Oversampled and undersolved: Depressive rumination from a predictive coding perspective
Rumination is a widely recognized cognitive deviation in depression. An integrative view that combines clinical findings on rumination with theories of mental simulation and cognitive problem-solving could help explain the development and maintenance of rumination in a computationally and biologically plausible framework. In this review, we connect insights from neuroscience and computational psychiatry to elucidate rumination as repetitive but unsuccessful attempts at mental problem-solving. Appealing to a predictive processing account, we suggest that problem-solving is based on an algorithm that generates candidate behavior (policy primitives for problem solutions) using a Bayesian sampling approach, evaluates resulting policies for action, and then engages in instrumental learning to reduce prediction errors. We present evidence suggesting that this problem-solving algorithm is distorted in depression: Specifically, depressive rumination is regarded as excessive Bayesian sampling of candidates that is associated with high prediction errors without activation of the successive steps (policy evaluation, instrumental learning) of the algorithm. Thus, prediction errors cannot be decreased, and excessive resampling of the same problems occur. This then leads to reduced precision weighting attributed to external, “online” stimuli, low behavioral output and high opportunity costs due to the time-consuming nature of the sampling process itself. We review different computational reasons that make the proposed Bayesian sampling algorithm vulnerable to a ruminative „halting problem”. We also identify neurophysiological correlates of these deviations in pathological connectivity patterns of different brain networks. We conclude by suggesting future directions for research into behavioral and neurophysiological features of the model and point to clinical implications.