AUTOMATIC LEARNING OF RULES - A PRACTICAL EXAMPLE OF USING ARTIFICIAL-INTELLIGENCE TO IMPROVE COMPUTER-BASED DETECTION OF MYOCARDIAL-INFARCTION AND LEFT-VENTRICULAR HYPERTROPHY IN THE 12-LEAD ECG
W. Kaiser et al., AUTOMATIC LEARNING OF RULES - A PRACTICAL EXAMPLE OF USING ARTIFICIAL-INTELLIGENCE TO IMPROVE COMPUTER-BASED DETECTION OF MYOCARDIAL-INFARCTION AND LEFT-VENTRICULAR HYPERTROPHY IN THE 12-LEAD ECG, Journal of electrocardiology, 29, 1996, pp. 17-20
The authors developed a computer program that detects myocardial infar
ction (MI) and left ventricular hypertrophy (LVH) in two steps: (1) by
extracting parameter values from a 10-second, 12-lead electrocardiogr
am, and (2) by classifying the extracted parameter values with rule se
ts. Every disease has its dedicated set of rules. Hence, there are sep
arate rule sets for anterior MI, inferior MI, and LVH. If at least one
rule is satisfied, the disease is said to be detected. The computer p
rogram automatically develops these rule sets. A database (learning se
t) of healthy subjects and patients with MI, LVH, and mixed MI + LVH w
as used. After defining the rule type, initial limits, and expected qu
ality of the rules (positive predictive value, minimum number of patie
nts), the program creates a set of rules by varying the limits. The ge
neral rule type is defined as: disease = lim(11) < p(1) less than or e
qual to lim(1u), and lim(2l) < p(2) less than or equal to lim(2u) and
... lim(nl) < p(n) less than or equal to lim(nu). When defining the ru
le types, only the parameters (p(1)...p(n)) that are known as clinical
electrocardiographic criteria (amplitudes [mV] of Q, R, and T waves a
nd ST-segment; duration [ms] of Q wave; frontal angle [degrees]) were
used. This allowed for submitting the learned rule sets to an independ
ent investigator for medical verification. It also allowed the creatio
n of explanatory texts with the rules. These advantages are not offere
d by the neurons of a neural network. The learned rules were checked a
gainst a test set and the following results were obtained: MI: sensiti
vity 76.2%, positive predictive value 98.6%; LVH: sensitivity 72.3%, p
ositive predictive value 90.9%. The specificity ratings for MI are bet
ter than 98%; for LVH, better than 90%.