Feature impact assessment: a new score to identify relevant metabolomics features in artificial neural networks using validated labels

Abstract

Introduction: Artificial Neural Networks (ANN) are increasingly used in metabolomics but are hard to interpret.

Objectives: We aimed at developing a feature impact score that is model-agnostic, simple, and interpretable.

Methods: Feature Impact Assessment (FIA) is calculated by varying combinations of features within their observed value range and checking for changes in prediction outcomes. FIA was implemented in R and tested on metabolomics datasets.

Results: FIA exceeded LIME and SHAP in selecting biologically meaningful features. Values were comparable across different ANN architectures.

Conclusion: FIA is a novel score ranking feature impact, helping interpreting ANN in the metabolomics field.

Keywords: Artificial intelligence; Artificial neural networks; Deep learning; Feature selection; Machine learning; Metabolomics.