CHI'16 Paper Presented by Qian Yang from Carnegie Mellon University. The presentation describes a field study investigating how to design better machine-learning-driven systems in support of better LVAD (left-ventricular assist device, the "heart pump") implant decision.
Apidays New York 2024 - The value of a flexible API Management solution for O...
Designing Machine Learning Driven Clinical Decision Support Tools
1. Qian Yang, John Zimmerman, Aaron Steinfeld
Lisa Carey, James F. Antaki
INVESTIGATING THE HEART PUMP
IMPLANT DECISION PROCESS
OPPORTUNITIES FOR
DECISION SUPPORT TOOLS TO HELP
CHI’16 Best Paper Honorable Mention
2. 2
The“HeartPump”
LVAD (leftventricular assist device),
implantable mechanical heart pump.
Adifficultend-of-lifedecision
• High-risk surgery and recovery
• Lifestyle change
• Critical implantwindow
The “Heart Pump” as DestinationTherapy
Source: www.mayoclinic.org
Background
3. 3
Many available decision-support tools (DSTs)
Clinicaldecisionsupporttools(DSTs):
Computersystemsthatusemedicalrecordsto
improvehealthcaredecision-making,efficiency,
patientsatisfactionandcompliance.
FunctionsDiagnosis generator, treatment
recommender, or prognosis predictor.
Outputsalerts, decisions, recommendations,
predictions, or considerations.
Example:Adecision-at-handsoftwareusing
predictivedatamining(Bellazzi&Zupan,2008)
Background
4. 4
Most fail in clinical practice
Number1reasonforfailureisthelackofHCIconsiderations
HCI literatures provide no design patterns or guidance.
Identifiedbarriers
Poor integration with clinical workflow
Not designed for collective nature of clinical work
No addedvalue perceived by clinicians
…
MissingfromLiterature
Clinician workflow
teamwork
needs
…
(Yang etal.2015)
Background
5. 5
Motivation
Emerging needs for HCI research
HCI research can help address these barriers
by integrating the richness of context
and redefining the role of DST technology in clinical practice.
(Yangetal.,2015)
6. 6
Investigate clinician decision-making in context
Scope
Clinician decision-making in advanced heart failure services of
implant hospitals, hospitals that provide heart pump implantation.
DataCollection
• 14-day field observation in 2 hospitals.
• 24 one-hour semi-structured interviews in 3 hospitals.
DataAnalysis
Affinity diagram, service blueprint.
1.
2.Identify opportunities for DSTs to help
Goals of Our Field Study
Methods
7. 7
F I N D I N G S
- Decision Landscape Overview
- Barriers for DSTs Coming to Effect
10. 10
Implantphysicians:
Decision is easy
Findings
Oral Medications
Other
Mechanical Support
HeartPump
“
”
Implant
Decision
We didn't knowwhat else to do.
Then that's the time that he gets
admitted for evaluation of LVAD.
(NursePractitioner,site3)
Intravenous
Medications
11. 11
Implantphysicians
Expressed no need for DSTs
Findings
For most cases, there arewell-established precedence.
For grey cases, physicians don’t think extra data are helpful.
Physicians do not use decision support tools.
They consult colleagues for decision support.
I can tell you who are really on the fringes.
But there is no data can guide this decision.“
”
(Cardiologist,site1)
12. 12
“
”
Three paths of patient journey
Home
Clinic/
Local
Hospital
Implant
Hospital
1)Theconsolidatedpath
(Cardiologist,site3)
He is a patient I’ve had 9 months to get know him, to
do test on, to follow… It’s hard to saywhat else Iwill
need. I had a lot of time to think through things.
Oral Medications
Other
Mechanical Support
Heart Pump
Implant
DecisionIntravenous
Medications
Implant Window
Findings
13. 13
“
”
Three paths of patient journey
Home
Clinic/
Local
Hospital
Implant
Hospital
2)emergencyroompath
(Cardiologist,site1)
We've got patients that come in here who are on
breathing tubes, and their families say go ahead.
And they wake up on a mechanical pump.
Heart Pump
Implant
Decision
Implant Window
Findings
14. 14
Implant Window
“
”
Three paths of patient journey
Home
Clinic/
Local
Hospital
Implant
Hospital
3)latereferralpath
The patient came hereveryvery sick. Hewas
progressing in the community. Didn’t get referred
here.
Heart Pump
Implant
Decision
(Cardiologist,site2)
Findings
15. 15
Decision breakdowns do not happen
whenfactoringpatientconditiontoimplantdecision
Instead, breakdowns happenwhen
• Upstream physicians missed implantwindow
• Upstream physicians delayed implant consideration
• Implant team has difficulties clarifying patients’
social and/or medical conditions
Findings
17. 17
Barriers for Decision-supportTool Adoption
Implications
• Attitudinal Barriers
• Need Barriers
• Informational Mismatch
• Environmental Barriers
18. 18
1) Embracing the Richness of Clinical Context
incl.physicalandsocialcontexts
Implications for DST design
• Need to minimize input of data due to clinicians’
frequent hand washing and lack of time spent in
front of a computer;
• Have to make an effort to reach and convince the
decision-makers
• i.e. through mid-levels or weekly meetings
19. 19
Implant Window
2) Decision Process as a Design Material
Animplantdecisioniscomposedofastringofsmallerdecisions.
Medication
Escalation
DestinationTherapy
Medication
Escalation
Treatment
Escalation
ClinicVisit
FrequencyAdjustments
Hospitalization
Implant
Decision
Medication
Adjustment
Clinic/
Local
Hospital
Implant
Hospital
Implications for DST design
20. 20
3) Blending Human and Machine Intelligence
Implications for DST design
• Support clinicians’ decisions, rather than make decisions
for them;
• Explore potentials for AI in clarifying and monitoring
patient condition as well as managing care escalations