Decile of Risk^{*} | (1) Logistic Regression | (2) Random Forest | (3) K-nearest Neighbor | (4) Neural Network | (5) JIA Category Alone |
---|---|---|---|---|---|

First (lowest risk) | 0.05 | 0.04 | 0.16 | 0.11 | 0.05 |

Second | 0.03 | 0.03 | 0.13 | 0.10 | 0.04 |

Third | 0.05 | 0.05 | 0.05 | 0.09 | 0.06 |

Fourth | 0.07 | 0.06 | 0.15 | 0.10 | 0.07 |

Fifth | 0.09 | 0.10 | 0.11 | 0.11 | 0.21 |

Sixth | 0.10 | 0.13 | 0.17 | 0.10 | 0.32 |

Seventh | 0.14 | 0.12 | 0.23 | 0.11 | 0.30 |

Eight | 0.21 | 0.22 | 0.28 | 0.22 | 0.34 |

Ninth | 0.43 | 0.45 | 0.29 | 0.41 | 0.36 |

Tenth (highest risk) | 0.91 | 0.88 | 0.51 | 0.74 | 0.37 |

C-index^{**} | 0.85 (0.80, 0.90) | 0.85 (0.82, 0.88) | 0.67 (0.59, 0.75) | 0.75 (0.71, 0.79) | 0.71 (0.65, 0.79) |

Maximum^{**} likelihood | −51 (−43, −59) | −58 (−49, −66) | −81 (−61, −100) | −107 (−99, −116) | −69 (−60, −77) |

Pearson statistic^{**} | 47 (34, 61) | 45 (37, 53) | 12 (3, 20) | 26 (18, 33) | 13 (7, 19) |

↵* Because the overall frequency of a severe disease course in the cohort was 20.5%, a method in which all deciles of risk had observed frequencies close to 20% has no predictive value, while a method in which deciles 1 to 8 had a frequency of 0% and deciles 9 and 10 had a frequency of 100% would be perfect.

↵** Numbers are the mean of 50 re-samples and (95% CI). The c-index of 0.85 for Model 1 reported in this table is the mean of 50 re-samples. When the final logistic model was refit to all the data the c-index was 0.87. JIA: juvenile idiopathic arthritis.