Table 1 TDNODE predictive performance of tumor dynamics using a 32-week observation window on the test set.

From: Explainable deep learning for tumor dynamic modeling and overall survival prediction using Neural-ODE

Treatment arm

Number of predictions for \(t \,> \,{w}_{i}\)

RMSE

(median 土 MAD)

R2 score

(median 土 MAD)

Arm 1: atezolizumab + carboplatin + paclitaxel

208

12.56 土 1.29

0.75 土 0.05

Arm 2: atezolizumab + carboplatin + paclitaxel + bevacizumab

214

7.65 土 0.93

0.93 土 0.01

Arm 3: carboplatin + paclitaxel + bevacizumab

79

4.34 土 0.31

0.98 土 0.01

All treatment arms

501

9.69 土 0.75

0.88 土 0.02

  1. We evaluated the predictive performance of TDNODE on the unseen portion of the test set. For each patient, we let the observation window \({w}_{i}\) = 32 weeks and only evaluate measurements collected at time values beyond \({w}_{i}\). Although TDNODE generates a continuous solution of predictions \(z\left(\cdot \right)\), the RMSE and R2 scores are calculated using the discrete set of predictions only at observation times with SLD measurements. The predictive performance across all treatment arms is shown in bold. Variability was measured via median absolute deviation (MAD).
  2. The bolded row indicates the model that gave rise to superior predictive performance.
Back to article page