BACKGROUND

METHOD

Theodora Chaspari, Ph.D., Professor of

Electrical Engineering, Texas A&M

Matthew Ahle,Graduate Student of Computational Data

Analytics, FIU

The Development of an mHealth Infrastructure for Child and Family Therapy

CONCLUSION

“Thismoodrecognizing

algorithmjustmightsaveyourrelationship.”– NBCNews

FEATURED

PATENTS

COLLABORATORS

• Couples carried smartphones and wore wearable sensors for one day• They provided hourly survey reports, including if and when they had conflict

1

4

6

9

10

Adela C. TimmonsUniversity of Southern California

FixedEffects B(SE)

Intercept(𝛾00k) 73.23(2.85)***Closeness(𝛾10k) -15.75(7.57)*

RandomEffects Estimate(SE)

Level1intercept(eijk) 197.79(19.67)

Level2intercept(u0jk) 114.15(48.96)

Level3 intercept(r00k) 10.91(14.96)

FixedEffects B(SE)

Intercept(𝛾00k) 73.35(3.89)***

Closeness(𝛾10k) -21.53(11.01)*

RandomEffects Estimate(SE)

Level1intercept(eijk) 638.07(63.19)

Level2Intercept(u0jk) 167.57(83.64)

Level3 intercept(r00k) 24.87(46.20)

Actual Closeness Predicted Closeness

CONTINUOUS HOURLY MOOD STATES

EXAMPLE OF IMPLEMENTED AUTOMATED MEASURES

Dating Aggression and Feelings of Closeness

RESULTS

Q Sensor Actiwave Nexus 5

PredictedActual

1. Timmons, A. C., Chaspari, T., Narayanan, S., & Margolin, G. (provisional patent filed March 2nd, 2018). A technology-facilitated support system for monitoring and understanding interpersonal relationships.

2. Timmons, A. C., Chaspari, T., Ahle, M., Narayanan, S., & Margolin, G. (provisional patent filed March 2nd, 2018). An expert-driven, technology-facilitated intervention system for improving interpersonal relationships.

EXAMPLES OF PREDICTED VS ACTUAL SCORES

TRY OUR PROTOTYPE ON YOUR MOBILE DEVICE

ALGORITHM DEVELOPMENT• Self-reports were used as the ground-truth criteria for

the machine learning algorithms• We started with easy-to-detect relationship states• We then recursively fed predicted scores into the new

models to improve the performance of later algorithms

WIREFRAME• Based on previous research and theory, we designed a

user interface for a mobile intervention system for use with families and couples

• After developing this measurement scheme, we built algorithms to automatically detect these constructs

THEORIZED COMPONENTS OF HEALTHY RELATIONSHIP FUNCTIONING

Connection Support Positivity

ACKNOWLEDGMENTSNSF Grant BCS-1627272 (Margolin, PI)NIH-NICHD R21HD072170 (Margolin, PI)SC CTSI (NIH/NCATS) 8UL1TR000130 (Margolin, PI)NSF GRFP DGE-0937362 (Timmons, PI)NSF GRFP DGE-0937362 (Han, PI)APA Dissertation Research Award (Timmons, PI)

Shrikanth Narayanan, Ph.D., Professor of

Electrical Engineering, USC

MAE = 1"∑ |yi − yi%| "&'( IMAE =

|(1) ∑ |yi.

+yi- |) +(1) ∑ |yi

+yi. |))/01 |)

/01

(1) ∑ |yi.

+yi- |))/01

×100

• Conflict, divorce, and interpersonal violence have far-reaching economic, psychological, and health costs for children and their families1,2

• Family fragmentation is estimated to cost US taxpayers $112 billion annually3

• We propose a framework for developing a mobile intervention system for improving relationship functioning in families and couples

• Our goal was to use these data to develop machine learning algorithms that could be used to automatically and passively detect theoretically-relevant relationship states via mobile technologies

DISCRETE HOURLY INTERPERSONAL EVENTS

Kappa = .97Sensitivity = .98Specificity = .99Accuracy = .99

Kappa = .90Sensitivity = .92Specificity = .99Accuracy = .96

Interacting Conflict

Person 1 Person 2

MAE = 14.54IMAE = 29.11%Correlation = .75

MAE = 11.40IMAE = 46.22%Correlation = .85

Gayla Margolin,Ph.D., Professor of Clinical Psychology,

USC

• Results provide initial proof-of-concept that it is possible to detect indices of relationship functioning in daily life with reasonable accuracy

• Next steps involve building and launching the fully functioning application system to detect important relationship processes in real-life settings as they naturalistically unfold