Pressor Gauge
William Sweeney
Insight Fellow
Predicting the need for life-saving blood pressure medication
Problem: planning for pressor administration
goals of care discussion
1 in 4 ICU patients need pressors
increased testing and overnight vigilance
Solution: assess need for pressors in the next 12 hrs
Probability of pressor need
web app:
Streamlit
Docker
pH, pO2, pCO2
creatinine, Cl, Na, K,HCO3, BUN, glucose
hemoglobin, hematocrit, white blood cell, platelets
blood gas
CHEM-7
CBC
PTT
coags
age, weight, sex
Labs +
patient info
Hours since hosp. adm.
Current prediction
Workflow
Training and validation
Random forest classifier
(~12k train, 80/20 CV,
2.5k holdout)
Predictions
& live feature ranking
Feature selection
& time-binning
Data Source
~61 thousand ICU stays
~28 million lab records
>50 GB
-sought physician input (use labs!)
-Shapley scores ranking
Local DB
Data Cleaning
~35,000
adult, non-repeat ICU stays
w/o obvious error
~15,000
ICU stays remain
~20,000
presssors < 12 hrs post admission ,
pressor lasts < 15 min,
~61,000
ICU stays
~11,000
no pressors
~26,000
under 18,
weight > 750 lbs (misrecorded),
multiple ICUs visits
~4,000
with pressors
drop
drop
final ICU candidates (still need time-binning)
Time-binning features
Positive class
pressor event
-6
-12
-18
-24
-30
-36
-42
-48 hrs
Labs are taken every ~6 hrs, but not regularly
samples
aggregate labs within each window
Negative class
After dropping missing values:
~12k training samples
~3k test
Feature Selection
Patient info (3)
CHEM7 (7)
Blood gas (3)
CBC (4)
Coags (1)
Liver function markers (4)
+
Lactate
+
anion gap,
+
tot. abnormal labs (proxy for overall health)
25 features
Impact: 8 hours advanced notice!
AUROC: .70, @ 83% recall, 60% FPR
Mean prediction time to pressor event: 8 hrs
Impact: Overnight Care and Goals of Care
have goals of care discussions
increase testing and plan overnight care
That’s 8 hours of advanced notice to
William Sweeney
Seeking new adventures!
Backup Slides
Distribution of Pressor Times
liver/kidney function
blood acidosis
hemoglobin & hematocrit
electrolytes
CV Hyperparameter Tuning
Max tree depth
Number of estimators
Learning curve
Model Comparisons
XGBoost: AUROC .71
Logistic Regression: AUROC .67
Cox PH Regression
Naive Bayes: AUROC .65