Edited by: Patrick O. Kanold, University of Maryland, USA
Reviewed by: Rustem Khazipov, Institut National de la Santé et de la Recherche Médicale, France; Paul Watkins, NIH/NINDS, USA
*Correspondence: Ileana L. Hanganu-Opatz, Developmental Neurophysiology, Neuroanatomy, University Medical Center Hamburg-Eppendorf, Falkrenried 94, 20251 Hamburg, Germany e-mail:
This article was submitted to the journal Frontiers in Neural Circuits.
†These authors have contributed equally to this work.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Flexible communication within the brain, which relies on oscillatory activity, is not confined to adult neuronal networks. Experimental evidence has documented the presence of discontinuous patterns of oscillatory activity already during early development. Their highly variable spatial and time-frequency organization has been related to region specificity. However, it might be equally due to the absence of unitary criteria for classifying the early activity patterns, since they have been mainly characterized by visual inspection. Therefore, robust and unbiased methods for categorizing these discontinuous oscillations are needed for increasingly complex data sets from different labs. Here, we introduce an unsupervised detection and classification algorithm for the discontinuous activity patterns of rodents during early development. For this, in a first step time windows with discontinuous oscillations vs. epochs of network “silence” were identified. In a second step, the major features of detected events were identified and processed by principal component analysis for deciding on their contribution to the classification of different oscillatory patterns. Finally, these patterns were categorized using an unsupervised cluster algorithm. The results were validated on manually characterized neonatal spindle bursts (SB), which ubiquitously entrain neocortical areas of rats and mice, and prelimbic nested gamma spindle bursts (NG). Moreover, the algorithm led to satisfactory results for oscillatory events that, due to increased similarity of their features, were more difficult to classify, e.g., during the pre-juvenile developmental period. Based on a linear classification, the optimal number of features to consider increased with the difficulty of detection. This algorithm allows the comparison of neonatal and pre-juvenile oscillatory patterns in their spatial and temporal organization. It might represent a first step for the unbiased elucidation of activity patterns during development.
Neuronal oscillations are a ubiquitous and robust phenomenon that is observed in various measures of brain activity. As such, they have received much attention as they provide effective means to time the firing of neurons (Fries et al.,
While the function and major underlying mechanisms of oscillatory activity within adult neuronal networks have been largely investigated and partially elucidated, much less is known about the activity patterns during neocortical maturation. Both human and animal research showed that coupling of neuronal networks in oscillatory rhythms emerges early during brain development (Anderson et al.,
To properly understand the function of these different patterns of early oscillatory activity it is mandatory to objectively characterize and distinguish them. This can be optimally achieved by using automatic analysis algorithms, as shown for human data (Vanhatalo et al.,
In this work, we describe a novel unsupervised method for detecting and classifying the different patterns of oscillatory activity in developing neocortices of anesthetized rodents. This method considered the root mean square (rms) of the recorded signal for the detection of early oscillatory events. The major features exhibited by these events were used for their classification based on a principle component analysis (PCA) and a fuzzy clustering algorithm. We calibrated the method for its reliability and yield on the prelimbic SB and NG from neonatal rats, the developmental stage at which these events are most clearly distinguishable. Therefore, we analyzed to which extent even single features may allow a robust classification. In the Results section we highlighted two applications of our method. First, we showed that the method performed equally well in classifying the neonatal activity in the primary visual cortex (V1) of the rat and the PFC of the mouse. With the developmental switch from discontinuous to continuous network oscillations in pre-juvenile rats (Brockmann et al.,
In the first part of this section we described the experimental paradigms and general analytical tools. In the second part we developed and calibrated a new three step unsupervised method for event classification. For this, we firstly showed how to initially detect discontinuous oscillatory events. We focused on the two major patterns of discontinuous oscillatory activity in the neonatal PL, the SB and NG. Subsequently, we quantified 11 features based on single oscillatory events, which captured the main characteristics of SB and NG. Finally, we described how to automatically cluster events into SB and NG oscillations based on a vector composed of these features. We concluded this section with a validation of the method by quantifying its classification performance by comparison to a manual classification of events in data from neonatal rats.
All experiments were performed in compliance with the German laws and the guidelines of the European Community for the use of animals in research and were approved (proposal number 94/08 and 111/12) by the local ethical committee. Pregnant Wistar rats were obtained at 14–17 days of gestation from the animal facility of the University Medical Center Hamburg-Eppendorf, housed individually in breeding cages with a 12 h light/12 h dark cycle and fed
Extracellular recordings were performed from the PFC (1.5–2.5 mm anterior to bregma suture, 0.1–1 mm from the midline) and V1 (0.5–1 mm anterior to lambda suture and 2–3 mm from the midline) of postnatal day (P) 7–12 male rat pups as well as from the PFC (0.5–0.7 mm anterior to bregma suture and 0.3 mm from the midline) of P9-10 mice using experimental protocols previously described (Brockmann et al.,
Simultaneous recordings of the LFP and multi-unit activity (MUA) were performed from the PFC using four-shank 32-channel Michigan electrodes (0.5–3 MΩ). The eight recording sites of each shank were separated by either 50 or 100 μm, while shanks were separated by 200 μm. The recording sites of multiple shanks covered the entire depth of the prefrontal sub-division PL (Van Eden and Uylings,
The data were split into two age groups containing the recordings from neonatal (P7-9) and pre-juvenile (P10–12) pups, respectively, (
Neonatal and pre-juvenile rats were deeply anesthetized with 10% ketamine (aniMedica, Senden-Bösensell, Germany)/2% xylazine (WDT, Garbsen, Germany) in NaCl (10 μ l/g body weight, i.p.) and perfused transcardially with 4% paraformaldehyde dissolved in 0.1 M phosphate buffer, pH 7.4. The brains were removed and postfixed in the same solution for 24 h. Subsequently, coronal slices were sectioned in the coronal plane at 100 μm and stored at −80°C.
For the reconstruction of DiI-labeled electrode tracks into the PFC, fluorescent Nissl staining was performed as previously described (Quinn et al.,
Data were imported and analyzed off-line using custom-written tools in Matlab software version 7.7 (Mathworks, Natick, MA). For the analysis of LFPs, the signals were low-pass filtered (<1500 Hz) using a third order Butterworth filter before reducing the sampling rate to 3255 Hz. All filtering procedures were performed in a manner preserving phase information. As previously demonstrated (Brockmann et al.,
Data in the text are presented as the mean ± standard deviation (
The short duration and variable frequency of oscillatory events covering multiple bands (from theta to gamma) required noise reduction in calculating the power spectra averaged across oscillatory events. To this end, we determined the power content in band-pass filtered versions of the recorded signal (Pfurtscheller and Lopes da Silva,
As spectral measure of correlation between two signals coherence was calculated from the cross-spectral density between the two signals and normalized by the individual power spectral density of each (Jerbi et al.,
Coherence coefficients were assessed for all pair-wise combinations of the LFP from one reference recording site (the deepest site on the shank closest to the midline, which was marked by X in Figures
As we previously reported, the patterns of population activity in the PL, as reflected in the LFP, critically depend on age (Brockmann et al.,
In a first step, we aimed to automatically identify the discontinuous oscillatory events using a threshold-based detection. According to our previous data (Brockmann et al.,
Here, we detected discontinuous oscillations in the neonatal PL as deflections of the rms of the band-pass (4–100 Hz) filtered signal (calculated in a sliding window of 200 ms) that exceeded a recording-specific threshold. To determine this threshold, a histogram of the rms was constructed from a 5 min-long time segment of each data set, which started 15 min after the onset of the recording (to exclude any type of boundary effect) and was confirmed by visual inspection to contain oscillatory events (Figure
Visual inspection of these discontinuous events led to the classification into SB and NG (Brockmann et al.,
The complexity and variability of the frequency and amplitude distributions observed across multiple discontinuous patterns of prelimbic activity requires long-lasting expertise for their characterization and classification by visual inspection. Consequently, the procedure of categorizing events as SB or NG might be biased by the researcher performing the analysis. Likewise, the high degree of variability between SB and NG makes them difficult to be described in a quantitative fashion. In the following, we therefore aimed at better characterizing these two types of events that constitute the discontinuous activity in the PL, in order to understand, which characteristics may be suitable for their differentiation.
We quantitatively assessed the characteristics that define the SB and NG extracellularly recorded in the PL of neonatal (P7–9) urethane-anesthetized rats (
Besides these slow oscillatory components, short episodes of fast low-amplitude oscillations superimposed the NG as nested events (Figure
The analysis above revealed that on average the quantitative differences between SB and NG arise (i) from their different durations and amplitudes, (ii) from additional beta/low-gamma cycles with high amplitude present only in NG events, and (iii) the presence of nested fast oscillations (i.e., HFOs) during NG events. In the following, we identified 11 features solely based on single oscillatory events that capture these properties. As reference, the features were visualized in examples shown in Figures
Duration and amplitude were quantified by five features. The event
The spectral composition of SB and NG was also quantified by a total of five features. First, the total raw power in the beta/low gamma band was calculated as the power content between 16 and 40 Hz normalized by the total power between 4 and 50 Hz (indicated as
As a final feature, we quantified the presence of nested HFO by evaluating the cross-frequency coupling by using the
The modulation index is obtained by normalizing the KL divergence to 1:
A uniform phase–amplitude distribution (
In summary, we calculated the following features: (1) duration of events, (2) maximum rms, (3) maximum negative peak, (4) maximum slope, (5) flatness, (6) beta/low gamma power, (7) mean iti, (8) N cycles, (9) N cycles > 10 Hz, (10) N cycles > 16 Hz, (11) modulation index. In Figures
We introduced an unsupervised classification method that distinguished SB and NG based on the combination of features. A feature vector for each oscillatory event was assembled by concatenation of its features. All resulting vectors were processed by principal component analysis (PCA) and the dimensionality of data was reduced to the first
The used fuzzy clustering approach was of practical advantage by providing a direct mechanism to control the number of events that remained UC when varying the threshold. However, the choice of clustering is not essential to the approach, since comparable classification could be obtained using standard unsupervised classifiers such as k-means clustering, however without identifying UC events (see Discussion).
In a first step we investigated which features contribute to the individual PCs. All subsequent analyses were separately performed by pooling data from the upper two or lower two shanks corresponding to the upper and lower PL, respectively, to account for possible layer-specific differences in the characteristics of oscillatory events (Figure
In a second step, we tested how well a partition based on these PCs is able to cluster the SB and NG. To this end, we performed a classification of all six datasets using the unsupervised method on all 11 features. This algorithm was initially based on a cluster analysis in space of the first
To quantify the results of the unsupervised classification, we defined two performance indices: (i) the reliability r = (TPSB+TPNG)/(TPSB+TPNG+FPSB+FPNG), which mirrored the output match of manual and unsupervised classification (ignoring UC events), and (ii) the yield y = (TPSB+TPNG+FPSB+FPNG+FPUC)/(total number of events), which reflected the fraction of events that were automatically assigned. Consistent with the observations reported above, for both upper and lower PL the reliability was high and independent of K. In contrast, the yield dropped as more PCs were used in the analysis, indicating an increasingly conservative classification (Figure
Since the unsupervised categorization was possible based on a linear classification using only PC1, we next assessed whether considering a single feature would lead to similar results. For this, we individually quantified the reliability r and the yield y of discrimination for each of the 11 features using a feature vector of length 1 (Figure
To investigate the reliability and yield as a function of both the number of features and the number of PCs, we performed the classification by varying the number
As described above we firstly developed and validated the unsupervised method for detection and classification of discontinuous patterns of oscillatory activity on prelimbic SB and NG from neonatal (P7–9) rats. We showed that a highly reliable classification of these events is obtained even when using single or few features only.
To decide whether the developed unsupervised algorithm is specie- and pattern-independent, we tested its performance on LFP recordings from the V1 of neonatal rats and the PFC of neonatal mice. In line with previous findings (Hanganu et al.,
At neonatal age the oscillatory events are easily distinguishable and, despite a high degree of inter-event variability, tend to exhibit clear features. In contrast, as the activity starts to switch toward continuous oscillatory rhythms at pre-juvenile age, SB and NG tend to become less obvious in their features and the classification becomes increasingly difficult (Figure
Unbiased classification of oscillatory events represents the prerequisite for characterization of their spatial and temporal organization over neocortical areas. We exemplified this application aiming at deciding whether SB and NG differently synchronize the developing PL. Oscillatory events recorded at multiple sites over the cortical depth were classified using the unsupervised method described above (Figures
Additionally, the unsupervised method for detection and classification of oscillatory events can be applied for unbiased quantification of the SB and NG properties at neonatal and pre-juvenile age (Figures
In this methodological study we have developed and explored the use of a novel unsupervised algorithm for detecting and classifying the patterns of oscillatory activity in the developing brain. This method involved three steps: the identification of oscillatory events using a threshold procedure, the extraction of up to 11 quantitative features from these events, and an unsupervised clustering of the resulting feature vectors corresponding to the events. We demonstrate that (i) the rms-based method reliably and almost independently of signal-to-noise ratio detects the periods with network oscillations; (ii) the discontinuous cortical activity of neonatal and pre-juvenile rodents can be characterized by 11 features that we defined on a per-event basis; (iii) due to the high degree of variability between events of the same type, the heuristically determined features differ in their ability to classify the oscillatory events (best features: maximum rms, maximum negative peak, maximum slope, number of cycles and modulation index); (iv) while for clearly distinguishable oscillatory events single features (e.g., amplitude-based) may be sufficient for a powerful linear classification, the inclusion of a large number of features improves the reliability of the algorithm without loss of the yield for manually difficult to detect and classify pre-juvenile oscillatory events. Finally, we demonstrated how the unsupervised characterization of discontinuous oscillatory patterns opens new perspectives for unbiased analysis of spatial organization in the developing brain.
Reliable detection and classification of SB and NG despite their inter-event variability represent the pre-requisite for assessing the spatial and temporal dynamics of early oscillatory activity in the PL. In this study, we implemented a method to automatically identify the two types of events using a PCA with subsequent unsupervised clustering. The method was shown to classify events at comparable precision as human observers across a number of different recording sessions without the need for session-specific parameter adjustments, e.g., a threshold on individual features. In addition to providing an objective measure to distinguish SB from NG, the PCA allowed to identify features of the events that best highlight their differences. The results revealed two characteristics that mainly differ between SB and NG. First, large maximal values for the rms are indicative of NG. These events tended to show rapid, short and large deflections of the amplitude, in particular during the start of the event. The maximum negative peak and the maximum slope of the rms were further, yet less characteristic amplitude-dependent signatures for NGs. Second, an increase in the number of cycles represented a powerful feature for the detection, as NGs typically exhibited a composition of not only slow (4–12 Hz), but also fast (16–40 Hz) oscillatory periods.
Feature vectors were clustered using a fuzzy clustering algorithm. Instead of assigning each event exclusively to one of the two postulated
For the practical implementation of the method the following two considerations appear as beneficial. First, multiple features should be used to construct the feature vectors. While for some data sets, even single features may already yield a good classification performance, the use of multiple features was shown to boost reliability, in particular for less well distinguishable data. The features identified to perform with highest performance are (starting from the best): maximum rms, maximum negative peak, maximum slope, number of cycles, number of cycles above 16 Hz, and Power LG. For recordings in upper PL, the modulation index was also a good classification measure. Second, the fuzzy clustering should be used as an approach to define UC events. This strategy allowed for a conservative identification of events by ignoring events that are too far apart from cluster centers in the classification space. It corresponds to a strategy in manual classification, where certain events are ignored if their type cannot be determined.
The relevance of an unsupervised and experimenter-unbiased algorithm is highlighted by the literature findings of the last decades. Extensive investigation of brain rhythms either by invasive (e.g., LFP) or non-invasive (e.g., EEG) methods identified several patterns of network oscillations, the properties of which significantly varied among reports. Their nomenclature is not less heterogeneous. In human preterm babies the multi-band discontinuous events including a low intrinsic frequency nested with fast activity have been termed according to their frequency (e.g., delta waves) or as spontaneous activity transients (Kostovic and Judas,
However, the present algorithm is based on a phenomenological procedure in the absence of knowledge regarding the underlying cause for different types of oscillatory episodes. We believe that this approach and the features may constitute a promising starting point for fine-tuning of the classification parameters. Nevertheless, even for the data presented in this study, our approach is not parameter-free. While the method allowed for a robust detection based on features that were easily extracted with a very low number of parameters, the details of the detection still depended on the precise choice of parameters and employed features. A certain fraction of false positives and false negatives compared to the ground truth identity of the events is unavoidable. Therefore, in practice one needs to adjust the sensitivity of the detection with respect to the analysis performed on the detected events. For example, if interested in estimating the absolute occurrence of a certain event type, a high classification yield is likely to be preferable to reduce the number of undetected events. In other cases the goal might be obtaining only a few events that clearly belong to the SB or NG category. For this, a low yield is desirable. A similar problem is faced when performing spike sorting, where data variability still prohibits the use a fully automatized, parameter-free procedure, and likewise suggests a close harmonization of sorting strategy and analysis (Pazienti and Grün,
Additionally, our results revealed that the automatically detected and classified patterns of oscillatory activity may be further used for synchrony analysis over distinct cortical areas at neonatal and pre-juvenile age. The SB with main frequency within theta-alpha band as well as the NG with nested gamma activity on the 4–12 Hz rhythm showed distinct properties in the upper vs. lower layers of the PL. Surprisingly, the coupling by synchrony of the theta–alpha and gamma rhythms of SB and NG was different and changed with age. While the coherence in 4–12 Hz was higher within than between layers, a column-like synchrony entrained the PL in gamma frequency band. In line with the increasing similarity between SB and NG, these coupling differences between low and high frequency activity diminished with age. The column-like synchrony is present in the pre-juvenile PL, although its pattern appears less precise than at neonatal age. This might be due to the reorganization of circuitry and pruning of connections (Changeux and Danchin,
This work was supported by the Emmy Noether-Program of German Research Foundation (Ha4466/3-1 to Ileana L. Hanganu-Opatz), Priority Program 1665 of the DFG (Ha4466/8-1 to Ileana L. Hanganu-Opatz and Michael Denker), Boehringer Ingelheim Fonds travel grant (to Nicole B. Cichon) and German Federal Ministry of Education and Research (01GQ0809 to Ileana L. Hanganu-Opatz).
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank Dr. Marco Brockmann for help with the development of multi-shank recordings from neonatal rats and Achim Dahlmann for excellent technical assistance.
The Supplementary Material for this article can be found online at:
1chronux.org. Available online at: