Incremental Predictive Value of Adding Past Blood Pressure Measurements to the Framingham Hypertension Risk Equation
The Whitehall II Study
Records of repeated examinations of blood pressure are increasingly available for primary care patients, but the use of this information in predicting incident hypertension remains unclear, because cohort studies with repeat blood pressure monitoring are rare. We compared the incremental value of using data on blood pressure history to a single measure as in the Framingham hypertension risk score, a validated hypertension risk prediction algorithm. Participants were 4314 London-based civil servants (1297 women) aged 35 to 68 years who were free from prevalent hypertension, diabetes mellitus, and coronary heart disease at baseline examination (the Whitehall II Study). Standard clinical examinations of blood pressure, weight and height, current cigarette smoking, and parental history of hypertension were undertaken on a 5-year basis. A total of 1052 incident (new-onset) cases of hypertension were observed in two 5-year baseline follow-up data cycles. Comparison of the Framingham risk score with a score additionally incorporating 5-year blood pressure history showed, at best, modest improvements in indicators of predictive performance: C statistics (0.796 versus 0.799), predicted:observed ratios (1.04 [95% CI: 0.95 to 1.15] versus 0.98 [95% CI: 0.89 to 1.08]), or Hosmer-Lemeshow χ2 values (11.5 versus 6.5). The net reclassification improvement with the modified score was 9.3% (95% CI: 4.2% to 14.4%), resulting from a net 17.1% increase in nonhypertensives correctly identified as being at lower risk but a net 7.8% increase in hypertensives incorrectly identified as at lower risk. These data suggest that, despite the net reclassification improvement, the clinical use of adding repeat measures of blood pressure to the Framingham hypertension risk score may be limited.
Preventive interventions can delay the onset of hypertension (systolic/diastolic blood pressure: ≥140/90 mm Hg).1–4 Current risk prediction tools to target preventive interventions at individuals with the highest risk of hypertension, such as the Framingham hypertension risk score,5,6 are on the basis of clinical data taken from a single examination. However, records of repeat blood pressure examinations are increasingly available for primary care patients. We, therefore, examined whether adding past blood pressure measurements to the Framingham hypertension risk algorithm actually improves its predictive power.
Population and Study Design
Data are taken from the Whitehall II Study, a large-scale prospective cohort study of 10 308 civil servants (6895 men and 3413 women) aged 35 to 55 years at the start of the study (Phase 1: 1985–1988).7 Since the Phase 1 medical examination, follow-up examinations have taken place approximately every 5 years: phase 3 (1991–1993), n=8104; phase 5 (1997–1999), n=6551; and phase 7 (2003–2004), n=6483.
The present analysis was on the basis of 2 history baseline follow-up screening cycles, each with 3 blood pressure examinations, the first for blood pressure history, the second for blood pressure at baseline, and the third for follow-up blood pressure (Figure). Participants were eligible for inclusion if they attended 3 consecutive screenings between phase 1 and phase 7. This resulted in 6210 and 5691 eligible participants at the 2 baseline phases, phase 3 and phase 5. At the baseline for both screening cycles, we successively excluded participants who were hypertensive or had a history of hypertension (n=1642 and n=1887 at phases 3 and 5, respectively), who had cardiovascular disease (n=75 and n=137), who had diabetes mellitus (n=39 and n=62), or who had missing data on any risk factors (n=313 and n=826). After these exclusions, 4141 participants at phase 3 and 2779 participants at phase 5 remained and formed the sample for the analyses.
Assessment of Risk Factors and Prevalent Disease
Assessment of risk factors has been described previously.6 Briefly, we measured systolic and diastolic blood pressures twice in the sitting position after 5 minutes of rest with the Hawksley random-0 sphygmomanometer (phases 1 to 5) and OMRON HEM 907 (Phase 7; hypertension risk prediction was not sensitive to the measure of blood pressure used).6 The average of each of the systolic and diastolic blood pressure readings was used. Current smoking and parental hypertension were self-reported. Weight was measured in underwear to the nearest 0.1 kg on Soehnle electronic scales. Height was measured in bare feet to the nearest 1 mm using a stadiometer with the participant standing erect with the head in the Frankfort plane. Body mass index was calculated as weight (in kilograms)/height (in meters) squared.
Prevalent coronary heart disease was defined using MONICA (Multinational Monitoring of Trends and Determinants in Cardiovascular Disease) project criteria,8 positive responses to questions about chest pain9 and physician diagnoses, evidence from medical charts, or positive ECG findings. Diabetes mellitus was defined as a fasting glucose ≥7.0 mmol/L, a 2-hour postload glucose ≥11.1 mmol/L (75-g oral glucose tolerance test), reported doctor-diagnosed diabetes mellitus, or use of diabetes mellitus medication.10
Assessment of Incident Hypertension
Hypertension was defined according to the seventh report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure (systolic/diastolic ≥140/90 mm Hg or use of antihypertensive medication).1 At both screening cycles, we determined incident (new cases) hypertension by the presence of hypertension at follow-up among participants free of this condition at baseline.
Participants were followed across the 2 screening cycles until incident hypertension or last study phase, whichever came first, contributing to a total of 6920 person examinations (each participant contributed observations to 1 or 2 person examinations; Figure). As in previous analyses,6 we selected at random 60% of these observations (0, 1, or 2 per participant) for a “derivation” data set and allocated the remaining 40% of the observations to a “validation” data set. We developed a risk prediction score on the basis of the derivation data, using the same variables as those used for the Framingham hypertension risk score and, additionally, records of systolic and diastolic blood pressures from the phase preceding the baseline. We identified significant predictors and interaction terms for incident hypertension in multivariable adjusted Weibull regression models for interval-censored data. To examine the robustness of this model, we repeated the analysis in a subcohort limited to the participants of the first data cycle only (ie, individuals with data on blood pressure history obtained from phase 1, other risk prediction components including the Framingham risk score at phase 3, and incident hypertension status at phase 5).
We calculated a risk prediction score (“the repeat-measure risk score”) for the validation data set from the β-coefficients obtained from the derivation data set and calculated the Framingham risk score using the β-coefficients derived in the Framingham study5 as described previously.6 We tested the performance of the repeat-measure risk score and the Framingham risk score in the validation data set using 3 methods: (1) discrimination on the basis of C statistics (1.0 indicates perfect discrimination and 0.5 indicates no discrimination); (2) the predicted:observed risk ratio calculations and calibration indicated by the Hosmer-Lemeshow χ2 statistics (<20 indicates good calibration); and (3) net reclassification index to examine whether prediction on the basis of the Framingham risk score was significantly improved after reclassification on the basis of the repeat-measure risk score.11
We then developed 2 alternative repeat-measure risk prediction scores in the derivation data set: the average blood pressure risk score and the “usual” blood pressure risk score. For the first algorithm, we calculated the average of the current and previous blood pressure measurements from different time points and entered this, instead of current and previous blood pressure measurements, in the risk prediction score. To obtain the latter score, we calculated usual systolic and diastolic blood pressures at the previous time point according to the following formula: UBPi=BPbm+[RDR×(BPbi−BPbm)], where UBPi refers to each participant’s usual blood pressure, BPbm to the average blood pressure in the population, RDR to the regression:dilution ratio, and BPbi to the participant’s blood pressure.12 We derived the regression:dilution ratio for a nonhypertensive population by using the mean values of the previous and current blood pressures, which were computed within quartiles of the previous blood pressure. The difference in mean blood pressure between the lowest and highest quartiles for the previous blood pressure (Δ1) and the current blood pressures (Δ3) were calculated and their ratio (Δ3:Δ1) used to estimate the regression:dilution ratio. We then entered usual blood pressure as a component of the risk prediction algorithm in addition the Framingham score variables. We tested the performance of using the average blood pressure and usual plus current blood pressure risk scores in the validation data set in a similar manner to that used for the repeat-measure risk score. All of the analyses were run with SAS version 9.2.
Table 1 presents clinical features for the baseline participants (those 4141 with phase 3 as the baseline and additionally those 173 whose first baseline was phase 5) and the derivation and validation data sets. As expected, the data sets were very similar. During the 2 examination cycles (median length from baseline to follow-up: 5.8 years), we recorded a total of 1052 incident hypertension cases.
Repeat-Measure Risk Prediction Score
The multivariable-adjusted Weibull β-coefficients for incident hypertension, on the basis of the derivation data set, showed a significant effect of blood pressure history on hypertension independent of the Framingham score components (please see Table S1 in the online Data Supplement at http://hyper.ahajournals.org), and this finding was replicated in a sensitivity analysis of participants from the first data cycle only (Table S2). The coefficients from the derivation data set were used to calculate the repeat-measure risk score for the validation data set.
The observed 5-year risk of incident hypertension was 13.1 per 100 (438 incident hypertension cases). The C statistic was 0.796 for the Framingham risk score and 0.799 for the repeat-measure risk score, indicating good discrimination for both. The agreement between the predicted and observed incidences of hypertension was also equally good for the Framingham risk score (predicted risk: 13.5 per 100; predicted:observed ratio: 1.04 [95% CI: 0.95 to 1.15]) and the repeat-measure risk score (12.8 per 100; 0.98 [95% CI: 0.89 to 1.08]). Hosmer-Lemeshow χ2 values of 11.5 for the Framingham score and 6.5 for the repeat-measure risk score were both <20, indicating good calibration.
Table 2 shows the reclassification of individuals between risk categories after replacing the Framingham risk score with the repeat-measure risk score. The net reclassification improvement was 9.3% (95% CI: 4.2% to 14.4%), suggesting that replacing the Framingham risk score with the repeat-measure risk score results in a statistically significant improvement in the prediction of incident hypertension. Repeating this analysis with other categorizations of risk led to similar results (for risk categories: <5%, 5% to 20%, and >20%: net reclassification index: 6.5% [95% CI: 2.2% to 10.8%]; for risk categories: <5%, 5% to 10%, and >10%: net reclassification index: 10.2% [95% CI: 6.7% to 13.8%]).
If the >20% predicted 5-year risk of developing hypertension category is used as the criterion to initiate preventive intervention, the risk prediction with repeat-measure score would lead to 20.2% (475 of 2347; Table 2) of the subjects unnecessarily targeted for preventive treatment compared with 22.4% (525 of 2347) using the Framingham score. Use of the repeat measure score would correctly predict 65.1% (285 of 438) of the incident hypertension cases, whereas the corresponding figure for the Framingham score is slightly greater (67.4%; 295 of 438). With a 10% predicted-risk threshold for the intervention, the corresponding figures for the repeat-measure score and the Framingham score would be 41.3% versus 48.5% (969 versus 1138 unnecessary treatments) and 84.2% versus 87.2% (369 versus 382 correctly targeted treatment).
Risk Prediction Score on the Basis of Average and Usual Blood Pressures
The multivariable-adjusted Weibull β-coefficients for incident hypertension, on the basis of the derivation data set, showed the effect of average blood pressure on hypertension (Table S3) to be stronger than those of blood pressure history and baseline blood pressure as separate terms (Table S1). However, the C statistic of 0.794 and the predicted:observed ratio of 0.96 (95% CI: 0.88 to 1.06) did not indicate superior predictive performance for the risk score on the basis of average blood pressure compared with the Framingham risk score or the repeat-measure risk score. The reclassification improvement of individuals between risk categories after replacing the Framingham risk score with the average blood pressure risk score was 5.8% (95% CI: 0.1% to 11.4%; Table S5), suggesting that replacing the Framingham risk score with the average blood pressure risk score results in a modest improvement in the prediction of incident hypertension. However, comparing this risk score, which incorporates average blood pressure, with the risk score that incorporates current and previous blood pressure as separate terms resulted in a reclassification improvement of −3.4% (95% CI: −7.0% to 0.1%; Table S5). This suggests that the explicit use of separate terms for current and previous blood pressures is more beneficial than the use of average blood pressure in the prediction of future hypertension risk. When using usual blood pressure, these terms in the risk score gave larger hazard ratios for incident hypertension (Table S4) than those on the basis of observed previous blood pressure (Table S1). However, the predictive performance of using usual blood pressure together with the current blood pressure in the risk score gave identical results, in terms of prediction, to those using observed using previous and current blood pressure. (Tables 2 and S5).
In this study of a large nonhypertensive British population, repeat measures of blood pressure independently predicted the risk of developing hypertension. However, information from repeat measures of blood pressure and use of average or usual blood pressures in the risk algorithm improved indices of calibration and the ability of the Framingham hypertension risk score to discriminate future hypertension events only marginally.
We observed a 9.3% improvement in reclassification of hypertension risk by using past blood pressure measurements in addition to the Framingham risk score variables. This improvement was a result of a 17.1% increase in nonhypertensives correctly identified as being at lower risk but also a 7.8% increase in hypertensives incorrectly identified as being at lower risk. Thus, the adoption of the repeated-measures risk prediction model would reduce any harm related to unnecessary preventive treatments (eg, waste of health care resources and adverse effects related to the treatment) but would increase missed prevention opportunities. The reduction in the number of unnecessary treatments was meaningful only when applying a low (10% rather than 20%) risk threshold for treatment, but it came with the cost of missing 2% to 3% of patients who actually develop hypertension.
This appears to be the first report estimating the clinical use of adding past blood pressure data to the Framingham hypertension risk score. Despite the statistically significant net reclassification improvement, our findings suggest that incorporating previous blood pressure records or estimates of average or usual blood pressure in the risk score provides relatively limited incremental value to the prediction of the development of hypertension.
Sources of Funding
This work was supported by the Medical Research Council; British Heart Foundation; Wellcome Trust; Health and Safety Executive; Department of Health; Agency for Health Care Policy Research, United Kingdom; John D. and Catherine T. MacArthur Foundation Research Networks on Successful Midlife Development and Socio-economic Status and Health; National Heart, Lung, and Blood Institute (R01HL036310) and National Institute on Aging (R01AG013196 and R01AG034454), National Institutes of Health; Academy of Finland, Finland; BUPA Foundation, United Kingdom; and European Science Foundation. G.D.B. is a Wellcome Trust Fellow. A.S.-M. is supported by a “European Young Investigator Award” from the European Science Foundation. M.G.M. is supported by a Medical Research Council Research Professorship. M.J.S. is supported by the British Heart Foundation.
- Received September 29, 2009.
- Revision received October 30, 2009.
- Accepted January 20, 2010.
Chobanian AV, Bakris GL, Black HR, Cushman WC, Green LA, Izzo JL Jr, Jones DW, Materson BJ, Oparil S, Wright JT Jr, Roccella EJ. Seventh report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure. Hypertension. 2003; 42: 1206–1252.
The Trials of Hypertension Prevention Collaborative Research Group. Effects of weight loss and sodium reduction intervention on blood pressure and hypertension incidence in overweight people with high-normal blood pressure: the Trials of Hypertension Prevention, phase II. Arch Intern Med. 1997; 157: 657–667.
He J, Whelton PK, Appel LJ, Charleston J, Klag MJ. Long-term effects of weight loss and dietary sodium reduction on incidence of hypertension. Hypertension. 2000; 35: 544–549.
Kivimäki M, Batty GD, Singh-Manoux A, Ferrie JE, Tabak AG, Jokela M, Marmot MG, Davey Smith G, Shipley MJ. Validating the Framingham Hypertension Risk Score: results from the Whitehall II Study. Hypertension. 2009; 54: 496–501.
Tunstall-Pedoe H, Kuulasmaa K, Amouyel P, Arveiler D, Rajakangas AM, Pajak A. Myocardial infarction and coronary deaths in the World Health Organization MONICA Project: registration procedures, event rates, and case-fatality rates in 38 populations from 21 countries in four continents. Circulation. 1994; 90: 583–612.
Rose GA, Blackburn H, Gillum RF, Prineas RJ. Cardiovascular Survey Methods. 2nd ed. Geneva, Switzerland: World Health Organization; 1982.
WHO. Definition, Diagnosis and Classification of Diabetes Mellitus and Its Complications. Geneva, Switzerland: World Health Organization; 1997.
Gasowski J, Li Y, Kuznetsova T, Richart T, Thijs L, Grodzicki T, Clarke R, Staessen JA. Is “usual” blood pressure a proxy for 24-h ambulatory blood pressure in predicting cardiovascular outcomes. Am J Hypertens. 2008; 21: 994–1000.