Within the IMAGINT Project, Firalis data analysis team headed by Dr. Peter Grass has created a multivariate Biomarker prediction suite consisting of a panel of diverse machine learning algorithms including partial least squares discriminant analysis, support vector machine, k-nearest neighbours, naïve bayes, logistic regression, random forest, CN2, classification trees, supervised principal component analysis and other non-linear projection methods.

A set of different model validation strategies has also been established to develop predictive tool for clinical application including n-fold cross-validation, leave-one-out validation, and different re-sampling schemes for calibration and test samples.

Each of these statistical learning approaches has its strengths and weaknesses, and the objective of the Firalis Biomarker prediction suite is to apply all methods in parallel and use a supervisor algorithm to integrate and benchmark the predictions of the individual prediction models. This approach is well known in the machine learning community and is called “Panel of Experts” approach.

In the next steps, Firalis will integrate the “Bayesian Latent Variable Model” developed by the laboratory of Prof. Ton Coolen at King College of London into the Firalis prediction suite and to compare its predictive performance with that of the standard machine learning approaches. In addition, Firalis is compiling a set of test cases of diverse numbers of samples/patients and variables including synthetic data, transcriptomics, proteomics and metabonomics data, clinical data like demographics, medication, morbidity, standard hematology, biochemistry and urinalysis parameters. Features extracted from imaging procedures will also be included. The objective of this benchmarking approach is to obtain an as unbiased as possible estimate of the predictive performance of the newly developed “Bayesian Latent Variable Model”.