Finding the balance between model complexity and performance: Using ventral striatal oscillations to classify feeding behavior in rats

Abstract

Binge eating is a disruptive and treatment-resistant behavior that complicates the management of obesity and comorbid psychiatric disorders. The distributed network that governs feeding behavior, particularly of palatable food, has a nexus point within the ventral striatum (VS). This important structure integrates information from energy homeostasis circuits of the hypothalamus and brain stem, cognitive control regions of the prefrontal cortex and central nodes of learning and memory. The VS is thus an ideal target for neuromodulation-based interventions (e.g., deep brain stimulation) for the treatment of appetitive disorders. Here, we extracted information about feeding behavior from VS oscillations in a rat model of binge eating -a critical step in the development of responsive neuromodulation systems for appetitive disorders. Local field potentials (LFPs) were recorded from bilateral nucleus accumbens (NAc) core and shell (regions within the VS) during binge sessions across varying conditions. LFP features (n=58) were used as predictors in statistical (logistic regression) and machine learning (lasso) models to predict the type and quantity of food consumed, classify rat behavior (feeding or not feeding) in real time, and predict the onset of future feeding bouts. The complexity of real-time models was then manipulated to determine the impact on model stability and performance.Performance of models was quantified (effect size, d) by comparison to the performance of models built from permuted data. Models were able to predict the amount of food eaten at baseline (mean absolute error, MAE = 2.6 ± 0.08 gm; d = 0.43), the increase in binge size following food deprivation (MAE = 0.8% ± 0.01%; d = 1.1) and the type of food eaten (accuracy = 59 ± 1%; d = 0.55). Other models were able to predict whether initiation of feeding was imminent (area under the curve, AUC = 0.81 ± 0.03; d = 2.68) up to 42.5 seconds before feeding began and classify current behavior as feeding or not-feeding (AUC = 0.87 ± 0.01; d = 3.27). For real-time models, an optimal balance between model complexity and performance was achieved using 3 LFP features with the largest contributions from features in the alpha and peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission.The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/241919 doi: bioRxiv preprint first posted online Jan. 2, 2018; high gamma frequencies. Similar models could be incorporated into algorithms used to improve treatment outcomes in adaptive neuromodulation systems for disorders of appetitive behavior.

Cannabis Vapour Exposure Alters Neural Circuit Oscillatory Activity In A Neurodevelopmental Model Of Schizophrenia: Exploring The Differential Impact Of Cannabis Constituents

“…Although cannabis use in schizophrenia is primarily observed in male patients 66 , previous studies have suggested that considerable sex differences exist in the oscillatory biomarkers associated with psychopathology 43 , warranting the assessment of sex as a biological variable in the measures assessed here. Future investigations will be performed using the same protocol to detect any changes in schizophrenia-like behaviours after acute and chronic cannabis vapour exposure using the NVHL rat, as past investigations indicate that THC-induced reductions in oscillatory activity relate to deficits in sensorimotor gating and working memory 36 67 ; we will also aim to identify which electrophysiological features are related to behavioural deficits 68 . Additionally, understanding how cessation of cannabis use impacts aberrant oscillatory activity may help further characterize the causal contribution of oscillatory activity to effects of cannabis vapour exposure on the symptoms of schizophrenia.…”

Sex differences in the ability of corticostriatal oscillations to predict rodent alcohol consumption

“…In order to link corticostriatal activity to biological sex or phase of estrous we used an unbiased machine learning approach similar to what we have published previously (30,31). We built predictive models using corticostriatal LFPs to classify rats by biological sex and female rats by phase of estrous.…”

“…We used a similar machine learning approach (as described above) to link corticostriatal activity to alcohol intake levels, except the outcomes were continuous (g/kg of alcohol consumed by each rat across each day) rather than binary. If the lasso indicated that information existed in the LFP signal, we implemented exhaustive single feature regressions using each LFP predictor to determine the relative information content of each feature, as we have previously described in detail (31).…”

A comprehensive characterization of rhythmic spiking activity in the rat ventral striatum

“…In the ventral striatum (vStr), a brain structure involved in the motivational control of behavior, several conditions are in place to suggest that oscillations can usefully inform studies of its function. First, the vStr LFP shows a full spectrum of rhythmic activity, including prominent gamma-band oscillations but also delta, theta, and beta-band activity, which are modulated in association with behavior and task variables and relate to local spiking activity (Leung and Yim, 1993; van der Meer and Redish, 2009; Berke, 2009; Kalenscher et al, 2010; Howe et al, 2011; Donnelly et al, 2014; Dejean et al, 2017; Dwiel et al, 2019). Manipulations of the dopaminergic and endocannabinoid system modify vStr LFP oscillations (Berke, 2009; Lemaire et al, 2012; Morra et al, 2012) and drug and disease states are associated with altered striatal LFPs (Dejean et al, 2017; Naze et al, 2018; Wu et al, 2018; Hultman et al, 2018).…”