Michael R. Pinsky, M.D., C.M., Dr.h.c., FCCP, MCCMProfessor of Critical Care Medicine, Bioengineering, Anesthesiology, Cardiovascular Diseases, and Clinical & Translational Sciences, Vice
Chair for Academic Affairs, University of Pittsburgh
Using Machine-Learning
Principals to Predict the
Future in Acute Care
Medicine
Michael R. Pinsky, MD, Dr hcDepartment of Critical Care Medicine
University of Pittsburgh
Systems Issues in Critical Illness
• Disease phenotype is a mixture of the process
and the hosts response to the process
• No two people are alike
• Healthcare systems are inherently complex
and inefficient
• Patient safety alerts often artifacts leading to
alarm fatigue and failure to rescue
• Goals of therapy are often unknown and
commonly change
A Modest Proposal
Make the patient the center
Personalized Medicine
• Unique personal goals and desires
– Defining start and stopping rules
• Unique expression of disease
– Defining threshold values for homeostasis
• Unique response to treatment
– Physiologic and regenerative reserve
How do we identify patterns of
health and disease?
Music Name that tune
Which instruments to listen to
(number of independent sensors)
How long to listen
(lead time)
Sampling frequency
(once a second, one a year)
Sometimes it is easy to predict
the future
Dynamic Changes
Box & Jenkins, 1976, p. 531
Situational Awareness
Variability is often a sign of health
It allows for adaptation
Loss of variability occurs with stress
Heart Rate Variability Indicies Predict Instability
1.5
2
2.5
3
3.5
4
4.5
024487296120
Heart
Rate
Va
riab
ilit
y I
nd
ex
Hours from Unstable Event
Unstable HRVI
Stable HRVI
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
12
13
14
15
16
17
18
19
20
21
22
024487296120
Mean
Res
pir
ato
ry R
ate
Hours from Unstable Event
Stable RRM
Unstable RRM
Mean Respiratory Rate Does Not
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Increasing Dimensionality
Improves Recognition
Faces
Health and Disease Defined as a
Time-Space Continuum
• In a static field of single point-in-time data health and disease can be separated in stochastic fashion using Artificial Neural Networkapproach to create a probabilistic equation:
Fused parameter VSI
• In a dynamic field of continuously changing but inter-related variables, Machine Learningdata-driven classification techniques:Principal Component Analysis, Support Vector Machines, K Nearest Neighbors, Random Forests, Naïve Bayesian Classifier
VSI
HR
RR
SpO2
BP
06:00 09:00 12:00 13:2911:00
Completely normal values
Hravnak et al. Arch Intern Med 168:1300-8, 2008
Vital Sign Trends Over 8 Hours in a SDU Patient
Mean time from 1st alert to Code = 6.3h
Of 40 patients who met CODE criteria
only 7 had codes called
Percentage of Patients in Each Phase
who Experienced a MET State
0
5
10
15
20
25
30
35
40
METall METmin METfull
% A
ll P
ati
en
ts i
n P
ha
se
Phase1
Phase 3
Hravnak et al. Crit Care Med 39:65-72, 2011
Cut total time alarms sounding >50%
Duration Patients in MET State
(for those who experienced it)
0
5
10
15
20
25
30
35
40
45
METall METmin METfull
Mea
n M
inu
tes
Phase 1
Phase 3
Hravnak et al. Crit Care Med 39:65-72, 2011
Alarm Fatigue: Is the Alert Real, Real
Important or Artifacts
• ECRI Institute’s Top Patient Safety Concern:
Health Data Integrity Failures
Using Machine Learning we could discriminate real
from artifact SpO2, RR & HR alerts >90% of the time
Hravnak et al. Crit Care Med 42:42, 2015
Health
Normal Homeostasis
Dis
ease
Sev
erit
y
Ad
apti
ve-
Mal
adap
tive
Res
pon
ses
Present Clinical
Threshold of Detection
Potential Threshold of
Pathologic Stress Detection
Threshold of Stress
Time
Early Detection of Disease Model
Identifying Hemodynamically
Unstable Patients
• What is the minimal data set needed to
predict instability: Monitoring parsimony
– Number of independent monitoring variables
– Lead time
– Sampling frequency
• What additional information will improve
specificity
• Monitoring response to therapy and define
end-points of resuscitation
Various methods to detect instability
• Anomaly detection
• Trigger alerts upon significant departure from the
envelope of expected variability
• Classification
• Classify current state of a patient as stable or unstable,
perhaps identify specific type of instability
• Regression
• Estimate the magnitude of instability as a function of
the extent of departure from stable behavior
Pinsky & Dubrawski. AJRCCM 190: 606-10, 2014
Identify Onset of Bleeding Earlier Train a multivariate regressive
model (Random Forest)
Leave one out
Guillame-Bert et al. Intensive Care Med 40: S287, 2014
N=47
Slide 19 Copyright © 2015 CMU Auton Lab
Balancing False Alert Rates with early Bleeding detectionComplete multivariate model vs. univariate detectors
This way towards the ideal performance
Individual vital signs
Holder et al. J Crit Care doi.org/10.1016/j.jcrc2013.07.028, 2013
Bleeding rate20 ml/min
Detect 3’ 40”False alerts 6 hr
Detect 3’ 40”False alerts 5 min
Slide 20 Copyright © 2015 CMU Auton Lab
Density of Data Affects Bleeding DetectionMultivariate models for various groups of measurements
[Sampled every 5 minutes]
[Sampled every 20 seconds]
[CVP + Arterial Pressure, FFT features, CVP, normalized using personalized baselines]
Holder et al. J Crit Care doi.org/10.1016/j.jcrc2013.07.028, 2013
Slide 21 Copyright © 2015 CMU Auton Lab
Knowing Baseline Important if Data SparsePersonalized baselines
Holder et al. J Crit Care doi.org/10.1016/j.jcrc2013.07.028, 2013
PITT Index
Probability of Event
Time to event<5 15
Type of Event
Cardiac
Volume
Vasomotor tone
VSI Index
3090
60
Predicting
Instability
Time and
Treatment
Severity Index
Future of Monitoring
• Monitor the monitors
• Treat the patients
• Robust artifact detection increases value of alerts
and reduces alarm fatigue
• Using fused parameters to define stability
• Scaling of monitoring devices and sampling
frequency will vary as patient conditions change
Thank YouCo-Investigators:
Artur Dubrawski, PhD
Auton Lab
Gilles Clermont, MD MS
Marilyn Hravnak, BSN, PhD