A Comparison of Automatically Extracted Quantitative EEG Features for Seizure Risk Stratification in Neonatal Encephalopathy.

PURPOSE: Seizures occur in up to 40% of neonates with neonatal encephalopathy. Earlier identification of seizures leads to more successful seizure treatment, but is often delayed because of limited availability of continuous EEG monitoring. Clinical variables poorly stratify seizure risk, and EEG use to stratify seizure risk has previously been limited by need for manual review and artifact exclusion. The goal of this study is to compare the utility of automatically extracted quantitative EEG (qEEG) features for seizure risk stratification. METHODS: We conducted a retrospective analysis of neonates with moderate-to-severe neonatal encephalopathy who underwent therapeutic hypothermia at a single center. The first 24 hours of EEG underwent automated artifact removal and qEEG analysis, comparing qEEG features for seizure risk stratification. RESULTS: The study included 150 neonates and compared the 36 (23%) with seizures with those without. Absolute spectral power best stratified seizure risk with area under the curve ranging from 63% to 71%, followed by range EEG lower and upper margin, median and SD of the range EEG lower margin. No features were significantly more predictive in the hour before seizure onset. Clinical examination was not associated with seizure risk. CONCLUSIONS: Automatically extracted qEEG features were more predictive than clinical examination in stratifying neonatal seizure risk during therapeutic hypothermia. qEEG represents a potential practical bedside tool to individualize intensity and duration of EEG monitoring and decrease time to seizure recognition. Future work is needed to refine and combine qEEG features to improve risk stratification.

Astrocytes as a mechanism for contextually-guided network dynamics and function.

Astrocytes are a ubiquitous and enigmatic type of non-neuronal cell and are found in the brain of all vertebrates. While traditionally viewed as being supportive of neurons, it is increasingly recognized that astrocytes play a more direct and active role in brain function and neural computation. On account of their sensitivity to a host of physiological covariates and ability to modulate neuronal activity and connectivity on slower time scales, astrocytes may be particularly well poised to modulate the dynamics of neural circuits in functionally salient ways. In the current paper, we seek to capture these features via actionable abstractions within computational models of neuron-astrocyte interaction. Specifically, we engage how nested feedback loops of neuron-astrocyte interaction, acting over separated time-scales, may endow astrocytes with the capability to enable learning in context-dependent settings, where fluctuations in task parameters may occur much more slowly than within-task requirements. We pose a general model of neuron-synapse-astrocyte interaction and use formal analysis to characterize how astrocytic modulation may constitute a form of meta-plasticity, altering the ways in which synapses and neurons adapt as a function of time. We then embed this model in a bandit-based reinforcement learning task environment, and show how the presence of time-scale separated astrocytic modulation enables learning over multiple fluctuating contexts. Indeed, these networks learn far more reliably compared to dynamically homogeneous networks and conventional non-network-based bandit algorithms. Our results fuel the notion that neuron-astrocyte interactions in the brain benefit learning over different time-scales and the conveyance of task-relevant contextual information onto circuit dynamics.

Propofol enhancement of slow wave sleep to target the nexus of geriatric depression and cognitive dysfunction: protocol for a phase I open label trial.

INTRODUCTION: Late-life treatment-resistant depression (LL-TRD) is common and increases risk for accelerated ageing and cognitive decline. Impaired sleep is common in LL-TRD and is a risk factor for cognitive decline. Slow wave sleep (SWS) has been implicated in key processes including synaptic plasticity and memory. A deficiency in SWS may be a core component of depression pathophysiology. The anaesthetic propofol can induce electroencephalographic (EEG) slow waves that resemble SWS. Propofol may enhance SWS and oral antidepressant therapy, but relationships are unclear. We hypothesise that propofol infusions will enhance SWS and improve depression in older adults with LL-TRD. This hypothesis has been supported by a recent small case series. METHODS AND ANALYSIS: SWIPED (Slow Wave Induction by Propofol to Eliminate Depression) phase I is an ongoing open-label, single-arm trial that assesses the safety and feasibility of using propofol to enhance SWS in older adults with LL-TRD. The study is enrolling 15 English-speaking adults over age 60 with LL-TRD. Participants will receive two propofol infusions 2-6 days apart. Propofol infusions are individually titrated to maximise the expression of EEG slow waves. Preinfusion and postinfusion sleep architecture are evaluated through at-home overnight EEG recordings acquired using a wireless headband equipped with dry electrodes. Sleep EEG recordings are scored manually. Key EEG measures include sleep slow wave activity, SWS duration and delta sleep ratio. Longitudinal changes in depression, suicidality and anhedonia are assessed. Assessments are performed prior to the first infusion and up to 10 weeks after the second infusion. Cognitive ability is assessed at enrolment and approximately 3 weeks after the second infusion. ETHICS AND DISSEMINATION: The study was approved by the Washington University Human Research Protection Office. Recruitment began in November 2022. Dissemination plans include presentations at scientific conferences, peer-reviewed publications and mass media. Positive results will lead to a larger phase II randomised placebo-controlled trial. TRIAL REGISTRATION NUMBER: NCT04680910.

Heterogeneous Forgetting Rates and Greedy Allocation in Slot-Based Memory Networks Promotes Signal Retention.

A key question in the neuroscience of memory encoding pertains to the mechanisms by which afferent stimuli are allocated within memory networks. This issue is especially pronounced in the domain of working memory, where capacity is finite. Presumably the brain must embed some "policy" by which to allocate these mnemonic resources in an online manner in order to maximally represent and store afferent information for as long as possible and without interference from subsequent stimuli. Here, we engage this question through a top-down theoretical modeling framework. We formally optimize a gating mechanism that projects afferent stimuli onto a finite number of memory slots within a recurrent network architecture. In the absence of external input, the activity in each slot attenuates over time (i.e., a process of gradual forgetting). It turns out that the optimal gating policy consists of a direct projection from sensory activity to memory slots, alongside an activity-dependent lateral inhibition. Interestingly, allocating resources myopically (greedily with respect to the current stimulus) leads to efficient utilization of slots over time. In other words, later-arriving stimuli are distributed across slots in such a way that the network state is minimally shifted and so prior signals are minimally "overwritten." Further, networks with heterogeneity in the timescales of their forgetting rates retain stimuli better than those that are more homogeneous. Our results suggest how online, recurrent networks working on temporally localized objectives without high-level supervision can nonetheless implement efficient allocation of memory resources over time.

Dynamical models reveal anatomically reliable attractor landscapes embedded in resting state brain networks.

Analyses of functional connectivity (FC) in resting-state brain networks (RSNs) have generated many insights into cognition. However, the mechanistic underpinnings of FC and RSNs are still not well-understood. It remains debated whether resting state activity is best characterized as noise-driven fluctuations around a single stable state, or instead, as a nonlinear dynamical system with nontrivial attractors embedded in the RSNs. Here, we provide evidence for the latter, by constructing whole-brain dynamical systems models from individual resting-state fMRI (rfMRI) recordings, using the Mesoscale Individualized NeuroDynamic (MINDy) platform. The MINDy models consist of hundreds of neural masses representing brain parcels, connected by fully trainable, individualized weights. We found that our models manifested a diverse taxonomy of nontrivial attractor landscapes including multiple equilibria and limit cycles. However, when projected into anatomical space, these attractors mapped onto a limited set of canonical RSNs, including the default mode network (DMN) and frontoparietal control network (FPN), which were reliable at the individual level. Further, by creating convex combinations of models, bifurcations were induced that recapitulated the full spectrum of dynamics found via fitting. These findings suggest that the resting brain traverses a diverse set of dynamics, which generates several distinct but anatomically overlapping attractor landscapes. Treating rfMRI as a unimodal stationary process (i.e., conventional FC) may miss critical attractor properties and structure within the resting brain. Instead, these may be better captured through neural dynamical modeling and analytic approaches. The results provide new insights into the generative mechanisms and intrinsic spatiotemporal organization of brain networks.

Precision data-driven modeling of cortical dynamics reveals idiosyncratic mechanisms underlying canonical oscillations.

Task-free brain activity affords unique insight into the functional structure of brain network dynamics and is a strong marker of individual differences. In this work, we present an algorithmic optimization framework that makes it possible to directly invert and parameterize brain-wide dynamical-systems models involving hundreds of interacting brain areas, from single-subject time-series recordings. This technique provides a powerful neurocomputational tool for interrogating mechanisms underlying individual brain dynamics ("precision brain models") and making quantitative predictions. We extensively validate the models' performance in forecasting future brain activity and predicting individual variability in key M/EEG markers. Lastly, we demonstrate the power of our technique in resolving individual differences in the generation of alpha and beta-frequency oscillations. We characterize subjects based upon model attractor topology and a dynamical-systems mechanism by which these topologies generate individual variation in the expression of alpha vs. beta rhythms. We trace these phenomena back to global variation in excitation-inhibition balance, highlighting the explanatory power of our framework in generating mechanistic insights.

A conceptual framework for astrocyte function.

The participation of astrocytes in brain computation was hypothesized in 1992, coinciding with the discovery that these cells display a form of intracellular Ca2+ signaling sensitive to neuroactive molecules. This finding fostered conceptual leaps crystalized around the idea that astrocytes, once thought to be passive, participate actively in brain signaling and outputs. A multitude of disparate roles of astrocytes has since emerged, but their meaningful integration has been muddied by the lack of consensus and models of how we conceive the functional position of these cells in brain circuitry. In this Perspective, we propose an intuitive, data-driven and transferable conceptual framework we coin 'contextual guidance'. It describes astrocytes as 'contextual gates' that shape neural circuitry in an adaptive, state-dependent fashion. This paradigm provides fresh perspectives on principles of astrocyte signaling and its relevance to brain function, which could spur new experimental avenues, including in computational space.

Computational modelling in disorders of consciousness: Closing the gap towards personalised models for restoring consciousness.

Disorders of consciousness are complex conditions characterised by persistent loss of responsiveness due to brain injury. They present diagnostic challenges and limited options for treatment, and highlight the urgent need for a more thorough understanding of how human consciousness arises from coordinated neural activity. The increasing availability of multimodal neuroimaging data has given rise to a wide range of clinically- and scientifically-motivated modelling efforts, seeking to improve data-driven stratification of patients, to identify causal mechanisms for patient pathophysiology and loss of consciousness more broadly, and to develop simulations as a means of testing in silico potential treatment avenues to restore consciousness. As a dedicated Working Group of clinicians and neuroscientists of the international Curing Coma Campaign, here we provide our framework and vision to understand the diverse statistical and generative computational modelling approaches that are being employed in this fast-growing field. We identify the gaps that exist between the current state-of-the-art in statistical and biophysical computational modelling in human neuroscience, and the aspirational goal of a mature field of modelling disorders of consciousness; which might drive improved treatments and outcomes in the clinic. Finally, we make several recommendations for how the field as a whole can work together to address these challenges.

Macroperiodic Oscillations: A Potential Novel Biomarker of Outcome in Neonatal Encephalopathy.

PURPOSE: Neonatal encephalopathy (NE) is a common cause of neurodevelopmental morbidity. Tools to accurately predict outcomes after therapeutic hypothermia remain limited. We evaluated a novel EEG biomarker, macroperiodic oscillations (MOs), to predict neurodevelopmental outcomes. METHODS: We conducted a secondary analysis of a randomized controlled trial of neonates with moderate-to-severe NE who underwent standardized clinical examination, magnetic resonance (MR) scoring, video EEG, and neurodevelopmental assessment with Bayley III evaluation at 18 to 24 months. A non-NE cohort of neonates was also assessed for the presence of MOs. The relationship between clinical examination, MR score, MOs, and neurodevelopmental assessment was analyzed. RESULTS: The study included 37 neonates with 24 of whom survived and underwent neurodevelopmental assessment (70%). The strength of MOs correlated with severity of clinical encephalopathy. MO strength and spread significantly correlated with Bayley III cognitive percentile (P = 0.017 and 0.046). MO strength outperformed MR score in predicting a combined adverse outcome of death or disability (P = 0.019, sensitivity 100%, specificity 77% vs. P = 0.079, sensitivity 100%, specificity 59%). CONCLUSIONS: MOs are an EEG-derived, quantitative biomarker of neurodevelopmental outcome that outperformed a comprehensive validated MRI injury score and a detailed systematic discharge examination in this small cohort. Future work is needed to validate MOs in a larger cohort and elucidate the underlying pathophysiology of MOs.

Central-positive complexes in ECT-induced seizures: Possible evidence for thalamocortical mechanisms.

OBJECTIVE: Central-positive complexes (CPCs) are elicited during electroconvulsive therapy (ECT) as generalized high-amplitude waveforms with maximum positive voltage over the vertex. While these complexes have been qualitatively assessed in previous literature, quantitative analyses are lacking. This study aims to characterize CPCs across temporal, spatial, and spectral domains. METHODS: High-density 64-electrode electroencephalogram (EEG) recordings during 50 seizures acquired from 11 patients undergoing right unilateral ECT allowed for evaluation of spatiotemporal characteristics of CPCs via source localization and spectral analysis. RESULTS: Peak-amplitude CPC scalp topology was consistent across seizures, showing maximal positive polarity over the midline fronto-central region and maximal negative polarity over the suborbital regions. The sources of these peak potentials were localized to the bilateral medial thalamus and cingulate cortical regions. Delta, beta, and gamma oscillations were correlated with the peak amplitude of CPCs during seizures induced during ketamine, whereas delta and gamma oscillations were associated with CPC peaks during etomidate anesthesia (excluding the dose-charge titration). CONCLUSIONS: Our findings demonstrate the consistency of CPC presence across participant, stimulus charge, time, and anesthetic agent, with peaks localized to bilateral medial thalamus and cingulate cortical regions and associated with delta, beta, and gamma band oscillations (depending on the anesthetic condition). SIGNIFICANCE: The consistency and reproducibility of CPCs offers ECT as a new avenue for studying the dynamics of generalized seizure activity and thalamocortical networks.

Primates chunk simultaneously-presented memoranda.

Though much research has characterized both the behavior and electrophysiology of spatial memory for single targets in non-human primates, we know much less about how multiple memoranda are handled. Multiple memoranda may interact in the brain, affecting the underlying representations. Mnemonic resources are famously limited, so items may compete for "space" in memory or may be encoded cooperatively or in a combined fashion. Understanding the mode of interaction will inform future neural studies. As a first step, we quantified interactions during a multi-item spatial memory task. Two monkeys were shown 1-4 target locations. After a delay, the targets reappeared with a novel target and the animal was rewarded for fixating the novel target. Targets could appear either all at once (simultaneous) or with intervening delays (sequential). We quantified the degree of interaction with memory rate correlations. We found that simultaneously presented targets were stored cooperatively while sequentially presented targets were stored independently. These findings demonstrate how interaction between concurrently memorized items depends on task context. Future studies of multi-item memory would be served by designing experiments to either control or measure the mode of this interaction.

Quiescence during burst suppression and postictal generalized EEG suppression are distinct patterns of activity.

OBJECTIVE: Periods of low-amplitude electroencephalographic (EEG) signal (quiescence) are present during both anesthetic-induced burst suppression (BS) and postictal generalized electroencephalographic suppression (PGES). PGES following generalized seizures induced by electroconvulsive therapy (ECT) has been previously linked to antidepressant response. The commonality of quiescence during both BS and PGES motivated trials to recapitulate the antidepressant effects of ECT using high doses of anesthetics. However, there have been no direct electrographic comparisons of these quiescent periods to address whether these are distinct entities. METHODS: We compared periods of EEG quiescence recorded from two human studies: BS induced in 29 healthy adult volunteers by isoflurane general anesthesia and PGES in 11 patients undergoing right unilateral ECT for treatment-resistant depression. An automated algorithm allowed detection of EEG quiescence based on a 10-microvolt amplitude threshold. Spatial, spectral, and temporal analyses compared quiescent epochs during BS and PGES. RESULTS: The median (interquartile range) voltage for quiescent periods during PGES was greater than during BS (1.81 (0.22) microvolts vs 1.22 (0.33) microvolts, p < 0.001). Relative power was greater for quiescence during PGES than BS for the 1-4 Hz delta band (p < 0.001), at the expense of power in the theta (4-8 Hz, p < 0.001), beta (13-30 Hz, p = 0.04) and gamma (30-70 Hz, p = 0.006) frequency bands. Topographic analyses revealed that amplitude across the scalp was consistently higher for quiescent periods during PGES than BS, whose voltage was within the noise floor. CONCLUSIONS: Quiescent epochs during PGES and BS have distinct patterns of EEG signals across voltage, frequency, and spatial domains. SIGNIFICANCE: Quiescent epochs during PGES and BS, important neurophysiological markers for clinical outcomes, are shown to have distinct voltage and frequency characteristics.

Detecting slow narrowband modulation in EEG signals.

BACKGROUND: We observed an unusual modulatory phenomenon in the electroencephalogram (EEG) of pediatric patients with acquired brain injury. The modulation is orders of magnitude slower than the fast EEG background activity, necessitating new analysis procedures to systematically detect and quantify the phenomenon. NEW METHOD: We propose a method for analyzing spatial and temporal relationships associated with slow, narrowband modulation of EEG. We extract envelope signals from physiological frequency bands of EEG. Then, we construct a sparse representation of the spectral content of the envelope signal across sliding windows. For the latter, we use an augmented LASSO regression to incorporate spatial and temporal filtering into the solution. The method can be applied to windows of variable length, depending on the desired frequency resolution. RESULTS: The sparse estimates of the envelope power spectra enable the detection of narrowband modulation in the millihertz frequency range. Subsequently, we are able to assess non-stationarity in the frequency and spatial relationships across channels. The method can be paired with unsupervised anomaly detection to identify windows with significant modulation. We validated such findings by applying our method to a control set of EEGs. COMPARISON WITH EXISTING METHODS: To our knowledge, no methods have been previously proposed to quantify second order modulation at such disparate time-scales. CONCLUSIONS: We provide a general EEG analysis framework capable of detecting signal content below 0.1 Hz, which is especially germane to clinical recordings that may contain multiple hours worth of continuous data.

Resolving and characterizing the incidence of millihertz EEG modulation in critically ill children.

OBJECTIVE: We analyze a slow electrographic pattern, Macroperiodic Oscillations (MOs), in the EEG from a cohort of young critical care patients (n = 43) with continuous EEG monitoring. We construct novel quantitative methods to quantify and understand MOs. METHODS: We applied a nonparametric bilevel spectral analysis to identify MOs, a millihertz (0.004-0.01 Hz) modulation of 5-15 Hz activity in two separate ICU patient cohorts (n = 195 total). We also developed a rigorous measure to quantify MOs strength and spatial expression, which was validated against surrogate noise data. RESULTS: Strong or spatially widespread MOs appear in both high clinical suspicion and a general ICU population. In the former, patients with strong or spatially widespread MOs tended to have worse clinical outcomes. Intracranial pressure and heart rate data from one patient provide insight into a potential broader physiological mechanism for MOs. CONCLUSIONS: We quantified millihertz EEG modulation (MOs) in cohorts of critically ill pediatric patients. We demonstrated high incidence in two patient populations. In a high suspicion cohort, MOs are associated with poor outcome, suggesting future potential as a diagnostic and prognostic aid. SIGNIFICANCE: These results support the existence of EEG dynamics across disparate time-scales and may provide insight into brain injury physiology in young children.

Synthesizing network dynamics for short-term memory of impulsive inputs.

Illuminating the mechanisms that the brain uses to manage and coordinate its resources is a core question in neuroscience. In particular, circuits and networks in the brain are able to encode, store and recall large amounts of information, in the service of a wide range of functionality. How do the various dynamical mechanisms within these networks allow for such coordination? We consider the specific problem of how the dynamics of networks can enact a representation of input stimuli that is retained over time, i.e., a form of short-term memory. We utilize modeling and control-theoretic methods to approach these questions, treating the state trajectory of a dynamical system as an abstract memory trace of prior inputs. The inputs impinge on the network via a variable gain, which is to be synthesized by optimization. In order to perpetuate these memory traces of stimuli, we propose that this gain is adapted to optimize: i) the error between a ground truth representation of stimuli and the encoding of them; as well as ii) overwriting of prior information. Optimizing over these central tenets of memory, we obtain a 'policy' for adapting the input gain that is dependent on the state of the network. This derived policy yields a recurrent neural network between the policy and the neural circuits, affirming existing theories that the prefrontal cortex may hold subnetworks dedicated to working memory while actively engaging with other neural subnetworks.

Control-theoretic integration of stimulation and electrophysiology for cognitive enhancement.

Transcranial electrical stimulation (tES) technology and neuroimaging are increasingly coupled in basic and applied science. This synergy has enabled individualized tES therapy and facilitated causal inferences in functional neuroimaging. However, traditional tES paradigms have been stymied by relatively small changes in neural activity and high inter-subject variability in cognitive effects. In this perspective, we propose a tES framework to treat these issues which is grounded in dynamical systems and control theory. The proposed paradigm involves a tight coupling of tES and neuroimaging in which M/EEG is used to parameterize generative brain models as well as control tES delivery in a hybrid closed-loop fashion. We also present a novel quantitative framework for cognitive enhancement driven by a new computational objective: shaping how the brain reacts to potential "inputs" (e.g., task contexts) rather than enforcing a fixed pattern of brain activity.

Macroperiodic Oscillations Are Associated With Seizures Following Acquired Brain Injury in Young Children.

PURPOSE: Seizures occur in 10% to 40% of critically ill children. We describe a phenomenon seen on color density spectral array but not raw EEG associated with seizures and acquired brain injury in pediatric patients. METHODS: We reviewed EEGs of 541 children admitted to an intensive care unit between October 2015 and August 2018. We identified 38 children (7%) with a periodic pattern on color density spectral array that oscillates every 2 to 5 minutes and was not apparent on the raw EEG tracing, termed macroperiodic oscillations (MOs). Internal validity measures and interrater agreement were assessed. We compared demographic and clinical data between those with and without MOs. RESULTS: Interrater reliability yielded a strong agreement for MOs identification (kappa: 0.778 [0.542-1.000]; P < 0.0001). There was a 76% overlap in the start and stop times of MOs among reviewers. All patients with MOs had seizures as opposed to 22.5% of the general intensive care unit monitoring population ( P < 0.0001). Macroperiodic oscillations occurred before or in the midst of recurrent seizures. Patients with MOs were younger (median of 8 vs. 208 days; P < 0.001), with indications for EEG monitoring more likely to be clinical seizures (42 vs. 16%; P < 0.001) or traumatic brain injury (16 vs. 5%, P < 0.01) and had fewer premorbid neurologic conditions (10.5 vs. 33%; P < 0.01). CONCLUSIONS: Macroperiodic oscillations are a slow periodic pattern occurring over a longer time scale than periodic discharges in pediatric intensive care unit patients. This pattern is associated with seizures in young patients with acquired brain injuries.

Enhancing task fMRI preprocessing via individualized model-based filtering of intrinsic activity dynamics.

Brain responses recorded during fMRI are thought to reflect both rapid, stimulus-evoked activity and the propagation of spontaneous activity through brain networks. In the current work, we describe a method to improve the estimation of task-evoked brain activity by first "filtering-out the intrinsic propagation of pre-event activity from the BOLD signal. We do so using Mesoscale Individualized NeuroDynamic (MINDy; Singh et al. 2020b) models built from individualized resting-state data to subtract the propagation of spontaneous activity from the task-fMRI signal (MINDy-based Filtering). After filtering, time-series are analyzed using conventional techniques. Results demonstrate that this simple operation significantly improves the statistical power and temporal precision of estimated group-level effects. Moreover, use of MINDy-based filtering increased the similarity of neural activation profiles and prediction accuracy of individual differences in behavior across tasks measuring the same construct (cognitive control). Thus, by subtracting the propagation of previous activity, we obtain better estimates of task-related neural effects.

Developing control-theoretic objectives for large-scale brain dynamics and cognitive enhancement.

The development of technologies for brain stimulation provides a means for scientists and clinicians to directly actuate the brain and nervous system. Brain stimulation has shown intriguing potential in terms of modifying particular symptom clusters in patients and behavioral characteristics of subjects. The stage is thus set for optimization of these techniques and the pursuit of more nuanced stimulation objectives, including the modification of complex cognitive functions such as memory and attention. Control theory and engineering will play a key role in the development of these methods, guiding computational and algorithmic strategies for stimulation. In particular, realizing this goal will require new development of frameworks that allow for controlling not only brain activity, but also latent dynamics that underlie neural computation and information processing. In the current opinion, we review recent progress in brain stimulation and outline challenges and potential research pathways associated with exogenous control of cognitive function.

Slow manifolds within network dynamics encode working memory efficiently and robustly.

Working memory is a cognitive function involving the storage and manipulation of latent information over brief intervals of time, thus making it crucial for context-dependent computation. Here, we use a top-down modeling approach to examine network-level mechanisms of working memory, an enigmatic issue and central topic of study in neuroscience. We optimize thousands of recurrent rate-based neural networks on a working memory task and then perform dynamical systems analysis on the ensuing optimized networks, wherein we find that four distinct dynamical mechanisms can emerge. In particular, we show the prevalence of a mechanism in which memories are encoded along slow stable manifolds in the network state space, leading to a phasic neuronal activation profile during memory periods. In contrast to mechanisms in which memories are directly encoded at stable attractors, these networks naturally forget stimuli over time. Despite this seeming functional disadvantage, they are more efficient in terms of how they leverage their attractor landscape and paradoxically, are considerably more robust to noise. Our results provide new hypotheses regarding how working memory function may be encoded within the dynamics of neural circuits.