Beyond the hype
Crossing the chasm with AI and
Patient Chronic Disease
Management
@timdaines
@quantumblack
quantumblack
#HXD2019
2All content copyright © 2017 QuantumBlack, a McKinsey company 2All content copyright © 2018 QuantumBlack, a McKinsey company
AI robots working as “co-workers” to
support core tasks of midwives
Human and Machine, not
Human Vs Machine
@timdaines @quantumblack #HXD2019
SOURCE:AZ Damiaan, 2nd March 2019
3All content copyright © 2017 QuantumBlack, a McKinsey company 3All content copyright © 2018 QuantumBlack, a McKinsey company
“Machines Treating Patients? It’s Already
Happening.”
Mabu, an AI robot that talks to heart
failure patients to help them manage
their disease
@timdaines @quantumblack #HXD2019
SOURCE: http://time.com/5556339/artificial-intelligence-
robots-medicine/
March 2019 – Alice Park – Time.com
4All content copyright © 2017 QuantumBlack, a McKinsey company 4All content copyright © 2018 QuantumBlack, a McKinsey company
Medication for large populations
Emergency room and GP own patient data
Doctor knows best
Doctor holds patient-data
Visit a hospital building or room
Big expensive medical equipment
’One Size’ fits all no longer works
@timdaines @quantumblack #HXD2019
SOURCE:www.shutterstock.com, 2nd March 2019
5All content copyright © 2017 QuantumBlack, a McKinsey company 5All content copyright © 2018 QuantumBlack, a McKinsey company
Personal to one or multiple chronic disease(s)
Machine intelligence reduced discovery costs
Patient knows their lifestyle
Patient generates their own data
Virtual visits and online monitoring
Patient owns the data
Personalising health with AI processes
SOURCE:www.shutterstock.com, 2nd March 2019
@timdaines @quantumblack #HXD2019
6All content copyright © 2017 QuantumBlack, a McKinsey company
Humans | Machine | Process
AI is very exciting technology and plays a
pivotal role for chronic disease
management ...
… the hard bit is designing for the vision
of how AI will play the role in caring for
the patient’s chronic disease, and the
interplay needed to support the
clinician’s way of working along the
patient pathway
SOURCE:www.shutterstock.com, 2nd March 2019
@timdaines @quantumblack #HXD2019
7All content copyright © 2019 QuantumBlack, a McKinsey company 7All content copyright © 2018 QuantumBlack, a McKinsey company
If we want AI in healthcare we need to design for predictive accuracy
Extending experiences beyond websites
and apps where the application of AI
needs to be explained
SOURCE:Garrett, J.J. (2002) Elements of User Experience, 2nd March 2019
@timdaines @quantumblack #HXD2019
8All content copyright © 2017 QuantumBlack, a McKinsey company
Service Plane
Connecting things in healthcare and
broadening to include….
….systems...
….platforms…
….data...
….and product-service ecologies
SOURCE:Cambridge Consultants, 2nd March 2017
@timdaines @quantumblack #HXD2019
9All content copyright © 2019 QuantumBlack, a McKinsey company 9All content copyright © 2018 QuantumBlack, a McKinsey company
Shape plane
Hunting for insights
Or
Needing alerts to interesting
insights
@timdaines @quantumblack #HXD2019
Are chronic disease patients…..
10All content copyright © 2018 QuantumBlack, a McKinsey company
SOURCE: QuantumBlack
Data Science, Engineering and Designer co-collaboration
LEGEND
QuantumBlack Data
Science Process
Business and End-User Input
Process Details
Initial modelling run
‘X’ number of iterations
Knowledge of
the process
Raw
data
Linked
raw data Data
Scientist
creates
variables/features
1,000s of
variables
500
variables
200
variables
ModellingSelect best
variables
from pool
Reduce
variables
20
variables
Business feedback
User Testing feedback
Important
independent
variables
UX Designer and
Data Scientist
Checks impact on
human behaviour,
results and iterates
Random Forest
OLSRegularisation
Forward
stepwise
GLMs
Mutual
Information
Variance
inflation
factors
Top data drivers
& quantification
of impact on
human behaviour
identified
Example Variable: ‘LAST TIME WORKED ON A DRUG DEVELOPMENT’
• User Experience Research
• First, we look at the list of employees who worked on a given batch of a
product
• For each employee, we find the most recent batch of the same product
and calculate the time passed
• The elapsed times are aggregated across all employees on the batch
User testing feedback
+
Business feedback
Panel FTE
Start End Product Batch IDTime Batch Activity
… … … … …… … …
Data
Engineer
finds where
the data lives
and builds
the service
platform
landscape
@timdaines @quantumblack #HXD2019
@timdaines
@quantumblack
quantumblack
Tim’s 101 lessons learnt
for designing AI into patient
pathways
4
#HXD2019
12All content copyright © 2019 QuantumBlack, a McKinsey company 12All content copyright © 2018 QuantumBlack, a McKinsey company
#1 Create a multi-disciplinary team to infuse a culture shift to design thinking
Bringing “old world” stakeholders who are just getting to grips with cloud, with the “new world” stakeholders flexing their AI muscles
Need to fully
understand
how AI works
to trust it
Predictive
performance in
real-life evaluation
trumps
interpretability
Patients
Regulators
Healthcare
professionals
Advocates of interpretability
Tech Engineers
Corporate
decision makers
Analytics
experts
Advocates of performance
C
h
a
s
m
@timdaines @quantumblack #HXD2019
13All content copyright © 2017 QuantumBlack, a McKinsey company
What's the right optimization model for the Human and the Machine?
@timdaines @quantumblack #HXD2019
Human Cognitive
understanding
(Interpretability)
AI and ML Complexity
(Predictive performance)
Neural
network
Support
Vector
Machines
Linear
regression
Decision
tree
Random
forest
HIGH
HIGH
LOW
14All content copyright © 2017 QuantumBlack, a McKinsey company
Designing a future adhering to medication
Interpretability discovery using eco-system mapping
Low
High
1.Chronic Disease Expertise
Medical Experts, Treatment lifestyle habits,
Medical Literature, Technology
2. Cluster the data
Diagnosis, treatment frequency, human
expertise
3. AI and Machine Learning
Learning complex non-linear
relationships between
everything.
@timdaines @quantumblack #HXD2019
SOURCE:Cambridge Consultants, 2nd January 2017
15All content copyright © 2017 QuantumBlack, a McKinsey company
What ‘jobs will be done’ by Human and Machine in this new ecosystem
@timdaines @quantumblack #HXD2019
#2 Map end-to-end experiences where explainable AI (XAI) is needed
Patient downloads
the app, which
guides them to
unbox the Insulin
port
•Patient downloads app
with HCP and leaflet
guidance
•New Smart Port and
CGM device given to
patient
Patient connects app
to Insulin port
seamlessly
•App and leaflet guide the
patient to set-up CGM
and Smart Port durables
(x4)
•Durables flash to
acknowledge connection
Patient uses
introducer needles to
insert soft cannulas
•HCP assists patient in
fitting CGM and Smart
Port consumables to
body, and explains
removal/ replacement
HCP shows patient
how to dose insulin
via port
•HCP explains role of
Smart Port
•HCP illustrates safe
dosing technique and
common pitfalls to avoid
Blood sugar level
appears in app and
web dashboard
•Patient opens app and
views blood sugar level
readings
•Proxied by interstitial
fluid readings
Patient doses
insulin via port
•Patient prepares insulin
and delivery method
(pen, needle, pump)
•Patient doses insulin (or
automated via closed-
loop)
App acknowledges
dose and confirms
accuracy
•Smart Port transmits
delivery volume to phone
in real-time
•Patient sees app
acknowledge dose and
confirm adherence/
Download and
unbox
Set-up and
connect
Fit to body Demo injection Optimizing sugar
intake (XAI)
Dose insulin Re-optimizing sugar
intake (XAI)
@timdaines @quantumblack #HXD2019
17All content copyright © 2018 QuantumBlack, a McKinsey company
Finding the fit end-to-end
1 Generalized linear mode
2 Least Absolute Shrinkage and Selection Operator
Applicability LowMediumHigh
e.g. in this case, a random forest model can achieve high
predictive power, while remaining relatively interpretable
Random Forest
Predictive Power
Scalability
Explainability (XAI)
Bayes Net Neural NetGLM SVM
@timdaines @quantumblack #HXD2019
Type of AI Model
1 2
18All content copyright © 2017 QuantumBlack, a McKinsey company
Tell Stories about what AI could enable…..
….sketch together….
……make prototypes together……
….role-play with a human ‘acting’ as the AI
with a human.
SOURCE:Jamescpai.com, 2nd March 2019
#3 Tell stories to imagine an future state where AI is not in our periphery
@timdaines @quantumblack #HXD2019
19All content copyright © 2017 QuantumBlack, a McKinsey company
Behaviour and data archetypes,
rather than personas
@timdaines @quantumblack #HXD2019
“Asthma is already a burden to
my life every single day. If you
are giving me a piece of
technology that I have to
manually connect to my cell
phone, it’s just going to add
more work.”
Female, 51 – Type II with mild visual impairment
20All content copyright © 2017 QuantumBlack, a McKinsey company
#4 Using the power of LEGO to visualise AI interpretability and performance
@timdaines @quantumblack #HXD2019
21All content copyright © 2017 QuantumBlack, a McKinsey company
Surface discussions around problems in behavioural terms
@timdaines @quantumblack #HXD2019
• What is the behaviour?
• Where does the behaviour
occur?
• Who is involved in
performing the behaviour?
Define metrics for better adherence
outcomes
Trust Transparency Accuracy
Challenges and Opportunities
Performance comes from
human + machine + process
@sg_williams
@quantumblack
23All content copyright © 2019 QuantumBlack, a McKinsey company 23All content copyright © 2018 QuantumBlack, a McKinsey company
Building trust within services powered
by explanatory artificial intelligence
(XAI)
XAI provides a transparent view of
what’s happening, they still need to be
standardized and systematically
assessed
Ethical challenges around the creation
of persuasive systems
Trust with AI
@timdaines @quantumblack #HXD2019
24All content copyright © 2019 QuantumBlack, a McKinsey company 24All content copyright © 2018 QuantumBlack, a McKinsey company
Measuring the effectiveness of AI
interventions
AI doesn’t give you a better truth
But data, advanced analytics models
and human actions give you a feedback
loop to determine behavioural changes
@timdaines @quantumblack #HXD2019
25All content copyright © 2019 QuantumBlack, a McKinsey company
Opportunity to use real-world data for drug development of chronic diseases
a
Target
identification
Target
validation
Preclinical trials
Drug to
the
market
Lead finding
(design/synthesis
of libraries and
screening)
Clinical
trials
Lead optimization
(ok
pharmacokinetic
Profile, toxicity,
mutagenicity)
Drug
Approval
SOURCE:QuantumBlack Data Science, 2nd March 2019
@timdaines @quantumblack #HXD2019
26All content copyright © 2018 QuantumBlack, a McKinsey company
Identifying what drug to make for a rare chronic disease
SOURCE: QuantumBlack
@timdaines @quantumblack #HXD2019
Thank you & We’re Hiring
@timdaines
@quantumblack
quantumblack.com/careers
careers@quantumblack.com
#HXD2019

Tim Daines, QuantumBlack

  • 1.
    Beyond the hype Crossingthe chasm with AI and Patient Chronic Disease Management @timdaines @quantumblack quantumblack #HXD2019
  • 2.
    2All content copyright© 2017 QuantumBlack, a McKinsey company 2All content copyright © 2018 QuantumBlack, a McKinsey company AI robots working as “co-workers” to support core tasks of midwives Human and Machine, not Human Vs Machine @timdaines @quantumblack #HXD2019 SOURCE:AZ Damiaan, 2nd March 2019
  • 3.
    3All content copyright© 2017 QuantumBlack, a McKinsey company 3All content copyright © 2018 QuantumBlack, a McKinsey company “Machines Treating Patients? It’s Already Happening.” Mabu, an AI robot that talks to heart failure patients to help them manage their disease @timdaines @quantumblack #HXD2019 SOURCE: http://time.com/5556339/artificial-intelligence- robots-medicine/ March 2019 – Alice Park – Time.com
  • 4.
    4All content copyright© 2017 QuantumBlack, a McKinsey company 4All content copyright © 2018 QuantumBlack, a McKinsey company Medication for large populations Emergency room and GP own patient data Doctor knows best Doctor holds patient-data Visit a hospital building or room Big expensive medical equipment ’One Size’ fits all no longer works @timdaines @quantumblack #HXD2019 SOURCE:www.shutterstock.com, 2nd March 2019
  • 5.
    5All content copyright© 2017 QuantumBlack, a McKinsey company 5All content copyright © 2018 QuantumBlack, a McKinsey company Personal to one or multiple chronic disease(s) Machine intelligence reduced discovery costs Patient knows their lifestyle Patient generates their own data Virtual visits and online monitoring Patient owns the data Personalising health with AI processes SOURCE:www.shutterstock.com, 2nd March 2019 @timdaines @quantumblack #HXD2019
  • 6.
    6All content copyright© 2017 QuantumBlack, a McKinsey company Humans | Machine | Process AI is very exciting technology and plays a pivotal role for chronic disease management ... … the hard bit is designing for the vision of how AI will play the role in caring for the patient’s chronic disease, and the interplay needed to support the clinician’s way of working along the patient pathway SOURCE:www.shutterstock.com, 2nd March 2019 @timdaines @quantumblack #HXD2019
  • 7.
    7All content copyright© 2019 QuantumBlack, a McKinsey company 7All content copyright © 2018 QuantumBlack, a McKinsey company If we want AI in healthcare we need to design for predictive accuracy Extending experiences beyond websites and apps where the application of AI needs to be explained SOURCE:Garrett, J.J. (2002) Elements of User Experience, 2nd March 2019 @timdaines @quantumblack #HXD2019
  • 8.
    8All content copyright© 2017 QuantumBlack, a McKinsey company Service Plane Connecting things in healthcare and broadening to include…. ….systems... ….platforms… ….data... ….and product-service ecologies SOURCE:Cambridge Consultants, 2nd March 2017 @timdaines @quantumblack #HXD2019
  • 9.
    9All content copyright© 2019 QuantumBlack, a McKinsey company 9All content copyright © 2018 QuantumBlack, a McKinsey company Shape plane Hunting for insights Or Needing alerts to interesting insights @timdaines @quantumblack #HXD2019 Are chronic disease patients…..
  • 10.
    10All content copyright© 2018 QuantumBlack, a McKinsey company SOURCE: QuantumBlack Data Science, Engineering and Designer co-collaboration LEGEND QuantumBlack Data Science Process Business and End-User Input Process Details Initial modelling run ‘X’ number of iterations Knowledge of the process Raw data Linked raw data Data Scientist creates variables/features 1,000s of variables 500 variables 200 variables ModellingSelect best variables from pool Reduce variables 20 variables Business feedback User Testing feedback Important independent variables UX Designer and Data Scientist Checks impact on human behaviour, results and iterates Random Forest OLSRegularisation Forward stepwise GLMs Mutual Information Variance inflation factors Top data drivers & quantification of impact on human behaviour identified Example Variable: ‘LAST TIME WORKED ON A DRUG DEVELOPMENT’ • User Experience Research • First, we look at the list of employees who worked on a given batch of a product • For each employee, we find the most recent batch of the same product and calculate the time passed • The elapsed times are aggregated across all employees on the batch User testing feedback + Business feedback Panel FTE Start End Product Batch IDTime Batch Activity … … … … …… … … Data Engineer finds where the data lives and builds the service platform landscape @timdaines @quantumblack #HXD2019
  • 11.
    @timdaines @quantumblack quantumblack Tim’s 101 lessonslearnt for designing AI into patient pathways 4 #HXD2019
  • 12.
    12All content copyright© 2019 QuantumBlack, a McKinsey company 12All content copyright © 2018 QuantumBlack, a McKinsey company #1 Create a multi-disciplinary team to infuse a culture shift to design thinking Bringing “old world” stakeholders who are just getting to grips with cloud, with the “new world” stakeholders flexing their AI muscles Need to fully understand how AI works to trust it Predictive performance in real-life evaluation trumps interpretability Patients Regulators Healthcare professionals Advocates of interpretability Tech Engineers Corporate decision makers Analytics experts Advocates of performance C h a s m @timdaines @quantumblack #HXD2019
  • 13.
    13All content copyright© 2017 QuantumBlack, a McKinsey company What's the right optimization model for the Human and the Machine? @timdaines @quantumblack #HXD2019 Human Cognitive understanding (Interpretability) AI and ML Complexity (Predictive performance) Neural network Support Vector Machines Linear regression Decision tree Random forest HIGH HIGH LOW
  • 14.
    14All content copyright© 2017 QuantumBlack, a McKinsey company Designing a future adhering to medication Interpretability discovery using eco-system mapping Low High 1.Chronic Disease Expertise Medical Experts, Treatment lifestyle habits, Medical Literature, Technology 2. Cluster the data Diagnosis, treatment frequency, human expertise 3. AI and Machine Learning Learning complex non-linear relationships between everything. @timdaines @quantumblack #HXD2019 SOURCE:Cambridge Consultants, 2nd January 2017
  • 15.
    15All content copyright© 2017 QuantumBlack, a McKinsey company What ‘jobs will be done’ by Human and Machine in this new ecosystem @timdaines @quantumblack #HXD2019
  • 16.
    #2 Map end-to-endexperiences where explainable AI (XAI) is needed Patient downloads the app, which guides them to unbox the Insulin port •Patient downloads app with HCP and leaflet guidance •New Smart Port and CGM device given to patient Patient connects app to Insulin port seamlessly •App and leaflet guide the patient to set-up CGM and Smart Port durables (x4) •Durables flash to acknowledge connection Patient uses introducer needles to insert soft cannulas •HCP assists patient in fitting CGM and Smart Port consumables to body, and explains removal/ replacement HCP shows patient how to dose insulin via port •HCP explains role of Smart Port •HCP illustrates safe dosing technique and common pitfalls to avoid Blood sugar level appears in app and web dashboard •Patient opens app and views blood sugar level readings •Proxied by interstitial fluid readings Patient doses insulin via port •Patient prepares insulin and delivery method (pen, needle, pump) •Patient doses insulin (or automated via closed- loop) App acknowledges dose and confirms accuracy •Smart Port transmits delivery volume to phone in real-time •Patient sees app acknowledge dose and confirm adherence/ Download and unbox Set-up and connect Fit to body Demo injection Optimizing sugar intake (XAI) Dose insulin Re-optimizing sugar intake (XAI) @timdaines @quantumblack #HXD2019
  • 17.
    17All content copyright© 2018 QuantumBlack, a McKinsey company Finding the fit end-to-end 1 Generalized linear mode 2 Least Absolute Shrinkage and Selection Operator Applicability LowMediumHigh e.g. in this case, a random forest model can achieve high predictive power, while remaining relatively interpretable Random Forest Predictive Power Scalability Explainability (XAI) Bayes Net Neural NetGLM SVM @timdaines @quantumblack #HXD2019 Type of AI Model 1 2
  • 18.
    18All content copyright© 2017 QuantumBlack, a McKinsey company Tell Stories about what AI could enable….. ….sketch together…. ……make prototypes together…… ….role-play with a human ‘acting’ as the AI with a human. SOURCE:Jamescpai.com, 2nd March 2019 #3 Tell stories to imagine an future state where AI is not in our periphery @timdaines @quantumblack #HXD2019
  • 19.
    19All content copyright© 2017 QuantumBlack, a McKinsey company Behaviour and data archetypes, rather than personas @timdaines @quantumblack #HXD2019 “Asthma is already a burden to my life every single day. If you are giving me a piece of technology that I have to manually connect to my cell phone, it’s just going to add more work.” Female, 51 – Type II with mild visual impairment
  • 20.
    20All content copyright© 2017 QuantumBlack, a McKinsey company #4 Using the power of LEGO to visualise AI interpretability and performance @timdaines @quantumblack #HXD2019
  • 21.
    21All content copyright© 2017 QuantumBlack, a McKinsey company Surface discussions around problems in behavioural terms @timdaines @quantumblack #HXD2019 • What is the behaviour? • Where does the behaviour occur? • Who is involved in performing the behaviour? Define metrics for better adherence outcomes Trust Transparency Accuracy
  • 22.
    Challenges and Opportunities Performancecomes from human + machine + process @sg_williams @quantumblack
  • 23.
    23All content copyright© 2019 QuantumBlack, a McKinsey company 23All content copyright © 2018 QuantumBlack, a McKinsey company Building trust within services powered by explanatory artificial intelligence (XAI) XAI provides a transparent view of what’s happening, they still need to be standardized and systematically assessed Ethical challenges around the creation of persuasive systems Trust with AI @timdaines @quantumblack #HXD2019
  • 24.
    24All content copyright© 2019 QuantumBlack, a McKinsey company 24All content copyright © 2018 QuantumBlack, a McKinsey company Measuring the effectiveness of AI interventions AI doesn’t give you a better truth But data, advanced analytics models and human actions give you a feedback loop to determine behavioural changes @timdaines @quantumblack #HXD2019
  • 25.
    25All content copyright© 2019 QuantumBlack, a McKinsey company Opportunity to use real-world data for drug development of chronic diseases a Target identification Target validation Preclinical trials Drug to the market Lead finding (design/synthesis of libraries and screening) Clinical trials Lead optimization (ok pharmacokinetic Profile, toxicity, mutagenicity) Drug Approval SOURCE:QuantumBlack Data Science, 2nd March 2019 @timdaines @quantumblack #HXD2019
  • 26.
    26All content copyright© 2018 QuantumBlack, a McKinsey company Identifying what drug to make for a rare chronic disease SOURCE: QuantumBlack @timdaines @quantumblack #HXD2019
  • 27.
    Thank you &We’re Hiring @timdaines @quantumblack quantumblack.com/careers careers@quantumblack.com #HXD2019