COMPUTER-ASSISTED INTERPRETATION
Computer programs for the interpretation of ECG are widely used.
In a large study, the ECGs of 1200 adult patients with known clinical conditions
were interpreted by cardiologists and by nine computer programs.[102]
The results demonstrated that, in some cases, computer interpretations of ECG were
correct and cardiologists were incorrect when their diagnoses were tested against
the clinical evidence. The opposite was twice as likely when the average computer
program and the average cardiologist were compared. The computer programs with the
best performance were almost as accurate as the best cardiologists in classifying
the ECGs of patients in seven common diagnostic groups. The study did not test the
computers' abilities, however, in two critical areas: acute myocardial ischemia
and arrhythmia recognition. Modern monitoring
equipment is highly computerized, and most of the physiologic information is manipulated,
analyzed, and stored in the digital format. An early step in data collection therefore
involves the conversion of analog signals (i.e., time-variable voltages or amplitudes)
into digital format with an analog-to-digital converter. Once in the digital format,
the physiologic information can readily be subjected to a variety of analyses. For
ECGs, the most common analyses, besides rate calculations, are related to arrhythmia
recognition and the detection of myocardial ischemia.
Arrhythmias
There is little doubt that during prolonged visual observation
of the ECG on the oscilloscope, certain arrhythmias go undetected. This was clearly
demonstrated by Romhilt and colleagues,[103]
who
showed that coronary care unit nurses failed to detect serious ventricular arrhythmias
in 84% of their patients. Computers have therefore been designed for the automatic
detection of arrhythmias in an attempt to increase the detection of abnormal rhythms.
Using an early preprocessing algorithm called AZTEC, a computer accurately detected
78% of ventricular ectopic beats. It measured QRS width, offset, amplitude, and
area to classify complexes into morphologic families.[104]
In a prospective evaluation of such a system, it was found that the computer accurately
detected 95.4% of VPBs but only 82.4% of supraventricular premature beats.
Other systems have depended on QRS recognition and cross-correlation
with stored QRS complexes.[105]
In cross-correlation,
each detected QRS complex is compared with a list of previously detected complexes.
If a complex does not correlate better than 0.9 with a previously stored complex,
it is considered to have a new configuration and is added to the list ( Fig.
34-28
). Certain points of the complex, such as the PR interval and ST
segment, are stored as a template for future comparison. Whenever a new complex
matches an existing template, it is averaged into that template, so that each template
represents a running average of all complexes of a particular configuration.[106]
Each template is defined as normal, abnormal, or questionable according to previously
defined criteria. The AHA has published several parameters that need to be tested
to permit a meaningful understanding of a system's values and limitations and to
permit reasonable comparison among systems.[25]
