integrating wearables and user interaction patterns to monitor mental health
TRANSCRIPT
Integrating Wearables and User Interaction
Patterns to Monitor Mental Health
Maulik R. Kamdar, Michelle J. Wu, Zeshan Hussain
Introduction
Noninvasive Quantitative Continuous
DSM-5 guidelines ✔ ✖ ✖
EEG ✖ ✔ ✖
Video Monitoring ✔ ✖ ✔
Our goal ✔ ✔ ✔
Problem
Aims
● Develop a web application that aggregates smart watch data and patient insights
● Provide proof of concept through preliminary data collection
● Extract underlying patterns in the data and demonstrate a method for predicting mental health status
Brain Health Platform
http://54.200.211.229/BrainHealth/index.php
Brain Health Platform
Brain Health Platform
Methods: Data Collection- 12 anonymized participants (Ages: 19-37)- Gear S watch (10 am - 5 pm)
- Light - Heart Rate (BPM, Peak-to-peak) - Accelerometer (acceleration, rotation) - Pedometer (total distance, speed, calories, step count)
- Brain Health Web Application (10 am, 1 pm and 4 pm).- Textual, subjective Insights- Keyboard Interactions (key press time, interkey latency, speed,
number of errors, number of presses - back keys, enter, Ctrl+Z)- mouse interactions logged (move speed, drag speed, clicks)
Methods: Data Analysis- Features
- mean, standard deviation & dominant frequency for smart watch data types (heart rate, light, steps, speed, distance, acceleration, rotation)
- total distance, steps, walk steps, run steps- mean & standard deviation for the distribution of interkey latencies of
each bigram and press times of each unigram- number of spelling errors, undos, backspaces- average mouse move/drag velocity, total move/drag distance
- Modeling- one training example for each hour of data collected- cross validation to select a model, with 20% of data withheld for
evaluation
Conclusions & Future Work- We developed an integrated smart watch and web application that can
collect a variety of passive data for a user as well as provide feedback to the user through visualizations.
- Through a pilot study, we were able to demonstrate the efficient integration of many data types, with the potential for use as a predictor of emotional state.
- In the future, we hope to address data quality issues by developing smarter imputation and noise filtering methods.
- In addition, we plan to extend this study to patients with neuropsychiatric disorders and continue to develop our models for predicting patient status.
Acknowledgements
- Pilot study participants- Samsung- Dave Stark & Team mHealth- Diego Calderon- Reviewers- Russ Altman, Steve Bagley, Tim Sweeney
Thank you!
[email protected] · [email protected] · [email protected]
“Exogenous data (behavioral, social, environmental) is overwhelming, but 70% determinants of health occur in these data types” - Rob Merkel, IBM (Big Data in Biomedicine Conference 2015)