Computational Algorithms for Predictive Health Assessment M. Popescu M. Skubic & J. Keller R. Koopman
Health Management & Informatics Electrical & Computer Engineering School of Medicine
University of Missouri , Columbia, MO, USA
More details @ www.eldertech.missouri.edu Funded by the NSF SHWB grant, award #: IIS-1115956
1. Popescu M, Chronis G, Ohol R, Skubic M, Rantz M, "An Eldercare Electronic Health Record System for Predictive HealthAssessment," IEEE Int. Conf. on Health Communication 2011, Columbia, MO, June 13-15, 2011, pp 193-196.
2. Rantz M, Marek K, Aud M, Tyrer H, Skubic M, Demiris G & Hussam A, "A Technology and Nursing Collaboration to Help OlderAdults Age in Place," Nursing Outlook, vol. 53, no. 1, pp. 40-45, January-February, 2005.
3. Z. Hajihashemi, M. Popescu, “Improving Health Pattern Recognition Using Smith Waterman Algorithm and NLP, AMIA FallSymposium, Washington Dc., Nov. 13-16 2013.
4. Z. Hajihashemi, M. Popescu,” Detection of Abnormal Sensor Patterns in Eldercare”, E-Health and Bioengineering Conference(EHB), 21-23 Nov. 2013, Iasi, Romania, 2013, pp. 1-4.
5. Z. Hajihashemi, M. Yefimova, M. Popescu, “Detecting Daily Routines of Older Adults Using Sensor Time Series Clustering”,submitted to EMBC 2014, Chicago, IL.
6. Z. Hajihashemi, M. Yefimova, M. Popescu, “A New Illness Recognition Framework Using Frequent Temporal Pattern Mining”,submitted to SmartHealthSys 2014, Seattle, WA.
Apartment
Wireless sensor network(bed, motion, stove and other sensors)
Computational
algorithms
Predictive
health
assessment(fall risk,
depression, UTI,
etc.)
Electronic
health records
(EHR)
Daily activity summaries obtained by annotation of FDA
or activity chunks using a home-grown nursing electronic
health record (EHR) system [1] and NLP [2] using a
bioinformatics approach (“guilt by association”)
Current trend
Aimed trend (with technology)
Functional Decline [2]
Time
1. Predictive health assessment framework
based on detection of missing frequent daily
activities [6] (as opposed to finding abnormal
patterns)
Unsupervised algorithms for finding FREQUENT
daily activities (FDA) [5], [6] (activities performed
at least once a day in a given period of time)
A bioinformatics motif finding algorithm (MEME)
A modified behavioral science algorithm (THEME)
based on time and symbol distribution in
sequences
TigerPlace aging-in-place facility, Columbia, MO
UserID SensorID Year Month Day Hour Minute Second
3 3 2005 10 5 12 34 38
3 2 2005 10 5 12 36 52
3 2 2005 10 5 12 37 04
3 2 2005 10 5 12 37 11
3 1 2005 10 5 12 37 26
3 1 2005 10 5 12 37 28
3 2 2005 10 5 12 37 32
3 2 2005 10 5 12 41 18
3 2 2005 10 5 12 41 11
3 2 2005 10 5 12 41 4
3 5 2005 10 5 12 42 40
3 5 2005 10 5 12 42 58
Discrete sequence representation
55 apartments monitored
2. Predictive health assessment framework
based on the detection of abnormal patterns (not
similar enough to previous patterns) [4]
Day chunking: find all activity chunks in a day
Compute a distribution of the activity chunk
similarity in a given time interval (2 weeks)
Send an alarm if a new activity is very
dissimilar to previous ones
Use a sensor sequence similarity, temporal
Smith-Waterman [3,4], to compute the similarity
between two activity patterns (sequences) T1, T2
Distribution of chunk similarities for 2 weeks
Hi0=H0j , iϵ[1,n] and jϵ[1,m]
Hij=max{0, Hi-1,j-1+ S(C1i, C2j), maxk≥1{ Hi-k,j - Wt), maxk≥1 { Hi, j-k - Wt}}
WΔt = g + c|t1i-t2j|
m}Min{n,
}Max{H)T,(TSimilarity
ij21
T1={(C11,t11), (C12, t12)…, (C1m,t1m)} T2={(C21,t21), (C22, t12)…, (C2n,t2n)}
Results: better than the typical Gaussian approach