Older adults live longer but have more chronic health issues

• Life expectancy +3 months every year since 1840

• 75% older adults report being in good to excellent health

• 70% report chronic health conditions

Growing older population need assistance leading to care gaps

• US health care system faces shortfalls in primary care physicians and paid caregivers

• 50% need or receive help with routine activities

• Baby boomers: 45M adults 65+ as of 2015

US Population Age Distribution

https://www.weforum.org/agenda/2016/03/167-years-of-the-us-age-demographic-in-one-chart

US Population Age Distribution

https://www.weforum.org/agenda/2016/03/167-years-of-the-us-age-demographic-in-one-chart

In-home sensor systems face significant challenges

1. Difficult to capture and label realistic, nature patterns of daily living

2. Passive sensor systems need to be minimally invasive

3. Low-cost and easy to deploy and maintain at large scales

Study Participants

Health BOOST, University of South Florida

• Recruitment

  • 55+ resident of The Villages
  • Living alone in pet-free home with internet access
  • Do not exhibit signs of cognitive impairment

• 14 total participants (7 ongoing)

  • 5,450 total days of observation

• 12 have completed at least 6 months

Participants considered for comparison of Activity Profiles

• 7 Participants

• Completed at least 196 days (~7 months):

  • Excluding vacations, maintenance days
  • At "home" for most of the day
  • Avoiding guests and visitors

• Similar floorplan and sensor layout

Sensor Layout & Coverage

Daily Activity Heatmap

Activity Profile Process

• Gather sensor event sequences

• Group into day-length event sequences

• Build activity profiles from day-event sequences

Sensor Event Sequences

Insert pauses for activity lulls

Pauses

Identify start of day

Day Start Marker: 5 sequential active events after 4am

Day Event Sequence

IiIAEFfFfA2AE1EGAgEAIiIAEIiIiOiIGAgEIIIAE1AE2AEAAE1A1AAEIiIFfFfAEA2AEAE1AEEAAEAEAA1AIiI2AEIiI3IIiI2

Day Event Sequence

IiIAEFfFfA2AE1EGAgEAIiIAEIiIiOiIGAgEIIIAE1AE2AEAAE1A1AAEIiIFfFfAEA2AEAE1AEEAAEAEAA1AIiI2AEIiI3IIiI2

Activity Profiles: Inspired by text mining & document classification

Similar to document categorization [Cavnar1994]

• Counted and ranked n-grams in category documents

• Compared n-grams from unknown document to category profiles

Activity Profile

• For m day-event sequences:

  • Count all n-grams from n_min to n_max
  • Calculate relative frequency p of each n-gram

Gather n-grams from day event sequence

Activity Profile: bag of n-grams representation of day event sequences

Activity Profile: bag of n-grams representation of day event sequences

Use KL Distance to measure distance between Activity Profiles

Comparing Activity Profile P to Daily Activity Profile Q

Kullback-Leibler Distance (Symmetric)

$$D(P\parallel Q)=\sum _{x\in \chi }\left(\left(P\left(x\right)-Q\left(x\right)\right)\log \frac{P\left(x\right)}{Q\left(x\right)}\right)$$

with back-off

$$\begin{array}{cc}P(t_k,d_j)& =\{\begin{array}{cc}\beta P\left(t_k\right|d_j)& \text{if}{t}_k\text{occurs in document}{d}_j\\ ϵ& \text{otherwise}\end{array}\\ \beta & =1-\sum _{k\notin d_i,k\in P}ϵ\end{array}$$

Identifying Occupants Using Activity Profiles

1. Do Activity Profiles capture occupant's unique patterns?

2. How many days of activity are needed?

3. Can we use a activity profiles as baseline as occupant ages?

Build Activity Profiles for each occupant from N day event sequences

Compare unknown single-day profile with all Activity Profiles to predict occupant

Experimental Setup

• 196 total days of data

• How well do Activity Profiles work?

  • Build profile for each occupant from N training days

  • Use profiles to identify occupant of test day

  • 7-fold modified cross validation:

    • 28 testing days (~ 1 month) per fold
    • 168 training days available for profile

• How many days are needed for quality Activity Profiles?

  • N profile days:

    • From 14 days to 3 months

Testing days from contiguous days

Profiles built from N sampled days

Samples are repeated for equal number of profile day trials

Profiles: 168 days (~6 months)

Test: 28 days

Results

Average Accuracy

Average Accuracy by Occupant

First 42 days as baseline profile
compared with remaining 154 days

Summary

Activity Profiles

• Accurately characterize smart home occupant daily activities

• Useful for long-term monitoring of older adults as they age

• Use simple, passive sensors emebedded in home environment

• Do not require expensive activity labelling for training

Future Work

Questions? gadenbuie@mail.usf.edu