As detailed above, chronic pain and neuropathic pain impact multiple organ systems. Advancing the value of pain biomarkers depends on (1) selection of measurements and metrics that are the most mechanistically valid and informative, and (2) combining the selected measurements such that they mechanistically and statistically maximize accurate classification. Advancement of measurement accuracy is vital and the subsequent steps of the approach are entirely contingent upon the success of this step. This literature for the domains discussed in this manuscript is too voluminous for a single review. In the above reviewed literature, we attempted principally to focus on which biological systems and which biomarkers should be the focus of measurement. For effective application of measurements of these domains it is important to discuss approaches for measurement selection.

In Figure 2, we provide a significantly abbreviated schematic of key available statistical approaches to handling multimodal datasets in building composite biomarkers. We have highlighted four general areas of statistics/machine learning: (1) feature reduction (225), (2) classification (226, 227), (3) regression (228), and (4) clustering (229). Feature reduction can occur during or prior to classification, regression, or clustering. Feature reduction primarily focuses on two primary approaches: (1) integration of measurements toward creation of a composite variable to simplify and enhance model performance, and (2) effective feature reduction through variable selection to use optimal variables. Thus, feature reduction can represent the effective combining of strong measurements to a meaningful and robust latent variable or elimination of unnecessary, or statistically weak, measurements. Some methods, such as random forest, has built in feature reduction (230). Classification methods are often utilized to build toward categorical variables, however methods like neural networks are also designed for predicting continuous variables (231). Regression models are often used for the prediction of continuous measures or in the case of canonical approaches this can be with multiple dependent variables predicted simultaneously (232). Finally, in the case where there is no existent or optimal category or variable that the biomarkers seek to predict unsupervised approaches can be useful. With all these approaches variables can either be approached as linear or non-linear, although transformations and feature reduction approaches can mitigate these differences. It is important, regardless of approach, to understand the biological mechanisms being modeled by defining a model that best reflects the underlying systems to optimize prediction.

Two key methods for statistical reduction of variables are (1) selecting top ranking variables and (2) creation of composite variables by factor or component-based analysis. Random Forest, as depicted in Figure 2, can be utilized to determine importance scores by evaluating the hierarchical functionality of a given variable as a bifurcator for optimizing classification (226). Random Forest is not alone in its utility to provide variable importance ranking but provides a nice mechanism for this analysis. The statistical creation of composite variables can be done through principal component analysis (or independent component analysis among other methods) such that novel values are calculated for a set of variables that account for large swaths of variance with a single value vector (233, 234). This can substantially increase the efficiency of a model and serve to highlight a robust latent feature.