For decades, neuroscientists have relied on a technique for reading out electrical “spikes” of brain activity in live, behaving subjects that tells them very little about the types of cells they are monitoring.
In a new study, researchers at the University of Tuebingen and MIT’s Picower Institute for Learning and Memory demonstrate a way to increase their insight by distinguishing four distinct classes of cells from that spiking information.
The advance offers brain researchers the chance to better understand how different kinds of neurons are contributing to behavior, perception and memory, and how they are malfunctioning in cases of psychiatric or neurological diseases.
Much like mechanics can better understand and troubleshoot a machine by watching how each part works as it runs, neuroscientists, too, are better able to understand the brain when they can tease apart the roles different cells play while it thinks.
“We know from anatomical studies that there are multiple types of cells in the brain and if they are there, they must be there for a reason,” said Earl Miller, Picower Professor of Neuroscience in the Department of Brain and Cognitive Sciences at MIT, and co-senior author of the paper in Current Biology.
“We can’t truly understand the functional circuitry of the brain until we fully understand what different roles these different cell types might play.”
Miller collaborated with the Tuebingen-based team of lead author Caterina Trainito, Constantin von Nicolai and Professor Markus Siegel, co-senior author and a former postdoc in Miller’s lab, to develop the new way to wring more neuron type information from electrophysiology measurements.
Those measures track the rapid voltage changes, or spikes, that neurons exhibit as they communicate in circuits, a phenomenon essential for brain function.
“Identifying different cell types will be key to understand both local and large-scale information processing in the brain,” Siegel said.
Four is greater than two
At best, neuroscientists have so far only been able to determine from electrophysiology whether a neuron was excitatory or inhibitory.
That’s because they only analyzed the difference in the width of the spike.
The typical amount of data in an electrophysiology study – spikes from a few hundred neurons -only supported that single degree of distinction, Miller said.
But the new study could go farther because it derives from a dataset of recordings from nearly 2,500 neurons.
Miller and Siegel gathered the data years ago at MIT from three regions in the cortex of animals who were performing experimental tasks that integrated perception and decision making.
“We recognized the uncommonly rich resource at our disposal,” Siegel said.
Thus, the team decided to put the dataset through a ringer of sophisticated statistical and computational tools to analyze the waveforms of the spikes.
Their analysis showed that the waveforms could actually be sorted along two dimensions:
how quickly the waveform ranges between its lowest and highest voltage (“trough to peak duration”), and how quickly the voltage changes again afterward, returning from the peak to the normal level (“repolarization time”).
Plotting those two factors against each other neatly sorted the cells into four distinct “clusters.” Not only were the clusters evident across the whole dataset, but individually within each of the three cortical regions, too.
For the distinction to have any meaning, the four classes of cells would have to have functional differences.
To test that, the researchers decided to sort the cells out based on other criteria such as their “firing rate” (how often they spike), whether they tend to fire in bursts and how variable their intervals are between spikes – all factors in how they participate in and influence the circuits they are connected in.
Indeed, the cell classes remained distinct by these measures.
In yet another phase of analysis the cell classes also remained distinguishable as the researchers watched them respond to the animals perceiving and processing visual stimulation.
But in this case, they saw the cells play different roles in different regions.
A class 1 cell, for example, might respond differently in one region than it did in another.
“These cell types are truly different cell types that have different properties,” Miller said. “But they have different functions in different cortical areas because different cortical areas have different functions.”
New research capability
In the study the authors speculate about which real neuron types their four mathematically defined classes most closely resemble, but they don’t yet offer a definitive determination.
Still, Miller said the finer-grained distinctions the study draws are enough to make him want to reanalyze old neural spiking data to see what new things he can learn.
One of Miller’s main research interests is the nature of working memory – our ability to hold information like directions in mind while we make use of it.
His research has revealed that it is enabled by a complex interplay of brain regions and precisely timed bursts of neural activity.
Now he may be able to figure out how different classes of neurons play specific roles in specific regions to endow us with this handy mental ability.
And both Miller’s and Siegel’s labs are particularly interested in different brain rhythms, which are abundant in the brain and likely play a key role for orchestrating communication between neurons.
The new results open a powerful new window for them to unravel which role different neuron classes play for these brain rhythms.
Journal information: Current Biology
Provided by Massachusetts Institute of Technology
Brain-computer interface (BCI) has become a hot topic of research as it is increasingly being used in gaming applications1 and in stroke rehabilitation2–7 for translating the brain signals of the imagined task into intended movement of the limb that has been paralyzed.
Moreover, BCI research is also being carried out to detect in advance that a person is going to suffer from a seizure attack so that they can be informed in order to prevent accident or serious injuries10–12.
Electroencephalography (EEG) signal obtained using non-invasive sensors have been widely used12,13 for these purposes due to its low cost, easy to use and that it does not require any surgery as required by invasive sensors.
BCI using non-invasive sensors are approaching their required technological advancements and translate neural activities into meaningful information that can be used to drive activity-dependent neuroplasticity or rehabilitation robots.
Although some promising results have been achieved, BCI for rehabilitation is still a new and emerging field.
Therefore, being able to classify the different tasks with greater accuracy using the EEG signal will not only be beneficial for gaming and rehabilitation but also help in better detection of diseases or abnormal behaviors such as seizure12,14, sleep apnea15, sleep stages16,17, and drowsiness18 detection.
Thus, developing a BCI system that can classify different types of EEG signals with high accuracy is highly desirable.
Common spatial pattern has been widely used for extracting the features from EEG signals for classification.
A poorly selected frequency band may contain unwanted or redundant information and will degrade the performance of the overall system.
The selection of the frequency bands plays a key role in extracting significant features and manually tuning the filters will be a difficult task.
These methods use multiple filter bands to filter the signal into different sub-bands and then utilize CSP for extracting the features. Some approaches proposed different methods of selecting the best sub-bands21,24,32,34 while other approaches considered various feature selection techniques20,28,31,35–37 using all sub-bands to achieve promising results. While appropriately using multiple sub-bands helped achieve improved performance, it also increased the computation complexity of the system38.
Few researchers have also considered directly improving the CSP algorithm39–43 for better performance. Other methods that have been proposed use wavelet packet decomposition23, empirical mode decomposition19,29, Riemannian tangent space mapping22,44, artificial neural networks40,45,46 and deep learning47,48.
Deep learning has recently gained widespread attention in the field of signal processing. However, it has not been fully explored for EEG signal classification.
In this study, we focus on subject-dependent approach and propose an Optimized CSP and LSTM based predictor named OPTICAL. An LSTM network is a recurrent neural network consisting of LSTM layers having the ability to selectively remember important information for a longer period and is mostly used for sequence prediction. As reported in our previous works30, to keep the computational complexity of the system low, OPTICAL uses a single Butterworth band-pass filter with cutoff frequencies of 7–30 Hz.
Promising 10 × 10-fold cross-validation results have been obtained using OPTICAL, which has been evaluated using the BCI Competition IV dataset 149 and GigaDB dataset50. OPTICAL showed improvement in the classification performance (achieving average misclassification rate of 17.48% and 31.81% for BCI Competition IV dataset 1 and GigaDB dataset, respectively) and can be beneficial in developing improved BCI systems for rehabilitation.
Apart from this, if applied appropriately, it might also help detect seizure, sleep apnea and sleep stages with greater accuracy.
The results obtained are superior compared to other competing methods.
Thus, we have shown that appropriately using LSTM network can help develop improved BCI systems. Almost all the related works19,20,23,26,31,34,35,51,52considered classification of MI tasks, which were limited to binary class MI EEG signal classification problem. However, it should be noted that real-time EEG signal contains noise and other activities (such as eye blinking, eyeball movement up/down, eyeball movement left/right, jaw clenching and head movement left/right), referred to as non-task related EEG signals.
Thus, it is important to show that the proposed approach will be able to perform well if the implementation takes place in real-time.
Therefore, we utilize the rest-state and non-task related EEG signals to show that the proposed method will perform well for real-time classification. For this purpose, we have utilized the one-versus-rest approach (as using the one-versus-rest approach yields substantially better results than using the multi-class classification) for classification of the multi-class MI tasks using the conventional CSP algorithm.
The GigaDB dataset, which also provides the recordings for the rest-state and other non-tasks related signals, has been used to show the effectiveness of OPTICAL for real-time implementation. For real-time implementation, we achieved an average misclassification rate of 17.78% over 52 subjects using GigaDB dataset.