y’all need to mess up your predictions …and ...aetna inc. not a whit. we defy augury. there’s...
TRANSCRIPT
Aetna Inc.
DEFY AUGURY?
Henry Wei, M.D. November 2014
Aetna Inc.
Not a whit. We defy augury. There’s a special providence in the fall of a sparrow.
If it be now, ’tis not to come. If it be not to come, it will be now. If it be not now, yet it will come—the readiness is all. Since no
man of aught he leaves knows, what is ’t to leave betimes? Let be.
‐
Hamlet, Act 5, Scene 2
2
ICD-10 : W61: Contact with birds (domestic) (wild) W61.99XA Other contact with other birds, initial encounter
Aetna Inc.
Whither Accuracy?
3
Aetna Inc.
“City vs
Highway Mileage”
Aetna Inc.
We all love this pyramid
5
Conventional Application
doesn’t require precision at
boundaries:1.Use fancy models to find the
highest‐risk populations2.Nurses line them up for calls
from outer space3.“Engagement”
levels
modest
“Truncated model”
note
in footnotes of tech
documentation
Pond full of Black Swans:Invisible to claims‐driven
models, yet full of risk.
Aetna Inc.
Clinically‐Derived Inputs
6
Medical Conditions
Acuity vs. Chronicity
Complexity of Care
Psychosocial
Groupers
Consumer Behavior
e.g. ACG, ERG, DxCG, CDPS, 3M
How sick is this population? Who are the sickest?
Aetna Inc.
Clinically‐Derived Inputs
7
Medical Conditions
Acuity vs. Chronicity
Complexity of Care
Psychosocial
Groupers
Consumer Behavior
e.g. ACG, ERG, DxCG, CDPS, 3M
How sick is this population? Who are the sickest?
“Actionable Insight”
=
Prioritized lists? That’s it?
Typical “dashboard”
graph from
100s of startups and well‐
established analytics companies
Aetna Inc. 8
Physician Decision‐Making (Provider Contracting)
Aetna Inc.
Clinically‐Derived Inputs
9
Medical Conditions
Acuity vs. Chronicity
Complexity of Care
Psychosocial
Groupers
Consumer Behavior
Add a drug
Add a test
Stop a drug
Do a diagnostic workup
Ask for more data
How sick is this population? Who are the sickest?
How well‐managed are they?What intervention are needed?
What will be the change in
utilization that results?Physician Decision‐Making
Aetna Inc.
Example: Stroke Prevention
Aetna Inc.
STROKE434.11
AVOIDEDSTROKE
ATRIAL
FIBRILLATION
427.31
Aetna Inc.12
Example of Clinically‐Derived Inputs Feature Extraction for Diabetes
Diagnose & Evaluate
Prevent & Monitor for Complications
Manage Disease & Complications Medication Safety
BMI>25 & Age >40, need diabetes test
High random blood sugar, need test
Gestational diabetes, need diabetes test
Metabolic syndrome, need treatment
Need Vaccines for Influenza, Pneumonia
Need Eye Exam
Need Foot Exam
Need Peripheral Artery Disease Test
Need A1c blood test
Need LDL cholesterol blood test
Need kidney damage urine protein test
Need kidney function blood test
>40, need aspirin
LDL >100, no meds, need statin therapy
LDL still high, on meds, need intensify
High A1c, no meds, need metformin
High A1c, on meds, need to intensify
Very high A1c, need insulin
Hypertension, need ACE inhibitor/ARB
Statins – liver, muscle damage warning
Aspirin & ulcer risk, need ulcer protection
Oral contraceptives – danger with diabetes
Hyperglycemia - Atypical Antipsychotics
Metformin danger conditions, need tests
Glitazones, liver danger, tests needed
Drugs that worsen blood sugar levels
ACE inhibitor/ARB side effects (e.g. potassium)
Kidney damage, need ACE inhibitor / ARB
Atypical Antipsychotics – Glucose Monitoring
Systemic steroids – Hyperglycemia screen
Autoimmune disorders – Diabetes screen
Aetna Inc.
EBM Gaps in Care: Cousins of Quality Measures
Patient Data
CareEngine Logic
Message• Relevant data• Applicable literature and
guideline(s)
• What data was found (claims for diabetes, labs for micro-albuminuria, no claims for ACE inhibitor or ARB)
• Literature and guidelines (American Diabetes Association recommendation)
Alert generation
Excluded• Already on medication
or equivalent• Contraindications to
medication• Similar CC sent
• Current ACE Inhibitor/ARB • Angioedema • Pregnancy• Hyperkalemia• Renovascular disease
Exception screening
Validated• Diabetes • Diabetic nephropathy• Age 18+
• ICD-9 claims for Diabetes• Diabetes medications or supplies or DME• Micro-albuminuria >=30 mL/day(Also with specific timeframe criteria)
Clinical confirmation
Aetna Inc.
Note on ICD code specificity & clinical data
ICD‐9
↓ICD‐10
↑ ↓HL7 + SNOMED
↓Mood Codes
14
Sensitivity vs. Specificity
EHR
Aetna Inc.
Clinically‐Relevant Outputs
15
Groupers Total Cost
Unit Volumes Unit Costsx
Aetna Inc.
Clinically‐Relevant Outputs
16
Groupers Total Cost
Inpatient
OP Procedures
ER
Specialty Visits
PCP Visits
Imaging
Lab Tests
Rx
Rates
Rates
Rates
Rates
Rates
Rates
Rates
Rates
Unit volume Unit cost
Aetna Inc. 17
InpatientAcute
InpatientNon‐Acute
OP
Procedures
& FacilityER
ROC (C‐stat) 0.82 0.84 0.79 0.84
PPV 53% 48% 59% 63%
Sensitivity 28% 23% 35% 60%
Specificity 96% 99% 95% 95%
% of overall spend ~39% 8.4% 3.8%
Aetna Inc. 18
Aetna Inc.
Shouldn’t our predictive models keep breaking down if we’re doing our job?
•
If Risk Adjustment models change
only modestly year‐over‐year…
are
we actually making progress?
•
Do we inadvertently remove the
most important predictor features /
variables because they’re so good
at reducing costs?
In other words, if we don’t believe
there are modifiable
risk factors…
then all of this healthcare business
is just an academic exercise.
Aetna Inc.
Real‐World Outcomes & High‐Dimensional Propensity Scoring: A Wish to the Predictive Modeling Community
20
Predict
assignment to
intervention
(e.g. logistic
regression)
Compare
intervention
group vs.
matched
“control”
Use PS as
adjustment
covariate,
stratification
tool, or identify
matches
How might we bring the strength of the predictive modeling community to
bear, with opportunity for massive automation & ways to solve for hidden
bias?
Aetna Inc.
A Final Note: Patient‐Centered Predictive Modeling
21