Transcript Predicting Clinically Relevant Events using Real

UPMC Critical Care
www.ccm.pitt.edu
Making Sense of Complex
Monitoring Signals
Michael R. Pinsky, MD, Dr hc
Department of Critical Care Medicine
University of Pittsburgh
Conflicts of Interest
• Michael R. Pinsky, MD is the inventor of a US patents “Use of aortic
pulse pressure and flow in bedside hemodynamic management ,”
“Device and system that identifies cardiovascular insufficiency” and
• “A system and method of determining a susceptibility to
cardiorespiratory insufficiency” owned by the University of Pittsburgh.
• Co-founder and stockholder of Intelomed
• Is or was* a medical advisor for:
–
–
–
–
Abbott Corporation*
Applied Physiology Ltd.
Arrow International*
Hutchinson Medical*
Edwards Lifesciences
Cheetah Medical*
väsamed*
LiDCO Ltd
• Is or was* receiving research funding from:
– Deltex Ltd*
– Pulsion Ltd*
– Edwards LifeSciences
• Michael R. Pinsky, MD is receiving research funding as Principal
Investigator from the NIH
– T32 HL07820, R01 HL074316, R01 NR013912 and UM1 HL120877
Monitoring Truth
No monitoring device, no matter how
accurate or insightful its data will
improve outcome,
Unless coupled to a treatment, which itself
improves outcome
Pinsky & Payen. Functional Hemodynamic Monitoring, pp 1-4, 2004
Pinsky & Payen. Crit Care 9:566-72 2005
Pinsky. Chest 123:2020-9, 2007
Three Primary Clinical Problems
• How to identify patients who are becoming
hemodynamically unstable before they
progress too far?
• How to determine the most appropriate
therapy to reverse the primary cause for
impending circulatory shock?
• How to you implement the most appropriate
therapy when individual training of care
givers and responses of patients vary?
Three Primary Clinical Problems
• How to identify patients who are becoming
hemodynamically unstable before they
progress too far?
• How to determine the most appropriate
therapy to reverse the primary cause for
impending circulatory shock?
• How to you implement the most appropriate
therapy with non-physician when individual
responses of patients and care givers vary?
Systems Issues in Critical Illness
• Disease phenotype is a mixture of the process
and the hosts response to the process
• No two people will present and respond the
same even if they have the same diagnosis
• Healthcare systems are inherently complex
– Monitoring, alerts, physical plants
– Bedside nursing, Physicians, Specialists
• Goals of therapy are often unknown and
commonly change
Why Blood Pressure Does Not Define
Cardiovascular Status
Heart failure
Hypovolemia
MAP = 80 surface
Pms
Sepsis
Eh
Why Cardiac Output Does Not Define
Cardiovascular Status
Heart failure
CO = 5 L/min surface
Pms
Eh
Hypovolemia
or Sepsis
Combining pressure and flow helps
Heart failure
MAP=80 &
CO=5 line
for Eh
CO=5
Pms
Hypovolemia
Eh
Separating Pms, Eh and SVR for
Hypovolemia, Heart Failure and Sepsis
Navagator 3-D Display
30
25
15
10
20
18
16
14
12
10
R
5
0
0.9
0.8
8
0.7
0.6
Eh
6
0.5
Eh
Septic Shock
Cardiogenic Shock
Hemorrhagic Shock
0.4
4
0.3
0.2
2
TP
Pms
20
Static Risk Prediction Scoring
• APACHE
– Acute physiologic and chronic health evaluation
• MEWS
– Medical early warning system
• LODS
– Logistic regression score
• SOFA
– Sequential organ failure assessment
Static Risk Prediction Scoring
Limitations
Retrospective
Usualy require manual data entry
Useful for predicting longer term outcomes
Useful for nurse and related resource
allocation
A Modest Proposal
Make the patient the center
• 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
Efferent Arm:
Medical Emergency Teams (MET)
Improve Acute Care
• System-wide ICU-based MET activation to evaluate
and treat patients at risk to develop adverse events
• Pre-defined MET activation criteria by non-MD staff
• Reduces adverse events Relative Risk Reduction:
– Stroke: 78%, Severe Sepsis: 74%, Respiratory failure: 79%
• Saves lives: post-op death decreased 37%
• Decreases costs: less ICU transfers, decreased LOS
• Bellomo et al. Crit Care Med 32: 916-21, 2004
But first one must identify these unstable patients
Electronic Automated Vital Sign Advisory
Reduces Morbidity in General Hospital Wards
Bellomo et al. Crit Care Med 40:2349-61, 2012
Effect of an Electronic Advisory
Alert in General Medical Wards
•
•
•
•
Alerts for vital signs outside of “safe” range
9617 patients before, 8688 patients after
MET activation not directly linked to alerts
Improved Survival with Advisory Alerts
Bellomo et al. Crit Care Med 40:2349-61, 2012
Medical Issues
• Identify circulatory insufficiency before secondary
tissue injury occurs
• Assess disease severity
• Accurately predict response to treatment
• Gauge adequacy of specific therapies
• Estimate improved predictions of disease severity
by additional biological measures (biosensors)
• Need to develop metrics to assess these challenges
• Functional Hemodynamic Monitoring
• Fused parameters (smart alarms)
• Machine learning-based pattern recognition
• Am J Respir Crit Care Med 190: 606-10, 2014
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 Neuronet approach to
create a probabilistic equation
• In a dynamic field of continuously changing
but inter-related variables, health and disease
can be defined by the differences their Lorenz
Attractors (ρ) independent of the actual
physiological variables raw values.
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 Neuronet approach to
create a probabilistic equation
• In a dynamic field of continuously changing
but inter-related variables, health and disease
can be defined by the differences their Lorenz
Attractors (ρ) independent of the actual
physiological variables raw values.
Background
MET activation is grouped around morning and afternoon
rounds, suggesting instability was missed at other times
DeVita et al. Crit Care Med 34:2463-78, 2006
An electronic integrated monitoring systems (BioSigns)
was developed to identify cardio-respiratory instability
using Neuronet analysis of existing ICU patient
behavior
Tarassenko et al. Br J Anaesth 97:64-8, 2006
Existing bedside monitor
Integrated Monitor (BioSign)
Added Information
Example of the BioSign™ screen developing a single
physiologic index value; alert threshold for Phase 1
was 3.0 based on prior ICU patient data
Phase 1 Results
333 patient admissions representing all patients, reflecting
18,692 hours continuous monitoring
All 7 MET activation events of respiratory and/or cardiac cause
were detected by BSI in advance of MET activation
Mean advanced detection time prior to MET activation was 6.3h
78% of METfull did not result in MET activation and <1/2 were
commented on in nursing notes
Cardio-respiratory deterioration was generally characterized by
progressive increases in BSI over time rather than step
increases
Most patients were stable during their SDU stay: 75%
Unstable patients were only unstable at most 20% of the time
Hravnak et al. Arch Intern Med 168:1300-8, 2008
BSI
HR
RR
Completely normal values
SpO2
BP
06:00
09:00
11:00
12:00
Example of Phase 1 METfull patient
who had MET Activation called at 13:29
13:29
Smart Monitoring Conclusions
Cardiorespiratory deterioration requiring MET
activation was uncommon but was preceded by
BioSign Index (VSI) values in danger range prior
to MET activation
Neither watching a central monitor nor intermitted
direct patient observation will measurably improve
identification of unstable non-ICU patients
Hravnak et al. Arch Intern Med 168:1300-8, 2008
Phase 3: Using BSI to Drive MET Activation
Clinical Algorithm
Percentage of Patients in Each Phase
who Experienced a MET State
% All Patients in Phase
40
35
30
25
Phase1
Phase 3
20
15
10
5
0
METall
METmin
METfull
Hravnak et al. Crit Care Med 39:65-72, 2011
Total Cumulative Duration Patients
in MET State
14,000
Phase 1
Phase 3
12,000
Total Minutes
10,000
*
*
*
8,000
6,000
* P < 0.05
4,000
2,000
0
METall
METmin
METfull
Hravnak et al. Crit Care Med 39:65-72, 2011
Duration Patients in MET State
(for those who experienced it)
Phase 1
45
Phase 3
40
Mean Minutes
35
30
25
20
15
10
5
0
METall
METmin
METfull
Hravnak et al. Crit Care Med 39:65-72, 2011
Using Neuronet Modeling to Look
into the Future
Using the VSI fusion parameter to
quantify instability
• VSImin occurred before METmin 80% of the time
with a mean advance time 9.4 ± 9.2 minutes and
correlated well (r=0.815, p < 0.001)
• VSIfull occurred 6.4 ± 1.5h before CODE called
Hravnak et al. Crit Care Med 39:65-72, 2011
Early Detection of Disease Model
Disease Severity
Adaptive-Maladaptive Responses
Health
Normal Homeostasis
Present Clinical
Threshold of Detection
Potential Threshold of
Pathologic Stress Detection
Threshold of Stress
Time
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 Neuronet approach to
create a probabilistic equation
• In a dynamic field of continuously changing
but inter-related variables, health and disease
can be defined by the differences their Lorenz
Attractors (ρ) and dynamical analysis
independent of the actual physiological
variable raw values
Chaos Theory and Biology
• Non-linearity
– Chaotic Behavior
– Fractal appearance
• Non-linear thinking can result in solutions
to otherwise unsolvable problems
• Benoît Mandelbrot & James A. Yorke
Complexity Theory
Self-Organizing Behavior
Complexity Theory
Self-Organizing Behavior
Thermal scan of Petri dish
during E. coli log growth at 37 C
With heating to 40 C, the thermal bands
driven to randomness
Tool: Variability Analysis
Comprehensive analysis of degree & character of
patterns of variation over intervals in time.
Carnivore Population Herbivore Population Plant Population
Population Phase Portraits for the IndividualBased Predator-Prey System
Flake GW. The Computational Beauty of Nature: Computer Explorations of Fractals, Chaos,
Complex Systems and Adaptation. MIT Press 1998.
Herbivore Population
Population Phase Portraits for the IndividualBased Predator-Prey System
Carnivore Population
Carnivore Population
Plant Population
Plant Population
Herbivore Population
Flake GW. The Computational Beauty of Nature: Computer Explorations of Fractals, Chaos,
Complex Systems and Adaptation. MIT Press 1998.
Solving common Solutions for Three Variables:
Plants, Herbivores, Carnivores
Plotted as
animal
change
Pinsky. Crit Care Med 38:S649-55, 2010
Identifying Hemodynamically Unstable
Patients using Machine Learning
• What is the minimal data set needed to
predict instability: Monitoring parsimony
– Number of independent monitoring variable
– Lead time
– Sampling frequency
• What additional information will improve
specificity
• Monitoring response to therapy and define
end-points of resuscitation
Heart Rate Variability of Trauma Patient Survival by analysis
of HR Complexity Reduces Data Set Size Requirements
Survivor
RR
Intervals
(sec)
Non-Survivor
Sample entropy: SampEn
Survivors
Survivors
Non-Survivors
Non-Survivors
Batchinsky et al. Shock 32:565-71, 2009
Heart Rate Variability of Trauma Patient Survival by analysis
of HR Complexity Reduces Data Set Size Requirements
Survivor
Non-Survivor
Approximate entropy: ApEn
Survivors
Survivors
Non-Survivors
Non-Survivors
Batchinsky et al. Shock 32:565-71, 2009
Heart Rate Variability of Trauma Patient Survival by analysis
of HR Complexity Reduces Data Set Size Requirements
ROC curves with
800, 200 or 100
beat long data sets
Batchinsky et al. Shock
32:565-71, 2009
Decreased Respiratory Rate Variability during mechanical
ventilation in associated with Increased Mortality
Ratio of first harmonic (H1)
to zero frequency (DC) by
Fourier Analysis
Gutierrez et al. Intensive Care Med 10.1007/s00134-013-2937-5
Decreased Respiratory Rate Variability during mechanical
ventilation in associated with Increased Mortality
Gutierrez et al. Intensive Care Med 10.1007/s00134-013-2937-5
Extracting Features from Data Streams
Input
Bedside monitor: heart rate, respiratory rate, SpO2 and
arterial pressure, CVP, etc.
Electronic health record: medications, co-morbidities, age
Processing
A large number of features is extracted from raw data:
Moving averages (uniform, triangular, exponential),
median, standard deviation, ranges, derivatives,
cumulative sums, hysteresis with various thresholds,
quality of signal, MACD derivatives, calendar events
(hours, days, etc.).
Output
Symbolic and scalar time sequences:
~ 7,400 types of secondary observations (features)
per each instance of time and patient
~ 32,300,000 such instances
Building a Model to Predict
Subsequent Instability
E19
E18
E17
E16
E15
E14
E13
E12…………E2
Event
Case
ALL NON-EVENT EPOCHS
No
Event
EARLY NON-EVENT
EPOCHS
LATE NON-EVENT EPOCHS
No
Stable
Control
Event
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Building Physiologic Trends
Observation window (up to 8 hours)
Trans-epoch variables (5-15-30-60-240-480)
Target
Window
(Event)
Intra-epoch features
E
E
483
243
E
16
E E
15 14
60m
E
13
E
12
E
11
E
10
E9
E8
E7
E6
30m
E5
E4
15m
E3
E2
E1
5m
4h
8h
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Epoch Specific Models for Cases
E
…
48
E
…
49
…
E
…
50
E
…
51
E
…
48
E
…
48
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
E
14
E
13
E
12
E
11
E
10
E
9
E
8
E
7
E
6
E
5
E
4
E
3
E
2
E
1
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Example of Model Predictors
across epochs
Epoch
HR
sdHR
hipHR
RR
regRR
acRR
SaO2
regSaO2 acSaO2
2
hr
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
3
hr
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
4
hr
sdhr
hiphr
regrr
acrr
5
hr
sdhr
hiphr
6
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
7
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
8
sdhr
hiphr
regrr
acrr
sao2
9
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
10
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
11
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
12
sdhr
hiphr
regrr
acrr
sao2
regsao2 acsao2
regsao2 acsao2
regsao2
Heart Rate Variability Index
4.5
Variability Index
4
3.5
3
Unstable HRVI
Stable HRVI
2.5
2
1.5
120
96
72
48
24
0
Hours from Unstable Event
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Choosing Appropriate Models
Area under the ROC Curves
Ridge Regression
Area
Area
Lasso Regression
event
event
Prediction Algorithm Before Event
Prediction Algorithm Before Event
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Mean Respiratory Rate Historic Trend
22
21
Mean Respiratory Rate
20
19
18
17
Stable RRM
Unstable RRM
16
15
14
13
12
120
96
72
48
24
0
Hours from Unstable Event
Ogundele et al. Am J Respir Crit Care Med 187: A5067, 2013
Improving Diagnosis Using Machine
Learning Modeling Approaches
• Machine learning tools to
– Define the relation between physiological variables as
either healthy or not
– Identify the “value” of adding additional measures
(monitoring) or extending lead time
• Artur Dubrawski, Carnegie Mellon University
• Gilles Clermont, University of Pittsburgh
Patterns of Cardiovascular, Humoral and Immunologic
Response to Hemorrhagic Shock
n=9
n=3
40 mmHg
30 mmHg
Zenker et al. J Trauma 63:573-80, 2007
Namas et al. Plus One Medicine 4:e8406, 2009
Gomez et al. J Surg Trauma 187:358-69, 2012
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
Example of Flow of Machine
Learning Analysis
Experimental setup
•
•
•
•
40 pigs bleed at 20 ml/min
Random Forest classifier and regressor
20-fold cross validation
Various groups of vitals, grouped by invasiveness and frequency
of observation:
• Intermittent, observed at 5 minute intervals (SysBP, DiaBP, MAP, HR,
PPV, SPV, SVO2)
• Beat-to-Beat (ditto)
• Waveform (central venous line allowed)
• Features include:
•
•
•
•
Common statistical operators (mean, std. dev., trends, etc.)
CVP + Art trackers
FFT projections
Normalization using personal baseline reference
Variability of PCA in Response to Hemorrhage
Video of 20 pig two derived feature PCA changes q1 min
Identify Onset of Bleed (20 ml/min)
Train a multivariate regressive
model (Random Forest)
47 pigs Leave one out
Predicted bleeding time (min). The bleeding begins at time = 0.
Mathieu Guillame-Bert & Andre Holder
20 ml/min Bleed All pigs-Leave one out
Predictable behavior
(No measurable VS changes
until 5-6 min)
Pre-bleed
blood
sampling
Some pigs are more predictable than others. Predicted bleeding of
the top 5 most predictable pigs with the lowest mean absolute error.
20 ml/min Bleed All pigs-Leave one out
Less Predictable behavior
(No measurable VS changes
until 3-4 min)
Predicted bleeding of the top 5 pigs with the highest
mean absolute predicting error.
20s moving average of pig #57
Anesthesia increased
Evaluation method
Beginning of bleeding
Number of false positives
Given a threshold
Time until detection
Evaluation method
Beginning of bleeding
Given a threshold
Time until detection
Number of false positives
Number of false positives
Time until detection
Evaluation method
Beginning of bleeding
last threshold
new threshold
Time until detection
Number of false positives
Number of false positives
Time until detection
Evaluation method
Beginning of bleeding
last threshold
new threshold
Time until detection
Number of false positives
Number of false positives
Activity
Monitoring
Operating
Characteristic
(AMOC)
Time until detection
Bleeding detection - AMOC
0
5
10
15
20
Bleeding time (min)
25
Tolerance for false alerts – Lockout threshold
Increasing lock out for alerts >5
min does not improve
performance
Primary Hemodynamic Data and PCA Reconstruction Error as a useful
detection signal for “Not Normal” during hemorrhage & resuscitation
[1/heartbeat]
HRV
MAP
PPV
SVI
Likelihood of
Error
Low Frequency
HR
SW
Vasopressor
added
Baseline (stable period)
Bleed
Volume
Time 
PCA Step III: Projection
Detection of instability induced by bleeding
[1/heartbeat]
Low Frequency
• A PCA model is trained from stable period data (green)
• It is tested during periods of slow bleeding (blue) and resuscitation
(orange)
HR
HRV
MAP
PPV
SVI
SW
Changing sensitivity of the obtained detector
affects latency of detection
and false alert probability
Time 
[1/heartbeat]
HR
CO
MAP
SvO2
Likelihood of
Error
Low Frequency
Primary Hemodynamic Data and PCA Reconstruction Error as a useful
detection signal for “Not Normal” during hemorrhage & resuscitation
Volume
Volume
Bleed
Vasopressor
added
Baseline (stable period)
Time 
AMOC: Complete model versus
Univariate Measures
• Complete model allows on average
3’40” of latency of detection at the
mean time between false alerts of 6
hours
• At the same speed of detection, the
best single-vital model triggers false
alerts every 5 minutes
Guillame-Bert et al. Intensive Care Med 40:S287, 2014
AMOC of Random Forest prediction of Not-Normal
Increasing sampling frequency and quality improves identification
Group 1: 5 min
Group 3:
waveform
Group 2:
Beat-to-beat
Reduced informativeness of
raw data reduces detection
power
Guillame-Bert et al. Intensive Care Med 40:S287, 2014
How much could we gain by personalizing the models?
Raw CVP
CVP normalized
using the
individual
pre-bleed data
Impact of personalized baselines
Not using personalized
baselines reduces
performance
•
Especially of models
using less frequent
observations
Estimating the amount of blood lost
ideal response for 20 ml/min
average
prediction
for 20 ml/min
average
prediction
for 5 ml/min
average
prediction
for 0 ml/min
Guillame-Bert et al. Intensive Care Med 40:S287, 2014
Lessons learned: Summary
1. Multivariate models can be more
powerful
2. Tradeoffs between invasiveness and
value of information can be quantified
Group 1: 5 min
Group 3:
waveform
3. Personalization is promising
Raw data
Normalized data
Group 2:
Beat-to-beat
4. Various types of instability can be
identified and characterized
Can We Predict Who Will
Crash If Stressed Beforehand?
• Aim:
Use pre-bleed period data to predict whether the pig
would eventually crash during bleeding
• Data:
• 24 non crash pigs, 4 crash pigs, Beat-to-beat data
• 11 vitals (Sys BP, Dia BP, MAP, PP, PPV, SV, SVV,
SPV, PPV, HR, HRV), each mapped onto four simple
derived variables (mean, stand deviation, slope, range)
 44 features aggregated over disjoint consecutive 20s
intervals, every 20s
• This data featurization approach is coarse and it leaves
room for potential further improvement
Dubrawski et al. Intensive Care Med 40:S288, 2014
Experiment setup (continued)
• Leave-One-Pair out cross validation:
• 24 non-crashes x 4 crashes = 96 (crash)—(non-crash) pairs
• We iteratively leave out one such pair as test set, and train ML
model on the rest of the data
• Assemble test results separately for crash and non-crash pigs
• Compute the average “crash” scores and their confidence limits
• Draw temporal plots of those stats over the pre-bleeding period
(truncating a few initial and a few ending minutes because of
substantial noise observed in raw data)
Dubrawski et al. Intensive Care Med 40:S288, 2014
Crash Prediction Prior to Stress
Summary of the results
Dubrawski et al. Intensive Care Med 40:S288, 2014
Crash Prediction Prior to Stress
Summary of the results
Dubrawski et al. Intensive Care Med 40:S288, 2014
Log scale ROCs for classifier based on raw score
at selected timestamps (5, 10, and 20 minutes)
Dubrawski et al. Intensive Care Med 40:S288, 2014
PITT Index
Probability of Event
<5
Cardiac
Volume
Vasomotor tone
Time to event
15
30
60
90
Type of Event
Severity Index
Predicting
Instability
Time and
Treatment
Future of Monitoring
• Monitor the monitors
• Scaling of monitoring devices and sampling
frequency as patient specifics define
• Using fused parameters to define stability
• Looking at monitors or intermittent direct
patient observation unlikely to improve
instability detection but very likely to
increase its undetection
Identifying Hemodynamically
Unstable Patients
• Phase One currently available
– Modified application of VSI input to identify
better who is sick now or about to be sick
• Phase Two coming soon to a health
information system near you
– Complexity Modeling of Health and Disease
– Identify the dynamical expression of
hemodynamic stability, instability and recovery
– Define the Architecture of instability
– Clarify the true Picture of health
Thank You