HLTHINFO 730 Healthcare Decision Support Systems Lecture 13

HLTHINFO 730 Healthcare Decision Support Systems Lecture 13: Monitoring Lecturer: Prof Jim Warren HLTHINFO 730 – Lecture 13 Slide #1

Monitoring • A few different domains – Critical care monitoring – reporting back to humans who will respond quickly – ‘Ubiquitous’ monitoring – getting data (probably over a long period of time) without being too obvious about it – Participatory monitoring – patients get a sense of engagement by participating in the medical record – ‘Coaching’ – the interaction is mostly about encouraging healthy behaviour HLTHINFO 730 – Lecture 13 Slide #2

Critical care systems • Classic app is ECG monitoring P - R interval 0. 12 - 0. 2 seconds (3 -5 small squares of standard ECG paper) QRS complex duration less than or equal to 0. 1 seconds (2. 5 small squares) Q - T interval corrected for heart rate (QTc) QTc = QT/ RR interval less than or equal to 0. 44 seconds See http: //www. nda. ox. ac. uk/wfsa/html/u 1105_01. htm HLTHINFO 730 – Lecture 13 Slide #3

Another view of the ECG • One heartbeat Particularly want to look out for lengthening Q-T HLTHINFO 730 – Lecture 13 Slide #4

Amplitude, Frequency, Phase Amplitude is ‘displacement’ (a distance) in a physical vibration and then is usually transformed to an electric current and is measured in voltage HLTHINFO 730 – Lecture 13 Slide #5

AM / FM • Can encode signals by changing (“modulating”) amplitude or frequency (or phase) of a carrier signal HLTHINFO 730 – Lecture 13 Slide #6

Basics of signal processing • Sampling frequency – Must take samples frequently enough – The Nyquist rate is twice the frequency of the highest frequency component of the signal – If there’s something higher frequency, then you’ll get aliasing – an incorrect interpretation of the signal HLTHINFO 730 – Lecture 13 Slide #7

Sampling in ECG • In ECG we have a lot of concern with interval lengths – Equipment commonly samples at 100 Hz (mobile devices) to 1000 Hz (high resolution) – At 100 Hz, due to the Nyquist rate, you miss any high-frequency features with a period of less than 0. 02 s (i. e. , 20 ms) (Period = 1 / frequency) – Moreover, at 100 Hz, you can be up to 10 ms late in seeing a rise or fall, and thus up to 20 ms inaccurate in estimate of an interval • Sampling requirements (now talking ECG or other apps) put demands on – the speed of your equipment to process – the bandwidth of your transmission (esp. in telemonitoring) – the size of your database (esp. for long-term monitoring) HLTHINFO 730 – Lecture 13 Slide #8

Signal classification • Algorithms can classify signals based on features of the signal – Might be straightforward (e. g. , time between lowest and highest amplitude – but keep in mind all those sampling errors!) – Signal can be mathematically transformed • Fourier transform – transforms from amplitude over time -> amplitude over frequency • We can then extract features from the transformed signal • Classifiers can then use whatever machine learning methods – Multiple regression, artificial neural networks, induced decision trees, etc. – Can classify the ‘system’ (e. g. , the patient’s heart) as being in any of a variety of states – And you can layer symbolic reasoning (production rules) and fuzzy logic on top of the signal-feature-based classifiers HLTHINFO 730 – Lecture 13 Slide #9

Fourier transform results • A sine wave is the pure ‘spike’ once Fourier transformed Time domain Frequency domain • Square waves and pulses make more complex patterns HLTHINFO 730 – Lecture 13 Slide #10

Markov model • Based on the ‘memoryless’ (or Markov) property (“M” either way!) – Your previous states say nothing; only need to think about current state and probability/rate of progression to other states from there e. g. , P(Bt+1 | At) = 0. 9 Can describe the system with a square matrix, Nx. N, where N is the number of states Again, only accurate if the system is memoryless with respect to those states Can use a series of low probability transitions to indicate that the system has changed (and throw an alert) HLTHINFO 730 – Lecture 13 Slide #11

Applications • ICU (esp. PICU) monitoring – Respiration, blood glucose, etc. – classify and alert on changes • Worn heart monitors – http: //www. nlm. nih. gov/medlineplus/news/fullstory_64123. html – Also, worn accelerometers for falls detection • ‘Smart’ homes – Monitor usage patterns of lights, water, refrigerator etc. and also track motion HLTHINFO 730 – Lecture 13 Slide #12

HLTHINFO 730 – Lecture 13 Slide #13

Discussion • Have you experienced any good (or not so good) automated monitors? HLTHINFO 730 – Lecture 13 Slide #14

Participatory Home Telemedcare • Home ECG, lung function, blood oxygen saturation, glucose, weight, BP • All with feedback so patient sees their state and their progress • Can, for instance, learn to deal with an asthma attack (possibly on phone to nurse) without called ambulance HLTHINFO 730 – Lecture 13 Slide #15

Reminders, life coaches • STOMP – txt messaging to quite smoking – “chewing gum for the fingers” – automated ‘friend’ to txt when craving – Plus staged supportive messages and monitoring • Significant quit effect (Maori and non-Maori at 6 months • Other obvious apps are exercise coaches, drug administration reminders and (esp. w. video phones) guides (e. g. , for insulin dosing or nebulizer spacer technique) HLTHINFO 730 – Lecture 13 Slide #16

What is a ‘care plan’ anyway? • Fundamental to monitoring or health promotion should be the notion of the care plan for a patient – What are our objectives (specified as goals and target values)? – What interventions do we have in place to achieve those objectives? – How often do we monitor status? – When do we plan to re-plan? HLTHINFO 730 – Lecture 13 Slide #17

Care plan model • We’ve created an information model for care plans (Khambati, Warren, Grundy and Hosking) HLTHINFO 730 – Lecture 13 Slide #18

Model (contd. ) HLTHINFO 730 – Lecture 13 Slide #19

Designing a care plan in the model HLTHINFO 730 – Lecture 13 Slide #20

Care plan in the model (contd. ) HLTHINFO 730 – Lecture 13 Slide #21

Automated interface generation • We’ve prototyped a process for generating multiple user interface implementations for an individual care plan around the care plan model HLTHINFO 730 – Lecture 13 Slide #22

Example interfaces • Part of a diabetes monitoring care plan being tailored in our care plan instantiation application HLTHINFO 730 – Lecture 13 Slide #23

Example interfaces Auto-generated interfaces are still a bit basic, but better than nothing • End-user Flash application compiled from Open. Laszlo HLTHINFO 730 – Lecture 13 Slide #24

“Your plastic pal that’s fun to be with” • Healthcare robots (or healthbots) are being considered to supplement human personnel – Particularly in low-intensity monitoring situations such as aged care – ‘Robot’ is from a Czech word for ‘to work’ • But many practical robots are actually more focused on being mobile sensor platforms and computer terminals • Real work robots are possible when fixed to an automotive assembly line, but not yet practical for dealing with people • Which doesn’t mean the Japanese aren’t trying… HLTHINFO 730 – Lecture 13 Slide #25

Robots that can lift and carry • Japanese RI-MAN (incidentally, that’s a doll it’s lifting) – still highly experimental HLTHINFO 730 – Lecture 13 Slide #26

Tele-presence healthbot • Much more common … and further along toward real-world use HLTHINFO 730 – Lecture 13 Slide #27

Robots for companionship • Gladys Moore, a resident at the NHC Healthcare assisted-living facility in Maryland Heights, Missouri, plays with AIBO, a robotic dog, in this undated handout photo. Researchers found that the robot dog was about as good as a real dog at easing the loneliness of nursing home residents in a study. HLTHINFO 730 – Lecture 13 Slide #28

Uo. A Health Robotics Centre • Working with ETRI (Korean Robotics Institute) – Looking at adapting an inexpensive robot for elder care – Combination of companionship and monitoring capabilities – Strong emphasis on speech interaction – More autonomous adjunct to human healthcare workers, rather than for tele-presence – Possibly supplement other smart home equipment Ultrasonic sensors to avoid bumping into HLTHINFO 730 – Lecture 13 Slide #29 things

Summary • Monitoring is a major class of health IT activity • It leads to the embedding of sometimes non-trivial artificial intelligence in devices (often with reliance on traditional signal processing) • Monitors may be overt or ubiquitous • They may engage the consumer – In fact, engaging the consumer may be the main point! • Monitoring implies the knowledge engineering of guidelines HLTHINFO 730 – Lecture 13 Slide #30