TBM performance assessment is based on a wide range of parameters from the encountered geology and the TBM itself, representing a complex system of machine-rock interaction. This complexity lends itself well to using machine learning and artificial intelligence methods for analysis and to develop predictive models. Nevertheless, prior to developing models, the selected feature set must first be optimized. In this chapter we present feature importance, feature engineering and feature selection techniques that are applicable to typical TBM datasets, including suitable tools for continuous, binary and/or categorical features (e.g. pre-processing, correlation, decision trees, random forest classifier, XGBoost), which, when combined, allow optimization of the features used to build predictive models. We illustrate the techniques as an iterative workflow using a case study of a gripper TBM in crystalline alpine terrain whose performance, as measured by the net advance rate, is affected by face instability. Within this non-rigid workflow feature engineering and feature importance co-evolve with insights from domain experts, rather than following a fixed linear sequence. Feature selection is applied more than once in the workflow as deeper insights are obtained with each subsequent step. We also show that considerable critical expert judgement is required when conducting feature optimization to address aspects of uncertainty, subjectivity, data quality and correlation of parameters, which can undermine the benefits of the optimization. The outcomes of the feature optimization are linked to conventional (non-ML) TBM performance analysis by incorporating established performance indices, such as tool-rock interaction indicators Penetration Index (PI) and Field Penetration Index (FPI), as engineered ratio features. We demonstrate, with this iterative feature optimization approach, that it is possible to reduce model complexity, improve interpretability and deepen the understanding of parameter relationships in TBM performance.