Customize your JAMA Network experience by selecting one or more topics from the list below.
Liu J, Hong Y, D'Agostino, Sr RB, et al. Predictive Value for the Chinese Population of the Framingham CHD Risk Assessment Tool Compared With the Chinese Multi-provincial Cohort Study. JAMA. 2004;291(21):2591–2599. doi:10.1001/jama.291.21.2591
Context The Framingham Heart Study helped to establish tools to assess coronary
heart disease (CHD) risk, but the homogeneous nature of the Framingham population
prevents simple extrapolation to other populations. Recalibration of Framingham
functions could permit various regions of the world to adapt Framingham tools
to local populations.
Objective To evaluate the performance of the Framingham CHD risk functions, directly
and after recalibration, in a large Chinese population, compared with the
performance of the functions derived from the Chinese Multi-provincial Cohort
Design, Setting, and Participants The CMCS cohort included 30 121 Chinese adults aged 35 to 64 years
at baseline. Participants were recruited from 11 provinces and were followed
up for new CHD events from 1992 to 2002. Participants in the Framingham Heart
Study were 5251 white US residents of Framingham, Mass, who were 30 to 74
years old at baseline in 1971 to 1974 and followed up for 12 years.
Main Outcome Measures "Hard" CHD (coronary death and myocardial infarction) was used as the
end point in comparisons of risk factors (age, blood pressure, smoking, diabetes,
total cholesterol, and high-density lipoprotein cholesterol [HDL-C]) as evaluated
by the CMCS functions, original Framingham functions, and recalibrated Framingham
Results The CMCS cohort had 191 hard CHD events and 625 total deaths vs 273
CHD events and 293 deaths, respectively, for Framingham. For most risk factor
categories, the relative risks for CHD were similar for Chinese and Framingham
participants, with a few exceptions (ie, age, total cholesterol of 200-239
mg/dL [5.18-6.19 mmol/L], and HDL-C less than 35 mg/dL [0.91 mmol/L] in men;
smoking in women). The discrimination using the Framingham functions in the
CMCS cohort was similar to the CMCS functions: the area under the receiver
operating characteristic curve was 0.705 for men and 0.742 for women using
the Framingham functions vs 0.736 for men and 0.759 for women using the CMCS
functions. However, the original Framingham functions systematically overestimated
the absolute CHD risk in the CMCS cohort. For example, in the 10th risk decile
in men, the predicted rate of CHD death was 20% vs an actual rate of 3%. Recalibration
of the Framingham functions using the mean values of risk factors and mean
CHD incidence rates of the CMCS cohort substantially improved the performance
of the Framingham functions in the CMCS cohort.
Conclusions The original Framingham functions overestimated the risk of CHD for
CMCS participants. Recalibration of the Framingham functions improved the
estimates and demonstrated that the Framingham model is useful in the Chinese
population. For regions that have no established cohort, recalibration using
CHD rates and risk factors may be an effective method to develop CHD risk
prediction algorithms suited for local practice.
The Framingham Heart Study has contributed to the identification of
risk factors for coronary heart disease (CHD)1-3 and
has developed multivariable functions to predict absolute CHD risk.4-7 Risk
reduction programs now focus on absolute risk of disease rather than on modification
of individual risk factors.8-11 The
Framingham prediction algorithms have been widely adopted to assess absolute
risk and guide the intensity of risk factor interventions.12-14 However,
since more than 99% of Framingham participants are of European descent, the
Framingham functions cannot be generalized to other populations without evaluation
of their appropriateness. Directly applying Framingham functions in some populations
overestimates CHD risk.7,15,16
Recalibrating Framingham functions can substantially improve predictive
ability and, thus, can be a useful approach to generalizing the Framingham
model to other populations. This approach has been evaluated in cohort studies
largely among ethnic groups within the United States,7 but
not in cohorts with different genetic, social, and cultural backgrounds. In
China, the Chinese Multi-provincial Cohort Study (CMCS) for cardiovascular
disease has conducted follow-up for 10 years. The Chinese experience provides
the opportunity to validate Framingham functions, either directly or after
recalibration, in a developing country with a significantly different genetic
and environmental background and substantially different profile of cardiovascular
disease.17 This research effort aims not only
to provide scientific evidence for the development of national guidelines
for CHD prevention in China but also to validate the implementation of Framingham
functions in populations having no ready cohorts to establish risk assessment
The CMCS Cohort. The CMCS was approved by the
Beijing Institute of Heart, Lung and Blood Vessel Diseases for the entire
duration and was conducted with informed consent from all participants. A
total of 30 121 Chinese participants aged 35 to 64 years were included
in the CMCS cohort, of whom 27 003 were recruited from 16 centers in
11 provinces of China in 1992 and 1993.18-22 Additionally,
3118 participants from Beijing were added to the cohort in 1996 and 1999.
A multistage sampling method was used. First, the centers were selected nonrandomly;
the major requirements were having taken part in the Sino-MONICA Project23 and being able to conduct the study. Twelve centers
(80.3% of participants) were in urban areas and 4 centers were in rural areas.
Next, a stratified random sampling for each sex and 10-year age group was
performed in each center for the baseline survey. The overall participation
rate was 82%. The proportions of workers, peasants, and intellectuals were
34%, 13%, and 44%, respectively. Others were housewives and salesmen, etc.
Prevalent cases (504 persons with clinical history of myocardial infarction
or angina pectoris) were excluded from the baseline examination. A face-to-face
follow-up interview to ascertain new CHD events was carried out at the end
of each year for these 27 003 participants from 1992 to 1995, and the
follow-up rate was 94%. From 1996 onward, 6 centers ceased follow-up because
of completion of that national research project. The remaining 16 552
participants in 10 centers of the 1992-1993 cohort and the 3118 participants
of the 1996-1999 cohort were continuously followed up until the end of 2002,
and the follow-up rate was 86% (65.3% of the total cohort).
"Hard" CHD events, comprising acute myocardial infarction, sudden death,
and other coronary deaths, were recorded. The events were diagnosed based
on symptoms, developments in electrocardiograms over time (using Minnesota
codes for up to 4 records), serum enzymes, and autopsy findings according
to the criteria of the World Health Organization (WHO) MONICA project.23,24 The diagnosis was made in 2 steps.
First, when a new case was found during follow-up, trained physicians in the
provincial collaborating unit were sent to visit the patient (or the relatives
in fatal cases), review the hospital records, and complete a standard event
form. Second, the form was sent to the collaborating center at the Beijing
Institute of Heart, Lung and Blood Vessel Diseases and reviewed by a group
of investigators. To ensure diagnostic validity, training was provided before
the study started, case-based tests were given every 2 years, and about 20%
of cases were verified by investigators from the collaborating center in Beijing.
The Framingham Cohort. All Framingham Heart
Study participants provided written informed consent for the study using forms
approved by the Boston University School of Medicine. This informed consent
has been in effect throughout the duration of the study, beginning in 1948.
As reported previously,7 Framingham Heart Study
participants were 5251 white US residents of Framingham, Mass, aged 30 to
74 years in 1971 to 1974 when they attended either the 11th examination of
the original Framingham cohort (2152 participants) or the initial examination
of the Framingham Offspring Study (3099 participants). Similar research protocols
were used in these 2 studies, and CHD patients identified at baseline were
excluded. Twelve-year follow-up was carried out for incidence of hard CHD
events, which included CHD death and myocardial infarction.
In the CMCS study, the baseline survey was conducted according to the
WHO-MONICA protocol for risk factor surveys.25 The
"smoking" variable comprised current smokers. Blood pressure (BP) was measured
in the right arm with a regular mercury sphygmomanometer. Diastolic BP was
defined as the beginning of Korotkoff phase 5. Two consecutive BP measurements
were performed and the mean value of the 2 readings was used. Fasting glucose
and total cholesterol (TC) levels were determined by the enzymatic method,
and high-density lipoprotein cholesterol (HDL-C) was measured by the phosphotungstic
acid/MgCl2 precipitation method. The methods used in the Framingham
Heart Study have been described elsewhere.6
This analysis focused on 6 major risk factors: age, BP, smoking, diabetes,
TC, and HDL-C. The predictive ability of the Framingham functions was assessed
in 2 ways. First, the discriminatory power of the 6 risk factors in predicting
CHD end points was assessed. Second, the calibration of the functions in predicting
degree of risk was assessed. To ensure comparability with the established
Framingham models, the stratification of BP, diabetes, TC, and HDL-C in the
CMCS model was defined according to the criteria used in the Framingham model.7 Specifically, hypertension was categorized according
to the Fifth Joint National Committee on Hypertension definitions.26 Diabetes was defined according to 1985 WHO criteria.27 In the CMCS, diabetes was diagnosed if the fasting
blood glucose level was at least 140 mg/dL (7.8 mmol/L) at the baseline examination
or by a previous clinical diagnosis. In 44% of cases, the diabetes diagnosis
was based on fasting glucose level, 42% on clinical history, and 14% on both
criteria. In the Framingham Heart Study, diabetes was defined as receiving
hypoglycemic treatment (49% of diabetes patients), casual blood glucose level
of at least 150 mg/dL (8.3 mmol/L) in the original cohort (24% of diabetes
patients), or fasting blood glucose level of at least 140 mg/dL (7.8 mmol/L)
at the initial examination of the Framingham Offspring Study participants
(27% of diabetes patients). The cut points for TC and HDL-C were based on
the National Cholesterol Education Program Adult Treatment Panel II (ATP-II),28 but we used more categories to test the association
between extremely low or high levels of serum cholesterol and CHD in different
populations. Specifically, TC of less than 200 mg/dL (5.18 mmol/L) in ATP-II
was further classified into less than 160 and 160 to 199 mg/dL, and TC of
at least 240 mg/dL (6.22 mmol/L) was further classified into 240 to 279 mg/dL
and at least 280 mg/dL; HDL-C of at least 35 mg/dL (0.91 mmol/L) was further
classified into 35 to 44, 45 to 49, 50 to 59, and at least 60 mg/dL.
Sex-specific Cox proportional hazards models were derived after testing
for the assumptions underlying its use. For each risk factor, the regression
coefficients for the CMCS and Framingham cohorts were compared using a 2-tailed z statistic, where z = (b[F] − b[C])/SE. The SE is the standard error of the
difference in coefficients, and SE = (SE[F]2 + SE[C]2)½. The b[F] and b[C] are the beta coefficients
of the CMCS model and the Framingham model, respectively. The SE[F] and SE[C]
are the standard errors of b[F] and b[C], respectively.
Because the relative risk (RR) of a risk factor is the exponential function
of its regression coefficient (RR = eβ), the z statistic was used to test the difference in RR between the 2 cohorts.
To put a heavier test on the functions, P<.10
was defined as significant.
The absolute 10-year risk of hard CHD was predicted with a Cox regression
model developed by Framingham investigators,7 where
P = 1 − S(t)exp(f[x,M]) and f(x,M) = β1(x1 − M1) + . . . + βp(xp −
Mp). Here, β1 . . . βpare the regression
coefficients, x1 . . . xp represent an individual's
risk factors, M1 . . . Mp are the mean values of the
risk factors in the cohort, and S(t) is the survival rate at the mean values
of the risk factors at time t (t = 10 years). Discrimination and calibration
were used to evaluate the predictive capabilities. The discriminatory power
of a model was assessed by the area under the receiver operating characteristic
curve (AUROC) or c statistic. A test developed by
Nam29 was used to compare the AUROCs of 2 models.
The second approach was calibration, which measured how closely the predicted
risk fit the actual risk. The CMCS participants were divided into deciles
of 10-year CHD risk predicted by the CMCS functions, the original Framingham
functions, and the recalibrated Framingham functions. The predicted and actual
risk in each decile were compared, and the difference was assessed by the
Hosmer-Lemeshow χ2 test. Values exceeding 20 indicate significant
lack of calibration (P<.01).7 SAS
software, version 6.12 (SAS Institute Inc), was used for all statistical analyses.
A total of 191 CHD events and 625 total deaths occurred during follow-up
in the CMCS cohort. In the Framingham cohort, 273 hard CHD events and 293
total deaths occurred. The 10-year CHD event rates were 1.5% for men and 0.6%
for women in the CMCS, without adjusting for age. The corresponding crude
incidence rates in Framingham men and women were 8.0% and 2.8%, respectively.
Numbers of participants, person-years of follow-up, and CHD events are shown
in Table 1 and Table 2.
Table 1 and Table 2 also display the risk factor levels at the baseline examination
of the 2 cohorts, unadjusted for age. In comparison with the Framingham cohort,
the smoking rate was higher in Chinese men but lower in Chinese women, the
prevalence of hypercholesterolemia and hypertension was lower in both sexes,
and the prevalence of reduced HDL-C was lower in Chinese men. Similar prevalence
rates for diabetes in men and women and for low HDL-C in women were observed
in the Chinese and US samples.
The β coefficients and RRs for major CHD risk factors were obtained
from Cox regression models for the 2 cohorts (Table 3). Major risk factors showed a similar relation to CHD in
both cohorts, except that smoking was inversely related to CHD risk in Chinese
women. This unexpected trend may be attributable to the low smoking rate among
CMCS women, with small numbers leading to unstable estimates. The coefficient
for age squared was significant in Framingham women, but this was not found
in Framingham men or CMCS participants. For most risk factor categories, the
magnitude of the RRs did not differ significantly. The few exceptions that
nearly reached statistical significance were that among CMCS men, age was
associated with a higher RR (P = .06) and TC of 200
to 239 mg/dL and HDL-C of less than 35 mg/dL were associated with lower RRs
(P = .07 for both); among CMCS women, smoking was
associated with a lower RR (P = .07). The effects
of BP on CHD risk were similar in both studies.
CMCS Functions. In the CMCS functions, the β
coefficients in the CMCS Cox model, mean values of the risk factors, and mean
incidence rates in the CMCS cohort were used. In the discriminatory analysis,
the AUROCs for men and women were 0.736 (95% confidence interval [CI], 0.696-0.776)
and 0.759 (95% CI, 0.699-0.818), respectively, showing good ability to distinguish
cases from noncases. In the calibration, the Hosmer-Lemeshow χ2 was
12.6 for men (P = .13) and 14.2 for women (P = .08), showing that the actual CHD rates in the CMCS
cohort were similar to the event rates predicted by CMCS functions (Figure 1).
Original Framingham Functions. In the original
Framingham functions, the β coefficients in the Framingham Cox model,
mean values of the risk factors, and mean incidence rates in the Framingham
cohort were used directly. In the discriminatory analysis, the original Framingham
functions separated cases from noncases in the CMCS cohort nearly as well
as the CMCS functions. The AUROCs were 0.705 (95% CI, 0.665-0.746) for men
and 0.742 (95% CI, 0.686-0.798) for women. However, in calibration, the original
Framingham functions statistically overestimated the event rates observed
in the CMCS cohort. The Hosmer-Lemeshow χ2 was 645.9 for men
(P<.001) and 147.6 for women (P<.001) (Figure 2). Larger
differences were observed in higher deciles. For example, in the 10th decile
in men, the predicted rate was 20% and the actual rate was only 3%.
Recalibrated Framingham Functions. In the recalibrated
Framingham functions, the β coefficients were taken from the Framingham
Cox model, but mean values from the CMCS cohort were used for the risk factors
and the mean incidence rates. Recalibration did not affect the discriminatory
ability but improved the calibration substantially, especially in women. The χ2 was 31.5 for men (P<.001) and 16.9 for
women (P = .03) (Figure 3). The largest difference between the actual rate and the
predicted rate after recalibration was 1.5% (in the 10th decile in men), compared
with the difference of 17% for the original Framingham functions.
The prevalence of body mass index of at least 25 was 33.5% in men and
33.9% in women. When body mass index (calculated as weight in kilograms divided
by the square of height in meters) was included in the CMCS model, the RR
for body mass index of 25 or higher was 1.29 for men and 1.68 for women, both
nonsignificant. Moreover, RRs for diabetes, TC, HDL-C, and BP were all reduced
after including body mass index. The AUROC had a nonsignificant increase (from
0.736 to 0.739 in men and from 0.759 to 0.763 in women) and the calibration
did not change significantly (data available from authors). Data on exercise
were obtained in the CMCS cohort. A total of 19.4% of men and 13.8% of women
reported physical activity (defined as physical activity regularly during
off hours at least once per week and lasting more than 20 minutes each time).
Those who reported less exercise also tended to have higher BP, higher glucose
levels, and lower HDL-C levels. After adjusting for these factors, the association
of exercise with CHD was not significant (data available from authors). Performance
of the functions were compared in urban vs rural residents. Performance of
the CMCS and recalibrated Framingham functions for urban vs rural men and
women were all very similar, with overlapping 95% CIs (data available from
To assess the effect of the portion of the study population that had
only 3 years of follow-up, a separate model was created after exclusion of
the participants who dropped out. The RRs, 10-year CHD rates, and prediction
capabilities did not differ from the current cohort. Nevertheless, the total
person-years of follow-up and CHD events were reduced and the 95% CIs for
some risk factor categories were wider after the exclusion (data available
In the present analysis, we tested the performance of the Framingham
functions in a large Chinese population, both directly and after recalibration,
and compared them with the usefulness of functions derived from the Chinese
cohort itself to determine absolute risk of CHD. Estimation of absolute risk
of CHD to treat and prevent CHD8-11 commonly
relies on prediction models derived from the experience of prospective cohort
studies. Although prediction algorithms developed by Framingham investigators
have been widely adopted to formulate clinical guidelines in the United States
and elsewhere,12-14 the
Framingham functions have overestimated CHD risk in some populations, leading
to concern that it is not appropriate to generalize the results to other populations.15,16,30 Framingham functions
have been recalibrated in some cohorts of ethnic groups in the United States,
resulting in marked improvement in predictive capability.17 If
this process is equally successful in other settings, it can be a useful approach
to apply the Framingham functions to other populations.
The previously reported lower CHD rates and risk factor levels in the
Chinese population compared with those in the United States were also observed
in this study.24,25 To show the
actual magnitudes in the 2 cohorts, the rates and risk factor levels were
not adjusted for age. The 10-year CHD rate of the Framingham cohort was 5
times that for the CMCS cohort. The CMCS men smoked more than their American
counterparts, and the prevalence rates for diabetes in both sexes and low
HDL-C in women were similar in the 2 samples, but all other risk factor levels
at the CMCS baseline examination were lower than those at the Framingham baseline.
Although CHD rates and risk factor levels in the 2 cohorts differed,
the differences between RRs in most of the risk factor categories were not
statistically significant. Homogeneity was observed for the relation of BP
with CHD, as reported by others.31 The unexpected
association of smoking in CMCS women may be attributable to the low smoking
rate and few CHD events, resulting in unstable estimates of risk of smoking.
The effects of diabetes appeared to be weaker in the CMCS, but the low prevalence
rates led to wider 95% CIs and nonsignificant results.
In the present analysis, only 6 major risk factors were included in
the risk prediction. Body mass index, family history of CHD, and exercise
were not included. Because the risk of obesity appears to be mediated through
TC, HDL-C, hypertension, and diabetes, the ATP-III did not include it as a
factor influencing treatment.14 Body mass index
was not significant after controlling for these other factors, nor was exercise.
Similar results were found in the Framingham study.6 Family
history was ascertained in the CMCS and in the original Framingham cohort
but was not included in the prediction model because its independent effect
is difficult to quantify. Familial influence on risk status is often mediated
through other major CHD risk factors, which run in families.
The RR indicates the importance of an individual risk factor, whereas
the absolute risk predicted by a mathematical model is crucial for effective
therapy of patients and populationwide prevention. In general, the direct
application of prediction models may be most appropriate for individuals who
resemble the cohort from which the model is derived (Box). For populations that have no ready cohorts, the adoption
of an established model provides an alternative way to assess absolute risk,
but validation of this application must be evaluated before incorporating
it into local practice and guidelines. It is not surprising that the CMCS
functions performed well in the cohort from which it was developed. The functions
derived from the CMCS were used to represent the "best" possible predictive
value against which the performance of the Framingham functions could be tested.
However, if the CMCS prediction functions are to be generalized across China,
the functions should be further tested in another validation cohort. In the
present analysis, only Framingham predictions were evaluated in this Chinese
cohort. Both the original and the recalibrated Framingham functions discriminated
between CHD cases and noncases in the CMCS cohort as well as did the CMCS
functions. Nevertheless, the discriminatory ability could not evaluate the
ability to estimate absolute risk of events. In the calibration analysis,
a systematic overestimation was observed when the original Framingham functions
were applied directly to the CMCS cohort, especially in the higher deciles.
The recalibration of the Framingham functions in the Chinese Multi-provincial
Cohort Study (CMCS) cohort for 10-year coronary heart disease (CHD) risk (P)
P = 1 − S(t)exp(f[x,M]) f(x,M) = β1(x1 − M1) + . . . + βp(xp −
where S(t) is the survival rate at the mean values of the risk factors at 10 years
β1 . . . βpare the regression coefficients;
x1 . . . xp represent an individual's risk factors; and
M1 . . . Mp are the mean values of the risk factors
in the CMCS cohort
(Table 1 for
men and Table 2 for women).
The functions used to calculate the absolute 10-year risk of CHD events
for CMCS men with the recalibrated Framingham functions are:
f(x,M) = 0.05 × (age − 47.4) + 0.73
× (smoking − 0.59) + 0.53 × (diabetes − 0.06) + 0.09
× (BPo − 0.36) + 0.42 × (BPhn −
0.13) + 0.66 × (BP1 − 0.19) + 0.90 × (BP2-4 − 0.10) − 0.38 × (TC160 − 0.27)
+ 0.57 × (TC200-239 − 0.22) + 0.74 × (TC240-279 − 0.06) + 0.83 × (TC280 − 0.03)
+ 0.61 × (HDL-C35 − 0.06) + 0.37 × (HDL-C35-44 − 0.22) + 0.00 × (HDL-C50-59 − 0.25)
− 0.46 × (HDL-C60 − 0.33), where age refers to
the individual's age in years, and other risk factor categories are binary,
entering 1 if the individual's value fits that certain category and 0 otherwise.
BP2-4 refer to optimal, high normal, hypertension stage 1, and
hypertension stage 2-4 blood pressure levels, respectively; TC160,
TC200-239, TC240-279, and TC280 refer to
total cholesterol of less than 160, 200-239, 240-279, and at least 280 mg/dL,
respectively; and HDL-C35, HDL-C35-44, HDL-C50-59, and HDL-C60 refer to high-density lipoprotein cholesterol
levels of less than 35, 35-44, 45-49, 50-59, and at least 60 mg/dL, respectively
(the categories used as references were not included in the equations because
their regression coefficients were 0). The equation for women can be created
similarly with the data in Table 2 and Table 3 and the mean survival rate of 0.9961.
A = ef(x,M).
P = 1 − S(t)A = 1 − 0.9895A, where 0.9895 is the mean survival rate, S(t), for men.
For example, the estimated absolute 10-year risk of CHD events for a
nonsmoking man aged 57 years, with TC of 222 mg/dL, HDL-C of 55 mg/dL, fasting
blood glucose of 95 mg/dL, and stage 1 hypertension is determined as follows:
f(x,M) = 0.05 × (57 − 47.4) + 0.73
× (0 − 0.59) + 0.53 × (0 − 0.06) + 0.09 × (0
− 0.36) + 0.42 × (0 − 0.13) + 0.66 × (1 − 0.19)
+ 0.90 × (0 − 0.10) − 0.38 × ( 0 − 0.27) + 0.57
× (1 − 0.22) + 0.74 × (0 − 0.06) + 0.83 × (0
− 0.03) + 0.61 × (0 − 0.06) + 0.37 × (0 − 0.22)
+ 0.00 × (1 − 0.25) − 0.46 × (0 − 0.33) = 0.8847.
A = ef(x,M) = e0.8847 = 2.4223.
P = 1 − S(t)A = 1 − 0.9895 2.4223 = 1 − 0.9748 = 0.0252.
Thus, the absolute 10-year risk of CHD events for this man is 2.52%.
Overestimation of CHD risk in Chinese persons could result in inappropriate
treatment. The estimation of 10-year absolute CHD risk for a nonsmoking man
aged 57 years with TC of 222 mg/dL, HDL-C of 55 mg/dL, fasting blood glucose
of 95 mg/dL, and stage 1 hypertension was 11.7% using the original Framingham
functions. However, the estimation from the CMCS functions was only 2.9%.
Thus, if the Framingham predictive tool was used, it would be necessary to
consider drug therapy according to ATP-III,14 but
it would not be necessary if the CMCS prediction was used.
From the public health perspective, the direct use of original Framingham
functions in Chinese guidelines is inappropriate, as that would lead to overestimation
of CHD risk and incorrect health resource allocation. For instance, by Framingham
estimates, the proportion of Chinese people whose 10-year CHD risk exceeded
10% was 9.9%, but the CMCS functions estimated that only 0.3% experience that
level of CHD risk. Thus, the burden of CHD in the Chinese population would
be overestimated if the original Framingham prediction were applied directly.
This study has several limitations. Although the CMCS cohort was nearly
6 times larger than the Framingham cohort, it had fewer CHD cases than the
Framingham Study because of lower CHD incidence. In addition, 34.7% of the
study participants were followed up for only 3 years. However, exclusion of
this group did not materially affect the results.
Our study oversampled individuals in urban areas compared with rural
areas. However, results of the functions were very similar in the 2 groups.
Finally, levels of risk factors increased in China during the past decades.32,33 These changes appeared to have limited
impact on the prediction capabilities. For example, the levels of most of
these risk factors were higher in the 1996-1999 cohort than those in the 1992-1993
cohort. However, the predictive functions with and without the 1996-1999 data
set were basically the same. Similarly, although population characteristics
changed in the Framingham cohort,34 the present
functions are still valid within US cohorts with more recent profiles of risk
The success of the recalibrated Framingham functions in this Chinese
cohort suggests that recalibration could be of great value for assessment
of CHD risk in other countries. For populations for which CHD risk factor
and event data are available, Framingham functions could also be recalibrated.
For example, several European countries that participated in the WHO-MONICA
project24,25 have data on risk
factor levels and CHD incidence rates. A recent study recalibrated Framingham
functions using cross-sectional data on risk factors and monitoring data on
CHD rates in a Spanish population.35
In conclusion, the original Framingham functions overestimate the risk
of CHD for CMCS participants, and they should not be directly incorporated
into estimates of CHD risk in China. Recalibration of the Framingham functions
corrects the overestimation and, thus, can be a useful approach for the generalization
of the Framingham model in other populations. For populations that have no
established cohort, recalibration may be an effective method to develop a
CHD risk prediction tool suited for local practice of CHD prevention.
Create a personal account or sign in to: