Machine learning classification of
ASA Physical Status
from preoperative data
and associations with outcomes
Zach Murphy, MD/MSE
Adult Cardiothoracic Anesthesia Fellow
University of Michigan
Disclosures
None
ASA Physical Status (ASA
-
PS)
1 Healthy
2 Mild systemic disease
3 Severe systemic disease
4 Constant threat to life
5 Moribund
6 Deceased organ donor
Inter-rater reliability is
moderate at best
Why should we care?
Accurate documentation
Research
Clinical care
Create ML model to classify ASA-PS 1-5
using preoperative data
Data
Adults in SDF 2018 – 2023
Patient/Nonpatient features
Classify ASA-PS 1-5
Model building
Variety of architectures
Hyperparameter optimization
Evaluation
Top-1 accuracy
Conformal prediction
Explainability (SHAP)
Undercoding
SDF: A large, broad dataset
9,486,382 cases
Train: 6,638,981 cases
Test: 2,847,401 cases
68
Institutions
Model building
Model
Patient features
Nonpatient features
Target
ASA-PS
1-5
Hyperparameter optimization
5-fold cross validation
Architectures
GLM — RF — XGBoost — K/ANN (HNSW)
TabNet — MedGemma 1.5 + LoRA
Age - Sex - Race - BMI
Preop MAP - HFRS
Elixhauser comorbidities
Holiday - Weekend - Outpatient
Service - CPT base units - E status
Intubated - Aline - GA - Staffing
Med school - Hospital size
Selected model:
Patient features only
68.1 %
Test set top-1 accuracy
Patient + Nonpatient features
70.7%
Test set top-1 accuracy
Conformal prediction
Conformal prediction
84.7%
Test set top-1 accuracy
Patient + Nonpatient features
On 24% of
test set
Explainability:
ASA-PS 1 ASA-PS 2
Explainability:
ASA-PS 3 ASA-PS 4
Explainability:
ASA-PS 5
Embedded biases
“Models are opinions
embedded in
mathematics.”
— Cathy O’Neil,
Weapons of Math
Destruction
Undercoding
Clinician assigned ASA-PS 2
Model assigned ASA-PS 3
“This patient has a combination of frailty,
uncomplicated hypertension, hypothyroidism,
cardiac arrhythmia, and liver disease.
Consider ASA-PS 3.”
# patients with X
assigned ASA-PS 1 or 2
Undercoding rate
for condition X
# patients with X
Takeaways & pearls
Model is only as
good as the data
Evaluation must be
contextual
Decision support,
not replacement
Use scalable &
parallel processing
Don’t discount “old”
architectures
ML evolves fast,
keep learning
Thank you!
Michael
Mathis
Michael
Burns
Thomas
Klumpner
Sachin
Kheterpal
Chad
Brummett
Emily
Mackay
Michelle
Busschots