A comparative evaluation of the generalised predictive ability of eight machine learning algorithms across ten clinical metabolomics data sets for binary classification

Abstract

IntroductionMetabolomics is increasingly being used in the clinical setting for disease diagnosis, prognosis and risk prediction. Machine learning algorithms are particularly important in the construction of multivariate metabolite prediction. Historically, partial least squares (PLS) regression has been the gold standard for binary classification. Nonlinear machine learning methods such as random forests (RF), kernel support vector machines (SVM) and artificial neural networks (ANN) may be more suited to modelling possible nonlinear metabolite covariance, and thus provide better predictive models.ObjectivesWe hypothesise that for binary classification using metabolomics data, non-linear machine learning methods will provide superior generalised predictive ability when compared to linear alternatives, in particular when compared with the current gold standard PLS discriminant analysis.MethodsWe compared the general predictive performance of eight archetypal machine learning algorithms across ten publicly available clinical metabolomics data sets. The algorithms were implemented in the Python programming language. All code and results have been made publicly available as Jupyter notebooks.ResultsThere was only marginal improvement in predictive ability for SVM and ANN over PLS across all data sets. RF performance was comparatively poor. The use of out-of-bag bootstrap confidence intervals provided a measure of uncertainty of model prediction such that the quality of metabolomics data was observed to be a bigger influence on generalised performance than model choice.ConclusionThe size of the data set, and choice of performance metric, had a greater influence on generalised predictive performance than the choice of machine learning algorithm.Electronic supplementary materialThe online version of this article (10.1007/s11306-019-1612-4) contains supplementary material, which is available to authorized users.

Metabolomics to Improve the Diagnostic Efficiency of Inborn Errors of Metabolism

“…With more substantial amounts of data being produced by these multiplex assays, machine learning tools facilitate reproducible and understandable models of prediction (classification and regression) [77]. These techniques take an entire metabolic snapshot of the metabolome and range from several to hundreds of analytes, to classify the sample in order to arrive at a diagnosis [54,56].…”

Migrating from partial least squares discriminant analysis to artificial neural networks: a comparison of functionally equivalent visualisation and feature contribution tools using jupyter notebooks

“…Second, model outcomes and resulting interpretations can affected by the quality of the input data. We have previously shown that PLS and ANNs show similar predictive ability, when using the same input data, and that sample size is an important determinant of model stability (Mendez et al 2019c). However, to our knowledge, an extensive comparison of different data cleaning (Broadhurst et al 2018), pre-treatment (van den Berg et al 2006), and imputation (Di Guida et al 2016;Do et al 2018) procedure options has not been performed for ANNs.…”

“…While true effectiveness of a model can only be assessed using test data (Westerhuis et al 2008;Xia et al 2013), for small data sets it is dangerous to use a single random data split as the only means of model evaluation, as the random test data set may not accurately represent the training data set (Mendez et al 2019c). An alternative is to use bootstrap resampling.…”

“…The computational libraries developed for this study require data to be converted to a standardised format using the tidy data framework (Wickham, 2014). This standardised format has been previously described (Mendez et al 2019b(Mendez et al , 2019c, and allows for the efficient reuse of these workflows for other studies. To demonstrate this, we include the application of the identical workflows and visualisation techniques to a second previously published dataset (Ganna et al 2016) as a supplementary document.…”

“…TensorFlow and PyTorch), and the general success within industry and other research fields, the reintroduction of ANNs warrants renewed investigation. We recently showed that ANNs have similar predictive ability to PLS across multiple diverse metabolomics data sets (Mendez et al 2019c). However, within the domain of metabolomics, if ANNs are to become a truly viable alternative to PLS it will be necessary to develop similar standardised and robust methods for data visualisation, evaluation, and statistical inference (Mendez et al 2019a).…”