Table 1 Model results achieved on test set, n = 5 356.

From: Development of risk models of incident hypertension using machine learning on the HUNT study data

Models

AUC (↑)

Scaled brier (↑)

ICI (↓)

ML

 XGBoost

0.795 [0.782, 0.808]

0.204 [0.181, 0.225]

0.016 [0.009, 0.025]

 Elastic regression

0.795 [0.781, 0.807]

0.204 [0.182, 0.223]

0.016 [0.009, 0.025]

 SVM

0.792 [0.779, 0.804]

0.198 [0.177, 0.217]

0.021 [0.012, 0.030]

 KNN

0.786 [0.772, 0.799]

0.186 [0.169, 0.202]

0.024 [0.015, 0.034]

 Random forest

0.778 [0.763, 0.791]

0.181 [0.157, 0.202]

0.017 [0.009, 0.027]

References

 Logistic regression

0.780 [0.766, 0.792]

0.181 [0.160, 0.201]

0.014 [0.007, 0.022]

 High normal BP rule*

0.656 [0.641, 0.670]

–

–

External

 Framingham risk model, original

0.786 [0.773, 0.799]

0.078 [0.037, 0.114]

0.115 [0.104, 0.125]

 Framingham risk model, recalibrated

0.786 [0.773, 0.799]

0.192 [0.170, 0.211]

0.010 [0.005, 0.017]

  1. Best observed mean performances are in [bold].
  2. Performance obtained applying the fitted models on the test set. Reported as mean and 95% confidence interval after bootstrapping. The symbols (↑) and (↓) indicates increasing or decreasing values as improved performance, respectively.
  3. *Scaled Brier score and ICI is omitted for ‘High normal BP rule’ as calibration is not meaningful when predictions are either 0% or 100% risk.
  4. AUC area under the receiver–operator curve, ICI integrated calibration index, KNN K-nearest neighbors, ML machine learning, SVM support vector machines, XGBoost eXtreme gradient boosting.
Back to article page