Assessment of Maximum Heart Rate Prediction Equations in Adults at Low and High Risk of Cardiovascular Disease.

Abstract

Maximum heart rate (HRmax) is commonly used to estimate exercise intensity. Since direct measurement of HRmax is not always practical, prediction equations were developed. However, most equations have not been properly validated in adults at low and high risk of cardiovascular disease (CVD). We sought to: 1) assess the accuracy of commonly used equations to predict HRmax among adults at low and high CVD risk and, 2) determine if SuperLearner (SL) modeling combining base machine algorithms could improve HRmax prediction. A total of 1208 participants (61.6 ± 7.3 yr; 62.7% male) were included. HRmax was measured during a maximal cardiorespiratory exercise test. Predicted HRmax was estimated using the following published equations: Fox, Åstrand, Tanaka, Gelish and Gulati, and a SL model. Bland-Altman analyses as well as performance indicators such as root mean squared error (RMSE) and Lin's Concordance Correlation Coefficient were performed. All predicted HRmax-derived equations were positively associated with measured HRmax (women: r = 0.31; men: r = 0.46, P ≤ 0.001) but to a greater extent using a SL model (women: r = 0.47; men: r = 0.59, P ≤ 0.001). Overall, all equations tended to overestimate measured HRmax, with a RMSE which varied between 10.4 and 12.3 bpm. Although the SL model outperformed other equations, with no significant difference between measured and predicted HRmax, RMSE remained high (11.3 bpm). Lack of accuracy was mainly observed among adults with low aerobic fitness and with CVD risk factors, such as obesity, diabetes, and hypertension. We showed that commonly used equations and the SL model have insufficient accuracy to predict HRmax among adults. The performance of the prediction equations varied considerably according to the population clinical characteristics such as the presence of CVD risk factors or a low aerobic fitness.