Table 2 Mean estimated C-indices (averaged over time) with estimated standard deviations, as obtained from training the DeepHit architecture on the simulated, CRASH-2, and SEER breast cancer data.

From: An imputation approach using subdistribution weights for deep survival analysis with competing events

Data

Type-1-rate

Type-2-rate

DeepHit\(^{1}\)

DeepHit\(^{1}\), no imp.

DeepHit\(^{2}\)

CRASH-2

\(4.9\%\)

\(78.3\%\)

78.17 ± 1.04

76.80 ± 4.96

78.18 ± 0.94

CRASH-2

\(14.9\%\)

\(68.3\%\)

80.14 ± 1.77

79.88 ± 2.01

80.05 ± 4.23

SEER

\(6.9\%\)

\(4.7\%\)

81.75 ± 3.46

81.80 ± 3.49

81.73 ± 3.34

Simulated

\(11.5\%\)

\(41.1\%\)

64.13 ± 0.75

62.58 ± 2.17

63.71 ± 0.96

Simulated

\(21.8\%\)

\(30.6\%\)

65.90 ± 0.69

64.59 ± 2.25

65.20 ± 3.26

Simulated

\(38.6\%\)

\(13.4\%\)

66.05 ± 0.47

64.97 ± 2.51

64.39 ± 6.26

  1. DeepHit\(^{1}\) = DeepHit architecture with one sub-network trained with the preprocessed input data; DeepHit\(^{2}\) = DeepHit architecture with two subnetworks; DeepHit\(^{1}\), no imp. = DeepHit architecture with one sub-network trained on the original input data (treating individuals with observed competing events as censored individuals). Best-performing methods are marked bold. Note that the C-indices must be compared within each row, as the datasets used for training were different in terms of size, censoring, and event rates across the rows. For CRASH-2, in the upper and the lower rows \(\epsilon = 1\) indicates death due to bleeding and death due to any recorded cause, respectively. The numbers in this table are obtained from the test datasets.
Back to article page