Linking Resident Behavior to Health Conditions in an Eldercare Monitoring System Mihail Popescu, PhD; Andrew Craver, MS, MPH Health Management and Informatics Dept. University Of Missouri School of Medicine Columbia, Missouri, USA http://eldertech.missouri.edu

Why eldercare monitoring? In US the number of people over the age 65 will more than double, from 40.2 million in 2010 to 88.5 million, by 2050 With older age there is a higher risk for multiple and more severe chronic conditions. Identifying a chronic condition early may lead to timely adjustment of supportive/health services to manage or reverse the functional decline. Sensor networks have emerged as a possible solution for detecting health changes

Clinical interpretation of sensor data is hard! Why: sensor data variety (many sensor types) and velocity (continuous monitoring)… When a abnormal sensor measurement is detected (value outside of some std from mean) an alert is sent to the clinician Currently, clinicians use a secure interface to review multiple data displays that may take 5-7 minutes per health alert. Our goal: provide more clinically meaningful and easier to interpret alerts to clinicians. How?: we are investigating a new data-to-knowledge methodology based on linguistic summaries and sequence data annotation

Project Location: Columbia, MO Monitoring systems deployed in 6 Americare long term care facilities, including TigerPlace An NLM funded project First apartment instrumented in 2005 with non-wearable sensors Currently, 40 apartments on line; target: 110 Sensors deployed: Motion sensors (PIR) in all rooms: measure the level of activity in the apartment Bed sensors (ballistocardiography- BCG): measure respiration rate, bed restlessness, heart rate Depth camera: measure gait speed and provide fall alerts To design a robust illness detection system, one efficient approach to health monitoring is to use sensor networks to collect information about the older adult’s activity. We deployed our integrated monitoring system in 47 TigerPlace apartments . we decided to use only non-wearable sensors for monitoring since they are unobtrusive and more acceptable by older adults. The monitoring has started in fall 2005. On average, we have two years of data for each resident. Various sensors have been deployed in each apartment: motion, radar, Microsoft Kinect, and bed which in this study we only used motion sensors and bed sensors. The bed sensor is able to measure bed motion, pulse and breathing.

Can health changes be linked to captured behavior? Non-wearable sensors can capture resident behaviors such as: Walking – walking speed, stride length, stride time, balance Bathroom (hygiene) activity- time in bathroom, number of times in bathroom General activity: amount of apartment in a certain time interval Sleep – restlessness, amount of sleep, quality of sleep, heart rate, respiration rate We don’t capture complex activities such as preparing meals or socialization Some health conditions may be linked to a specific behavior Pacing (walking back-and-forth) may linked to dementia Frequent bathroom visits during the night may be due to urinary tract infection (UTI) Decreased walking speed may linked to depression or arthritis However, more specific relations between behavior and health changes are needed for our monitoring system

Resident behavior captured by our sensors Apartment activity (overall motion) Motion sensors Number of bathroom visits Amount of time spent in bed Bed sensor Rate of respiration while in bed Restlessness while in bed Pulse rate while in bed Walking speed and stride length Depth sensor a

Survey to link behaviors to health conditions To guide the development of our data-to-knowledge methodology we employed a survey of clinicians involved in eldercare. Goal: determine the main behavioral patterns most relevant in evaluating several health conditions in older adults. Sample: we surveyed 22 clinicians (11 physicians, 5 registered nurses and 6 licensed practical nurses). All participants were familiar with our monitoring system

Survey instrument Consisted of a 7-item questionnaire Rate how useful they would find a set of sensor measurements when evaluating or treating health conditions in elderly. Use a1-4 scale to assess the usefulness of each behavior and sign: 1= Not Useful; 2 = Somewhat Useful; 3 = Useful; 4 = Very Useful. In processing: score (3,4) were considered as useful, (1,2) as not useful We assessed 12 health conditions: depression, dementia, mental status change, chronic obstructive pulmonary disease (COPD) exacerbation, chronic health failure (CHF) exacerbation, atrial fibrillation, transient ischemic attack (TIA)/stroke, fall risk, hypo/hyperglycemia, urinary tract infection (UTI), pneumonia and pain. The survey was performed using Qualtrix

Results Proportion of clinicians finding 3 resident behaviors (signs) important for assessing a given health condition Condition Behavior (sign) 1 Behavior (sign) 2 Behavior (sign) 3 Atrial Fibrillation Pulse: 91% Respiration: 50% Restlessness: 36% CHF Exacerbation Pulse: 82% Respiration: 82% Restlessness: 55% COPD Exacerbation Respiration: 96% Pulse: 68% Restlessness: 50% Dementia Restlessness: 68% Overall Motion: 59% Time in Bed: 59% Depression Time in Bed: 91% Overall Motion: 82% Restlessness: 59% Fall Risk Stride length: 100% Bathroom Motion: 82% Overall Motion: 68% Hypo/Hyperglycemia Bathroom Motion: 32% Restlessness: 28% Overall Motion: 24% Mental Status Change Time in Bed: 82% Restlessness: 77% Pain Stride length: 82% Respiration: 73% Pulse: 73% Pneumonia Respiration: 86% Pulse: 59% Overall Motion: 41% TIA/Stroke Stride length: 55% Pulse: 46% Respiration: 27% Urinary Tract Infection Bathroom visits: 96% Restlessness: 46% Overall Motion: 46%

Behavior trend direction and time span Our intention is to be able to write (fuzzy) rules using behavior trend (similar to Haimowitz&Kohane – 1994) of the form: IF “Bathroom motion” has increased and “Bed restlessness” has increased for the last 5 days THEN UTI is possible While for physiological variables the trend is easy to evaluate, it is not so for behavioral ones We surveilled only 4 clinicians (the most experimented ones) for trend and time span for the 12 conditions.

Behavioral trends and time span for UTI Clinician Bathroom Motion (0.96) Restlessness (0.46) Overall Motion (0.45) Time span Clinician1 increase either Days/weeks Clinician2 Probably increase  ? Clinician3 Increase or stable 1 day Clinician4 Increase 10 days TREND summary Increase/varies 1 day - 10 days

Behavioral trends and time span for PAIN Clinician Stride (0.81) Respiration (0.73) Pulse (0.73) Timeframe Clinician1 Either Increase Most likely increase Few days Clinician2 Decrease Unchanged or increase  ? Clinician3 Decrease or unchanged (depends on location of pain, i.e. legs (but could decrease with opioid pain meds) 1 day Clinician4 1-24 hrs TREND

Conclusions and future work We tried to identify the link between 12 health conditions in elderly and behavioral variables measured by our monitoring system While for 9 of the conditions surveilled we obtained relevant results, for other three 3 (TIA, hypo/hyperglycemia and dementia) our monitoring system doesn’t seem able to provide useful information The pilot survey for behavior trend direction and time span for each condition produced conflicted information As we accumulate more data in our data warehouse, we should be able to apply data mining techniques (association rules) to detect the link between behavioral variables and clinical conditions.

Thank you! This work has funded by the National Library of Medicine grant #R01LM012221.