Summary talk of the underlying philosophy, guiding principles, targeted behavior change products, and process of agile science for creating, optimizing, repurposing, and curating tools and evidence.
1. @ehekler, ehekler@gmail.com www.agilescience.org keynote @ #ISRII2017 11
Eric Hekler, PhD
Associate Professor, Arizona State University
Associate Professor, University of California, San Diego
(Dec 2017 onward)
2. Thank you!
2
@ehekler
• Predrag (Pedja) Klasnja, John Harlow,
Elizabeth Korinek, Sayali Phatak, Bill
Riley, Daniel Rivera, Mathew Buman,
Kevin Patrick, Bob Evans, Cesar Martin,
Jennifer Huberty, Marios Hadijamichael
• Linda Collins & MOST
• The Robert Wood Johnson Foundation
• DISCLAIMER: I am a scientific advisor to:
eEcoSphere, Proof Pilot, & Omada Health,
HopeLab, Sage Bionetworks
6. Crisis of methods
6
“Most scientists are not trained
today on the basics of
epistemology or logic… We
need to go back to work on the
basics.”
-Dr. Arturo Casadevall
Johns Hopkins Bloomberg School
of Public Health
7. @ehekler 77
Supply & demand problem
Evidence
Demanded
Evidence
Supplied
Yes No
Yes Success Problem
No Need Success
McNie, Parris, Sarewitz, 2016, Research Policy
8. Patients: What do I do now?
8
https://pixabay.com/p-690128/?no_redirect
15. User’s needed evidence (based on the question)
@ehekler 15
Sub-Group IdiosyncraticContext-bound
Universal
Sub-
Group
Evidence generated (based on study designs used)
Context-
Bound*
Idiosyncratic*
Universal
*Largely considered “noise”
16. What is causality? How do we infer it?
• Cause preceded effect
• Cause related to effect
• No alternative explanations
17. Can a cause/effect occur without a human?
https://commons.wikimedia.org/wiki/File%3AIf_a_tree_falls_in_the_forest.jpg
21. Pre-conditions
• When, where, for whom,
and in what state will a
given intervention produce
the desired outcome.
21
@ehekler
Hekler et al. 2016, AJPM
26. Complex interventions vs. modules
From perfect “packages”
Flickr - Paul Swansen=
To repurposable pieces
Flickr - Benjamin Esham
@ehekler
www.agilescience.org
30. Proximal outcomes of the module
Shortest timescale for measuring a meaningful effect
@ehekler
Walk within
30min of
prompt
Prompt
to Walk
Steps/
Day
National
Guidelines
(PA/WK)
Cardiovascular
Fitness (vO2)
CVD
Proximal Outcomes
(often skipped/ignored)
www.agilescience.org
34. Tools of personalization
“Learning” adjustments when previous evidence does not match
@ehekler
Eng. (Adaptive Control) CS (e.g., reinforcement learning)
35. Tools of personalization
“Learning” adjustments when previous evidence does not match
@ehekler
Measure
success
towards
goal
Results
Self-experimentation
Goal +
Plan
Implement for
1 week
36. User’s needed evidence
@ehekler 36
Sub-Group IdiosyncraticContext-bound
UniversalWhat should I (or my client) do now in this context
to produce the desired outcome(s)?
37. Using evolution as a
model for the scientific
process.
Agile Science
process
37
@ehekler www.agilescience.org
49. Optimization criteria
49@ehekler Hekler et al. under review
• Initiation “Set-point”
– 10,000 steps/day, on average per week for 22 out of 26 weeks OR
– +3,000 steps/day, on average per week relative to baseline for 22 out
of 26 weeks
• Maintenance set-point
– Same steps set-point
– 0 interactions with participant, except use of wearable device
54. Modularizing
1) Cutting out pre-conditions
• When, where, for whom,
and in what state will a
given intervention produce
the desired outcome.
54
@ehekler
Hekler et al. 2016, AJPM
57. Science of matching/generalization
• Does it remain true across variations among
other people, places, times, treatments?
• Is it predictive of the future for that same
person/unit of study?
Shadish, Cook, & Campbell, 2002
60. Science of matching
60
Meaningful variations
In hormone replacement
therapy
Meaningful
Definitions of
Success
Meaningful clusters
of people, place, times
(i.e., niches)
61. Pragmatic clinical trials?
61
• Implementation science
• Scaling up and scaling out
• Connection?
– ACTS?
– Others?
– LOVE TO HEAR YOUR THOUGHTS!
63. The Human Behaviour-Change Project
A Collaborative
Award funded
by the
Participating
organisations
@HBCProject
www.humanbehaviourchange.org
Adapted from Susan Michie/slides: http://www.ucl.ac.uk/human-behaviour-change
64. Human Behavior-Change Project
computer science
information science
behavioural science
Ontology of
behaviour change
interventions
How can we
organise the
evidence?
Extracting and
interpreting the
evidence
What does the
evidence show?
Making the evidence
accessible at scale in
real time
How can we make the
evidence usable?
Adapted from Susan Michie/slides: http://www.ucl.ac.uk/human-behaviour-change
67. Using
• “The big question”
What works,
compared with what,
how well,
with what degree of exposure,
for whom,
in what settings
with what behaviours,
and why?
67Adapted from slides from Robert West; http://www.ucl.ac.uk/human-behaviour-change
68. Summary
68@ehekler
• Goal:
– Knowledge accumulation to support behavior change
• Problems
– Evidence created vs. needed
– Complex causal problem vs. simple causal philosophy
• “Building blocks”:
– Modules
– Computational models
– Decision policies
– Tools for personalization
• Activities in the process:
– Creating
– Optimizing
– Repurposing
– Curating
69. Open
questions
69@ehekler
• What does science look like when
people are different, context
matters, and things change?
• What about citizen-led science?
• What does a 21st cent. scientist do?
– Science of matching
– Empower citizens/practitioners
• How might funding look different?
Any time a person asks, “what do I do now?” and they don’t have good evidence to help them make a decision, I think of that as science not supplying the information that’s being demanded. Sadly, I think, it’s also true that scientists are creating evidence that nobody really wants.
Why does this supply and demand problem exist?
Professionals still focus on “on average” science (even, it appears, with many precision medicine efforts)
Professionals need to move towards studying the utility of personalization algorithms
Creators, users, and participants of resources into different stages of the process. All of it is driven by a decision though.
Professionals still focus on “on average” science (even, it appears, with many precision medicine efforts)
Professionals need to move towards studying the utility of personalization algorithms
Creators, users, and participants of resources into different stages of the process. All of it is driven by a decision though.
How does this help with supply and demand? Well, modularizing the evidence makes it so that any time someone’s looking for tools to, for example, help someone walk, the module tools can be used as part of there intervention. This is like the difference between Lego pieces vs. the things created from Legos, like this Volve car; intervention modules are much more likely to be repurposable, precisely because they are small and scoped.
Then, all you need are instructions, such as models and algorithsm to figure out how to package them together for potentially novel uses, which becomes possible when the evidence is more about the modules.
How does this help with supply and demand? Well, modularizing the evidence makes it so that any time someone’s looking for tools to, for example, help someone walk, the module tools can be used as part of there intervention. This is like the difference between Lego pieces vs. the things created from Legos, like this Volve car; intervention modules are much more likely to be repurposable, precisely because they are small and scoped.
Then, all you need are instructions, such as models and algorithsm to figure out how to package them together for potentially novel uses, which becomes possible when the evidence is more about the modules.
Our first secret weapon. Modules!
Modules and APIs are the backbone of the digital economy. They all have the basic form of Inputs Process and Output. For example, when you use Google Maps, you put in addresses of where you are and where you want to go, Google does it’s magic process, and out comes directions.
The cool thing is that Google Maps was built to do that scoped task very well, but not necessarily anything else. This scoping of its purpose makes it so that Google Maps can be used across the internet to help Yelp, companies, universities, and others find their way.
We think there’s great opportunity to think much more deeply about how to modularize health interventions, including behavioral interventions and disease management strategies but even more advanced tools like the components of medical devices.
To do this though, requires a different type of science that is built around modules and their use in the real-world, not multicomponent complex interventions.
How does it work?
Based on this, we need to move more into an open discussion in which we explore lots and lots of different ideas if we really want to understand which ones are best.
Sadly, science, particularly behavioral science doesn’t really have the sort of “maker” culture that would allow us. As such, a key emphasis.
Based on this, we need to move more into an open discussion in which we explore lots and lots of different ideas if we really want to understand which ones are best.
Sadly, science, particularly behavioral science doesn’t really have the sort of “maker” culture that would allow us. As such, a key emphasis.
Based on this, we need to move more into an open discussion in which we explore lots and lots of different ideas if we really want to understand which ones are best.
Sadly, science, particularly behavioral science doesn’t really have the sort of “maker” culture that would allow us. As such, a key emphasis.
Based on this, we need to move more into an open discussion in which we explore lots and lots of different ideas if we really want to understand which ones are best.
Sadly, science, particularly behavioral science doesn’t really have the sort of “maker” culture that would allow us. As such, a key emphasis.
Professionals still focus on “on average” science (even, it appears, with many precision medicine efforts)
Professionals need to move towards studying the utility of personalization algorithms
Creators, users, and participants of resources into different stages of the process. All of it is driven by a decision though.
My colleagues and I have been developing process to counteract the supply and demand problem and the resource transfer problem and, by extension, speed the pace of health sciences.
We call it Agile Science.
How does it work? Well, we’ve got three secret weapons: modules, modeling evolution as a process, and fostering co-creation and early-and-often sharing across a community.
Getting this to generalize and expand it’s reach occurs through careful curation, which, based on evolution analogy, can be considered niche expansion Just like how Facebook was first tested in Harvard, then Stanford, and what not and slowly grew out, we picture a delicate back and forth between evaluating and curating to see how broad a given “niche” is foreach modular tool and packaging of tools are to enable more rapid an thoughtful uses of tools created.
Overall, we see this process as giving us the flexibility and the resources we feasibly need for fostering a healthy ecosystem, which is essential for fostering a culture of health.
I’ve been calling this alternative process agile science, which I’ll jump into briefly here.
Thankfully, there has been great movement away from that classic pipeline and particularly the use of a randomized trial of interventions with multiple components in it, to other strategies that are more mirrored on strategies from engineering. Central to this work is a careful understanding of how to develop the evidence around the components of the intervention, with the assumption being htat the components will be more repurposable. SO, for example, Linda Collins has been pioneering the use of fractional factorial study designs to run interventions with multiple components but with a methodology that supports understanding of how the components and how they interact might function.
Blackbox modeling is the first step in the system id analyses. We used goals, points, and some of our self-report measures as inputs to predict daily steps in this procedure.
The primary interest here is to fit the data regardless of a particular structure of the model. So this is not considering the SCT model structure when conducting the analyses
Typically a trial and error process where you estimate the parameters of various structures and compare results
Minimal knowledge of the structure is used – so used an autoregressive model structure. (consistent estimation with probability of 1)
What we have been currently doing as part of the blackbox modeling is finding the best fitting model for all participants...as I mentioned earlier, there are various ways to go about this and we carried out an exhaustive search looking over every possible ARX structure (output and input lags), and this is a trial and error process...
Used all combinations of cycles for estimation and validation, and then obtained
We have been trying to find ties to the statistical methods we use in the social sciences for this process..such as checking assumptions. To try and bring structure into interpreting and choosing the best models for each participants in a way that they are also reliable. This is our first pass at this...
In choosing these models, we have looked at the best average validation fits (using roughly 50-50 estimation/validation), and cross-correlations between the inputs.
We looked at cross-correlations amongst the inputs to try and use only orthogonal signals. So we have removed those signals that were highly correlated to choose the most parsimpnious models.
We also tried to maintain inter-rater reliability by having two different individuals go over the model-choosing process.
We will be able to properly validate these models only when they enter a controller/ when we do the semi-physical modeling which uses the SCT model structure.
Orthogonal inputs
Autoregressive
A portion of the
Only 1 participant below 10% model fit, suggesting “good enough” model fit for 95% of our sample
For all combinations of cycles as esti and vali. Sets, you chose the best ARX structure (most predictive) for that combination.
Model fit per cycle (in the validation set), and then average over that.
Based on this, we need to move more into an open discussion in which we explore lots and lots of different ideas if we really want to understand which ones are best.
Sadly, science, particularly behavioral science doesn’t really have the sort of “maker” culture that would allow us. As such, a key emphasis.