, Benzekry), biorXiv, 2018.
, Conclusions and perspectives ? Machine learning (random survival forest) performed better than the mechanistic model for pure prediction (c-index 0.69 vs 0.62), with similar performances as classical Cox regression
, But mechanistic modeling provides biological and clinical insights that ML does not: ? Ki67 correlates with proliferation rate !
,
, ? prediction of the invisible metastatic state at diagnosis ? potential for personalized adjuvant therapy
, This is a first attempt of a mechanistic, individual-level, predictive metastatic model. A lot remains to be done! ? Refinement to well-established breast cancer molecular subtypes ? Further investigations to refine the modeling to improve the predictive power ? Predictive power to be confirmed in external data sets Percentage of missing values in each variable
, Missing covariate values were imputed using an iterative algorithm based on random forests (missForest R package)