Table 2 Model discrimination performance for prediction of progression to Cambridge Prognostic Group 3 (CPG3) event.

From: Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate cancer

Method

Prediction Time

Evaluation Time

3 years

5 years

Cox (Standard)

From baseline

0.796 ± 0.03

0.786 ± 0.04

+1 yr F/up data

0.760 ± 0.04

0.704 ± 0.06

+2 yr F/up data

0.728 ± 0.10

0.701 ± 0.07

+3 yr F/up data

0.673 ± 0.09

0.728 ± 0.09

Landmarking Cox

From baseline

0.796 ± 0.03

0.786 ± 0.04

+1 yr F/up data

0.765 ± 0.03

0.717 ± 0.05

+2 yr F/up data

0.764 ± 0.04

0.745 ± 0.05

+3 yr F/up data

0.701 ± 0.15

0.745 ± 0.08

Dynamic-DeepHit-Lite

From baseline

0.778 ± 0.05

0.789 ± 0.06

+1 yr F/up data

0.795 ± 0.05

0.740 ± 0.07

+2 yr F/up data

0.780 ± 0.08

0.754 ± 0.09

+3 yr F/up data

0.794 ± 0.11

0.816 ± 0.08

  1. Time-dependent concordance indices are used and compared to standard Cox model using baseline variables only, landmarking and the Dynamic-DeepHit-Lite (DDHL) method. Prediction time refers to the period over which data was collected: at baseline and +1 to 3 years after starting active surveillance. Evaluation time is the follow-up period over which events were predicted. The results shown are averaged over 5 random training/testing splits.
Back to article page