Post on 09-Mar-2018
transcript
Data-Driven Healthcare: Visual Analytics for Exploration and Prediction of Clinical Data Adam Perer IBM Research
Patient Clinician
Patient
Electronic Medical Record Databases
Clinician
Thousands or Millions of Patients • •
•
10+ Years of Data Per Patient Tens of Thousands of Features • • • • •
Demographics Diagnoses Labs Procedures Claims
Unstructured Physician Notes
Patient
Search and Analysis
Electronic Medical Record Databases
Clinician
Thousands or Millions of Patients • •
10+ Years of Data Per Patient Tens of Thousands of Features • • • • •
Demographics Diagnoses Labs Procedures Claims
• Unstructured Physician Notes
Patient
Electronic Medical Record Databases
Expert
ise vi
a Inter
actio
n Clinician
Thousands or Millions of Patients • 10+ Years of Data Per Patient •
•
Tens of Thousands of Features •••••
Demographics Diagnoses Labs Procedures Claims Unstructured Physician Notes
Patient Clinician
Patient Clinician
Patient Clinician
outline
•
•
Visual tools for exploring clinical data support unearthing insights from clinical records
• CareFlow
Beyond exploration, clinical researchers often want predictions, too.
• Coquito, Prospector
CareFlow
IBM WatsonHealth
electronic medical records
Time
Patient #1
A B C
Time
Patient #1
AntihypertensiveFebruary 7, 2016
A B C
electronic medical records
A B C Patient #1
AntihypertensiveFebruary 7, 2016
Beta Blockers February 28, 2016
Time
electronic medical records
A B C Patient #1
AntihypertensiveFebruary 7, 2016
Beta Blockers February 28, 2016
Diuretics April 1, 2016
Time
electronic medical records
Time
Patient #1
A B C
A B D Patient #2
Patient #3
A D B
Patient #4
A C D B
electronic medical records
Time
Patient #1
A B C
A B D Patient #2
Patient #3
A D B
Hospitalized
Well-Managed
Well-Managed
Well-Managed Patient #4
A C D B
electronic medical records
Heart designed by Catherine Please from The Noun Project
heart failure
• Potentially fatal disease that affects 2% of adults in developed countries
•
•
•
Difficult to manage
No systematic clinical guidelines fortreating Heart Failure
Presence of co-morbidities affectstreatment recommendations.
population • Hundreds of Thousands of Patients
diagnosed with Congestive Heart Failure
EMR Database
Demographics
Treatments
Lab T ests
Diagnoses
Symptoms
aggregation •
•
Start with target patient
Find similar patients
• Using our similarity analytics on relevant data
• Features include medications, symptoms, and diagnoses, and lab tests
• Align all patients by disease diagnosis
What are the treatment pathways
after diagnosis?
aggregation
Diagnosis Date
[A,B,C]
[A,B]
[A,C]
[A]
[B] [ ]
[C] [B,C] Average outcome = 0.4Average time = 10 daysNumber of patients = 10
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
Hospitalized Managed
Hospitalized Managed
[Diuretics]
[ ]
[t1 ,t
3 ][D
iuretics,B
eta Blockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
Hospitalized Managed
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[Cardiontonics]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers][D
iureticds,A
ntianginal][Cardiontonics,
Antihypertensive]
Hospitalized Managed
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive][t
1 ,t3 ]
[Diuretics,
Beta B
lockers][D
iureticds,A
ntianginal]
[ ]
[Diuretics]
[Cardiontonics]
[Cardiontonics,A
ntihypertensive]
Hospitalized Managed
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
[Cardiontonics,A
ntihypertensive][Cardiontonics,
Antihypertensive]
[ ]
[t1 ,t
3 ]
[Diuretics]
[Diuretics,
Beta B
lockers]
[Cardiontonics]
[Diureticds,
Antianginal]
Care Pathways of 300 similar patients
Optimal Care Pathway among 300 similar patients
clinical researchers:
clinical researchers:
clinical researchers:
clinical researchers:
clinical researchers:
the role of visualization
in prediction
what can visualization do?
START
Cohort Definition
coquito
END
Model Interpretability
prospector
coquito cohort queries with iterative overviews
Cohort Construction 1
Josua Krause, Adam Perer, and Harry Stavropolous. SupportingIterative Cohort Construction with Visual Temporal Queries. IEEE Visual Analytics Science and Technology (VAST 2015).
defining cohorts • Typically, defining cohorts is a slow process:
•
•
•
First, medical researchers define requirements.
Then, Technologists write SQL queries and deliver them to medical researchers.
But, often too many patients or too few patients, and the process must restart.
defining cohorts with coquito
drag and drop constraints with immediate feedback
and hints for query refinement
supports using complex temporal logic
support for multiple queries side-by-side (for
cases and controls)
coquito lessons ▪ Easy and interactive query formulation lets domain experts explore the
data
▪
▪
Visible intermediate results provide critical feedback
Hints for query refinements are helpful in improving queries
predictive model prospector Model Interpretability
Josua Krause, Adam Perer, and Kenney Ng. Interacting with Predictions: Visual Inspection of Black-box Machine Learning Models. ACM Conference on Human
Factors in Computing Systems (CHI 2016). San Jose, California. (2016).
typical predictive model report Typically simply a list of top features and their weights
Why? Difficult to summarize complex models
Issues One cannot interpret how the values of each feature impact the prediction
One cannot interact with the model to test hypotheses
partial dependence
diabetes diagnoses
bmi
glucoselevel
teeth
eyebrows
prediction
0 0 1 1 87 1 7 0
30 30 30 30 30 30 30 30
N N N N N Y N Y
Y Y Y N ? N N N
120 100 140 0 200 140 160 0
H S S S S S S S 7/8 are predicted sick
partial dependence
15 20 25 30 35
partial dependence
localized inspection
7
22
N
N
160
diabetes diagnoses
bmi
eyebrows
teeth
glucoselevel
0.95prediction
localized inspection
7
20
N
N
140
diabetes diagnoses
bmi
eyebrows
teeth
glucoselevel
0.45prediction
32
33
•
•
•
predicting onset of diabetes for 4000 patients
4 month long term case study with 5 data scientists
stories of visualization-driven insights in the p aper
take-aways
Clinical Data is complex and messy.
Exploratory visual analytics tools fill a much needed gap.
However, exploratory tools alone do not address their predictive desires.
There is a strong role for visualization in predictive tasks.
Adam Perer [ papers and videos at http://perer.org ]
CareFlow (CHI 2013)
COQUITO (VAST 2015)
Prospector (CHI 2016)