Conference PaperPDF Available

In Silico Oncology: Evaluating the Predictability of Acute Lymphoblastic Leukemia Patients’ Response to Treatment Utilizing a Multiscale Oncosimulator Model in Conjunction with Machine Learning Methods

A preview of the PDF is not available
... Training of the regression models was based on pathway-aggregated [9] gene expression profiles. Moreover, the first results of an extended, fully automated and complete workflow for estimating and predicting the parameter values of the ALL Oncosimulator were presented at the VPH2016 conference [10]. Based on the previous efforts, a detailed presentation and a more thorough study of the Hybrid ALL Oncosimulator is provided in this paper. ...
... In step (c), a model intended to predict ALL Oncosimulator parameters values (regression model) is trained for the Train set. Its parameters are optimized by an internal k-fold cross-validation procedure (k in -CV in figure 3) again using the caret package [34] in R. Based on the previous experience [4,10], the random forests algorithm [35] has been selected for assessment. Moreover, the Weighted k-nearest neighbours (k-NN) [36] algorithm was also evaluated. ...
... As can be seen in figure 4 and in agreement with [4,10], prednisone good responders tend to have higher mean CKR PRED values compared to prednisone poor responders. This finding further supports the validity of the parameter estimation/adaptation procedure and of the ALL Oncosimulator model as a whole. ...
Efficient use of Virtual Physiological Human (VPH)-type models for personalized treatment response prediction purposes requires a precise model parameterization. In the case where the available personalized data are not sufficient to fully determine the parameter values, an appropriate prediction task may be followed. This study, a hybrid combination of computational optimization and machine learning methods with an already developed mechanistic model called the acute lymphoblastic leukaemia (ALL) Oncosimulator which simulates ALL progression and treatment response is presented. These methods are used in order for the parameters of the model to be estimated for retrospective cases and to be predicted for prospective ones. The parameter value prediction is based on a regression model trained on retrospective cases. The proposed Hybrid ALL Oncosimulator system has been evaluated when predicting the pre-phase treatment outcome in ALL. This has been correctly achieved for a significant percentage of patient cases tested (approx. 70% of patients). Moreover, the system is capable of denying the classification of cases for which the results are not trustworthy enough. In that case, potentially misleading predictions for a number of patients are avoided, while the classification accuracy for the remaining patient cases further increases. The results obtained are particularly encouraging regarding the soundness of the proposed methodologies and their relevance to the process of achieving clinical applicability of the proposed Hybrid ALL Oncosimulator system and VPH models in general.
ResearchGate has not been able to resolve any references for this publication.