Session Item

Radiomic studies typically involve a large number of features that makes it difficult for a single institution or even multiple centres in a distributed learning setting to meet the adequate number of patients for statistically significant outputs for survival analysis. In these cases, feature selection methods can be very helpful to reduce the study variables to a smaller group, including only the most outcome-relevant ones, thus loosening the constraint on the cohort size. The aim of this work is to develop and validate, both on simulated and on real world data, a new algorithm performing feature selection for Cox Proportional Hazard survival analysis in a privacy-preserving setting through distributed learning.

Lasso regularization was chosen as feature selection method due to its capability of gradually reducing, during model training, the absolute values of the coefficients of less significant covariates, setting to 0 the coefficients of non-significant or redundant features. The algorithm for the distributed Cox regression with Lasso regularization was implemented in Python 3.7; we also designed and developed an automated testing & validation platform to generate survival data, simulate the federated learning and compare the result of the new algorithm with those of any given ground truth solution. We assessed the algorithm performances calculating the mean absolute error (MAE) between its regression coefficients and those obtained from Python centralised state-of-the-art algorithm from the “sksurv” library.

The first step of the validation process involved testing the algorithm on several simulated survival datasets with a number of covariates ranging from 10 to 100 varying the regularization parameter (α) from 0.1 to 0.7. Secondly, we used for further validation a real dataset of 20 radiomic features extracted from RT-planning CT scans of patients affected by Lung Cancer using Overall Survival as model outcome.

We report, in Table 1, 17 tests, 15 of which were done on simulated datasets and two on real world data. For the simulated data, MAE was overall lower than < 0.01. For the real world data testing, the algorithm was trained on 189 lung-cancer patients with two different regularization parameters. Real world data had coefficients in the distributed model, for both values of α, slightly more different than the centralised ground truth resulting in a larger MAE compared to the simulated data, but overall lower than 0.03. As expected, for both types of data, the number of excluded variables (0-coefficients) increases together with α.

To the best to our knowledge, this is the first implementation of a Cox Proportional Hazard model with Lasso features selection for federated learning. In the near future, the model will be tested in a non-simulated distributed setting on a real multi-institution retrospective study.