Improve generalizability of survival models via low-rank decomposition

Tao Xu

Abstract:

Introduction:
Clinical predictive models can help to stratify patients with certain risk profiles, thus reducing sample size for clinical trials. With the comprehensive collection of data on disease trajectory and biomarker data, historical clinical trials of the same disease provide a high quality data source for building such models. However, heterogeneity between trials poses additional challenges in model performance and interpretation.

Methods:
We adapted a low-rank decomposition based method (Piratla et al., 2020) for survival modelling, which can extract a common component and trial-specific components when building models across multiple trials. We compared this method with two other common modelling approaches: 1) pooling data from multiple trials; 2) stratifying data by trials and tested their performance on a common hold-out dataset.

We evaluated the methods both in simulated datasets, with different levels of variation in the correlations between the covariates and the outcome, and also in real clinical trials where the model was trained on seven non-squamous NSCLC patient trials and tested on a squamous NSCLC trial.

Results:
In the real trial data, the model trained with the proposed method performed better than the pooled and stratified approaches by a delta C-index between +0.02 to +0.05. With the simulated datasets, where variations were introduced to the associations between covariates and the outcome while the marginal distribution of the covariates and outcome were maintained, the models built with low-rank decomposition retained good performance, while the other methods lost performance.