Transcript Guest Lecture notes from Prof Jim Warren

Where to put the Intelligent
Systems in healthcare?
Jim Warren
Professor of Health Informatics
http://www.cs.auckland.an.nz/~jim/
Potential roles
• Support at the point of care
– Classic ‘decision support system’ (we prefer not to
say ‘expert system’!)
• Support for the health consumer
– As such they may or may not be a ‘patient’
• Support for ‘Population Health’
– Identify big trends, clusters and at-risk groups
Why expert systems are hard
• Any decision algorithms will have an error rate
– Sum of Type I and Type II errors, each with different
consequences
– And if you’re too cautious, you’ll cause ‘alert fatigue’
Focus on adherence
• ‘Adherence’ is the extent to which a patient
undertakes a therapeutic regimen in the
manner directed by their healthcare provider
– Sometimes called ‘compliance’
– Ideally it emerges from good communication and
‘concordance’
• Medications
process at a
typical elderfocused
community:
Lots of players
– Doctors (GP,
specialist)
prescribe
– Pharmacists
dispense
– Residents
take their
medication
as directed
– Some live as
couples
– May be
helped by
nurse
– Supported by
family and
caregivers
Let’s add a robot!
That’ll help, right?
So what about the elder-robot situation?
• High need for adherence promotion
– Elderly usually have complex medication regimens
(often over 5 long-term medications)
• High risk
– Medication misadventure is a frequent cause of
hospitalisations in the elderly
• But patient empowerment / autonomy are
critical issues
– Elderly are already struggling on these issues and now
we’re introducing a large anthropomorphic machine
into their lives
“Charlie” – started life as a vacuum cleaner, then became ‘Cafero’
Patient data
Based on extracts from the prescribing software uploaded to a robot server
(Robogen) with stored profiles of patient details and preferences
Must avoid ‘intimidation’
Take your
medication now
or you will be
EXTERMINATED!!!
Must avoid robot-as-’snitch’
Take that pill
NOW or I’m telling
all the doctors and
nurses that you’re
ignoring their
advice
Solution 1: Give options
• Create an ‘out’ for taking the medication now
– Allow defer (like a snooze alarm)
• Accept that patient doesn’t want to take this
dose
– Make turning down the offer a ‘normal’ path
• Make informing healthcare providers of a
missed dose an option
– Can sound a lot more positive that way
Solution 2: seek understanding
• Don’t pretend to know everything
– Lots of reasons the robot could be wrong when
prompting to take a medication
• Offer a pick list of a few common reasons for
non-adherence
Balancing adherence and safety
Solution 3: provide more help
• Offer information about the medications
– Currently text, but can expand to video
• Check about side-effects
– Most medications have a couple common side-effects
that account for most of the trouble
– We throw in periodic checks on these, alternating
between general and situated questions
• Also provide safety through physiological
monitoring
– Indeed I don’t think we would’ve gotten research ethics
approval to leave people alone with the robot without
this feature!
Customised questions to probe
undesirable symptoms
One of these questions would randomly appear at the end of each session
Weaving the social network
• Our goal is to empower
– Provide the prescribing physician with more detail
about the adherence challenges facing their elderly
patients
– Step the patient through taking their medications in a
knowledgeable fashion
• We believe it’s a healthy alternative to taking them out of
the loop with a dispensing machine
• Has met with good acceptance in small trials so
far
In the end we went small
iRobiQ from Yujin Robotics South Korea
Many features: e.g. 8 is a floor detection sensor
9 is a bumper sensor
iRobiQ taking a blood pressure
Quality Issues in discharge summaries
(Makes them hard to use for everybody, but especially for patients and their families!)
•
•
•
•
•
•
•
•
•
Irrelevant information
Missing results/ interpretation
Information hard to access
Information written in other section
Poor formatting
•
•
•
•
Relevant
Results
Follow
up
Advice
to
Patient
Advice
to
GP
No useful synopsis
Brief or incomplete advice
Missing information
Use of abbreviation and medical
jargon
Incomplete follow up advice
No useful information
Missing important information
Information is written in other
section
• Insufficient information
• Unclear goals
• Incorrect, incomplete and
missing advice
• Information written in other
section
IT-based Remediation Plan
SemAssist – intelligent agent on top
of summary authoring environment
Writing
Support
Interactive Decision Support
Recommend Advice to
Patient
Critique Advice to Patient
Patient
Support
Semantic Annotation
Synonym Provision
Reading
Support
Hyperlink to Explanatory
Material
SemLink – based on near-future where
patient reads discharge summary online
Writing Support
(Ontology Model)
Patient
Age, Sex
hasHighRiskDischargeMedication
Anticoagulant
Warfarin
Cardiovascular
ACE Inhibitor
• Marevan
• Warfarin
• Warfarin Sodium
Glyceryl
Trinitrate
hasPatientInformation
High Risk Discharge
Medication
Digoxin
Anti-Infective
Analgesic
Corticosteroid
Amoxicillin
Acetaminophen
Prednisone
Quinine
Ibuprofen
Morphine
Medication
Category
Layer
Patient Information
• Abnormal Urine
• Severe Headache
• Loose Stool
• Hemoptysis
• Easy Bruising
• Gum Nose Bleeding
Warfarin Patient Action
• Prescribed Dose Adherence
• Avoid Alcohol
• Avoid Salicylates
• Avoid Green Leafy
Vegetables
• Avoid Cranberry Juice
• Avoid Vitamin K Dairy
Products
• Dietary Change
• Start Medication
• Stop Medication
Follow up
hasFollowup
Warfarin Side Effect
Patient Action
hasPatientAction
hasSideEffect
Side Effect
Warfarin Follow up
• Blood Test
• INR
Medication
Information Layer
Consumer health search engines used in
Hyperlinking for reading support
Search Engine Name
Available from
MedlinePlus
http://medlineplus.gov/
Healthline
http://www.healthline.com/
WebMD
http://www.webmd.com/
Mayo Clinic
http://mayoclinic.com/health/
Patient UK
http://www.patient.co.uk/
Medsafe
http://www.medsafe.govt.nz/Consumers
Family Doctor
http://www.familydoctor.co.nz/
Yahoo! Health
http://www.health.yahoo.net/
Health.com
http://www.health.com/
MedTerms
http://www.medterms.com/
HealthEngine
http://www.healthengine.com.au/
Yahoo Kids
http://kids.yahoo.com/reference/dictionary/
MedlinePlus/medical dictionary
http://www.merriam-webster.com/medlineplus/
WordNet
http://wordnet.princeton.edu/
SemLink in Portable Document
Format (PDF)
Skin and Subcutaneous Tissue
Infection (SSTI)
• Common cause of avoidable hospitalisation in
New Zealand
– Often caused by staph and strept bacteriological
infections
– Treatable with antibiotics but delay or nonadherence can cause complications leading to
hospitalisation
• Can we detect households with recurrent
SSTIs in children based on the electronic
medical record (EMR) in General Practice?
Ontology model
SSTI identification
flowchart
• 36% of individuals age 20 or
under had recurrent SSTI
from 4 Auckland practices
serving substantial Māori
and Pacific population
• 65% of cases identified by
notes (not diagnosis code or
lab test)
– Hence, needed NLP
Conclusion
• Exciting applications of AI technology to
improve healthcare delivery
• The AI is not always delivered as an explicit
‘agent’
– May be hidden as a subtle part of the user
interface or in a detection algorithm
• Range of users and settings
– Point of care, health consumer, population health