Objective: To compare the performance of predicting incident cardiovascular disease (CVD) using past, current, and usual systolic (SBP), separately and in combination, using large-scale routine electronic health records (EHR). Design and method: Using data from UK primary care linked EHR, we applied a landmark cohort study design involving 80,964 patients aged 50 years (derivation cohort = 64,772, validation cohort = 16,192) who, at study entry, had recorded SBP, no prior CVD, and no previous antihypertensive or lipid-lowering prescriptions. We estimated past SBP (mean, time-weighted mean and variability of SBP recorded up to 10 years prior to baseline) and usual SBP (correcting current values for past time-dependent SBP variability). We used Cox regression to estimate hazard ratio (HR), and applied Bayesian analysis within a machine learning framework in developing and validating models. Discrimination (area under the curve [AUC]) and calibration metrics were used to evaluate predictive performance of models. The outcome was incident CVD (first hospitalisation for or death from coronary heart disease or stroke/transient ischaemic attack). Results: Elevated past, current and usual SBP were separately and independently associated with increased incident CVD risk. Per 20-mmHg rise in SBP was associated with an increased risk for current SBP, with a HR (95% credible interval [CI]) of 1.22 (1.18–1.30), but the corresponding rise in risks were higher for past SBP (mean and time-weighted mean) and usual SBP (HR ranging from 1.39 to 1.45). The AUC for a model that included current SBP, sex, smoking, deprivation, diabetes and lipid profile was 0.750 (95% CI 0.716 to 0.810); adding past SBP mean, time-weighted mean or variability to this model increased the AUC (95% CI) by 0.005 (0.002–0.007), 0.005 (0.002–0.006) and 0.003 (0.003–0.003), respectively (all models showed good calibration). Small improvements in AUC were observed when testing models on the validation cohort data, in sex-specific analyses, or by using different landmark ages (40 or 60 years). Conclusions: Using multiple BP recordings from patients’ EHR showed stronger associations with incident CVD than a single BP measurement, but their addition to multivariate risk prediction models had negligible effects on model performance.