Table 2 Precision and recall at various class decision thresholds for each model

From: Greater value add from electronic health records than polygenic risk scores for predicting myocardial infarction in machine learning

Metric

Threshold

NG2 NN

NG2 LR

NG1 NN

NG1 LR

NG2 + PRS NN

NG2 NN

NG2 + PRS LR

NG2 LR

NG1 + PRS NN

NG1 NN

NG1 + PRS LR

NG1 LR

Recall (SE)

0.5

0.849 (0.011)

0.847 (0.012)

0.715 (0.017)

0.713 (0.015)

0.742 (0.015)

0.731 (0.021)

0.682 (0.025)

0.685 (0.024)

0.6

0.804 (0.013)

0.801 (0.012)

0.545 (0.016)

0.530 (0.018)

0.567 (0.022)

0.538 (0.032)

0.481 (0.025)

0.471 (0.021)

0.7

0.744 (0.021)

0.734 (0.016)

0.345 (0.014)

0.329 (0.014)

0.349 (0.044)

0.289 (0.039)

0.261 (0.021)

0.243 (0.023)

0.8

0.649 (0.034)

0.632 (0.030)

0.154 (0.020)

0.143 (0.015)

0.122 (0.030)

0.073 (0.019)

0.084 (0.015)

0.072 (0.014)

Precision (SE)

0.5

0.045 (0.001)

0.044 (0.001)

0.056 (0.002)

0.054 (0.002)

0.045 (0.002)

0.044 (0.002)

0.049 (0.001)

0.046 (0.002)

0.6

0.049 (0.002)

0.048 (0.002)

0.071 (0.002)

0.069 (0.002)

0.056 (0.003)

0.054 (0.003)

0.060 (0.003)

0.057 (0.003)

0.7

0.055 (0.002)

0.053 (0.003)

0.093 (0.004)

0.092 (0.005)

0.073 (0.007)

0.072 (0.006)

0.076 (0.006)

0.072 (0.006)

0.8

0.063 (0.004)

0.061 (0.004)

0.127 (0.012)

0.128 (0.011)

0.096 (0.013)

0.097 (0.017)

0.098 (0.012)

0.091 (0.012)

  1. Values in bold are those for which the multi-modal NG + PRS model and the counterpart NG model are significantly different. This was calculated using an unpaired two-sample t-test at alpha = 0.05 with the mean and standard deviation values across the ten trials.
  2. SE standard error, NG non-genetic, PRS polygenic risk score, LR linear model (logistic regression), NN nonlinear model (neural network), NG1 non-genetic feature set 1 (small-scale established risk factors), NG2 non-genetic feature set 2 (NG1 features plus large-scale clinical diagnoses).
Back to article page