Introduction: Being able
to predict when a person would have a myocardial infarction (MI)
would have multiple implications: (1) that we understand what
leads to MI's, (2) that we have a means to identify patients who
are at risk for having immanent MI's and that we could help them
to delay or avoid an MI, and (3) that we could give patients accurate
feedback about the results for them of risk factor modification.
The goal of this research was to develop a systems dynamics computer
simulation model that would predict MI's.
Methods: A model created
to simulate the course of patients' cardiovascular health and
disease, beginning at birth and predicting time in months at which
eventual coronary artery occlusion and myocardial infarction (MI)
would occur. The model was developed on a cohort of patients with
acute myocardial infarction using such variables as cholesterol;
systolic and diastolic blood pressure; sex; menstrual status;
triglyceride levels; HDL and LDL cholesterol levels; glucose levels;
family history of age at initial myocardial infarction; smoking
status and amount; exercise frequency, duration, and intensity;
activity level of the patient's lifestyle; amount of dietary fat
consumed; weight and height; hostility; levels of psychosocial
stress; and a number of other variables known to influence the
rate of development of coronary artery disease. Psychosocial data
was obtained about stress, support and depression. Lifestyle data
was obtained about exercise, activity levels, diet, alcohol, caffeine,
cigarettes and drug consumption.
A second group of patients interacted with the model
to study how patients would react to the feedback provided and
whether or not they would change risk factors. Patients were given
specific predictions with graphs about when they would be expected
to have a heart attack and predictions of how much longer we would
expect them to live with risk factor modification. They were permitted
to query the computer regarding how other changes they might make
would affect their longevity. For example, patients could stop
smoking and begin exercising at the present time, re-run the simulation,
and determine how this behavior change in the present would affect
the time before they could expect to have a myocardial infarction
(heart attack; crise cardiac). Patients could ÒbargainÓ
with the simulation to determine what combination of risk factors
they were willing to change to obtain specific lengthening of
time before they could expect a myocardial infarction.
Results:
Sample characteristics.
Of the 44 patients used for the initial validation testing, 14
were women and 11 were Hispanic. Systolic blood pressures ranged
from 79 to 176 (mode = 130). Diastolic pressures ranged from 44
to 110 (mode = 80). Cholesterol ranged from 118 mgm/dl to 373
mgm/dl (mode = 222 mgm/dl). Triglycerides ranged from 56 mgm/dl
to 710 mgm/dl (mode = 139 mgm/dl). Random blood glucoses ranges
from 80 mgm/dl to 724 mgm/dl (mode = 132 mgm/dl). Activity levels
ranged from 0 to 6 (mode = 0). Total CK levels ranged from 102
to 6200 I.U. (mode = 712 I.U.). Time of initial myocardial infarction
ranged from 432 months to 989 months (mode = 635 months). Pack
years of smoking ranged from 0 to 210 (mode of 0). Body mass index
ranged from 0.0127 to 0.0483 (mode = 0.0254). Table 2 shows the
breakdown in MI Outcome Index by race.
Comparisons to conventional statistics.
1. Cluster analysis. Both the Joins and the K-means procedures were used. Despite attempting all possible permutations of variables, the best classification obtained was 13 correct assignments out of the 22 test cases (59.1% accuracy). This was achieved by using all independent variables with the single linkage method (nearest neighbor). The distance metric was the Pearson correlation coefficient. This compares to 88% correct classification with dynamic systems modeling.
2. Multiple regression analysis. A multiple
regression model predicted 26.6% of the variance (p = 0.018).
A borderline significant interaction of race and sex was noted
with white males faring the worst. The most significant and best
predictors were triglyceride and cholesterol levels.
Dynamic Systems Modeling.
The model correctly classified 88% of patients. The false positive
rate was 6% and the false negative rate was 13%. The false positive
patients included a 64 year old Hispanic female with a blood glucose
of 529 mgm/dl, cholesterol of 373 mgm/dl and triglycerides of
323 mgm/dl. The assumption of 20 years of chronicity may not have
been accurate for this patient or she may have had an unmeasured
protective factor operable. The first false negative was a 43
year old white, post-menopausal female with all normal indicators
except for 25 pack-years of cigarette smoking. The second false
negative was a 42 year old Hispanic male with all normal indicators
except for a blood pressure of 155/105 and a cholesterol level
of 234 mgm/dl.
An unidentified risk factor was isolated in the learning
sample and on closer inspection was found to be family history
of early MI. It became prominent when age was greater than 57
years, blood pressure normal, smoking less than 2 packs per day,
body mass index greater than 0.0177 and cholesterol levels between
166 and 239 mgm/dl, with glucose levels between 100 and 210 mgm/dl
and triglycerides between 105 and 185 mgm/dl for men. When these
parameters were exceeded, the genetic effects appeared to be sufficiently
overshadowed as to be rendered non-operable. It was operable in
women with a positive history when age was greater than 5o, blood
pressure normal, non-smoking was true, race was white and cholesterol
was greater than 200 mgm/dl with all other indices normal.
Results of Interaction with the Model:
21 patients who interacted with the model were followed for an
average of 5.9 mos. (range 5 to 20 mo). The patients were seen
at 4-8 week intervals to update data and to interact with the
model. Nineteen reduced some aspect of risk behavior; 19 increased
exercise and/or activity (90%); 4 increased HDL cholesterol (21%);
8 decreased LDL cholesterol (38%); 12 decreased BP (58%). Patients'
acceptance of the systems dynamics model was high. They preferred
what they learned from the model to the standard of medical care,
which consists of lectures from the physician to change behavior
(Òstop smoking, lose weight, exercise) or else. In the
standard medical approach, Òor elseÓ is poorly specified
and patients can question whether or not the information presented
applies to them or not. Patients being able to use the systems
dynamics model to evaluate their own personal risks and benefits
and make choices based upon data presented rather than the usual
lecture approach of their physician.
Sample Output. Figure 1 shows the prediction for a patient with many risk factors. The time (in months) at which line 1 (Time of MI) changes slope represents the time at which an MI is predicted to occur (about 520 months or age 43 and 4 months).
Figure 2 shows the graph for a patient with no risk
factors and for whom no MI is predicted at even 1,000 months (age
83).