next up previous
Next: Discussion and conclusions Up: The Long-Term ST Database: Previous: Introduction

The LTST database

The LTST DB contains 24-hour ambulatory records selected from Holter recordings obtained in routine clinical practice settings in Europe and in the United States between 1994 and 2000. Contributions were made from the BIDMC, ICP, Brigham and Women's Hospital (Boston), and Duke University Medical Center (Durham). The records were collected to model real-world clinical conditions as far as possible with or without known coronary artery disease while containing significant number of ischemic and non-ischemic ST events. Analog records were made using standard AECG recorders. Since AECG recorders preserves frequency content in the signals typically up to 30 Hz, or to 45 Hz in best cases, we digitized the records at 128 or 250 samples per second per channel depending on the scanning system (Marquette MARS, ICR7200, Oxford Medilog 4-24, REMCO LP 103, ZYMED) with the resolution of 12 bits. After resampling the records to unique sampling frequency of 250 samples per second and adjusting amplitude scale to 200 ADC units per mV, the records were preprocessed [1]. Trends of the derived

\begin{figure}%% episode representation
\begin{picture}(330,85)(-10,-5) % (0,0)
...tion is subtracted. For the legend refer to table 1.

features, original ECG data and clinical informations formed the basis for selecting the records. Each selected record contains significant transient ST segment episodes corresponding to known ischemia (ischemic ST episodes), significant non-ischemic heart-rate related transient ST episodes, significant non-ischemic ST events due to axis shifts (postural changes), or significant non-ischemic ST events due to changes in QRS conduction. Some of the records contain arrhythmias such as atrial and ventricular ectopy, atrial fibrillation, and/or atrio-ventricular and intraventricular conduction defects. Other records were selected to include examples of baseline ST displacement resulting from conditions such as hypertension, ventricular dyskinesia, and effects on medications. We also included a number of 24-hour records with proven acute myocardial ischemia such as effort, resting, unstable, mixed, or Prinzmetal's angina originally recorded at the ICP from which the 2-hour excerpts in the European Society of Cardiology ST-T Database (ESC DB) were obtained. The LTST DB contains sixty-five 24-hour two lead and fifteen three lead ambulatory records with significant ST events annotated by human experts. The records also provide individual QRS and rhythm annotations made by an expert Holter scanning technician using a Marquette MARS system. Each record is accompanied with detailed and compact clinical summary including: age and sex; Holter report; report on other clinical investigations (if performed) such as ventriculography, ECG stress test, thallium positron emission tomography, echocardiography, stress echo, coronary angiography; relevant information on patient conditions what includes history, medications, history of arterial hypertension, previous myocardial infarction, left ventricular function, cardiomyopathy, valve disease, electrolyte disturbances, hypercapnia, intraventricular conduction block, baseline ECG, final diagnosis; and technical information about the record including recorder, leads, date and starting time of recording.

During development of the LTST DB, it became obvious that non-ischemic ST events such as sudden axis shifts, slow changes in QRS axis, QRS conduction changes and slow drifts of ST level in general appear without regularity, and are thus forming mixtures of non-ischemic ST events. For these reasons, the expert annotators established improved annotation protocol. ST events were defined and annotated independently in each channel. Initially, the protocol requires manual identification of the isoelectric and J points simultaneously in all ECG leads throughout the records. The basis for annotating ST events in each ECG lead was the ST level function (see figure 1), which was defined as change of ST segment amplitude over time measured 80 ms after the J point, or 60 ms after if heart rate exceeds 120 bpm. ST segment amplitude measurements and positions of the isoelectric and J point were obtained on time-averaged (16 seconds) heart beats derived for each ``clean'' beat which passed the preprocessing phase. ST level function typically varies widely in amplitude due to drifts, position changes, changes in conduction, intermittent QRS conduction changes, heart rate changes, and ischemia. Since non-ischemic ST events could also cause significant ($>$ 50 $\mu$V) shifts in ST level function, the annotating cardiologists manually tracked the ST segment level to eliminate these non-ischemic ST changes. The resultant ST reference function (defined as piecewise linear function between the knot points as annotated throughout the record by local-reference annotations in the ST level function) approximates the ST reference level and was after that subtracted from the original ST level function to form the ST deviation function. To successfully annotate ST events, annotators considered ST level and ST deviation functions, original ECG signals, time series of QRS complex and ST segment KL coefficients, and clinical information about the patient (final diagnosis, other investigations, patient history). Annotators also

\begin{figure*}%% 89 488 563 571 40mm 31.1 mm 2.96mm
..., XIL, EI) is
also bounded by two local references.

established few characteristic categories relating to time-domain ST segment and QRS complex morphology describing their temporal change, origin or nature: During establishing the ST reference level, the annotators first set simultaneously for each ECG lead the global reference annotation (for the annotation codes see table 1) representing the first stable five-minute interval after the beginning of the record without significant variation in the ST level (basal interval). Annotators tracked the ST segment level by a sequence of local references. They tracked everything but ischemic and heart-rate related ST episodes, and changes due to noises. Individual ST episodes of both types, or salvos or sequences of these episodes, were preceded and ended by a local reference. The ST segment level was tracked in the cases of drift, or in the cases of non-ischemic change in ST segment morphology which had to be accompanied by simultaneous change in QRS complex morphology and also evident in the time course of QRS complex KL coefficients. Changes of ST level function tracked were significant ($>$ 50$\mu$V) or not. Any significant sudden-step change of ST level function which was accompanied by simultaneous sudden-step change in QRS complex morphology was bounded by a local reference before and after the step change and was annotated as significant axis shift or significant conduction change according to its nature. Figure 2 shows an example of tracking the ST segment level when two significant ST shifts and significant ST episode are present. Other ST events were annotated in the ST deviation function. In order to be annotated, transient ST episode had to be significant satisfying following criteria: Significant ST episodes accompanied by non-ischemic heart-rate related change in ST segment morphology were annotated as significant heart-rate related ST episodes, while those accompanied by ischemic change in ST segment morphology as significant ischemic ST episodes. Sometimes significant axis shifts or conduction changes appeared within significant ST episodes. In these cases, they were not tracked out, but annotated within the episodes. Sometimes significant ST episodes appeared due to noisy ST intervals as bumps of ST deviation function. Those shorter were annotated as noisy events at their extrema, while those longer as unreadable intervals. Longer intervals with all heart beats rejected during preprocessing as noisy were annotated as unreadable intervals as well.

Annotating procedure consisted from three phases. The first phase consisted from deriving ST level functions on the basis of manually adjusted positions of the isoelectric level and J point in time-averaged (16 seconds) heart beats which corresponded to frequent manually set ``dummy'' annotations along the records [1]. These positions estimated correct positions of the isoelectric and J point for the corresponding heart beats in the centers of averaging windows. The positions of the isoelectric level and J point for the rest of clean heart beats were calculated by interpolating the two positions at dummy annotations along the records. In the second phase, annotators set local references to estimate ST reference functions and set annotations indicating significant ST shifts. During the third phase, the ST reference functions were reviewed/corrected, significant ST episodes were annotated automatically in the ST deviation function, and after that manually verified/corrected. Annotators reached the agreement on the annotations during joint meetings.

Identification and classification of transient ST episodes was accomplished by expert cardiologists using SEMIA (semiautomatic), version 3.0 (mainly developed by the fifth author), a special purpose graphic event-driven user interface and signal-processing tool designed specially for this project. It provides detailed insights into the data at multiple resolutions, examines data at any point, presents the real ECG waveforms and trends of ST and QRS functions, and supports manual and automatic annotation of the records by cardiologists at different sites interacting via the Internet and without paper tracings. New features of version 3.0, as compared to previous version [1], are following: up to three ECG channels, user selects ``lead'', ``data'' and ``KL'' windows to form his/her own interface, dynamic menus not permitting an annotator to go out of the required protocol and stay consistent, several fast modes simplifying repeatable operations, memorizing operations in the mouse buttons, and enhanced menus regarding automatic deletion and changing attributes of annotations.

The LTST DB record files are in the WFDB format and contain detailed clinical information for the subjects, waveform data, true QRS annotations, and ST annotations. Each record is composed by header file sNxxxx.hea (where N is 2 or 3 describing the number of ECG signals and xxxx is the record number), signal file sNxxxx.dat, ARISTOTLE's QRS annotation file sNxxxx.ari, QRS annotation atruth file sNxxxx.atr, ST annotation atruth file, sNxxxx.sta, and ST segment measurements atruth file sNxxxx.16a. Header files describe the format of the signal files and contain technical information about the records, comments of expert annotators, and detailed and compact clinical summaries for the subjects. ARISTOTLE's QRS annotation files contain automatically derived QRS annotations and heart-beat fiducial points which were used during preprocessing and annotating the records. QRS annotation atruth files contain individual QRS and rhythm annotations made by expert Holter scanning technician. ST annotation atruth files contain ST annotations (see table 1), while ST measurements atruth file contain measurements obtained

... level in $\mu$V, {\em dddd}: ST deviation in $\mu$V.

on average heart beats. These measurements were attached back to individual heart beats in the centers of averaging windows. An annotation corresponds to each clean beat and contains: ST amplitude measurements at the points J+80(60)ms, J+0ms, J+20ms, J+40ms, J+60ms, J+80ms, J+100ms, and J+120ms; positions of the isoelectric level and J point relative to the ARISTOTLE's fiducial point; and the number of heart beats left and right to the center beat included into the average beat.

next up previous
Next: Discussion and conclusions Up: The Long-Term ST Database: Previous: Introduction