABILITY
Tolerability How Positive (0-10)
How Negative (0-10)
SZ 8.9 (1.0) 1.0 (1.4)
CN 8.5 (1.2) 1.4 (1.2)
Strauss et al, in prep: Acceptability
Summary:• Compliance is high• The method is highly tolerable• Minimally disruptive• Non-completion due to unavailability or missing prompt• High negative sx, older, less educated, lower IQ can complete surveys
Study Authors/Year # Surveys per Day/# Days
% Adherence
Brenner et al., 2014 6/7 98.1%
Visser et al., 2018 4/6 90.2%
Ben-Zeev et al., 2012 6/7 97.7%
Granholm et al., 2013 4/7 72.1%
Moran et al., 2016 4/7 80%
Sanchez et al., 2014 4/7 80.6%
Oorschot et al., 2013 10/6 80%
Granholm et al., 2019 7/7 85%
Strauss et al., 2019 4/6 90.2%
Palmier-Claus et al., 2012 6/7 82%
Edwards et al., 2018 7/6 71%
Summary 5.6 surveys, 6.6 days 84.3%
Moran, Culbreth, Barch, 2017
Correlation w/ # surveys: r = -.07; # days: r = -.12
ACTIVE DIGITAL PHENOTYPING: SURVEY VALIDITY
Moran, Culbreth, Barch, 2017
Associated with clinical interview based measures of the same construct and reward processing mechanisms underlying the symptoms
ACTIVE DIGITAL PHENOTYPING: DATA FROM VIDEO• Videos recorded concurrently with
each survey
• “Give a step by step description of what you did in the last hour. Please talk for a full 30 seconds"
• Submitted to automated facial and vocal analysis
Acoustic Analysis Lexical Analysis Semantic Analysis
Blunted AffectAlogia Emotional
Experience Motivated Behavior Mood Symptoms(Mania, Depression) Disorganization
Cognition
Facial Expression, Gesture, Eye movement Analysis
ACTIVE DIGITAL PHENOTYPING: DATA FROM VIDEOSignal is highly variable over time: But shows consistency for context/time:
Patients Versus Controls: Modelling Group Differences:
10-folds Training Validation:Predicting Group status
339 control cases271 patient cases
Average Accuracy = 74%
Cohen et al., under review)
Highly correlated with Disorganization Self-Harm
ACTIVE DIGITAL PHENOTYPING: DATA FROM VIDEOBNSS: Blunted Affect
222 control cases49 patient cases
R2 = .33Average Accuracy = 82%
BNSS: Alogia243 control cases28 patient cases
R2 = .36Average Accuracy = 91%
EMA Paranoid Delusions207 control cases15 patient cases
R2 = .18Average Accuracy = 92%
EMA Anxiety180 control cases42 patient cases
R2 = .12Average Accuracy = 82%
Cohen et al., in prep)
NLP: Speech to Text and Semantic Analysis can be automated (> 80% accuracy)
NLP-based measures of Cognition Measures of Cognitive State
ICC (3 sessions): Mean: 0.68
r with Neuropsych Measures 0.60
ACTIVE DIGITAL PHENOTYPING: COGNITION
Holmlund et al., 2019
PASSIVE DIGITAL PHENOTYPING: ACCELEROMETRY
▪ Change in ACL was passively measured on the phone.
▪ Participants wore an Empatica wristband, which passively collected ACL as a constant force in g, which was later converted to m/s2.
▪ Both measures of ACL were collected over x, y, and z axes, from which a magnitude score was calculated, using:
SZ
ACL = 𝑥2+𝑦2+𝑧2
3
Representative Data Across 24Hrs Equipment
ACCELEROMETRY
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
Phone Mean Phone SD Band Mean Band SD
Co
he
n's
d E
ffec
t Si
ze
ACL Variable
9.9
10.0
10.1
10.2
10.3
10.4
10.5
10.6
10.7
Resting Not Resting
Ph
on
e M
ean
Activity Type
SZ CN
Validity: At Rest vs Not Cohen’s d Effect Sizes
ACL Band Mean w/ Negative Symptoms: r = -.56, p<.001
Strauss et al in prepControl: n = 54; Schizophrenia: n = 54
PASSIVE DIGITAL PHENOTYPING: GEOLOCATIONVariable Abbreviation(s) Definition
Home time Amount of time spent at home
Distance change Δd Distance traveled in meters from previous sample
calculated by Haversine formula
Distance from
home
Δmh Meters from home for each sample calculated by
Haversine formula
Stationary location
clusters
NC Number of distinct geographical locations sampled
Location variance LV Variance within locations traveled to
Entropy ENT Amount of movement and randomness in locations
Transition time* TT Time spent moving between distinct locations
GEOLOCATION
Cohen’s d Effect Sizes
Depp et al 2019Strauss et al under reviewControl: n = 54; Schizophrenia: n = 54Control (n
= 56)Schizophrenia (n = 86)
Z, P-Value Cohen’s D
Median Distance Traveled
23.8 (17.6)
12.3(10.4)
2.5, <.001 0.80
Median Distance from home
19.8 (16.6)
8.1 (9.0)
2.3, <.001 0.88
% Samples at Home
51.1% (0.38)
74.4% (0.25)
1.9, p<.001 0.72
CAINS Negative Symptoms
Median Distance Traveled -.35***
Median Distance from home -.35***
% Samples at Home 0.29**
VALIDITY SUMMARY
• The active and passive digital phenotyping measures demonstrate validity via:• Group differences between clinical and control groups
• Correlations with symptom, functional outcome, and cognitive measures from the same constructs measured via standard clinical instruments
• Correlations between active and passive digital phenotyping measures that are temporally proximal
Sensitivity to Change in Clinical Trials
• Bell et al (2020) Blended EMA/I for Coping with Voices in Psychosis
Sensitivity to Change in Clinical Trials
• Bell et al (2020) Blended EMA/I for Coping with Voices in Psychosis
Sensitivity to Change in Clinical TrialsMoore et al., 2016: Anxiety & Depression
Sensitivity to Change in Clinical TrialsMunsch et al., 2009: Binge Eating
SUMMARY
▪ DP is a reliable and valid objective measure of psychiatric symptoms and functional outcome
▪ Objective measures overcome several limitations of traditional clinical rating scales
▪ More contextually sensitive and temporally dynamic than clinical rating scales.
▪ Combining active and passive methods may be most powerful approach
▪ Near-continuous recordings and large numbers of samples enhance probability of detecting treatment effects via enhanced power, offering promise for clinical trials.
▪ Evidence for sensitivity to change in clinical trials
LIMITATIONS AND ISSUES TO BE RESOLVED
• Do these tools measure something meaningful to those with psychiatric conditions?
• Industry-ready, compliant, validated tools
• Norms and scaling
• Automated scoring
• Data analysis
• Data management
• Compliance with FDA regulations (audit trails and security)
• Optimal use of phone vs other technologies (e.g., band)
ACKNOWLEDGMENTS & DISCLOSURES
Acknowledgments
FUNDING• NIMH
• R01-MH116039• R21-MH112925• K23-MH092530• NARSAD Young Investigator
CAN Lab Research Team
• Ian Raugh (Grad student)
• Cristina Gonzalez (lab manager)
• Sydney James (coordinator)
• Katie Visser (Grad student)
• Lisa Bartolomeo (Grad student)
Collaborators• Alex Cohen, Ph.D.• Brian Kirkpatrick, MD• Eric Granholm, PhD
Thank you!