Exploring hemodynamic measurements from the Tromsø Study for prediction of cardiovascular disease using traditional statistical models and machine learning approaches

Abstract

Abstract In Norway, NORRISK2 is the government-recommended risk model for predicting an individual’s 10-year probability of getting cardiovascular disease (CVD). This study aims to investigate the potential for improvement of CVD prediction by using hemodynamic measurements from a non-invasive beat-to-beat blood pressure monitor, taken as part of pain sensitivity assessment with the cold-pressor test (CPT) during the Tromsø6 Study (2007–2008). Using 6694 recordings, ultra-short-term pulse rate variability (PRV) and baroreflex sensitivity (BRS) obtained during the CPT were added as additional variables into the existing NORRISK2 survival model (extended model). In addition, the time-series data was used in a machine learning (ML) model without the NORRISK2 background variables. Both models were compared to a recalibration of the original NORRISK2 model. The predictions from the recalibrated NORRISK2 model and the ML model were then combined with logistic regression. The statistical models performed similarly on the test set, with an area under the receiver operating characteristic (AUROC) of 0.8 (95% CI: 0.71–0.86), 0.79 (0.71–0.85) and 0.77 (0.69–0.84) (original, recalibrated and extended NORRISK2, respectively). The ML model using only hemodynamic measurements obtained a test set AUROC of 0.73 (0.67–0.80). Combining the NORRISK2 and ML model did not increase the AUROC. Adding ultra-short-term PRV and BRS derived from Tromsø6 did not improve the prediction of the NORRISK2 model either. Although with lower accuracy, the beat-to-beat time series of hemodynamic variables from a CPT had a significant (p < 0.01) ability to predict future CVD without any other person-specific data.