Transcript Re-engineering Computational Research to Improve Medical Care

Re-engineering Computational
Research to Improve Medical Care
Peter Szolovits
Prof. of EECS & HST
CSAIL
September 24, 2003
How to Help Stop Screw-ups in
Medical Care
Re-engineering Computational
Research to Improve Medical Care
Peter Szolovits
Prof. of EECS & HST
CSAIL
September 23, 2003
How to Help Stop Screw-ups in
Medical Care
Re-engineering Computational
Research to Improve Medical Care
Peter Szolovits
Prof. of EECS & HST
CSAIL
September 23, 2003
What to
do when
success
fails
Outline
•
•
•
•
Medical Informatics vision 30 years ago
AI Contributions
Lack of impact
Current medical hot topic: quality
improvement
• New needs/research opportunities
“Medicine and the Computer:
The Promise and Problems of Change”
--W. B. Schwartz, NEJM 1970
– Ever-expanding body of
knowledge, limited memory
– Physician shortage and
maldistribution
• Computer as an “intellectual”, “deductive” tool
– Improve medical care: 2nd opinion, error monitor
– Separate practice from memorization
– Allow time for human contact; different personalities in
medicine — the “healing arts”
Practice of Medicine is …
• Art
– Learning by apprenticeship
– Individual variation & creativity
• Science
– Baconian “hypothetico-deductive reasoning”
• Engineering
– Systems to reduce failure, optimize care
Consider the following:
• Middle-aged woman complains of severe pedal
edema (foot swelling), which is neither painful or
erythematous (red), symmetric (both feet),
pitting, lasting for weeks.
• She drinks heavily, has jaundice, painful
hepatomegaly (enlarged liver), …
• … 50 other facts from lab, physical exam, etc.
• Conclusions: Cirrhosis, hepatitis and portal
hypertension; possible constrictive pericarditis
Reasoning Tasks
•
•
•
•
Diagnosis
Prognosis
Therapy
…Management
observe
data
decide
patient
therapy
initial presentation
information
plan
diagnosis
Medicine provided challenges for
AI, and AI responded
• Probabilities
Bayes nets, qualitative probabilistic networks, partiallyobservable semi-Markov decision processes, …
• Temporal patterns and uncertainty
Temporal belief nets, temporal constraints, …
• Spatial localization
 {vision, not reasoning}
• Causality, physiology and pathophysiology
Feedback models, multi-level models, …
• Combinatorial explosion of hypotheses
Symptom clustering, theories of abduction
• Modularity
 Rule-based systems, …
Diagnostic
Reconstruction
weak
heart
fluid therapy
low
cardiac
output
heart
failure
digitalis
effect
Long, Reasoning about State
from Causation and Time in a
Medical Domain, AAAI 83
norm
high
10
9
8
7
retain
high
water
blood
volume
lose
low
definite cause
possible cause
possible correction
(not all shown)
present
? norm
low
?
loss
low
present
present
present
present
retain
past
diuretic
effect
edema
6 5 4
3
2 1
?
edema
blood volume
water
cardiac output
heart failure
weak heart
diuretic effect
diuretic
0
now
future
So why aren’t computers in your
medical life today?
• 7-minute doctor’s visit
– We forgot about $$$, workflow, usability,
technophobia, …
• Medical records still primitive
– We forgot about needing data…
• Paper, thus inaccessible
• English text, thus incomprehensible
• Unsuccessful investments in health IT
– We don’t know how to turn quality$
Current Challenges/
Opportunities
• 44-98,000/year die in
hospitals from medical errors,
at least ½ preventable (IOM)
• Cost of health care growing
without bounds
– GM spends more on health
than steel
• Aging population  chronic
health care
IOM “To Err is Human” report
• NY state (30,000 cases) and Colorado/Utah
(15,000 cases) studies of randomly selected
hospital discharges: Adverse events occur in
2.9-3.7% of hospitalizations
–
–
–
–
–
50% minor, temporary injuries
7-14% result in death
2.6% result in permanent disabling injury
53-58% preventable
28% due to negligence (failed to meet reasonable
standard of care)
Problems
Process Errors
• Majority of errors do not result from individual
recklessness, but from flaws in health system
organization (or lack of organization).
• Failures of information management are common:
–
–
–
–
illegible writing in medical records
lack of integration of clinical information systems
inaccessibility of records
lack of automated allergy and drug interaction
checking
Suboptimal performance
everywhere
% of ideal candidates who received Rx for AMI by hospital type
Intervention
Community Academic
ASA
80%
90%
ACE
58%
62%
Beta Blockers
36%
48%
Reperfusion
55%
60%
JAMA, Sept 2000
Why?
• In the absence of facts, opinion prevails
(85% of healthcare)
- T. Clemmer, M.D.
• “A Thousand Doctors, A Thousand Opinions”
- French proverb
• “We practice healthcare as if we never wrote anything
down. It is a spectacle of fragmented intention.”
- L. Weed, M.D.
• Healthcare is labor intensive and information bereft
- B. Hochstadt, M.D.
• “Until clinician’s are paid by the word and not by the
procedure, medical records will remain unsupported,
unmanageable and of limited value.”
- I. Kohane, MD, PhD
Computerized Clinical Decision
Support
• Reference
– Bates DW et al. A randomized trial of computer-based
intervention to reduce utilization of redundant laboratory tests.
Am J Med 1999 Feb;106(2):144-50
• Aim
– To determine the impact of giving physicians computerized
reminders about apparently redundant laboratory tests.
• Methods
– Randomized trial of giving physicians immediate feedback
upon ordering of tests via computer order entry system vs. no
feedback
Computerized Clinical Decision Support:
necessary but not sufficient
to overcome opinion
• Results
– 939 apparently redundant lab tests among 77,609
ordered on 5700 intervention Pts and 5886 control
Pts.
– In intervention group, 300 of 437 tests (69%) were
cancelled in response to alerts. Of 137 overrides, only
41% justified on chart review.
Nevertheless:
– In control group, 51% of ordered redundant tests were
performed vs. 27% in intervention group. (P<.001)
Short-term solutions
If computers can capture even some of what goes
on, they can help avoid errors, assure consistency:
“One-rule” expert systems:
– If you’re about to prescribe a lethal dose of medicine,
don’t!
Guidelines: routine methods for routine care
– E.g., remember x-ray after appendectomy
– Ready surgical team when doing balloon angioplasty
Workflow integration
– E.g., persistent paging for critical situation
The communication space
• is the largest part of the health system’s
information space
• contains a substantial proportion of the
health system information ‘pathology’
• is largely ignored in our informatics
thinking
• is where most data is acquired and
presented
How big is the
communication space?
• Covell et al. (1985): 50% info requests are
to colleagues, 26% personal notes
• Tang et al (1996): talk is 60% in clinic
• Coiera and Tombs (1996,1998): 100% of
non-patient record information
• Safran et al. (1998): ~50% face to face,
EMR ~10%, e/v-mail and paper remainder
What happens in the
communication space?
• Wilson et al. (1995): communication errors
commonest cause of in-hospital
disability/death in 14,000 patient series
• Bhasale et al. (1998): contributes to ~50%
adverse events in primary care
• Coiera and Tombs (1998): interrupt-driven
workplace, poor systems and poor
practice
ER communication study
• Medical Subject #4
– 3 hrs 15 min observation
– 86% time in ‘talk’
– 31% time taken up with 28 interruptions
– 25% multi-tasking with 2 or more
conversations
– 87 % face to face, phone, pager
– 13 % computer, forms, patient notes
Implications (Coiera)
• Clinicians already seem to receive too
many messages resulting in:
– interruption of tasks
– fragmentation of time, potentially leading to
inefficiency
– potential for forgetting, resulting in errors
Communication options
• We can introduce new:
– Channels eg v-mail
– Types of message eg alert
– Communication policies eg prohibit sending an email organisation-wide
– Communication services eg role-based call
forwarding
– Agents creating or receiving messages eg web-bots
for info retrieval
– Common ground between agents eg train team
members
Communication channels
• Synchronous:
– face to face, pager, phone
– generate an interrupt to receiver
• Asynchronous:
– post-it notes, e-mail, v-mail
– receiver elects moment to read
Hijacking Administrative Computing
• Referrals and Authorization – major pain
NEHEN
Membership,
Oct. 2001
Contract
Affiliates
Non-Member
Payers with
Secondary
Connectivity
Solutions
 BC/BS of
Massachusetts
Additional Members
 Massachusetts
Medicaid
 Medicare
The New England Healthcare EDI Network (NEHEN LLC) is a consortium of
payers and providers in Massachusetts.
Oct.
1997
Feb.
1998
Initial
discussions
Apr.
1998
Oct.
1998
Pilot
commences
Commitment
in
principle
Nov.
1999
Incorporation
as
NEHEN LLC
Eligibility live
at founding
members
• Current membership
represents
–
–
–
–
Dec.
1999
40 Hospitals
Over 7,500 licensed beds
Over 5,000 affiliated physicians
~2 million covered lives
(not including Medicare and
Medicaid)
Feb.
2000
Jun. Jul.
2000 2000
Seventh
and eighth
members join
Sixth
member
joins
Two
affiliates
join
Specialty
referrals
live
Jan.
2001
Apr.
2001
Ninth
and tenth
members join
Claim status
inquiry pilot
commences
Summer
2001
Sep.
2001
Members
12-14 join
Eleventh
member
joins
Referral
auth and
inquiry pilot
• Expanding membership
interest
– Additional integrated delivery
networks
– Smaller payers
– Smaller community/specialty
hospitals
– Multi-specialty practices and
their business partners (i.e.,
third-party billing companies,
practice management software
vendors)
– State agencies and task forces
NEHENlite and Integrated
Options
Intranet version –
NEHENLite
– Use when integrated EDI is
unavailable in core system
– Supports ad hoc business
processes like collections
– Provides means of acquiring early
experience with process change (in
parallel with core system
integration)
– Extends functionality to outlying
practices and business processing
areas
Integrated version –
IDX, Meditech, Eclipsys,
others
– Preferred method for workflow
improvement in core business
processes
– Avoids double-keying / re-keying
– Eases distribution and reduces
training requirements for registration
clerks, billing clerks, etc.
Interactive submission
and review
– Eligibility
•
At point of registration or scheduling (or
both)
Batch submission and
review
– Eligibility
•
– Referral Submission
•
•
Complete online form rather than paper
form and submit directly to plan
Response usually not required real-time
(can be asynchronous)
•
Submit all appointments scheduled
for the next day and “work” the 2030% of problem cases (patient not
found, wrong date of birth, patient
inactive, etc.)
Can be used in conjunction with and
in addition to real-time request at
point of registration or scheduling
(i.e., no-cost double-checking)
Real-Time and Batch
Alternatives
– Claim Status Inquiry
•
Efficiency tool for billing and collections
– Claim Status Inquiry
•
Submit inquiries for all claims more
than 10 days old and review the
results
NEHENLite – Specialty
Referral Submission
NEHENLite – Claim Status
Inquiry
nMesh
• Add clinical details to referral transactions
• Integrate with patient’s own records
• Research foci:
– Scale
– Confidentiality
– Usability
Current
Opportunities
• Involve the patient
– Most concerned, knowledgeable, representative,
motivated, and inexpensive
• Life-long active personalized secure health
information system (Guardian Angel)
– Persistent over lifetime (PING project)
– communication channel among patient, provider,
community
– expert guidance, education
• Home health
– Non-intrusive “intensive care”
DCCT: Diabetes Control and
Complications Trial (’83-93)
• Lowering blood glucose reduces risk:
– Eye disease: 76% reduced risk
– Kidney disease: 50% reduced risk
– Nerve disease: 60% reduced risk
• Elements of Intensive Management in the DCCT
– Testing blood glucose levels 4 or more times a day
– Four daily insulin injections or use of an insulin pump
– Adjustment of insulin doses according to food intake and
exercise
– A diet and exercise plan
– Monthly visits to a health care team composed of a physician,
nurse educator, dietitian, and behavioral therapist.
New England Journal of Medicine, 329(14), September 30, 1993.
Home Care for Chronic Illness
• Who else?
• Treatment titration
– E.g., heart disease, renal dialysis
• Compliance nagging
• Instrumentation: “walking ICU”
Long-term
• Genomic Medicine
– Human phenome project to learn clinical
correlates of gene expression
– Customized interventions/drugs
– Customized decision making
• But, how to get the clinical data?
Clinical data
Autonomous Witness
• Natural language (and speech)
understanding
• Knowledge representation standards
for what is understood
• Perceptually aware systems
– See, hear, record and present data
– “Real” autonomous health agent
• Don’t forget communication!
Automated messages
• Notification - that an event has occurred:
– Alert (push)- draws attention to an event
determined to be important eg abnormal test
result, failure to act
– Retrieve (pull) - return with requested data
– Acknowledgment (push or pull) - that a
request has been seen, read, or acted upon
Effects of notification systems
• Channel effect: shift existing events from
synchronous to asynchronous domain,
reducing interruption
• Message effect: generate new types of
events in the asynchronous domain,
increasing message load, demanding
time, and creating a filtering problem
• potential to either harm or help
Interpretation 1 - communication is
replaceable
• Problem is size and nature of communication
space i.e. need to shift to formal information
transactions
• Implies a 1:1 hypothesis i.e. communication
tasks replaceable with computational tasks
• Strong hypothesis (100% replacement) a matter
of debate
Interpretation 2 - the necessity of
communication
• Size of communication space is natural and
appropriate
• Communication tasks are ‘different’
• Reflects informal and interactive nature of most
conversations
• Problem lies with the way we support those
tasks, either ignoring them or shoe-horning them
into formal IT solutions
Choosing Channels
• Highly grounded conversations need
– low bandwidth
– frequent small updates
• Poorly grounded conversations need
– high bandwidth
– prolonged initial priming exchange
• Building common ground should be
specifically supported e.g. shared
information objects, images, designs
– Students & Colleagues
Thanks
• Esp. Zak Kohane
– Collaborators
•
•
•
•
Finally, back to
the fun:
reasoning!
http://medg.lcs.mit.edu
Children’s Hosp.
Tufts/NEMC
Harvard Med
BU