Francisco J. Flores

Objective. Rapid and accurate detection of electrographic seizures is critical for both clinical diagnosis and neuroscience research. Although seizure identification is commonly performed in the time domain, analysis in the time–frequency domain provides a more comprehensive representation of seizure characteristics. Approach. In this study, we present a 3D convolutional neural network (CNN) that incorporates a trainable continuous wavelet transform (CWT) layer, enabling adaptive time–frequency feature learning directly from raw EEG. To address common data challenges, we augment the 3D CNN for pre-training with contrastive learning, comparing contrastive predictive coding against bidirectional contrastive learning. Main results. On single-channel and multi-channel data, the standard 3D CNN outperformed both a 2D CNN with pre-computed CWT and a 1D CNN that processes raw signals, achieving > 95% accuracy down to 0.5-second segments. Contrastive pre-training improved the performance of the 3D CNN in both low-data and class-imbalance scenarios. Further experiments involving band-pass filtering and temporal shuffling revealed that classification is driven primarily by low-frequency patterns and statistical features rather than temporal dependencies. The proposed framework also maintained > 90% accuracy with moderate noise and downsampling applied to inputs, as well as when cross-subject generalization was evaluated using held-out subjects. Significance. We show that a 3D CNN with a trainable CWT layer and contrastive pre-training enables accurate sub-second seizure detection and effectively mitigates data limitations common in clinical settings. This work demonstrates that time–frequency embedding within CNNs, augmented by self-supervised pre-training, offers a promising path toward architectures for sub-second seizure detection in the presence of practical limitations of real-world scenarios.

Rapid and accurate detection of electrographic seizures is critical for both clinical diagnosis and neuroscience research. Although seizure identification is commonly performed in the time domain, analysis in the time-frequency domain provides a more comprehensive representation of seizure characteristics. In this study, we present a 3D convolutional neural network (CNN) that incorporates a trainable continuous wavelet transform (CWT) layer, enabling adaptive time-frequency feature learning directly from raw EEG. To address common data challenges, we augment the 3D CNN for pre-training with contrastive learning, comparing contrastive predictive coding (CPC) against bidirectional contrastive learning (BiCL). On single-channel and multi-channel data, the standard 3D CNN outperformed both a 2D CNN with pre-computed CWT and a 1D CNN that processes raw signals, achieving > 95% accuracy down to 0.5-second segments. Compared to the standard 3D CNN, the 3D CNN with BiCL pre-training showed superior performance in both low-data and class imbalance scenarios. Further experiments involving band-pass filtering and temporal shuffling revealed that classification is driven primarily by low-frequency patterns and statistical features rather than temporal dependencies. The proposed framework also maintained > 90% accuracy with moderate noise and downsampling applied to inputs, as well as when cross-subject generalization was evaluated using held-out subjects. We show that a 3D CNN with a trainable CWT layer and BiCL pre-training enables accurate sub-second seizure detection and effectively mitigates data limitations common in clinical settings. This work demonstrates that time-frequency embedding within CNNs, augmented by self-supervised pre-training, offers a promising path toward architectures for sub-second seizure detection in the presence of practical limitations of real-world scenarios.

Background Deep brain stimulation of central thalamus (CT-DBS) has potential for modulating states of consciousness, but it can also trigger spike-wave discharges (SWDs). Objectives To report the probability of inducing SWDs during CT-DBS in awake mice. Methods Mice were implanted with electrodes to deliver unilateral and bilateral CT-DBS at different frequencies while recording EEG. We titrated stimulation current by gradually increasing it at each frequency until an SWD appeared. Subsequent stimulations to test arousal modulation were performed at the current one step below the current that caused an SWD during titration. Results In 2.21% of the test stimulations (10 out of 12 mice), CT-DBS caused SWDs at currents lower than the titrated current, at currents as low as 20 μA. Conclusion Our study found a small but significant probability of inducing SWDs even after titration and at relatively low currents. EEG should be closely monitored for SWDs when performing CT-DBS in both research and clinical settings.

Given a continuous-domain signal that can be modeled as the superposition of localized events from multiple sources, the goal of Convolutional Dictionary Learning (CDL) is to identify the location of the events—by Convolutional Sparse Coding (CSC)—and learn the template for each source—through a Convolutional Dictionary Update (CDU) step. In practice, because we observe samples of the continuous-domain signal on a discrete grid, classical CSC methods can only produce estimates of the locations of the events on this grid, which degrades the performance of the CDU step. We introduce a CDL framework that significantly reduces the errors arising from performing the estimation on the grid. Specifically, we construct an expanded dictionary that comprises, not only discrete shifts of the templates, but also variants shifted by a non-integer amount and smoothly interpolated, that enable CSC and CDU to operate at a finer resolution than that of the original sampling grid. We term this approach CDL with grid refinement. For CSC, we develop a novel computationally efficient algorithm, termed Convolutional Orthogonal Matching Pursuit with an interpolated dictionary (COMP-INTERP).We use simulated data to compare COMP-INTERP to state-of-the-art CSC algorithms for estimating off-the-grid events, and demonstrate that it 1) achieves a competitive level of accuracy, and 2) is one order of magnitude faster. For CDU, we derive a novel procedure to update the templates given sparse codes that can occur both on and off the sampling grid. We also show that 3) dictionary update with the overcomplete dictionary yields more accurate templates. Finally, we demonstrate the competitive performance of the algorithms in two applications, namely spike sorting and super-resolution microscopy.

The precise identification of loss of consciousness (LOC) is key to studying the effects of anesthetic drugs in neural systems. The standard behavioral assay for identifying LOC in rodents is the Loss of Righting Reflex (LORR), assessed by placing the animal in the supine position every minute until it fails to right itself. However, this assay cannot be used when the rodents are head-fixed, which limits the use of powerful techniques such as multi-electrode recordings, in vivo patch clamp, and neuronal imaging. In these situations, an alternative way to assess LOC is needed. We propose that loss of movement (LOM) in whiskers and paws of head-fixed animals can be used as an alternative behavioral assay in head-fixed animals. Unlike LORR, LOM in whiskers and paws is much harder to detect by visual inspection. Therefore, we developed a method to automatically assess for LOM of whiskers and paws in head fixed rodents during in vivo patch clamp recordings. Our method uses an algorithm based on optical flow and point-process filtering which can be run on images acquired on regular cameras at low frame-rates. We show that the algorithm can achieve at least comparable accuracy in detecting LOC when compared with consensus among human observers, as well as improved precision when compared with individual observers. In the future, we aim to to expand the method to detect more behavioral end-points during anesthesia such as paradoxical excitation. Eventually, we hope to enable multi-modal anesthesia studies, which incorporates behavioral and neurophysiological data.

Spike sorting refers to the problem of assigning action potentials observed in extra-cellular recordings of neural activity to the neuron(s) from which they originate. We cast this problem as one of learning a convolutional dictionary from raw multi-electrode waveform data, subject to sparsity constraints. In this context, sparsity refers to the number of neurons that are allowed to spike simultaneously. The convolutional dictionary setting, along with its assumptions (e.g. refractoriness) that are motivated by the spike-sorting problem, let us give theoretical bounds on the sample complexity of spike sorting as a function of the number of underlying neurons, the rate of occurrence of simultaneous spiking, and the firing rate of the neurons. We derive memory/computation-efficient convolutional versions of OMP (cOMP) and KSVD (cKSVD), popular algorithms for sparse coding and dictionary learning respectively. We demonstrate via simulations that an algorithm that alternates between cOMP and cKSVD can recover the underlying spike waveforms successfully, assuming few neurons spike simultaneously, and is stable in the presence of noise. We also apply the algorithm to extra-cellular recordings from a tetrode in the rat Hippocampus.

The activities of groups of neurons in a circuit or brain region are important for neuronal computations that contribute to behaviors and disease states. Traditional extracellular recordings have been powerful and scalable, but much less is known about the intracellular processes that lead to spiking activity. We present a robotic system, the multipatcher, capable of automatically obtaining blind whole-cell patch clamp recordings from multiple neurons simultaneously. The multipatcher significantly extends automated patch clamping, or 'autopatching’, to guide four interacting electrodes in a coordinated fashion, avoiding mechanical coupling in the brain. We demonstrate its performance in the cortex of anesthetized and awake mice. A multipatcher with four electrodes took an average of 10 min to obtain dual or triple recordings in 29% of trials in anesthetized mice, and in 18% of the trials in awake mice, thus illustrating practical yield and throughput to obtain multiple, simultaneous whole-cell recordings in vivo.

Target controlled infusion (TCI) of intraveneous anesthetics can assist clinical practitioners to provide improved care for General Anesthesia (GA). Pharmacoki-netic/Pharmacodynamic (PK/PD) models help in relating the anesthetic drug infusion to observed brain activity inferred from electroencephalogram (EEG) signals. The parameters in popular population PK/PD models for propofol-induced GA (Marsh and Schnider models) are either verified based on proprietary functions of the EEG signal which are difficult to correlate with the neurophysiological models of anesthesia, or the marker itself needs to be estimated simultaneously with the PD model. Both these factors make these existing paradigms challenging to apply in real-time context where a patient-specific tuning of parameters is desired. In this work, we propose a simpler EEG marker from frequency domain description of EEG and develop two corresponding PK/PD modeling approaches which differ in whether they use existing population-level PK models (approach 1) or not (approach 2). We use a simple deterministic parameter estimation approach to identify the unknown PK/PD model parameters from an existing human EEG data-set. We infer that both approaches 1 and 2 yield similar and reasonably good fits to the marker data. This work can be useful in developing patient-specific TCI strategies to induce GA.

We analyzed subthreshold activity preceding spikes in hippocampus and barrel cortex of awake mice. Aperiodic voltage ramps extending over tens to hundreds of milliseconds consistently precede and facilitate spikes, in a manner dependent on both their amplitude and their duration. These voltage ramps represent a “mesoscale” activated state that gates spike production in vivo.