%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% -*- Mode:TeX -*-
\documentclass[twocolumn]{cinc}
\usepackage{epsfig}

\newcommand{\1}{\v{c}} \newcommand{\2}{\v{C}}
\newcommand{\3}{\v{s}} \newcommand{\4}{\v{S}}
\newcommand{\5}{\v{z}} \newcommand{\6}{\v{Z}}
\newcommand{\7}{\aa}   \newcommand{\8}{\`{e}}
\newcommand{\9}{\~{n}} \newcommand{\Q}{\'{c}}

\begin{document}

\bibliographystyle{cinc}

\title{Development of a Long Term Database for Assessing\\
the Performance of Transient Ischemia Detectors}

\author{F Jager$^{1}$, GB Moody$^{2}$, A Taddei$^{3}$, G Antolic$^{4}$,
M Zabukovec$^{1}$, M Skrjanc$^{1}$, M Emdin$^{3}$, and RG Mark$^{2}$\\
\ \\
$^{1}$ Faculty of Computer and Information Science, Ljubljana, Slovenia\\
$^{2}$ Harvard-M.I.T. Division of Health Sciences and Technology, Cambridge,
MA, and\\
Cardiology Division, Beth Israel Hospital, Boston, MA, USA\\
$^{3}$ CNR Institute of Clinical Physiology, Pisa, Italy\\
$^{4}$ Department of Cardiology, University Medical Center, Ljubljana, Slovenia
\vspace{-1em}}

\maketitle

\begin{abstract}
We have begun to develop a new annotated long term ambulatory 
ST-T database. The aim of the database is to be a reference
set containing a number of well documented ischemic ST episodes,
axis-related non-ischemic ST episodes, episodes of slow ST level
drift and mixed episodes to support development and
evaluation of detectors capable of accurate differentiation
of ischemic and non-ischemic ST events, as well as basic research
into mechanisms and dynamics of ischemia.  We discuss selection
criteria, define the events of interest, and describe the annotation
procedure.
\end{abstract}
 
\section{Introduction}

Ambulatory electrocardiographic (AECG) monitoring is widely used for
analysis of transient ST-segment and T-wave changes compatible with
ischemia.  Most AECG instruments do not attempt to distinguish between
ischemic and non-ischemic ST and T changes, however, because of a lack
of standard definitions of transient ST-T events and knowledge about
their meaning.

In order to study these events, and to evaluate and compare automated
methods for their detection and interpretation, the ICP group in Pisa
defined diagnostic criteria for transient ST and T changes, and a
protocol for annotating them \cite{ltst:TA-88}.  This group took the
leading role in the development of the European Society of Cardiology
ST-T Database (ESC DB) \cite{ltst:TA-92}, which was the first generally
available set of well-characterized, representative ECG recordings
with documented ischemic and non-ischemic ST and T changes.  The ESC
DB has proven to be an invaluable tool for designers and evaluators of
automated ischemia detectors.  Its availability has stimulated
extensive research and publication in this field during the past
several years, including recognition algorithms based on
time-domain analysis,
% \cite{ltst:TA-95}
the Karhunen-Lo\8ve Transform (KLT),
%\cite{ltst:JF-92,ltst:PL-96}
neural networks,
% \cite{ltst:RS-94,ltst:NM-96}
and fuzzy logic.
% \cite{ltst:JP-95}

The ESC DB contains 90 two-hour, two-channel ambulatory records with
368 documented transient ischemic ST episodes, but only 11 non-ischemic ST
episodes.  Non-ischemic ST episodes, which are of no clinical interest
per se, account for many of the false positives of automated ischemic
ST detectors.  Thus it is particularly important to understand these
events and to define their distinctive characteristics in order to
improve detector performance.  The small number of non-ischemic
episodes in the ESC DB does not permit exhaustive study of these
differences, however.

Furthermore, our previous study on characterization of transient ST
segment changes in the ESC DB\cite{ltst:JF-95}, revealed two additional
types of important ST events.  We found three cases of ``mixed episodes''
(non-ischemic episodes containing ischemic episodes within), and 17
cases of significant ($>$100$\mu$V) slow drift of ST deviation level
(15 of which also contain ischemic episodes).  We also
described striking and varied temporal patterns of transient ischemic
ST changes.  These observations provoke questions regarding the
relationships between these patterns and the underlying mechanisms
that are responsible for ischemia.  We cannot answer these questions
definitively, however, since we are not able to observe more than a
handful of repetitions of each pattern in the two-hour segments of
the ESC DB.

We have therefore begun to prepare a new, long-term, ambulatory ST database,
in order to support the development and evaluation of detectors
capable of more accurate differentiation of ischemic and non-ischemic
ST changes, and to provide more examples of non-ischemic episodes,
episodes of slow ST level drift and mixed episodes.

\section{Methods}

The Long Term ST Database (LTST DB) is being developed by the joint
efforts of our research groups in Ljubljana, Pisa, and Cambridge.  It
is planned to contain up to 70 annotated 2-channel records, each 24
hours in duration, obtained from AECG recordings.  These will include
approximately 30 of the 90 24-hour recordings from which the two-hour
excerpts in the ESC DB were obtained.  The records have been selected
to represent ``real world'' data as much as possible, while
documenting significant numbers of ischemic and non-ischemic ST
events.  The annotation protocol is compatible with that developed for
the MIT-BIH Arrhythmia and ESC Databases, but we
have extended it to permit more detailed descriptions of non-ischemic ST
events.

We obtain accurate human annotations of ST events using
special-purpose interactive editing software developed by the FCIS
group in Ljubljana, and using the general-purpose WAVE software system
developed by the second author\cite{ltst:GM-90}.  Each record also includes
a compact clinical summary, with technical information about the
recording as well as relevant clinical information (e.g.,
electrolytes, medications, and pathology).  When complete, the
database and associated utility software will be published on CD-ROMs
in the standard MIT-BIH format, as also used for the ESC DB.

The recordings chosen for the original ESC DB were selected to include
examples of baseline ST displacement resulting from conditions such as
hypertension, ventricular dyskinesia, and effects of medication.  From
these recordings, we will include in the LTST DB those originally recorded by
the ICP group.

In addition, we are selecting new AECG recordings from those obtained
in routine clinical practice at Boston's Beth Israel Hospital (BIH)
and at the ICP. Each selected recording must contain significant
($>$100$\mu$V) transient ST segment episodes corresponding to known or
suspected ischemia, significant non-ischemic ST episodes, significant
slow ST level drift, or mixed episodes. Recordings containing
combinations of these events are preferred.

Both sets of analog recordings have been made using standard AECG
recorders
% manufactured by Applied Cardiac Systems, Del Mar Avionics,
% ELA Medical, ICR, Marquette Electronics, Oxford Instruments, Remco
% Italia, and Reynolds Medical
(the model of recorder used is documented in each case).  The analog
outputs of the playback units are passed through antialiasing filters
and digitized.  Since none of the AECG recorders preserves frequency
content in the signals above about 45 Hz in the best cases, and closer
to 30 Hz in typical cases, we digitize the records at 125 samples per
second per channel.  There is simply no additional information to be
gained from using a higher sampling frequency for these
recordings. The resolution is 12 bits, and the amplitude scale is 200
ADC units/mV for all signals.

As for the ESC DB, we defined \emph{ST deviation} as a change in ST
level relative to a reference level.  Since some recordings exhibit
fixed ST depression relative to the isoelectric level (due to prior
infarcts, for example), it is not meaningful to define the
significance of \emph{transient} ischemic change in terms of \emph{ST
amplitude} (ST level relative to the isoelectric level) in these
recordings.  We identify the reference ST level by searching for a
five-minute interval without significant variation in ST level as near
as possible to the beginning of the record.  Within this interval, a
reference beat is selected and annotated for each ECG lead.  The ST
levels of these beats become the reference ST levels.

We define and annotate events independently on each channel, retaining
the ESC DB's definition of \emph{significant ST episodes}:

\begin{itemize}
\item
An episode begins when the magnitude of the ST deviation first exceeds
50$\mu$V.
\item
The deviation must reach 100$\mu$V or more throughout a continuous interval
of at least 30 seconds.
\item
The episode ends when the deviation becomes smaller than 50$\mu$V,
provided that it does not exceed 50$\mu$V in the following 30 seconds.
\end{itemize}

Any significant event in the LTST DB must meet these criteria.  The
events of interest are ischemic episodes, non-ischemic episodes and
episodes of slow ST level drift. \emph{Ischemic episodes} typically
exhibit a distinctive triangular pattern of ST deviation over time.

Based on our previous studies, we defined characteristics for
\emph{non-ischemic episodes} resulting from position-related (postural)
changes in the cardiac electrical axis:

\begin{itemize}
\item
A non-ischemic episode must exhibit a stable ST deviation level of less than
200 $\mu$V throughout.
\item
The episode must begin or end (or both) with a significant concurrent axis
shift.
\end{itemize}

Axis shifts are best observed in time series of QRS morphology features.

Slow ST level drift is the most difficult event to recognize,
especially if no other ischemic episodes are present.  Drift may
result from slow (non-postural) changes in the cardiac electrical
axis, effects of medication on repolarization, or effects of changes
in heart rate on repolarization.  Since the cumulative effect of drift
over periods ranging from 10 minutes to several hours may amount to a
significant change of ST level (100$\mu$V or more), it cannot be
ignored.  Drift episodes are best identified from ST trend plots.
Based on our previous studies, and the data at hand, we identify a
{\em drift episode} as a significant ST episode that meets any of the following
criteria:

\begin{itemize}
\item
It contains one or more significant ischemic or non-
ischemic episodes within.
\item
It is longer than the ischemic episodes in its neighborhood,
and neither the temporal pattern of ST deviation nor the changes in ST
morphology resemble those of the ischemic episodes.
\item
It appears due to rate-related ST-T changes.
\end{itemize}

In clinical practice, there is usually evidence independent of the ECG
to support a diagnosis of ischemia.  Hence it is likely that criteria
such as those described above will miss \emph{events of borderline
significance} that would be considered ischemic in light of additional
non-ECG evidence.  To account for these events, we also annotate
episodes for which the maximum ST deviation is nearly 100$\mu$V, and
which meet all of the other criteria for ischemic episodes.  These are
annotated as borderline plus or minus (respectively, with or without
ST morphology change). Another category of events of borderline
significance (borderline minus) is that satisfying the criteria
described above but without ST morphology change.

Various significant {\em mixed} (compound) episodes require special
treatment.  This category includes non-ischemic episodes containing
ischemic ``sub-episodes,'' and drift episodes containing ischemic or
non-ischemic sub-episodes. Figure 1 schematically shows such a mixed
episode. The general trend of ST deviation during a mixed episode is
more or less stable (typically between 100 and
200$\mu$V).  While the boundaries of the mixed episode itself are
determined by ST deviations relative to the reference ST level from
the beginning of the record, those of the sub-episodes are
defined within the context of the mixed episode by local ST deviations
relative to a local reference beat.  This local reference is selected and
annotated immediately after the beginning of each mixed episode.
Each ST annotation contains both the ST deviation relative
to the initial (or local) reference, and the ST amplitude (deviation
relative to the isoelectric level).  Except for reference beat
annotations, each ST annotation also contains information about the
type of episode to which it belongs.

\begin{figure} %episode representation
% \centerline{\psfig{figure=ee.ps,scale=135}}
\begin{picture}(330,85)(0,0)
\unitlength 0.25mm
%
\put(0,80){\vector(0,1){30}}
\put(10,100){\scriptsize ST deviation level}
%
\put(10,50){\line(1,0){20}}
\put(10,49.1){\line(1,0){20}}
\put(20,55){\line(0,-1){35}}
\put(17,7){\scriptsize R}
\put(40,50){\line(1,0){240}}
\put(290,50){\line(1,0){40}}
\put(40,49.1){\line(1,0){41}}
%
\put(0,50){\tiny 0}
%
\put(80,50){\line(0,1){25}}
\put(80.7,49.1){\line(0,1){25.9}}
\put(80,55){\line(0,-1){20}}
\put(77,22){\scriptsize (N}
\put(100,80){\line(0,-1){60}}
\put(95,7){\scriptsize LR}
\put(125,80){\line(0,-1){45}}
\put(122,22){\scriptsize AN}
\put(300,55){\line(0,-1){20}}
%\put(297,22){\scriptsize N)}
\put(294,22){\scriptsize N)}
\put(308,55){\line(0,-1){35}}
\put(307,7){\scriptsize IR}
%
\put(50,70){\line(1,0){40}}
\put(55,50){\vector(0,1){20}}
%
\put(30,75){\tiny 100$\mu$V}
%
\put(80,75){\line(1,0){70}}
\put(80,74.3){\line(1,0){70}}
\put(150,75){\line(1,-1){40}}
\put(150,74.1){\line(1,-1){40}}
\put(160,58){\line(0,1){37}}
\put(157,103){\scriptsize (}
%
\put(190,35){\line(1,1){40}}
\put(190,34.1){\line(1,1){40}}
\put(190,25){\line(0,1){70}}
\put(187,103){\scriptsize A}
\put(220,58){\line(0,1){37}}
\put(219,103){\scriptsize )}
%
\put(230,75){\line(1,0){50}}
\put(230,74.3){\line(1,0){50}}
%
\put(290,75){\line(1,0){10}}
\put(290,74.3){\line(1,0){10}}
\put(300,75){\line(0,-1){25}}
\put(299.3,75){\line(0,-1){25.9}}
\put(299.3,49.1){\line(1,0){30}}
%
\put(140,75){\vector(0,-1){10}}
\put(135,65){\line(1,0){95}}
\put(130,82){\tiny 50$\mu$V}
\put(240,75){\vector(0,-1){20}}
\put(155,55){\line(1,0){95}}
\put(250,62){\tiny 100$\mu$V}
%
\put(310,70){\line(1,0){20}}
\put(315,50){\vector(0,1){20}}
\put(306,75){\tiny 100$\mu$V}
%
\end{picture}
{\bf Figure 1.}\ Schematic representation of a mixed
episode (a non-ischemic episode containing an ischemic one).
The shortened annotations shown here [R, LR, IR, (N, AN, (, A, ), and N)] are
defined in table 1.
\end{figure}

\begin{table}[t]
\begin{tabbing}
\=xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx\=xxxxxxxxxxxxxxxxxxxxxxxxxxx\kill
\  \> {\textbf Code} \> {\textbf Meaning} \\
\  \> \  \> \  \\
\  \> ($\:\:\:\:\!$[D$\mid$N]$\:\:$ST$\:$x$\:\pm\:$50,$\:$aaaa \> {\em Beginning} \\
\  \> $\:$A$\:$[D$\mid$N]$\:\:$ST$\:$x$\:\pm\:$dddd,\ aaaa$\:$[,$\:$s] \> {\em Extremum} \\
\  \> \ $\:\:\:\:$[D$\mid$N]$\:\:$ST$\:$x$\:\pm\:$50,$\:$aaaa$\:\:\:$) \> {\em End} \\
%\  \> \  \> \  \\
\  \> R$\:\:\:\:\:$ST$\:$x,$\:$aaaa  \> {\em Reference} \\
\  \> LR$\:\:$ST$\:$x,$\:$dddd,\ aaaa  \> {\em Local reference} \\
\  \> IR$\:\:\:$ST$\:$x,$\:$aaaa  \> {\em Initial reference} \\
\  \> \  \> \  \\
\end{tabbing}
{\bf Table\ 1.}\ ST annotation codes used for the database.
[D$\mid$N]: type of episode (D: drift, N: non-ischemic, none: ischemic);
x: lead number (0 or 1); dddd, aaaa: ST deviation and amplitude
in $\mu$V; s: subtype ($+$: borderline plus, $-$: borderline minus)
\end{table}

ST annotations are made manually with reference to the ECGs and to
trend plots of heart rate and QRS and ST morphologic features.  The
trend plots\cite{ltst:JF-95} are produced using ARISTOTLE\cite{ltst:GM-82} for
QRS complex detection and classification, followed by removal of
baseline wander using a cubic spline approximation and subtraction
technique, low-pass filtering by a 6-pole Butterworth filter (with a
cut-off frequency of 55 Hz), and extraction of ST and QRS morphology
features with heart rate.  Next, the Ljubljana group's
software for interactive ST analysis is used for rejecting abnormal
beats and their neighbors, filtering of the feature time series
%\cite{ltst:JF-92}
(see figure 2, center trace), resampling, and smoothing (figure 2, top
trace).  This approach makes use of the representation power of the
KLT series, while compact trend plots assure accurate detection of
important as well as subtle events in the series \cite{ltst:JF-95}.  Events
are visually detected after the final preprocessing step.  Annotations
are made directly on the trend display (at user-selected scales from 2
minutes to 24 hours) after noise detection in the KLT space, with
reference to the original ECG signals in the region of interest
displayed using the WAVE system.

In the final phase of annotating, the exact locations of the ST
annotations are determined by visual comparison of each clean beat of
the original ECG signals in the region of an ST event with the
reference beat at high resolution (see figure 2 at bottom).  This
requires manual determination of the isoelectric level
and J point for each beat under consideration.  As for the ESC DB, the
ST segment amplitude and ST deviation level (according to the
reference beat) are measured 80ms after the J point (or 60ms after the
J point if the heart rate exceeds 120 bpm).  The database will also
include semi-automated measurements of ST amplitude at both J+60ms and
J+80ms for each beat.
     
As of August, 1996, we had collected and digitized 50 24-hour records:
30 at BIH, 20 at the ICP. These 50 records have been preprocessed and
15 of them annotated by the FCIS group, and subsequently
verified and corrected by a cardiologist. These 15 records contain
179 ischemic, 40 non-ischemic, 7 drift, and 10 mixed episodes.

The Ljubljana group's interactive ST analysis software is being
developed in parallel with the database.  Although the current version
has been used in the annotation of 15 records, planned improvements
should permit greater efficiency in the remaining work.

\begin{figure*}[t]
{\centering\epsfig{file=figures/fig.ps,width=\linewidth}}
{\centering\epsfig{file=figures/ff1.ps,width=.15\linewidth}}\hfill
{\centering\epsfig{file=figures/ff2.ps,width=.15\linewidth}}\hfill
{\centering\epsfig{file=figures/ff3.ps,width=.15\linewidth}}\hfill
{\centering\epsfig{file=figures/ff4.ps,width=.15\linewidth}}\hfill
{\centering\epsfig{file=figures/ff5.ps,width=.15\linewidth}}\hfill
{\centering\epsfig{file=figures/ff6.ps,width=.15\linewidth}}

{\bf Figure\ 2.}\ Example of a mixed episode (record 20612).  The
center and top traces show ST deviation over 24 hours, before and
after smoothing.  The sequence of beats below, shown as displayed by
the FCIS annotation editing software, illustrates the evolution of the
ST segment.  From left to right: the reference beat, the beginning of
the drift episode, the local reference beat, the extremum of the drift
episode, the beginning of the ischemic episode, and the extremum of the
ischemic episode.  Dotted lines mark $\pm 100 \mu$V.
\end{figure*}

\section{Discussion and conclusions}

Analysis of transient ST events is considerably more complex than was
believed before the development of the ESC DB.  The publication of the
ESC DB has given researchers a tantalizing view of temporal patterns
in ST change that are as yet poorly understood but are likely to be of
future clinical interest.  At the same time, the ESC DB shows us
examples of ST changes that confound most automated analysis
techniques.  These events appear to have the principal characteristics
of ischemic ST episodes, yet on closer examination are clearly
non-ischemic in nature.

We are developing a new long-term ST database as a complement to the
ESC DB.  It is important to observe that the LTST DB is not intended
as a replacement for the ESC DB; its goals are different, and (because
of its far greater size) it is not practical to annotate the LTST DB
beat-by-beat as was done with the ESC DB.  What we hope to accomplish
is to better represent the wide variety of ``real-world'' data,
including many more examples of mixed and non-ischemic episodes, and
to permit researchers to study lengthy examples of quasi-periodic
and other temporal patterns in ST change \cite{ltst:JF-95}.  The LTST DB is
intended to support the development of improved algorithms to differentiate
ischemic from non-ischemic ST events, and (by its size) to permit more
reliable prediction of clinical performance from first-order performance
statistics.

\begin{flushleft}
{\bf Acknowledgment}
\end{flushleft}
{\small
This publication is based on work sponsored
by the U.S.-Slovenian Science and Technology Joint Fund
in cooperation with the Department of Health and Human Services,
U.S.A. and the Ministry for Science and Technology of the Republic of
Slovenia under Project number 95-158. Beside, it was sponsored
by the Dean's Funds of the University of Ljubljana. We are also grateful for the
support of the BIH Arrhythmia and Bioengineering Laboratories, and we
especially wish to thank Diane Perry for her help in obtaining recordings.
}
\bibliography{bib/ltst}

\vspace{-2em}
\begin{correspondence}
Franc Jager\\
Faculty of Computer and Information Science\\
Tr\5a\3ka 25 / 1001 Ljubljana / Slovenia\\
Tel./Fax: +386-61-1768-362/1264-647\\
franc@manca.fri.uni-lj.si
\end{correspondence}

\end{document}