Machine learning and mechanistic modeling for prediction of metastatic relapse in early-stage breast cancer

Abstract

Predicting the probability of metastatic relapse for patients diagnosed with early-stage breast cancer is essential for decision of adjuvant therapy. Current predictive models rely on biologically-agnostic, statistical models. Using the breast cancer database from the Bergonié Institute (n=1057 patients), we investigated the potential use of machine learning algorithms for predicting 5-years metastatic relapse. Furthermore, we developed a novel prediction tool based on an experimentally validated mechanistic model of metastatic development. Preliminary results indicate a higher positive predictive value when using this biologically-based model when compared to the alternative algorithms.

On Fair Performance Comparison between Random Survival Forest and Cox Regression: An Example of Colorectal Cancer Study

“…The motivation of this paper is an observation we have made on existing literature: that when RSF is compared to Cox regression, many papers conclude that the two have similar performance [20,21], while detail concerning the error calculation (on the Cox regression side) is not provided [22][23][24][25][26]. Some comparisons are even hard to judge because k-fold CV is introduced to RSF [27], or models with different number of independent variables are compared [28]. Steele et al (2018) and Zhang et al (2019) [29,30], k-fold CV is used in evaluating other models including Cox regression, which may be another example of unfair comparison.…”

Bayesian combination of mechanistic modeling and machine learning (BaM3): improving personalized tumor growth predictions

“…Even if the latter become more and more accurate, the black-box nature of the prediction process limits its transparency and accountability [33]. The urge for interpretability advocates for novel approaches, able to shed light on the machine decision processes [34,35,36]. In this work, we presented a novel method (BaM 3 ) to couple mathematical modeling and nonparametric regression in a Bayesian framework.…”

Current challenges in metastasis research and future innovation for clinical translation

“…Data gathered from preclinical and clinical studies of metastasis dynamics are underpinning the implementation of mathematical modelling and artificial intelligence to build computational predictors of metastatic burden and therapy response [88]. These mathematical models also confer the opportunity to infer the actual stage of progression for a patient's tumor at diagnosis, to predict their likely burden of disseminated disease and to assess their risk of developing overt metastases [89]. Importantly, mathematical models have validated the non-linear relationship between primary tumor size and survival, indicating that metastatic propensity is not simply a function of tumor size but that tumor intrinsic, extrinsic and host factors all contribute to metastatic progression [90].…”