The LTST DB is described in:
Franc Jager, Alessandro Taddei, George B. Moody, Michele Emdin, Gorazd Antolic, Roman Dorn, Ales Smrdel, Carlo Marchesi, and Roger G. Mark. Long-term ST database: a reference for the development and evaluation of automated ischaemia detectors and for the study of the dynamics of myocardial ischaemia. Medical & Biological Engineering & Computing 41(2):172-183 (2003). [HTML] [PDF]
Please cite this publication when referencing this material.
The [LTST DB] was also posted on Physionet.
The Long-Term ST Database is a tool for development and evaluation of automated systems to detect transient ST-segment changes in electrocardiograms, and for supporting basic research into the mechanisms and dynamics of transient myocardial ischemia.
Development of the Long-Term ST Database begun in 1994 and was completed in 2002. The LTST DB includes 86 long-term ECG recordings of 80 human subjects with: ischemic ST episode annotations, heart-rate related ST episode annotations, annotations of ST-segment level shifts due to body position changes and due to conduction changes, signal quality annotations, ST level measurements, and beat-by-beat QRS annotations.
The individual recordings of the Long-Term ST Database are between 21 and 24 hours in duration, and contain two or three ECG signals each sampled at 250 samples per second with 12-bit resolution over a range of ±10 millivolts.
For each recording, the second digit in the record name (2 or 3) indicates the number of ECG signals. Next three digits in the record name indicate the subject number, while the last digit indicates the record number for the subject.
Each record consists of a (text) header (.hea) file, containing detailed clinical information for the subject; a (binary) signal (.dat) file, containing the digitized ECG signals; and several (binary) annotation files, identifiable by suffix:
|.ari||automatically-generated heart beat annotations using ARISTOTLE arrhythmia detector|
|.atr||manually corrected heart beat annotations|
|.16a||automatically-generated, manually-corrected ST-segment level measurements obtained on average heart beats (computed over a 16-second moving window) updated for each normal non-noisy heart beat|
|.sta||ST-segment episode annotations, protocol A (Vmin = 75 µV, Tmin = 30 s) (see below)|
|.stb||ST-segment episode annotations, protocol B (Vmin = 100 µV, Tmin = 30 s)|
|.stc||ST-segment episode annotations, protocol C (Vmin = 100 µV, Tmin = 60 s)|
These files are in the WFDB format.
Initially, ECG signals were preprocessed by ARISTOTLE arrhythmia detector and Karhunen-Loeve (KL) transform coefficients of ST segment and QRS complex were derived for each heart beat. After that, abnormal heart beats, their neighbors, and noisy heart beats were rejected. The .16a files contain "fine" ST-segment level measurements with non-equidistant time indexes. These measurements were obtained on the average heart beats (16-second window) which were derived for each normal and non-noisy heart beat. Frequent average heart beats were chosen throughout the recordings and the locations of the PQ junction (the isoelectric level) and the J point were marked. These locations were then linearly interpolated and the ST-segment level measurements were obtained at the point J + 80(60) milliseconds depending on heart rate. The ST-segment level measurements were then resampled at a constant rate of 0.5 samples per second and smoothed by 7-point moving average filter to form a "raw" ST level function at equidistant time indexes for each ECG signal. Similarly, time series of ST segment and QRS complex KL coefficients (those of normal and non-noisy heart beats) were smoothed (15-point moving average), resampled at a constant rate of 0.5 samples per second, and further smoothed (9-point moving average). The raw ST level function in each ECG signal was used to construct a baseline ST level function and an ST deviation function. Local reference ST annotations were set in the ST level function to estimate the baseline ST level function. This ST reference function was then subtracted from the ST level function to form the ST deviation function. (The ST level function, ST reference function, and ST deviation function for each ECG lead are stored in the .stf files, while time series of KL coefficients in the .klt.zip files.) ST episodes were then identified independently for each ECG signal, based on its ST deviation function and on these criteria:
Three sets of ST episode annotations are provided, since differing criteria may be appropriate depending on the application. Annotators differentiated between ischemic ST episodes and heart-rate related ST episodes. The .sta, .stb, and .stc ST annotation files also contain manual annotations of sudden ST-segment level shifts (>50 µV) due to body position changes (axis shifts) and QRS conduction changes, and annotations on signal quality. For detailed explanation, see annotation protocols, annotation codes, and references below. The time indexes of the annotations in the .ari annotation files are fiducial points determined by ARISTOTLE arrhythmia detector, indexes in the .atr files are fiducial points determined by the Holter scanning device during scanning the records, annotation indexes in the .a16a files are ARISTOTLE's fiducial points, but NOTE, shifted for 16 milliseconds (4 signal samples) forward (right), while the time indexes of the annotations in the .sta, .stb, and .stc ST annotation files are time indexes as obtained in the raw, equidistant, ST deviation functions.
For each record, the numbers of ST episodes and their durations as determined by each of the three sets of criteria are summarized in an additional text file (with suffix .cnt). This file also contains numbers of normal and abnormal heart beats according to beat annotations in the .atr file.
Development of the Long-Term ST Database was an inter-institutional and international effort coordinated by Prof. Franc Jager of the Faculty of Computer and Information Science, University of Ljubljana, Ljubljana, Slovenia. Other investigators include: Roman Dorn, PhD, and Ales Smrdel, MSc, of the Faculty of Computer and Information Science, Ljubljana; Dr. Gorazd Antolic of the University Medical Center, Ljubljana; Drs. Alessandro Taddei and Michele Emdin of the CNR Institute for Clinical Physiology, Pisa, and Prof. Carlo Marchesi of the University of Firenze, Firenze, Italy (the creators of the European ST-T Database); and Dr. Roger Mark and George Moody of the Massachusetts Institute of Technology (the creators of the MIT-BIH Arrhythmia Database), Cambridge, MA, USA, and the Beth Israel Deaconess Medical Center, Boston, MA, USA. The project was supported by Medtronic, Inc. (Minneapolis, MN, USA) and Zymed, Inc. (Camarillo, CA, USA). Development of the Long-Term ST Database began in 1995 and was completed in 2002. We thank all who contributed to this project; further details are here.
Several sources contributed recordings to the Long-Term ST Database:
Other recordings were collected at the CNR Institute of Clinical Physiology, Pisa, Italy; at the Beth Israel Deaconess Medical Center, Boston, MA, USA; and at the Brigham and Womens Hospital, Boston, MA, USA.
The annotation of the Long-Term ST Database was performed using SEMIA, a graphic annotation editor and signal-processing tool written by the group in Ljubljana for this purpose. SEMIA provides an interactive graphical user interface to a semi-automated algorithm for measurement of ST levels. SEMIA's displays provide the annotator with a global view (at different resolutions) of the ST level, ST reference, and ST deviation functions and heart rate, a close-up view of individual heart beat waveforms, and a view of the temporal course of ST segment and QRS complex KL coefficients. SEMIA supports: manual adjustment of heart beat fiducial points, manual tracking of the ST reference level, and manual or automatic annotating of ST-segment events according to selected criteria. Each recording was reviewed independently by expert annotators using SEMIA at each of the three sites (Ljubljana, Pisa, and Cambridge). Participants met several times annually to obtain the consensus reference annotations.
A series of SEMIA screenshots illustrates the annotation process. (Use your browser's Back button to return to this page after following the links to these screenshots in the next paragraph.)
The first task faced by the expert annotators was to mark the locations of the PQ junction and the J point, based on 16-second averaged cardiac cycles chosen at frequent intervals throughout the recordings. These marks served as guideposts for the automated ST-segment level measurement algorithm that performed the next step. Manually determined locations of the PQ junction and the J point were linearly interpolated, ST-segment level measurements were obtained, and these measurements were resampled and smoothed in each ECG signal to form the raw ST level function. The experts then examined the ST level measurements and waveforms in each ECG signal in order to locate and to mark a set of local reference points (marked as LR in the upper panel of the figure). These were used to construct a piecewise linear baseline ST level function (ST reference function), which may vary over time as a result of body position changes or other factors unrelated to ischemia, especially in subjects with prior myocardial infarctions. Axis shifts reflect body position changes, and are usually most apparent in the QRS complexes (note the changes in the temporal course of QRS complex KL coefficients, KL1 - KL5, in the lower panel of the figure). By contrast, when ischemic or heart-rate related ST changes occur, they are most apparent in the temporal course of ST segment KL coefficients (see the lower panel in this screenshot). Local references were placed before and after each such episode, and the episodes were annotated next. During this process, the expert annotators had the option of viewing either the ST level function or the ST deviation function (formed by subtracting the ST reference function from the uncorrected ST level function), as shown in the upper panels of the two screenshots.
(The LTST DB and other materials are also available on Physionet (http://www.physionet.org/physiobank/database/ltstdb/).)
Franc Jager and Miha Amon have contributed two additional sets of time series computed from the ST segments of each normal and non-noisy beat in the database. In each case, they provided time series computed separately for each ECG lead.
For the first of these contributions, in 2009, Miha and Franc calculated coefficients of normalized and non-normalized Legendre orthonormal basis functions. The Legendre orthonormal-transform coefficient time series are in the subdirectory legendre.
In 2011, Miha and Franc derived new single-lead KL basis functions for the ST segments, and used them to compute normalized and non-normalized KL coefficients. The single-lead KL coefficient time series are in the subdirectory kl-single.
Derivation of the Legendre orthonormal-transform normalized and non-normalized coefficient time series, derivation of new single-lead KL basis functions for the ST segments, and derivation of normalized and non-normalized KL coefficient time series is described in reference 6 below.
Name Last modified Size Description
Parent Directory - lt96/ 2014-06-11 14:16 - CinC 1996 paper lt00/ 2014-06-11 14:17 - CinC 2000 paper lt03/ 2014-10-17 10:18 - MBEC paper figures/ 2014-06-11 14:19 - SEMIA screenshots and figures kl-single/ 2015-11-20 15:56 - karhunen-loeve coefficients legendre/ 2015-11-11 10:05 - legendre orthonormal coefficients RECORDS 2014-10-17 10:20 602 list of records notes/ 2014-10-17 10:18 - notes on records ltstdb/ 2014-10-17 10:18 - records tables/ 2014-10-17 10:18 - tables subset/ 2014-10-17 10:19 - time series trends/ 2014-10-17 10:20 - trend plots kl-single-uncentralized/ 2015-11-18 14:37 - uncentralized karhunen-loeve coefficients