Transcript Example 2 - Dr. Marcia Testa

Clinical Investigation and
Outcomes Research
Analysis of Physiologic and
Pharmacologic Data
Marcia A. Testa, MPH, PhD
Department of Biostatistics
Harvard School of Public Health
1
Objective of Presentation
• Introduce analytical methods for the
special case where biomedical data are
collected during a session which contains:
– repeated observations over time
– numerous, frequently sampled data points
– measures collected over a relatively short
interval of time (several hours or days) within
one session
– commonly, more measures per session per
subject, than subjects overall
2
Intensively Sampled Data
• Data collected during a physiology,
monitoring or pharmacologic study over
several hours or days with measurement
every 1, 5, 10, 15, 30 or 60 minutes, or as a
continuous function
• Each session may be repeated at weekly or
monthly intervals to investigate the effects of
interventions as part of clinical trials or
treatment assessment, and to correlate
session summary parameters with clinical
events, morbidity and mortality
• In physiologic research, these data are often
referred to as “complex physiologic signals”
3
Why Study Signals?
ECG
BP
Physiologic signals and time
series reveal aspects of health,
disease, biotoxicity and aging
not captured by static
measures.
Time = 2 seconds
Raw (original) signals are of interest as means of




 developing new biomarkers
 measuring parameters of known interest
 developing new insights into basic
mechanisms of human physiology
4
Physiologic Response
Periodic Functions
Time (minutes)
Response may represent a periodic function such as this
graph of interstride intervals for a patient with Huntington’s
disease, or a smooth function in response to a stimulus
5
such as oral drug administration.
Plasma concentration
Smooth Functions
14
Ka = Absorption Constant
Ka/Ke=10
12
Ke = Elimination Constant
Ka/Ke=1
10
Ka/Ke=0.1
8
Ka/Ke=0.01
6
4
2
0
0
Oral Drug
5
10
TIME (hours)
15
20
6
Intensive Data: Cardiology Studies
• Continuous recording:
ECG is recorded
continuously during the
entire testing period.
• Event monitor, or loop
recording: ECG is
recorded only when the
patient starts the
recording, when
symptoms are felt.
7
A Complex Signal Dataset
Physiologic time series,
such as this series of
cardiac interbeat (RR)
intervals measured over
24 hours, can capture
some of the information
lost in summary
statistics.
Data from the NHLBI Cardiac Arrhythmia Suppression
Trial (CAST) RR Interval Sub-study Database
8
Example 1: Heart Rate
Dynamics
Pathology can affect
physiologic recordings in
unexpected and interesting
ways.
Analysis of complex signals
can extract information hidden
in data. Figure shows shows
the instantaneous heart rates
of four subjects. The plot of
heart rate (beats/min) versus
time (min) is called a
tachogram.
Of the four tachograms
shown, only one signal is
from a healthy person. Can
you tell which it is?
9
Excessive regularity
Healthy heart rate
Excessive regularity
Uncorrelated Randomness
In A and C we can see a rather
periodic signal, with low
variability of its values. In case
C, there is a pattern of periodic
oscillations (1/min), which is
associated with CheyneStokes breathing.
The healthy record B is
characterized by a rather
rough and ‘patchy’
configuration, attributed to
fractal properties of the heart
rate signal.
The breakdown of such
behavior (fractal dynamics)
can lead to either excessive
regularity (A &C) or
10
uncorrelated randomness (D).
A
Example 2: Ambulatory
ECG
Schedule of study
events is shown in panel
A.
B
Panel B shows inhospital activity
schedules on the two
activity days. AEM
indicates ambulatory
ECG monitoring.
Vertical arrows represent
timing of venous
sampling.
11
Example 2: Rates of Ambulatory Ischemia
– Bar Graphs and Polynomial Regression
Regular Activity Day
Parker JD, Testa MA, Jimenez AH, Tofler GH, Muller JE, Parker JO and Stone PH. Morning increase in
ambulatory ischemia in patients with stable coronary artery disease: Importance of physical activity
and increased cardiac demand. Circulation 1994;89:604-614.
12
Example 2: Rates of Ambulatory Ischemia
– Bar Graphs and Polynomial Regression
Delayed Activity Day
Parker JD, Testa MA, Jimenez AH, Tofler GH, Muller JE, Parker JO and Stone PH. Morning increase in
ambulatory ischemia in patients with stable coronary artery disease: Importance of physical activity
and increased cardiac demand. Circulation 1994;89:604-614.
13
A Regular Activity Day
B Delayed Activity Day
Example 2: Ambulatory
ECG
Bar graphs show
frequency of episodes of
ambulatory ischemia
during therapy with
placebo and nadolol on
the two activity days.
Panel A, Regular activity
day;panel B, delayed
activity day.
14
Example 2: Minute by Minute Heart Rate
Placebo
Nadolol
Parker JD, Testa MA, Jimenez AH, Tofler GH, Muller JE, Parker JO and Stone PH. Morning increase in
ambulatory ischemia in patients with stable coronary artery disease: Importance of physical activity
and increased cardiac demand. Circulation 1994;89:604-614.
15
Example 2: Minute by Minute Heart Rate
Placebo
Nadolol
Parker JD, Testa MA, Jimenez AH, Tofler GH, Muller JE, Parker JO and Stone PH. Morning increase
in ambulatory ischemia in patients with stable coronary artery disease: Importance of physical 16
activity and increased cardiac demand. Circulation 1994;89:604-614.
Example 3: Continuous Glucose
Monitoring in Diabetes
Continuing Glucose
Monitoring Systems
Each colored line represents 5-minute glucose samples
for a different day of the week.
17
Intensively sampled data can arise from
many sources during the same clinical study
Continuing Glucose
Monitoring Systems
Glucose Meter
E-Diary
18
Example 4: Pharmacokinetics
• Pharmacokinetics provides good general
framework for the family of models which
involves extracting parameters
representative of biological processes
– Drug absorption, distribution, metabolism and
excretion
– Intensity and duration of therapeutic and toxic
effects of many drugs are closely related to
their biological availability and disposition
19
Example 4: Plasma Concentration of Drug
after Oral Administration
5
Elimination phase
4
3
2
1
Absorption phase
0
-1
-10
0
10
20
30
40
Time in Hours
20
Steps in Analysis
1. Collect raw signal data (e.g., heart rate, glucose,
plasma concentration) and transfer to relational
database for estimation of parameters
2. Estimate signal parameters (e.g., heart rate
variability, glucose variability, pharmacokinetic rate
constants) using analytical programs
3. Use estimated parameters as dependent measures
for prediction of health outcome or mortality
(Exposed vs Unexposed), or determine how
treatment (e.g., beta blocker) changes signal and
how that change impacts health outcome, clinical
event or mortality (Experimental vs. Control)
21
Disease
Hypertension
Authors
Guzzetti
1991
Langewitz
1994
Saul 1998
Heart failure
NYHA III &IV
Cardiomyopathies
Sudden death-heart attack
Ventricular arrhythmias
Study
Population
Methods
49 with
hypertension
versus 30 controls
34 with
hypertension vs
54 controls
Autoregressive
modeling (AR)
Fast Fourier
transformation
(FFT)
25 with heart
failure vs 21
controls
10 with heart
failure vs 10
controls
12 with heart
failure
Statistical
methods
4 minutes FFT
FFT and
statistical methods
Counihan
1993
104 patients with
myo-cardiopathy
FFT and
statistical methods
Algra 1993
193 survivors vs
230 controls
22 survivors vs 22
controls
Statistical
methods in 24
recordings
Autoregressive
modelling in 24
hour Holter
Biknley 1991
Townend
1992
Huikuri 1992
Huikuri 1993
18 patients with
ventricular
fibrillation
Autoregressive
modelling in 24
hour Holter
recordings
Clinical
Findings
 LF in hypertension,  HF
component and loss of
circadian variation
(both studies)
Low HRV
 HF ( 0,1 Hz)
 LF/HF
↑ HRV with treatment with
inhibitors of converting
activation enzyme (ACEs)
 HF ( 0,1 Hz)
↓HRV induces ↑in mortality
by a factor of 2.6
↓ HF in survivors
↓ of all HRV components
before the arrhythmic episode
22
Estimating HRV Parameters
Hear Rate Variability (HRV)
Adapted from Goldberger AL. Fractals dynamics in physiology: Alterations with disease and aging. PNAS 2002; 99:
2466-2472, downloaded from www.physionet.org.
23
HRV: Time-Domain Methods
• Based upon beat-to-beat or RR
intervals
– SDRR: standard deviation (SD) of RR
intervals over 24 hours
– SDARR: SD of average RR intervals
calculated over short periods ( 5 mins)
– RR50: number of pairs of successive RRs
that differ by more than 50 minutes.
24
HRV: Frequency-Domain Methods
• Fast Fourier transform
• High Frequency band (HF) between 0.15
and 0.4 Hz. HF is driven by respiration and
appears to derive mainly from vagal activity
(parasympathetic nervous system).
• Low Frequency band (LF) between 0.04
and 0.15 Hz. LF derives from both
parasympathetic and sympathetic activity and
has been hypothesized to reflect the delay in
the baroreceptor loop.
25
HRV: Frequency-Domain Parameters
• Fast Fourier transform
• Very Low Frequency band (VLF) band between
0.0033 and 0.04 Hz. The origin of VLF is not well
known.
• Ultra Low Frequency (ULF) band between 0 and
0.0033 Hz. The major background of ULF is day–
night variation and therefore is only expressed in 24hour recordings.
• The ratio of low-to-high frequency spectra
power(LF/HF) has been proposed as an index of
sympathetic to parasympathetic balance of heart rate
fluctuation, but this is controversial because of the
lack of understanding of the mechanisms for the LF
26
component.
HRV: Non-linear Methods
• Poincaré plot. Each data point represents a
pair successive beats, the x-axis is the
current RR interval, while the y-axis is the
previous RR interval.
• HRV is quantified by fitting mathematically
defined geometric shapes to the data.
• Other methods used are the correlation
dimension, nonlinear predictability, point wise
correlation dimension and approximate
entropy.
27
Poincaré plot
The abscissa represents
the RR interval of the
current normal beat and
ordinate represents the RR
interval of the succeeding
normal beat.
An ellipse is fitted to the
data points and the
Poincaré plot indices are
calculated by estimating
the short diameter (SD1),
the long diameter (SD2)
and the ratio of the short
and long diameters
(SD1/SD2 ratio) of the
fitted ellipse
28
Pharmacokinetic Processes
• Liberation – the release of the drug from its dosage
form
• Absorption – the movement of drug from the site of
administration to the blood circulation
• Distribution – the process by which the drug diffuses
or is transferred from intravascular space to
extravascular space (body tissues)
• Metabolism – the chemical conversion of drugs into
compounds that cab be eliminated
• Excretion – the elimination of unchanged drug or
metabolite from the body via renal, biliary, or
pulmonary processes.
29
Elimination Constant
First order elimination, rate is proportional to concentration.
The elimination rate constant Kel represents the portion of
the drug eliminated per unit time.
30
Elimination Constant
(Log scale)
The slope of the line of the concentration plotted on the
log scale correlates with Kel.
Kel = ln(Peak/Trough)/time (P-T))
31
First Order Process
T = 0, C = 100
dC
Loss from 1 to 2 is
proportional to C
L(2, 1)
dt
First order
rate constant
SIDE A
SIDE B
COMP 1
COMP 2
32
Calculation of Parameters
33
How do you estimate parameters?
There are several software
packages that can be used to
estimate parameters – such as
those from
www.adinstruments.com as
34
shown here.
How do I estimate parameters?
There are several software
packages that can be used to
estimate parameter – such as
those from
www.adinstruments.com as
35
shown here.
Pharmacokinetic
Analysis
Software
Several different
packages may be
used.
e.g.,(shown)
http://www.summitpk.com/
36
www.physionet.org
NIH has a data archive
and free software.
37
What is Physionet?
• NIH-sponsored Research (Harvard, BU,
McGill) established in 1999
• Freely available physiologic data and
open-source software
• PhysioBank: 4000 recordings of
digitized physiologic signals and time
series, over 40 databases
• PhysioToolkit: Open source software
38
Physionet Tutorials and Data
http://www.physionet.org/tutorials/hrv/
39
Continuous Glucose Monitoring
(CMG)
40
Continuous Glucose Monitoring
(CGM)
41
Data for Sample Patient – 4 Days
Is the “mean” the best way to
summarize these data?
42
Data for Sample Patient – Session Week 12-- there are many
parameters that could be estimated for each subject
43
Summarize the Raw Data
• The individual daily curves should be
summarized to obtain signal parameters
meaningful to the research objectives
• Examples
–
–
–
–
–
Mean, Max, Minimum for each day
Percent > 180 mg/dl (hyperglycemia)
Percent < 36 mg/dl (severe hypoglycemia)
Intraday standard deviation (glucose variability)
Area above and below defined thresholds
44
Simple Numeric Transformations
/BREAK=Patient_ID by CGMS_num by Date by Nocturnal
/Sensor_Glucose = NU(Sensor_Glucose)
Data Reduction from
/Sensor_1 = MEAN(Sensor_Glucose)
/Sensor_2 = MEDIAN(Sensor_Glucose)
1000’s to only 15
/Sensor_3 = SD(Sensor_Glucose)
measures per subject
/Sensor_4 = MIN(Sensor_Glucose)
– all representing a
/Sensor_5 = MAX(Sensor_Glucose)
/Sensor_6 = PGT(Sensor_Glucose 140)
different parameter of
/Sensor_7 = PLT(Sensor_Glucose 70)
the CGMS profile
/Sensor_8 = PGT(Sensor_Glucose 180)
curve
/Sensor_9 = PLT(Sensor_Glucose 60)
/Sensor_10 = PLT(Sensor_Glucose 50)
Code Shown – using
/Sensor_11 = PGT(Sensor_Glucose 300)
/Sensor_12 = MEAN(Sens_gluHI)
functions from a
/Sensor_13= MEAN(Sens_gluLO)
common statistics
/Sensor_14 = SD(Sens_gluHI)
package or Excel.
/Sensor_15= SD(Sens_gluLO)
45
More Sophisticated Modeling Techniques:
Fourier Series
Start with a sine wave:
Build a model using Fourier
Series
The theory of Fourier series lies in the idea that most signals, can be
represented as a sum of sine waves
46
CGM Daily Measures
•
•
•
•
•
•
Mean Glucose (24-hour, day-time, nocturnal)
Mean Glucose Standard Deviation
Mean amplitude glucose excursions (MAGE)
Low blood glucose index (LBGI)
High blood glucose index (HBGI)
AUC of BG < 70 mg/dL (3.9 mmol/L) and < 50
mg/dL (2.8 mmol/L)
• Nocturnal hypoglycemia – measures < 36,
50, or 70 mg/dL during late night and early
morning (sleep time)
47
CGM Post-Prandial Measures
•
•
•
•
•
•
•
•
•
•
•
Some summary
Meal Interval Start Glucose
parameters may be
Meal Interval Start Time
in response to
Pre-Meal Insulin Dose
meals.
Meal Type
Glucose (C0 (mg/dl), Time (0)
Glucose Cmax (mg/dl), Glucose Tmax (min),
Glucose (Cmax - C0), Glucose (Tmax - T0),
Glucose Cmin (mg/dl - trough)
Glucose Tmin (min)
Glucose (Cmax – Cmin )
Glucose Upstroke (Appearance Rate)
Glucose Downstroke ( Elimination Rate)
48
Data for Sample Patient
• The patient had three sessions of
continuous glucose monitoring with each
session lasting several days.
• Below are the overall mean glucoses for
each of the sessions
Case
100000.0
100000.0
100000.0
Week
0
12
24
Initials Mid Interval Date Glucose
XYZ
XYZ
XYZ
20-OCT-2009
15-JAN-2010
06-APR-20`0
185.33
133.63
133.90
49
Graph of Mean Glucose at Weeks 0, 12
and 24 for Patient 100000
190
180
Mean of Sensor_Glucose
170
160
150
140
130
120
1
2
3
CGMS-Session
0
12
24
Weeks
50
Number of Glucose Values
15 patient feasibility study
Each patient is measured
during 3 session (Week 0, 12
and 24). Each session lasts r
3 – 5 days with measures
taken every 5 minutes
yielding a maximum of 288
values per day.
Clinic 1 ID 200000’s
Clinic 2 ID 400000’s
What is the mean glucose,
glucose variability and hyper
and hypoglycemia
parameters for the subjects at
Week 12?
There are a total of 13,050
glucose measures for 15
patients.
51
15 patient feasibility study
The 13,050 glucose measures
for these 15 patients are
reduced to 4 summary
parameters for each patient -yielding 60 summary parameters
in total for the 15 patients.
Summary Parameters
1. Mean Glucose
2. Glucose Variability (SD
Glucose)
3. Percent values > 140 mg/dL
(hyperglycemia)
4. Percent values < 70 mg/dL
(hypoglycemia)
52
Glycemia CGM Parameter
Estimates at Week 12
Here we summarize the parameters for the 15
subjects.
In the next session we will learn how to construct
confidence intervals and develop different
hypotheses for these measures.
53
Summary
• Identified the types of clinical research
studies requiring analytical methods for
complex data signals and parameter
estimation
• Reviewed various analytical techniques and
software packages for obtaining clinical
physiology and pharmacologic methods
• Introduced examples in cardiology (HRV) and
CGM (diabetes) where such techniques are
useful
54