Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1 diabetes mellitus.
ABSTRACT The aim of this study was to test whether a new heart rate variability (HRV) complexity measure, the Point Correlation Dimension (PD2i), provides diagnostic information regarding early subclinical autonomic dysfunction in diabetes mellitus (DM). We tested the ability of PD2i to detect diabetic autonomic neuropathy (DAN) in asymptomatic young DM patients without overt neuropathy and compared them to age and gendermatched controls.
HRV in DM type 1 patients (n=17, 10 female, 7 male) aged 12.931.5 years (duration of DM 12.4±1.2 years) was compared to that in a control group of 17 healthy matched probands. The RR intervals were measured over 1h using a telemetric ECG system.
PD2i was able to detect ANS dysfunction with p=0.0006, similar to the best discriminating MSE scale, with p=0.0002.
The performance of PD2i to detect DAN in asymptomatic DM patients is similar to the best discriminative power of previously published complexity measures.
The PD2i algorithm may prove to be an easy to perform and clinically useful tool for the early detection of autonomic neuropathy in DM type 1 patients, especially given its minimal data requirements.

Conference Paper: PD2i Heart Rate Complexity Measure can Detect Cardiac Autonomic Neuropathy: an Alternative Test to Ewing Battery
Ahsan H Khandoker, Daniel N Weiss, James E Skinner, Jerry M Anchin, Md Hasan Imam, Herbert F Jelinek, Marimuthu Palaniswami[Show abstract] [Hide abstract]
ABSTRACT: This study evaluates the usefulness of a new heart rate variability (HRV) complexity measure, the Point Correlation Dimension (PD2i), derived from short term ECG recordings, as a screening tool for Cardiac autonomic neuropathy (CAN). The PD2i was developed to measure complexity in nonstationary data with some tolerance for background noise. ECG recordings during supine rest were acquired from diabetic subjects with CAN (CAN+) [10 subjects] and without CAN (CAN) [33 subjects] and analyzed. PD2i indices (mean, standard deviation, minimum and maximum) were used for analyzing HRV signals of all subjects. Significantly reduced (p < 0.01) PD2i indexes were found in CAN+ group, which could be a practical diagnostic and prognostic marker.Computing in Cardiology, Hangzhou, China; 09/2011
Page 1
Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy
in patients with type 1 diabetes mellitus
James E. Skinnera, Daniel N. Weissa,⇑, Jerry M. Anchina, Zuzana Turianikovab, Ingrid Tonhajzerovab,
Jana Javorkovac, Kamil Javorkab, Mathias Baumertd, Michal Javorkab
aVicor Technologies, Inc., Boca Raton, FL, USA
bDepartment of Physiology, Jessenius Faculty of Medicine, Comenius University, Martin, Slovakia
cClinic of Children and Adolescents, Martin Teaching Hospital, Martin, Slovakia
dSchool of Electrical & Electronic Engineering, The University of Adelaide, Adelaide, Australia
a r t i c l ei n f o
Article history:
Accepted 18 December 2010
Available online xxxx
Keywords:
Autonomic neuropathy
Heart rate variability
Complexity
Diabetes mellitus
Nonlinear methods
PD2i
Point Correlation Dimension
Algorithms
a b s t r a c t
Objective: The aim of this study was to test whether a new heart rate variability (HRV) complexity
measure, the Point Correlation Dimension (PD2i), provides diagnostic information regarding early
subclinical autonomic dysfunction in diabetes mellitus (DM). We tested the ability of PD2i to detect
diabetic autonomic neuropathy (DAN) in asymptomatic young DM patients without overt neuropathy
and compared them to age and gendermatched controls.
Methods: HRV in DM type 1 patients (n = 17, 10 female, 7 male) aged 12.9–31.5 years (duration of DM
12.4 ± 1.2 years) was compared to that in a control group of 17 healthy matched probands. The R–R
intervals were measured over 1 h using a telemetric ECG system.
Results: PD2i was able to detect ANS dysfunction with p = 0.0006, similar to the best discriminating MSE
scale, with p = 0.0002.
Discussion: The performance of PD2i to detect DAN in asymptomatic DM patients is similar to the best
discriminative power of previously published complexity measures.
Conclusions: The PD2i algorithm may prove to be an easy to perform and clinically useful tool for the
early detection of autonomic neuropathy in DM type 1 patients, especially given its minimal data
requirements.
? 2010 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights
reserved.
1. Introduction
Diabetic autonomic dysfunction, with or without neuropathy, is
one of the least understood complications of diabetes mellitus
(DM). The neuropathy has a negative impact on the survival and
quality of life (Vinik and Erbas, 2001) as it is associated with fatal
and nonfatal cardiovascular events (Liao et al., 2002; Whang and
Bigger, 2003), ischemic cerebrovascular events (Toyry et al.,
1996) and overall mortality (Wheeler et al., 2002). Cardiovascular
autonomic neuropathy (CAN) may be the most clinically important
form of diabetic autonomic neuropathy (Vinik et al., 2003) because
of its link to arrhythmic death. Early detection of subclinical auto
nomic dysfunction in diabetic patients is, therefore, of vital impor
tance for risk stratification and management for the prevention of
serious adverse events (Schroeder et al., 2005).
A study on young DM patients by Javorka and associates
(Javorka et al., 2008) indicated that new measures of heart rate
variability (HRV) that assess its complexity, as opposed to its
statistical parameters and power spectra, provide additional diag
nostic information regarding early subclinical autonomic dysfunc
tion. Specifically, normaltonormal R–R intervals were assessed in
DM patients and in their matched controls by the more routine
linear stochastic algorithms (Standard Deviation, SDNN; and
power spectrum, HF/LF) and the new nonlinear complexity
algorithms: Multiscale Entropy (MSE), compression entropy and
various symbolic dynamic measures. The two types of measures,
linear and nonlinear, were found to be uncorrelated with one an
other, an indication that they were sensitive to different types of
information in the data, and the nonlinear ones were far more sta
tistically significant between the DM and control groups, an indica
tion that the information they captured was more discriminating.
The early complexity and entropy algorithms (Hausdorf Dimen
sion, Correlation Dimension, Largest Lyapunov Exponent) required
data with low noise that remained stationary over long data
lengths, requirements not met by physiological data (Schreiber,
1999). Nonstationarity and noise are problems for nonlinear
algorithms, as any noise in the data becomes amplified and a
nonstationarity in the data obfuscates the algorithmic result
(Elbert et al., 1994).
13882457/$36.00 ? 2010 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.
doi:10.1016/j.clinph.2010.12.046
⇑Corresponding author. Tel.: +1 561 995 7313.
Email address: dweiss@vicortech.com (D.N. Weiss).
Clinical Neurophysiology xxx (2011) xxx–xxx
Contents lists available at ScienceDirect
Clinical Neurophysiology
journal homepage: www.elsevier.com/locate/clinph
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046
Page 2
The Point Correlation Dimension (PD2i) was developed to
measure complexity in nonstationary data with some tolerance
for background noise. As a complexity measure it determines the
degrees of freedom, or a number of independent variables operat
ing at each point in time to produce the data and will track those
changes with only a small (4%) error (Skinner et al., 1994). That
unique feature is why, in comparison to other complexity and
entropy algorithms developed to work on shorter data lengths,
the minimum PD2i value was far more predictive of future lethal
arrhythmogenesis in highrisk cardiac patients (Skinner et al.,
1993, 2008, 2009). In addition, in contrast to the other nonlinear
methods that require long, relatively noisefree ECG data, the
PD2i algorithm has been found to perform effectively on even rel
atively short, noisy data (Batchinsky et al., 2008, 2009, 2010).
Greater detailed explanation of the PD2i algorithm has been
published previously (Skinner et al., 1993).
For the present study, PD2i analysis was performed on the same
dataset used by Javorka et al. to discriminate between young DM
patients and their agematched controls. That study had used a
variety of complexity and traditional algorithmic measures to
assess the R–R intervals of each subject. The hypothesis now tested
is that the PD2i algorithm is at least as good in its discrimination
between the DM and controls as the MSE algorithm, where the
latter algorithm, at least at one scaling length, was found to be
the best discriminator with the smallest pvalue.
2. Methods
The same patient cohort as used in Javorka and associates was
reanalyzed. Briefly, there were 17 patients with type 1 DM (10
women, 7 men) aged 12.9–31.5 years (mean ± SEM was 22.4 ±
1.0 years), mean duration of disease 12.4 ± 1.2 years, and a control
group consisting of 17 healthy gender and agematched subjects
(mean age: 21.9 ± 0.9 years). Only one diabetic patient showed
clinical symptoms of autonomic dysfunction (orthostatic intoler
ance); mechanisms other than cardiac autonomic neuropathy were
not excluded. The Michigan Neuropathy Screening Instrument, a
history and physical assessment (foot sensation), did not reveal
peripheral neuropathy in any patient, although one subject showed
borderline values of suspected neuropathy. Physical examination
(predominantly foot inspection) showed excessively dry skin in
one subject. No other abnormalities were observed. Additionally,
vibration sensation was tested using a graduated tuning fork
(128 Hz) applied to the dorsum of the patient’s large toe. A reduced
vibration sensation was found in one subject. Ankle reflexes were
bilaterally present in all subjects. At the end of the physical exam
ination, standard monofilament sensation testing was performed
at a pressure of 10 g on 10 separate places on both feet. All DM
patients showed correct responses to these stimuli.
The clinical data for both groups are given in Table 1. Seven sub
jects (four in the control group, three in the DM group) were mild
smokers with daily number of cigarettes smoked being less than
five. All subjects were instructed not to use substances which influ
ence activities of the cardiovascular system (caffeine, alcohol) and
to refrain from smoking 12 h before examination.
Table 1
Values are presented as medians (interquartile range); pvalues were obtained using
Mann–Whitney Utest. Clinical data: matched controls (CON); patients with type1
diabetes mellitus (DM).
Clinical data CON DM
p
Age (years)
Body mass index (kg m?2)
Plasma glucose (mmol?1)
HbAlc(%)
Systolic BP (mmHg)
Diastolic BP (mmHg)
Duration of DM (years)
Age at DM diagnosis
Ankle reflexes
Toe vibration sensitivity
Neuropathy screen
22.1 (20.3–24.2)
20.8 (19.5–23.6)
4.9 (4.6–5.1)
4.7 (4.5–5.0)
124 (113–132)
74 (64–76)
–
–
Normal in 17/17
Normal in 17/17
Negative in 17/17
23.3 (20.8–24.1)
21.8 (21.4–24.6)
8.6 (6.4–24.6)
9.2 (8.6–9.9)
117 (116–122)
71 (68–77)
12.9 (10.4–14.2)
9.4 (8.2–11.6)
Normal in 17/17
Normal in 16/17
Negative in 17/17
ns
0.033*
0.001*
0.001*
ns
ns
–
–
ns
ns
ns
*Significant (p < 0.05) betweengroup differences.
Fig. 1. Nonstationary data (upper) made by linking sine (S), Lorenz (L), Henon (H) and random (R) subepochs of 1200 data points each and their corresponding accepted PD2i
(lower). (A) Simulated R–R data are adjusted to equal the mean of the actual RR scores of the subjects (900 ms). (B) Simulated R–R are divided by 4. PD2i values are exactly the
same for (A) and (B). Each 1200 point mean PD2i for the different subepochs of data are within 4% of the known D2 values for infinite data length, as previously reported by
Skinner et al. (1994).
2
J.E. Skinner et al./Clinical Neurophysiology xxx (2011) xxx–xxx
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046
Page 3
2.1. Study protocol
As described in Javorka, all subjects were examined over a 1h
period under standardized conditions in a quiet room between
the hours of 8 and 12 AM. The subjects were instructed to lie com
fortably in a supine position and not to speak or move unnecessar
ily. To ensure the subject’s quiet state the data recording was
supervised by an examiner and a nurse. The subjects were rested
in the supine position for 20 min before the heart rate recording
started, allowing the cardiovascular system to reach the equilib
rium, i.e., a quasistationary condition. The patients and controls
were asked to neither move nor speak during data acquisition.
The VariaCardio TF4 device (Sima Media, Olomouc, Czech
Republic) was used to measure continuous beattobeat heart rate
(R–R intervals) recorded at a sampling frequency of 1000 Hz, using
a bipolar thoracic ECG lead. All ECG traces were visually scanned
for artifacts.
2.2. Data analysis
The Vicor PD2i Analyzer (Vicor Technologies, Inc., Boca Raton,
FL, USA) was used to calculate all PD2i values from the dataset
of R–R interval provided to Vicor in a coded blinded manner by
Dr. M. Javorka. All calculations were made without the knowledge
of the clinical data. Following a mathematical proof by Takens
(1981), Grassberger and Procaccia (1982) developed a computer
ized algorithm called the Correlation Dimension and abbreviated
as D2. This algorithm measures the number of degrees of freedom
in the data stream and is accomplished by a reconstruction from
the data themselves. The PD2i, or ‘‘Point Correlation Dimension,’’
is an algorithm based on the same D2 model, but with some
notable differences, (Skinner, 1998a,b) which allows it to estimate
the degrees of freedom as a function of time (i) but, unlike D2, it
can be used in the data with any number of nonstationarities in
it – the PD2i tracks the changes in the degrees of freedom in time.
2.3. Statistics
The statistics used to examine the PD2i results were Student’s
ttest for independent samples to determine the alpha level and
the Fleiss table (A3) for determining the power (Fleiss, 1981). For
the statistics used for examining the results of the other measures
see Javorka et al. (2008).
3. Results
3.1. Patient characteristics
Table 1 shows the clinical data for the DM and control groups.
Note that the DM patients are young (mean age is 22 years) and
do not show signs of neuropathy (Michigan Neuropathy Screen
and ankle reflexes are normal in 17 of 17 DM subjects and toe
sensitivity to vibration is normal in 16 of 17 DM subjects).
3.2. PD2i analysis
Fig. 1 shows the accepted PD2i values for finite nonstationary
subepochs of simulated R–R interval data with 1200 data points
each, adjusted to a mean for normal R–R intervals (900 ms). The
mean degrees of freedom for each subepoch are known (D2, using
long data length) and the mean for each subepoch was found to be
within 4% of the known published values. Thus, the software used
Fig. 2. Reason for dividing each R–R datapoint by 4. In (A) the standard deviation (Std, arrow) of the 20 point mean is NOT less than the noise tolerance level, whereas in (B),
where the R–Rs are divided by 4, the Std (arrow) is less than the noise tolerance level. The three horizontal lines are the mean and plus and minus the standard deviation for
the 20 point window.
J.E. Skinner et al./Clinical Neurophysiology xxx (2011) xxx–xxx
3
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046
Page 4
to calculate the PD2i values (Vicor Technologies PD2i Analyzer)
reproduces previously published results.
Fig. 2A shows the raw R–R data for a typical control subject. The
standard deviation for a 20 heartbeat window (Std, arrow) with a
horizontal mean is larger than that for the noisetolerance level
of the PD2i algorithm (i.e., Std < ± 5 integers (Skinner, 1998a,b,
2007b; Skinner et al., 1993)). In Fig. 2B each R–R interval was
divided by 4, as this amplitude reduction brought the standard
deviation (Std) below the noisetolerance level.
Fig. 3 shows the accepted PD2i results for the typical control
and DM subjects in which the R–R intervals were divided by 4.
Note in Fig. 3B the nonstationary increase in the R–R values toward
the end of the 3200 point data period. Even with this nonstationa
rity, the PD2i continued to show many values below the criterion
line of 2.0 (horizontal bar). This criterion line was determined by
an ROC plot that maximized Sensitivity and Specificity for the 34
subjects. The control subject (Fig. 3A) showed, in contrast, all
accepted PD2i points to be above the criterion line.
Table 2 shows the primary results of the Minimum PD2i for the
DM and control groups. These highly significant results (p < 0.0006,
ttest) are adequately powered (1 ? b > 90%). All PD2i were
assessed in R–R data from the previous study in which the R–R
intervals were divided by 4 to reduce noise.
Table 3 shows the pvalues for the minimum PD2i and those
resulting from the other HRV measures found in the previous study
to be statistically significant. The analyses were of the normalto
normal R–R intervals made from the 1h ECG recordings for each
of the subjects in the DM (n = 17) and matched the control
(n = 17) groups. The minimum PD2i for the first 1000 R–R intervals
(approximately 15 min) correlated well with those for the 3200
R–R intervals (approximately 1 h); the Pearson Correlation Coeffi
cient was equal to 0.81.
Table 4 shows the contingency table results for the PD2i Test.
TheSensitivity andSpecificity
significant. Using the Fleiss (1981) tables, the alphalevels
(a 6 0.01)were foundtobe
1 ? b > 90%).
were highand statistically
adequatelypowered (with
Fig. 3. The R–Rs (upper) and corresponding accepted PD2i (lower) are shown for a typical control (CON) and diabetic (DM) subject. (A) All accepted PD2i are above the 2.0
level (horizontal bar). (B) The majority of the accepted PD2i is below the 2.0 level. It is the minimum accepted PD2i found anywhere in the 3200 data points that is the score
used for the various subjects (Min PD2i).
Table 2
Minimum PD2i for DM and control groups. DM = Dia
betic patients CON = controls.
Min PD2i
DM**
Min PD2i
CON**
0.50
0.50
0.64
0.69
0.84
0.98
0.99
1.24
1.35
1.36
1.38
1.72
1.79
3.01
1.81
1.81
1.87
0.96
1.22
1.23
1.34
1.38
2.05
2.26
2.27
2.50
2.53
2.53
2.61
2.65
2.75
3.06
3.17
3.30
**p 6 0.0006.
4
J.E. Skinner et al./Clinical Neurophysiology xxx (2011) xxx–xxx
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046
Page 5
4. Discussion
The major result of our study is that the minimum PD2i mea
sure is able to successfully detect autonomic dysfunction in young
diabetic patients, and far earlier than by the Ewing maneuvers. The
PD2i has the same order of magnitude for its pvalue as that of
Multiscale Entropy Scale 3 (but not Scales 1, 2 or 4), which had
previously performed the best. Since MSE is known to require
relatively long data lengths and noisefree data (Batchinsky et al.,
2008, 2010), whereas PD2i does not, this suggests that PD2i may
be a clinically useful tool, which can be more easily applied to
realworld ECG data than other algorithms for the detection of
autonomic dysfunction.
When testing a null hypothesis it is not permitted to use the
pvalue as a discriminability measure; it is either significant or
nonsignificant according to a given a priori acceptance level. In
the current cases, however, each pvalue is not being used to test
a null hypothesis, so its use as a discriminability measure is
acceptable.
Without an a priori hypothesis, the pvalues for the consecutive
Multiscale Entropy (MSE) measures should perhaps be averaged;
there are no data concerning which scale is to be used for the
coarsegraining of successive R–R intervals, and it is not appropri
ate to choose the best scale after the analyses are made. For a scale
of one, however, the score is the same as Approximate Entropy
(Sample Entropy), a measure which has been shown not to discrim
inate between the control and DM groups (Bettermann et al.,
2001). Eliminating Scale 1 still leaves an average pvalue of
p = 0.001 for the three significant MSE scales.
Neither the traditional linear stochastic measures (time and
frequency domains)northe
measures that were significant (Table 3) had the same order of
pvalues as those for the PD2i or MSE Scale 3. These other
measures may perhaps lose too much critical information in the
way they collapse the data or alternatively they may depend on
longer data length. For example, in a previous comparison of
PD2i to other algorithms (ApEn, Detrended Fluctuation Analysis,
1/f Spectra, SDNN, HF, LF), it was noted that in their original
publications each of the other algorithms used a 24h data length
(Holter tapes) and perhaps that is why they were statistically
significant; but still these other algorithms were not significant
in comparison to the PD2i when all algorithms analyzed the same
set of 1000 datapoint files. The high correlation in the current
study between the minimum PD2i for 1000 and 3200 datapoint
files suggests that the PD2i may also capture the DM disease state
within smaller data lengths.
The contingency table in Table 4 is the more conventional way
of determining the medical significance of the PD2i Test, and it
turns out to be highly statistically significant (p 6 0.0006, Fischer’s
Exact test) with a high Sensitivity (94%), Specificity (71%), Negative
Predictive Value (92%), Positive Predictive Value (76%) and Relative
Risk. Such high statistical significance has also been found for the
PD2i (p = 0.00016) in discriminating among Emergency Depart
ment chestpain patients who will and will not manifest lethal
arrhythmias within 1year of followup (Skinner et al., 2009).
The early discovery by our laboratory that the same cerebral
loci that control vulnerability to lethal arrhythmogenesis also
control heart rate (Skinner, 2007a) is the rationale behind our
original use of HRV analysis as a way to detect and stratify risk
of arrhythmic death. The link between cardiovascular events and
diabetes (Toyry et al., 1996; Wheeler et al., 2002) leads to the
current rationale that the PD2i may be sensitive to cardiac risk in
early DM patients. But the PD2i is reduced in DM patients without
signs of cardiac effects (this study) and in cardiac patients without
nonlinearsymbolicdynamics
Table 3
Use of the probability (p) value for the ttest (twotailed, equal variances) as a measure of discriminability between the DM group and the matched control group (i.e., a measure
of the mean difference/joint variances) for the various traditional and complexity algorithms found to be statistically significant in the same dataset (for more details see Javorka
et al., 2008).
Algorithm
pvalue#
Type of measure
Linear stochastic models
Mean NN
SDNN
RMSSD
LF
HF
p = 0.049
p = 0.009
p = 0.006
p = 0.016
p = 0.011
Statistical property (time domain)
Statistical property (time domain)
Statistical property (time domain)
Oscillatory property (frequency domain)
Oscillatory property (frequency domain)
Nonlinear complexity models
PD2i min
p = 0.0006 Degrees of freedom (number of independent variables, as a function of time; insensitive to nonstationarity)
Forbidden words
Shannon Entropy
Renyi Entropy 4
Renyi Entropy 0.25
Compression Entropy
p = ns
p = 0.073
p = 0.026
p = 0.099
p = 0.0013
Symbolic dynamics, very low probability words (<0.001)
Symbolic dynamics, all words (low to high)
Symbolic dynamics, high probability words
Symbolic dynamics, low probability words
Symbolic dynamics, similar to data compression
Multiscale Entropy (S1)
Multiscale Entropy (S2)
Multiscale Entropy (S3)
Multiscale Entropy (S4)
Multiscale Entropy (S5)
p = ns
p = 0.001
p = 0.0002
p = 0.002
p = ns
Irregularity, conditional probability (ApEn, SampEn)
Irregularity, conditional probability, 2 R–Ri grains
Irregularity, conditional probability, 3 R–Ri grains
Irregularity, conditional probability, 4 R–Ri grains
Irregularity, conditional probability, 5 R–Ri grains
#Unpaired ttest results for PD2i algorithms: the twotailed pvalue for the PD2i Test equals 0.0006; the mean of the DM group minus the control group equals ?0.9018,
where the 95% confidence interval of this difference is from ?1.3870 to ?0.4165 (t = 3.7856, df = 32, standard error of difference = 0.238). All other pvalues were calculated
the same way, using different algorithmic outputs.
Table 4
Contingency table for the PD2i Test (Positive = Min PD2i < 2, Negative = Min
PD2i P 2) to predict disease state (DM = Positive, Control = Negative). Sensitivity,
Specificity, Negative Predictive Value, Positive Predictive Value and Relative Risk are
statistically significant (p < 0.01) by the Fisher Exact Test for contingency tables.
FN = false negative FP = false positive TN = true negative TP = true positive.
Disease (+) Disease (?)
5
12
Total
Test (+)
Test (?)
Total
16
1
21
13
17 17
Sensitivity = TP/(TP + FN) = 94%.
Specificity = TN/(TN + FP) = 71%.
Negative Predictive Value = TN/(TN + FN) = 92%.
Positive Predictive Value = TP/(TP + FP) = 76%.
Relative Risk = TP/FN ? (TP + FN)/(FN + FP) = 16.
J.E. Skinner et al./Clinical Neurophysiology xxx (2011) xxx–xxx
5
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046
Page 6
signs of DM (Skinner, 2007b), so the two variables are doubly
dissociated and are thus not likely to be causally related. The
common feature of both variables is that they are associated with
the reduction of PD2i to low levels (PD2i 6 1.4 for arrhythmic
death; PD2i 6 2.0 for DM) and with high discriminability from
the controls (p = 0.00016 for arrhythmic death; p = 0.0006 for DM).
The reduced PD2i has been interpreted as the result of cooper
ation (phase coherency) among the various independent regulators
of the heartbeats that lie in the brain (Skinner et al., 2009). These
cerebral regulatory centers each have their own visceral afferents
but the same medullary autonomic efferents and each is an inter
connected sensorymotor loop that ranges hierarchically from the
heart (intrinsic cardiac nervous system) through the brainstem
(baroreflex and respiratory reflex) and hypothalamus (temperature
and pH regulation) to the amygdalae and mesofrontal cortices
(cerebral defensive mechanisms) (Skinner et al., 1993). Together
these interconnected loops, through their integrated autonomic
projections, are thought to underlie the normal sinus rhythm and
its degrees of freedom. This autonomic output may change in a
nonstationary manner from momenttomoment. These phasic
and/or tonic autonomic changes appear to control cardiac vulner
ability (Skinner et al., 1993, 2009) and would now appear to play
some as yet unknown role in DM.
The results show the PD2i algorithm is at least as good as or
better than MSE Scale 3 in detecting autonomic dysfunction in
DM patients without overt, symptomatic neuropathy.
The current study was performed on a small group of young DM
patients without neuropathy and their age and gendermatched
controls and may require additional validation. The intention was
to explore the PD2i as a measure of HRV complexity along with
the others used in our previous study. Based on the current data
we cannot conclude whether or not complexity measures allow a
detection of autonomic dysregulation earlier than conventional
HRV measures. One might speculate, however, that due to the
reduction in HRV complexity, which is more pronounced than
the reduction in HRV magnitude, the loss of complexity might
precede affects on HRV itself.
5. Conclusion
In summary, the results clearly show that the PD2i Test (min
PD2i 6 2.0) is able to detect autonomic dysfunction very early in
diabetic patients. Given the algorithm’s ease of use and usability
with even short, noisy ECG data, it may prove to be a valuable
clinical tool.
Acknowledgments
This study was supported by the project of Centre of Excellence
Perinatological Research No. 26220120016, Grant VEGA No.
1/0064/08, Grant VVZ MSMT 0021622402 and by the Australian
Research Council (Grant # DP0663345). M.B. holds an Early Career
Researcher fellowship from the Health faculty of the University of
Adelaide. J.E. Skinner, J.M. Anchin and D.N. Weiss are employees of
Vicor Technologies, Inc., and are shareholders.
for
References
Batchinsky AI, Salinas J, Skinner J, Weiss D, Wade C, Holcomb J, Cancio LC.
Exploration of heart rate complexity to determine the need for lifesaving
interventions in combat casualties. J Crit Care 2008;23:265–6.
Batchinsky A, Skinner J, Salinas J, Necsolu C, Nguyen R, Waker K, et al. Does heart
ratecomplexity analysis work on realworld data? The point correlation
dimension of the heart rate and prehospital lifesaving interventions.
Circulation 2009;120:S1441–2.
Batchinsky AI, Skinner J, Necsoiu C, Jordan BS, Weiss D, Cancio LC. New measures of
heartrate complexity: effect of chest trauma and hemorrhage. J Trauma
2010;68(5):1178–85.
Bettermann H, Kroz M, Girke M, Heckmann C. Heart rate dynamics and
cardiorespiratory coordination in diabetic and breast cancer patients. Clin
Physiol 2001;21:411–20.
Elbert T, Ray WJ, Kowalik ZJ, Skinner JE, Graf KE, Birbaumer N. Chaos and
physiology. Physiol Rev 1994;74:1–47.
Fleiss JL. Statistical methods for rates and proportions. 2nd ed. New York (NY): John
Wiley & Sons; 1981. p. 260–80.
Grassberger P, Procaccia I. Characterization of strange attractors. Phys Rev Lett
1982;50:346–9.
Javorka M, Trunkvalterova Z, Tonhajzerova I, Javorkova J, Javorka K, Baumert M.
Shortterm heart rate complexity is reduced in patients with type 1 diabetes
mellitus. Clin Neurophysiol 2008;119:1071–81.
Liao D, Carnethon M, Evans GW, Cascio WE, Heiss G. Lower heart rate variability is
associated with the development of coronary heart diseases in individuals with
diabetes: the atherosclerosis risk in communities (ARIC) study. Diabetes
2002;51:3524–31.
Schreiber T. Interdisciplinary application of nonlinear time series methods. Phys
Rep 1999;308:1–64.
Schroeder EB, Chambless LE, Liao D, Prineas RJ, Evans GW, Rosamond WD, et al.
Diabetes, glucose, insulin, and heart rate variability: the atherosclerosis risk in
communities (ARIC) study. Diabetes Care 2005;28:668–74.
Skinner, JE. PD2i electrophysiological analyzer. US Patent and Trademark Office.
Number 5720294; 1998a.
Skinner, JE. PD2i electrophysiological analyzer. US Patent and Trademark Office.
Number 5709214; 1998b.
Skinner JE. New paradigms in heart–brain medicine: nonlinear physiology and
statedependent proteomics. Cleve Clin J Med 2007a;74:S79–85.
Skinner JE. Role of the brain in ventricular fibrillation and hypertension: from
animal models to early human studies. Cleve Clin J Med 2007b;74:S73–8.
Skinner JE, Pratt CM, Vybiral T. A reduction in the correlation dimension of
heartbeat intervals precedes imminent ventricular fibrillation in human
subjects. Am Heart J 1993;125:731–43.
Skinner JE, Molnar M, Tomberg C. The point correlation dimension: performance
with nonstationary surrogate data and noise. Integr Physiol Behav Sci
1994;29:217–34.
Skinner JE, Meyer M, Dalsey WC, Nester BA, Ramalanjaona G, O’Neil BJ, et al. Risk
Stratification for Arrhythmic Death in an Emergency Department Cohort: a new
method of nonlinear analysis of RRintervals. Therap Clin Risk Manag
2008;4:1–9.
Skinner JE, Meyer M, Nester BA, Geary U, Taggart P, Mangione A, et al. Comparison
of linearstochastic and nonlinear deterministic algorithms in the analysis of
15minute clinical ECGs to predict risk of arrhythmic death. Therap Clin Risk
Manag 2009;5:671–82.
Takens F. Detectingstrangeattractors
1981;1125:99–106.
Toyry JP, Niskanen LK, Lansimies EA, Partanen KP, Uusitupa MI. Autonomic
neuropathy predicts the development of stroke in patients with noninsulin
dependent diabetes mellitus. Stroke 1996;27:1316–8.
Vinik AI, Erbas T. Recognizing and treating diabetic autonomic neuropathy. Cleve
Clin J Med 2001;26:928–44.
Vinik AI, Maser RE, Mitchell BD, Freeman R. Diabetic autonomic neuropathy.
Diabetes Care 2003;26:1553–79.
Whang W, Bigger Jr JT. Comparison of the prognostic value of RRinterval variability
after acute myocardial infarction in patients with versus those without diabetes
mellitus. Am J Cardiol 2003;92:247–51.
Wheeler SG, Ahroni JH, Boyko EJ. Prospective study of autonomic neuropathy as a
predictor of mortality in patients with diabetes. Diabetes Res Clin Pract
2002;58:131–8.
in turbulence. LectNotes Math
6
J.E. Skinner et al./Clinical Neurophysiology xxx (2011) xxx–xxx
Please cite this article in press as: Skinner JE et al. Nonlinear PD2i heart rate complexity algorithm detects autonomic neuropathy in patients with type 1
diabetes mellitus. Clin Neurophysiol (2011), doi:10.1016/j.clinph.2010.12.046