Why did my patient develop
acute kidney injury?
Using machine learning to understand
peri-operative health outcomes
Karandeep Singh, MD, MMSc
Assistant Professor of Learning Health Sciences, Internal Medicine, Urology, and Information
University of Michigan
Disclosures
• No relevant disclosures or conflicts of interest
To answer this question, we will:
• Describe the role of risk models within learning health systems
• Discuss ways that risk can be quantified
• Understand how risk models are operationalized to improve quality of care
• Discuss issues with attribution for peri-operative outcomes
• Demonstrate how machine learning “model explanations” can be used to
explain contributions of different factors (e.g., vital signs, surgeon) to risk in
individual cases
My background
• Nephrologist
• Masters in biomedical informatics
• Research lab focuses on Machine Learning for Learning Health
Systems
• Member of the MUSIC Collaborative Quality Initiative (urology)
• Co-chair of Clinical Intelligence Committee at U-M health
system
• Member of Michigan Artificial Intelligence Advisory Board
convened by MEDC and the Center for Automotive Research
What is a learning health system?
• Any system focused on:
• Improving people's health
• Through continuous cycles of knowledge
discovery and implementation of best
practices
• And doing this at scale
Infrastructure
Knowledge
to Practice
Practice to Data
Data to
Knowledge
Friedman C, Rubin J, Brown J, Buntin M, Corn M, Etheredge L, Gunter C, Musen M, Platt R,
Stead W, Sullivan K, Van Houweling D: Toward a science of learning systems: a research agenda
for the high-functioning Learning Health System. J. Am. Med. Informatics Assoc. 43–50, 2014
What does learning health system
require?
• A community engaged in continuous
learning through repeated learning
cycles
• Learning cycles
• Practice to Data
• Data to Knowledge
• Knowledge to Practice
• Infrastructure: IT, governance, ethics,
policy
Infrastructure
Knowledge
to Practice
Practice to Data
Data to
Knowledge
What is a learning health system?
• The Collaborative Quality Initiatives are
exemplars of learning health systems
• A community engaged in continuous
learning that crosses organizational
boundaries
• Rich data collection through manual and
automated mechanisms
• Ability to learn from multi-institutional data
• Ability to implement interventions at scale
ASPIRE
What is the role of risk models in
learning health systems?
At a patient level
• Select the best diagnostic or
treatment approach
• Example: Mallampati score
• Counsel patients on prognosis
At a system level
• Allocate resources more
efficiently
• Example: Michigan Medicine birthing
center delivery volume model
• Early warning systems
• Example: Michigan Medicine sepsis pilot
• Identify areas of improvement
• Example: comparing observed versus
expected risk
How can we quantify risk?
• Decision tree models
• Regression models
• Machine learning models
How can we quantify risk?
• Decision tree models
Decision tree models
Advantages
• Easy to interpret and implement
• Can model simple interactions between predictor variables
Disadvantages
• Rarely the most accurate
How can we quantify risk?
• Regression models
Regression models
Advantages
• Easy to interpret
• Not too difficult to implement
• Sometimes the most accurate
Disadvantages
• Will not capture non-linear relationships unless you use
“polynomials” or “splines”
• Will not capture interactions unless explicitly included
How can we quantify risk?
• Machine learning models
• Refers to many types of algorithms
Random forest or gradient
boosted decision trees
Support vector machines
Neural networks (or deep learning
if multiple hidden layers)
Machine learning models
Advantages
• Often the most accurate
• Can capture non-linear relationships
• Can capture interactions between variables
• Neural networks can handle imaging and signal waveform data
Disadvantages
• Difficult to implement (and share)
• May “overfit” the training data if not carefully trained
How can risk models be operationalized to
improve quality of care?
• At the point of care
• Shared decision-making tools or decision aides
• Clinical decision support alerts
• Changing default selections for order sets
• Population health management
• Run the model at a fixed interval
• Use it to:
• Find patients who need immediate attention (early warning systems)
• Allocate resources efficiently by prioritizing the “sickest” or “modifiable risk”
patients
• Identify areas where observed outcomes worse than expected risk
Michigan Urological Surgery Improvement Collaborative
A consortium of urologists and urology practices
throughout the State that aims to improve the quality
and cost-efficiency of urologic care provided to
patients in Michigan
Our Goal: Make Michigan #1 in Urologic Care
Example of shared decision-making:
askMUSIC
Random forest
model
Example of clinical decision support
ALERT
Sorry, I know you are giving a talk right
now but your version of Windows is slightly
out of date. I went ahead and installed the
latest version but now your computer
needs to be restart. Press OK to restart.
OK Cancel
• Will not show an
example here to
minimize further
alert fatigue
Changing default selections for order sets
Olson J, Hollenbeak C, Donaldson K, Abendroth T, Castellani W: Default
settings of computerized physician order entry system order sets drive
ordering habits. J. Pathol. Inform. 6: 16, 2015.
Changing default selections for order sets
Olson J, Hollenbeak C, Donaldson K, Abendroth T, Castellani W: Default
settings of computerized physician order entry system order sets drive
ordering habits. J. Pathol. Inform. 6: 16, 2015.
Example of population health mgmt:
State Innovation Model
• Use a risk model to identify highest emergency department
utilizers over the next 6 months in Livingston and Washtenaw
Counties
• Run this model every 2 months
• Update the model with new data every 2 months
• Enroll highest risk patients in care coordination intervention
Example of population health mgmt:
State Innovation Model
Model version 1
• UM data received: 9/8/2017
• No Trinity Health data
• Created 11 predictors
• Trained on 185,597 patients
• RMSE: 0.364 visits
• Run on 188,089 patients
• Results shared on: 11/1/2017
Model version 2
• UM data received: 11/10/2017
• TH data received: 11/29/2017
• Created 56 predictors
• Trained on 303,514 patients
• RMSE: 0.511 visits
• Run on 311,962 patients
• Results shared on: 1/25/2017
Model version 3
• Awaiting next round of data
• Geocoded addresses
to latitude + longitude
Model version 4
• Add billing codes
….
Model version 11
For context:
Total population of
2 counties =
538,683 people
Example of population health mgmt:
Identify areas of improvement
• Find where observed outcomes worse than expected risk
• Why did my patient develop AKI?
Provider AKI Mechanism
Anesthesiologist
Hemodynamics, fluid management, diuretics, glycemic control
Surgeon /
Proceduralist
Direct
injury, physiologic insult, nephrotoxins
Other
Pre
-existing CKD, comorbid conditions, lifestyle factors
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
Attribution of AKI varies by case type
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
Contributors
to AKI
100%
Colectomy AAA Repair Colonoscopy Nephrectomy
Anesthesia
Surgeon
Other
Anesthesia
Surgeon
Other
Surgery type
Anesthesia
Endoscopist
Other
Risk adjustment is the “usual” way to
attribute risk
• What is it?
• Method to more accurately assess performance, accounting for baseline risk
• Why do we need it?
• Establishes basis for comparison across providers/institutions with varying
baseline risk
• Isolates component of outcome attributable to the anesthesiologist
• How does it work?
• Compares a provider’s observed performance to what was expected
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
How does risk adjustment work?
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
• Using Risk Adjustment:
– Incidence of AKI:
–Provider A 20%
–Provider B 3%
Case Type
Performed
# Cases observed
to
have AKI
Provider A
100
AAA repairs
20
Provider B
100 colonoscopies
3
How does risk adjustment work?
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
• Using Risk Adjustment:
– Incidence of AKI:
–Provider A 20%
–Provider B 3%
– Comparing provider’s observed performance to what was expected:
Case Type
Performed
# Cases observed
to
have AKI
# Cases
expected
to have AKI
Provider A
100
AAA repairs
20
50
Provider B
100 colonoscopies
3
2
How does risk adjustment work?
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
• Using Risk Adjustment:
– Incidence of AKI:
–Provider A 20%
–Provider B 3%
– Comparing provider’s observed performance to what was expected:
–Provider A 0.40
–Provider B 1.50
Case Type
Performed
# Cases observed
to
have AKI
# Cases
expected
to have AKI
Observed /
Expected
(O/E) Ratio
Provider A
100
AAA repairs
20
50
20/50 = 0.40
Provider B
100 colonoscopies
3
2
3/2 = 1.50
How is “expected” risk calculated?
• Use variables to develop a risk model considering:
• Patient characteristics
• Demographics: age, gender, BMI
• ASA status
• Comorbidities: renal insufficiency, HTN, HF, diabetes, CAD, liver disease, etc.
• Labs: hemoglobin, creatinine
• Surgical characteristics
• Procedure type (anesthesia CPT code)
• Emergent / elective
• Center characteristics
• Type of hospital
Source: Michael Mathis, MD ASPIRE talk 7/20/2018
Observed/expected ratio is what
matters
Provider 1
High risk patients
High risk procedures
Provider 2
Low risk patients
High risk procedures
Provider 3
High risk patients
Low risk procedures
Provider 4
Low risk patients
Low risk procedures
Median
0%
100%
O
O
O
O
E
E
E
E
O/E Ratio
AKI Rate (%)
Observed/expected ratio is what
matters
Provider 1
High risk patients
High risk procedures
Provider 2
Low risk patients
High risk procedures
Provider 3
High risk patients
Low risk procedures
Provider 4
Low risk patients
Low risk procedures
Median
0%
100%
Provider 1
Provider 2
Provider 3
Provider 4
O/E Ratio
1.0
0
2.0
O
O
O
O
E
E
E
E
O/E
O/E
O/E
O/E
O/E Ratio
AKI Rate (%)
Why not settle for risk adjustment?
• Risk adjustment is useful at an anesthesiologist level
• Helpful for identifying “underperforming” anesthesiologists
• Not helpful for attributing risk at a case level
• If you want to identify which AKI outcomes were preventable
from the perspective of the anesthesiologist
• Cannot do it with risk adjustment alone
Why did my patient develop AKI?
• To answer this question, we need to know the
contributions of each factor to the patient’s risk
Can we generate this for each surgery?
Yes, we can.
If we know which cases have a large
anesthesiologist contribution to AKI risk,
We can focus our QI efforts and learning on
those patients.
Contributors
to AKI
My patient
Anesthesia
Surgeon
Other
• First, we need to measure AKI risk
factors that are cleanly attributable
to a provider
• Ideally, some should be modifiable
Variables
• Anesthesiologist
• fluids administered, intraoperative BP,
electrolytes (modifiable)
• Surgeon
• skill, prior rate of AKI, whether an artery
was lacerated (operative report)
• Patient factors
• Comorbidities, presence and severity of
CKD
• Other
• surgery type
How do we attribute risk at a case
level?
• Next, we need to measure contribution of each variable to the
case risk
• In a linear model, this is straightforward
• Let’s say AKI risk is a number
• If AKI risk = 0.5 x anesthesiologist’s years of experience + 0.1 x
surgeon’s years of experience, then:
• For anesthesiologist with 10 years experience and surgeon with 10 years of
experience
• Anesthesiologist’s contribution is 5/(5+1) or 83%, surgeon’s contribution 17%
How do we attribute risk at a case
level?
• Case-level attribution is more difficult for machine learning
models
• The contribution of each variable is not fixed – it depends on the
other variables for that case
• Why does this matter?
• Machine learning models often outperform regression models for
modeling risk
• What good are they if we can’t attribute risk for individual cases?
How do we attribute risk at a case
level?
• These methods generate “model explanations”
• Shapley values
• Locally interpretable model explanations (LIME)
Machine learning can attribute risk
using 2 recently described methods
Shapley values
• A way to calculate how much to pay each person
when people are working together to earn money
• Originally described in 1953 by Lloyd Shapley
• Shapley won the Nobel prize for economics in 2012
• Shapley values rediscovered by machine learning
community in 2013
• Variables are “working together” to generate prediction
Machine learning can attribute risk
using 2 recently described methods
LIME
• A less computationally intensive way of generating
concise “model explanations” for an individual data
point in a dataset
Machine learning can attribute risk
using 2 recently described methods
Shapley and LIME in action
• Let’s use these methods on MUSIC data
• We will fit a random forest model to predict 3-month continence
following prostatectomy
• For a handful of cases, we will take a look at the contributions of
individual predictors
• Think about whether the cases are worth investigating (from a
surgeon’s perspective)
Example: Why did my patient develop
incontinence following prostatectomy?
• Average risk of incontinence
at 3 months for all patients is
34%
• Model predicts 37% chance
of incontinence
• Patient is continent at 3
months
• Is this worth investigating?
0 = average risk
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 16% chance
of incontinence
• Patient is continent at 3
months
• Is this worth investigating?
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 30% chance
of incontinence
• Patient is incontinent at 3
months
• Is this worth investigating?
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 30% chance
of incontinence
• Patient is incontinent at 3
months
• Is this worth investigating?
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 52% chance
of incontinence
• Patient is incontinent at 3
months
• Is this worth investigating?
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 70% chance
of incontinence
• Patient is incontinent at 3
months
• Is this worth investigating?
Example: Why did my patient develop
incontinence following prostatectomy?
• Model predicts 70% chance
of incontinence
• Patient is incontinent at 3
months
• Is this worth investigating?
Cases are worth investigating when…
• Model gets it wrong
• Surgeon or practice is primary contributing factor to high risk
One last word about Shapley vs. LIME
Shapley is gaining in popularity
• In the US finance sector and in the UK more broadly, the law
requires that any model with financial implications has to be
interpretable
Shapley values have an “additive property”
• If you can determine which variables are anesthesiologist-
related, you can add up their contributions to get the
anesthesiologist contribution
• Cannot do this with LIME
Where do we go from here?
• There are machine learning models of
peri-operative AKI described in the
literature but not openly available
• I am currently participating in a VA
contract where we are developing an
in-hospital AKI early warning system
(PI: Michael Heung, MD)
Where do we go from here?
• Need to define which variables are attributable to the
anesthesiologist versus other factors
• Should operative reports be included in the model?
• Either extracted variables or the whole report as a sequence of text
• Hypotension wouldn’t necessarily be attributed to the anesthesiologist
in the setting of major surgical bleeding
• Potential for collaboration with surgical CQIs
