Upload
selina-walle
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Behavioural Intervention Technologies (BITs) for Depression
David C. Mohr, Ph.D.Center for Behavioral Intervention
Technologieswww.cbits.northwestern.edu
Northwestern UniversityFunded by NIMH P20 MH090318© Northwestern University & David C. Mohr
www.cbits.northwestern.edu
DepressionMohr, et al, Ann Behav Med. 2006;32:254-258; Mohr, et al, J Clin Psychol. 2010;66(4):394-409.)
• 12 month prevalence rates for mood disorders are 9.5% and 6.7% for Major Depressive Disorder.(Kessler Arch Gen Psychiatry, 2005. 62: 617-27)
•In 2 surveys in primary care at UCSF & Northwestern (≈950) o More than 50% of patients reported at least one substantive barrier.o More than 75% of depressed patients (PHQ>10) reported barriers
•Most barriers are structuralo Costo Lack of available serviceso Time constraintso Participation restriction (e.g. disability/symptom interfering)o Caregiving responsibilitieso Other non-structural issues include
o Stigmao Negative impressions of therapy/therapistso Lack of motivation
www.cbits.northwestern.edu
Telephone Administered Psychotherapy Mohr DC, et al Clin Psych: Sci Prac. 2008;15(3):243-253.
T-CBT vs. F2F-CBTNon-inferiority trial
F2F T-CBTpMean Mean
Age 47±13.5 47±12.6 0.87
% % Gender 0.71 Female 78.4% 76.7% Race 0.63 African American 24.0% 24.3% White 65.3% 60.1% More than one race 8.0% 12.2% Other 2.7% 3.4% Ethnicity 0.76 Not Hispanic or Latino 13.0% 14.2% Education 0.57 High School 8.6% 12.3% Some college 25.3% 24.5% Bachelor’s Degree 39.5% 33.7% Advanced Degree 26.5% 29.5%
•Trial Characteristics• 325 Patients recruited
from Primary Care• MDD + Ham-D≥16• 18 session of CBT• PhD therapists
supervised by Beck Institute
• Non-inferiority Margin: d=0.41
www.cbits.northwestern.edu
PHQ-9 OutcomesMohr DC, et al. JAMA. 2012;307:2278-2285.
www.cbits.northwestern.edu0 4 9 14 180.00
2.00
4.00
6.00
8.00
10.00
12.00
14.00
16.00
18.00
20.00
FaceTele
p = 0.89ES = -0.02; 90% CI (-0.20, 0.17)
Week
Attrition• Total Attrition (p = 0.02)
• F2F-CBT 32.7%• T-CBT 20.8%
• Failure to Engage (> 4 sessions; p = 0.005)• F2F-CBT 13.0%• T-CBT 4.3%
• Failure to Complete (p = 0.46)• F2F-CBT 20.0%• T-CBT 16.6%
www.cbits.northwestern.edu
Internet Penetration in the US
Computer access to the Internet continues to increase:
• 19-29 y olds ≈ 92%• 30-49 y olds ≈ 87%• 50-64 y olds ≈ 79%• 65+ ≈ 42%
> 60% have broadband access
www.cbits.northwestern.edu
Among 495 Primary Care Patients from Northwestern’s GIM who want
psychological/behavioral intervention
Interest in Web-based Treatment
252 (51.5%)
Definitely Want Would consider Not interested
56 (11.5%) 181 (37.0%) 252 (51.5%)
www.cbits.northwestern.edu
Mohr DC, et al. Ann. Behav. Med. 2010;40:89-98.
Meta-Analyses of Web-based Intervention for DepressionAndersson, G., & Cuijpers, P. (2009). Cogn Behav Ther, 38(4), 196-205.
• For depression, web-based interventions performed modestly well (d =.41).
• Standalone interventions for depression have small effects (d = .18) while those guided by a therapist/coach email have larger effects (d = .61).
• Attrition is very high, particularly for standalone treatments.
• Human involvement is important
• Lay persons performed as well as mental health professionals in providing support for web-based treatments (Titov N, et al. PLoS One. 2010;5(6):e10939)
www.cbits.northwestern.edu
www.cbits.northwestern.edu
moodManager
• Based on CBT principals, teaching behavioral activation and cognitive restructuring.
• Learning modules (10-15 min).– 7 core lessons + 12 optional lessons addressing
comorbidities (anxiety, anger, interpersonal difficulties, etc.)
• Interactive tools (2-4 min) that help patients implement learning. Tools are scaffolded, so that additional components are added on as patient becomes proficient.
• Coach interface that can track patient utilization on the site, observe work performed, and receive alerts (e.g. for suicidality).
cbits.northwestern.edu
www.cbits.northwestern.edu
Patient Dashboard
www.cbits.northwestern.edu
www.cbits.northwestern.edu
www.cbits.northwestern.edu
www.cbits.northwestern.edu
ThoughtDiary
www.cbits.northwestern.edu
www.cbits.northwestern.edu
www.cbits.northwestern.edu
www.cbits.northwestern.edu
www.cbits.northwestern.edu
Mood RatingsOver Time
www.cbits.northwestern.edu
Coach Interface
TeleCoaching Model:Supportive Accountability
Mohr DC, et al. Supportive Accountability.: A model for providing human support for Internet and ehealth interventions. Journal of Medical Internet Research. 2011;13:e30
Adherence
MotivationIn- Extrinsic
Communication “Bandwidth”
Human Support
Accountability
Legitimacy benevolence, expertise
Bond
www.cbits.northwestern.edu
TeleCoach
• Based on Supportive Accountability• First call was a 25 min “engagement session”.• Weekly calls of 5-10 minutes• Calls were designed to engage patient when
logins were low, create accountability, and reinforce adherence.
• Does not require mental health specialization to administer.
www.cbits.northwestern.edu
moodManager Pilot
• Patients with current MDD are recruited from primary care
• moodManager + S/A TeleCoaching (12 weeks)• moodManager alone (12 weeks)• Wait list control (6 weeks)
www.cbits.northwestern.edu
Descriptive Statistics101 patients enrolled
www.cbits.northwestern.edu
Age: M = 48.2
Female: 71%
Marital Status: • Single: 34%• Married/Partnered: 50%• Divorced: 16%
Race• African American: 34%• Caucasian:
59%• Other:
7%
Education• High School: 3%• Some college: 24%• Bachelor’s Degree: 34%• Advanced Degree: 38%
Employment• Unemployed: 27%• Disabled: 16%• Employed: 57%
www.cbits.northwestern.eduBaseline Week 6 Week 12
0
2
4
6
8
10
12
14
16
18moodManager
Coached mMmM onlyWLC
PHQ
-9
p=.01
www.cbits.northwestern.eduBaseline Week 6 Week 12
0
2
4
6
8
10
12
14
16
18moodManager
Coached mMmM onlyWLC
PHQ
-9
p=.06 p=.01
www.cbits.northwestern.eduBaseline Week 6 Week 12
0
2
4
6
8
10
12
14
16
18moodManager
Coached mMmM onlyWLC
PHQ
-9
p=.06
p=.97
p=.01
www.cbits.northwestern.edu
Percent of patients reaching full remission (PHQ-9>5)
0%
5%
10%
15%
20%
25%
30%
35%
Week 6 Week 12
Guided iCBT
Unguided iCBT
WLC
Adherence
• Mean cumulative logins were significantly greater for Coached moodManager (p = .02).• Coached mM: M=16.8, Range = 0-112• mM only: M=8.3, Range = 1-27
• Approximate point for 50% attrition• Coached mM: 7th week• mM only: • 2nd week
www.cbits.northwestern.edu
Supportive Accountability Measurement
www.cbits.northwestern.edu
Items• ___ notices when I use moodManager• ___ is aware of how I have completed the tools in moodManager• ___ expects I will use moodManager regularly• ___ will think less of me if I don’t use moodManager• It would bother me if ____ thought less of me• If I use moodManager less than expeced, I feel like I need to give
___ reasons why.
Reliability: Cronbach’s alpha = .79
Administered at week 6.• Correlation with # logins: r = 0.45
Onward – Peer Networking for Cancer Survivors
• Once we begin to understand the principles of why coach involvement improves outcomes and adherence, we can apply the principles in new contexts, such as online social groups.
• Approximately 1/3rd of cancer survivors experience clinically significant levels of depression after completion of treatment.
• Online treatments, primarily online support groups, are very popular among cancer survivors.
• Could peer interactions in an online intervention be designed to promote supportive adherence?
www.cbits.northwestern.edu
www.cbits.northwestern.edu
Promoting AccountabilityIn Peer Groups
Promoting AccountabilityIn Peer Groups
www.cbits.northwestern.edu
www.cbits.northwestern.edu
0
2
4
6
8
10
12
Baseline Week 4 Week 8
HA
DS -
Dep
ress
ion
Onward Pilot Trial17 cancer patients with depressive symptoms
randomized (p= 0.12)
Onward (d=1.27)
moodManager-C (d=0.89)
Onward Adherence
• Mean Logins• Onward 14.5
• moodManager-C 8.0
www.cbits.northwestern.edu
Mobile Interventions
Two basic components of behavioral interventions have been tested in mobile phones
• Components similar to Internet interventions, including didactic content and interactive tools
• Ecological Momentary Intervention (EMI). Users can log information about their situation or current state. Personalized messages are sent to assist the user with problems. (Patrick K, et al. J Med
Internet Res. 2009;11e1)
www.cbits.northwestern.edu
General Model of e-Health Activity & Information Flow
Initiate Use
Data Input(Logging activities)
Feedback(Graphs Insight)
Behavioral Prescriptions(Instructions Behavior Change)
www.cbits.northwestern.edu
Context Awareness
• Context awareness refers to the idea that computers can both sense, and react based on their environment.
• Example: Body area networks (BANs) can transmit data from sensors which can be used to monitor health conditions.– Sensors can include ECG, oxygen saturation, accelerometers,
etc.– BANs have been used with a variety of health states (contexts)
such as diabetes, asthma, and health-related behaviors such as physical activity.
• Mobile phones today include numerous sensors that can potentially be harnessed to understand patient states.
www.cbits.northwestern.edu
Context Inference System4) The Machine Learner is “trained” using EMA (queries)
How context sensing works1) An mobile phone PASSIVELY collects 10000s of
RAW data points of in-phone “probes”
41.8322 N, 87.6513° W250 m, 70% accurate, Wifi, Currently Near 4
Wifi Networks [Linksys, VultureWeb, Bob’s HouseNet], more…
Called for 3 minutes @ 2:30PMText @ 1 PM
Email @ NoonMore…
RotationAcceleration
VelocityGravity
Gyroscope
2012-09-12 04:38:36Z
Device ID: 33:AD:4F:C3:3F:B2:D1:11
2) Participants are prompted to engage in ACTIVE (self-reported) DATA COLLECTION during a wide range of instances:
1. Randomly
2. Whenever they want
3. Retrospective ratings (to capture contexts that when current ratings are not feasible, such as driving, or to protect privacy)
4. User-Scheduled Intervention Moments
5. Anomaly Detection
6. Classification Model Validation…
How context sensing works
www.cbits.northwestern.edu
3) Building a Model Starts with “Pairing”
PASSIVESensor Data
(3PM)
ACTIVEAssessment
Data (eg Happy)
(~3:03 PM)
Time
PASSIVESensor Data
(3:01PM)
PASSIVESensor Data
(3:02PM)
PASSIVESensor Data
(3:03PM)
PASSIVESensor Data
(3:04PM)
PASSIVESensor Data
(2:59PM)
A “pair”
www.cbits.northwestern.edu
Now we build a set of pairs
LOTS OF SENSOR
DATA
MOOD = 6 out of 10
LOTS OF SENSOR
DATA
MOOD = 3 out of 10
LOTS OF SENSOR
DATA
MOOD = 7 out of 10
LOTS OF SENSOR
DATA
MOOD = 1 out of 10
LOTS OF SENSOR
DATA
MOOD = 10 out of 10
Training Set
www.cbits.northwestern.edu
And build it…
LOTS OF SENSOR
DATA
MOODASSESS-MENTS
Training Set
Lots of pairs
MachineLearner
MOODMODEL
Is sent to
Which crunches
them together
and makes
Can use a variety of analytics:• Decision Tree• Naïve Bayes• Logistic Regression
www.cbits.northwestern.edu
4) We make educated guesses by continuously monitoring probes with our models
PASSIVESensorData
(8AM Wed)
Time
WedModel
Accuracy37%
MOOD PREDICTION 6 out of 10
They must be:
ACCURATEENOUGH
INTERVENTIONWORTHY
PASSIVESensorData
(8AM Tues)
Tues Model
Accuracy20%
MOOD PREDICTION 4 out of 10
PASSIVESensorData
(8AM Thur)
ThursModel
Accuracy58%
MOOD PREDICTION 2 out of 10
PASSIVESensorData
(8AM Fri)
FriModel
Accuracy84%
MOOD PREDICTION 1 out of 10
NO NO NO YES!
5) When the model is accurate enough and pertinent, we intervene OR assess…
PASSIVESensor Data
(930AM)
ModelAccuracy
.84
MOOD PREDICTION 1 out of 10
www.cbits.northwestern.edu
Field Trial
• Aim: – Conduct initial feasibility study– Beta-test the context awareness system
• Mobilyze Intervention (8 weeks)– Weekly didactic information provided to support behavioral activation– Interactive tools (activity monitoring and scheduling, managing
avoidance)– Weekly 5-10 minute calls to support adherence and obtain usability
info• 8 patients with MDD enrolled
– 7 women– Mean age: 37 (range 19-51)– 6 had comorbid anxiety disorders
www.cbits.northwestern.edu
Mobilyze!• 1 participant dropped out due to problems
with the phone.• Satisfaction was reasonable (M=5.7 on a 1-7
likert scale)
Depression OutcomesWeek 1 Week 8 p
PHQ-9 17.1 ± 3.8 3.6 ± 4.1 <.0001
QIDS 13.8 ± 2.1 3.4 ± 3.1 <.0001
MDE 8 (100%) 1 of 7 (14.3%) <.01
www.cbits.northwestern.edu
Context Awareness
• Patients were asked to train the phone by answering 11-15 questions, 5 times/day for the first few weeks. (They could also answer the questions without prompting)
www.cbits.northwestern.edu
Accuracy of Prediction ModelsContext % Accuracy 95% CI
Location 60.3 43.2 – 77.2
Alone in the Immediate Vicinity 80.1 76.2 – 84.5
Friends in the Immediate Vicinity 90.8 84.3 – 95.7
Alone in the Larger Environment 72.6 61.0 – 82.8
Miscellaneous People in the Larger Environment 90.9 83.8 – 97.3
Having a Casual Conversation 66.1 54.0 – 77.6
Not Conversing 64.5 58.4 – 70.3
Accuracy for many models (e.g. mood) was not better than chance
www.cbits.northwestern.edu
Sample Learning Algorithm
www.cbits.northwestern.edu
Sample Learning Algorithm
www.cbits.northwestern.edu
Problems
• Multiple technological problems, for example– Battery drainage and connectivity issues– Poor sensor data quality– Assessment problems (e.g. low variability)– Usability problems– And…
www.cbits.northwestern.edu
Data does not equal information
LOTS OF SENSOR DATA 41.8322 N, 87.6513° W
250 m, 70% accurate, fWifi, Currently Near 4 Wifi Networks [Linksys,
VultureWeb, Bob’s HouseNet], more…
MOST RAW DATA DOESN’T HAVE MUCH MEANING
=
www.cbits.northwestern.edu
HAPPINESS MODEL v1AccelerometerY
HappinessBetween 8-10
AccelerometerX
Happiness Between 4-6
Longitude
Happiness Between 6-8
Day of Week
Happiness Between 0-2
Happiness Between 2-4
Accuracy: 32%
Thur
sday
Mon
day
>87.
654
<=87
.654
>9.2
3<=
9.23
<=0.
1324
1>0
.132
41
www.cbits.northwestern.edu
41.8322 N, 87.6513° W250 m, 70% accurate, Wifi, Currently Near 4
Wifi Networks [Linksys, VultureWeb, Bob’s HouseNet], more…
MOST RAW DATA DOESN’T HAVE MUCH MEANING
So we build FEATURES to help usunderstand the raw data better
• Use domain specific expertise (e.g. How can we use this to give us more information about a person’s behaviors?)
• Location sensors might be coded as distance from key locations (home, work)
• Days of week might be coded as work days/non-work days
becomes Distance from Work
Distance from Home
www.cbits.northwestern.edu
LOTS OF SENSOR
DATA
PASSIVE (SELF-
REPORT) ASSESSMENT
Training Set
Lots of pairs
MachineLearner HAPPINESS
MODEL
Is sent to
Which crunches
them together
and makes“SMART” FEATURES
Each are now validated independently and the most accurate is used:• Decision Tree• Naïve Bayes• Logistic Regression• Support Vectors• Neural Network
www.cbits.northwestern.edu
HAPPINESS MODEL v2
Distance from Home
Happiness Between 8-10
Happiness Between 6-8
Day of Week
Happiness Between 4-6
AccelerometerY
Happiness Between 0-2
Happiness Between 2-4
Accuracy: 82%
<4.2
>=4.
2
Sa, S
unM
,Tu,
W,T
h,F>
7 m
iles
<7 m
iles
www.cbits.northwestern.edu
• Work with the plain old telephone suggests we can extend care into people’s environments, increasing the reach of behavioral care.
• But, if users don’t use it, it doesn’t work.• Human support models are needed that are effective and
scalable. • The most profound technologies are those that disappear.
They weave themselves into the fabric of everyday life until they are indistinguishable from it. Weiser M.. Sci. Am. 1991;265(3):94-104
Summary
www.cbits.northwestern.edu