Feature-Selection Based Cognitive Control

To the Editor: Improper cognitive interpretation of emotional events is a common attribute in depression and anxiety disorders. Hence, strategies causing adaptive interpretations of emotional states are now a part of treatment. “Cognitive control,” the ability to use cognitive and active coping strategies to modify emotional reactions, is in the core of attention in recent research. Cognitive control includes reinterpreting an emotional stimulus to adjust emotional reactions. Successful reappraisal of emotional events appears to be related to the abstract rule-like representations of the prefrontal cortex (PFC).^1,2 PFC representations can be studied from the perspective of the pattern-recognition domain, containing two important notions: “feature-extraction” and “feature-selection.” Feature-extraction is a dimension-reduction procedure to find a small number of features that are informative. The goal of feature-selection, also called Feature Subset Selection, is to select a yet-smaller subset of the extracted features that are considered to be the most informative.³ Feature selection may be performed by using a “search algorithm” that explores feature space to select candidate subsets, reevaluating at each step an “objective function” to evaluate these candidates, and returning a measure of their goodness, until a “stopping criterion” is reached. From an algorithmic perspective, employing causality in designing feature-selection methods may enhance interpretations. The goal of “causal feature-selection” is to uncover causal relationships, with purposes as: 1) Prediction: predicting future data; 2) Data understanding: finding a model of underlying data production mechanisms; 3) Manipulation: predicting the consequence of actions by manipulating the system; and 4) Counterfactual prediction: given that a specific outcome was observed, predicting what would have happened if a different action had been taken.⁴ In our opinion, PFC representations are features extracted from environmental information (actions and emotional events in daily life). Because it is impossible to process these inputs simultaneously, procedures are needed to select some inputs and discard others. Those procedures that correspond to feature subset-selection in the pattern-recognition domain are called “attention.” We can benefit from strategies in feature subset-selection to find effective cognitive coping strategies to treat improper interpretation of events. Reinterpretation of emotional states as an approach in cognitive control corresponds to reevaluation of “objective function” for candidate feature subsets. For a successful reappraisal, objective function can be chosen from the goals of causal feature-selection. For example, we have usually interpretations for life events, which produce an emotion in us. Using causal feature-selection, we can reevaluate (reinterpret) our negative emotions until a stopping-criterion, such as changing the current emotion, is reached. This reevaluation can be done by using an algorithmic objective function with four steps: 1) Prediction: conscious and proper understanding of current emotion; 2) Data understanding: determining the underlying mechanisms of the current emotion interpretation system; 3) Manipulation: predicting the consequences of current interpretation in the case of an imaginary different situation or event; and 4) Counterfactual prediction: predicting what emotion would have emerged if a different interpretation had been taken. Surely, these hypothetical interventions can be a part of treatment, if validated in clinical trials.