Distributed and temporally organized patterns of spiking activity in cortical circuits reflect the processing of sensory information, the execution of cognitive tasks and the generation of motor commands. Large-scale recordings have revealed the presence of brief (50–200 ms) multi-neuronal spiking sequences in cortical networks. The brief sequences appear as rapid increases in the multi-unit firing rate and they are generated endogenously or in response to an external stimulus (Luczak et al., 2009Luczak et al., 2007; Luczak et al., 2009; Pasquale et al., 2017; Pastalkova et al., 2008; Riehle et al., 1997). The omnipresence of these activity patterns in the cortical mantle suggests that they constitute the basic elements of the cortical neural code (Harris, 2005; Luczak, 2015).

Multi-neuronal sequential firing patterns have been extensively studied in the hippocampus, in large part because of the comprehensive characterization of a well-defined spatial code (O’Keefe and Nadel, 1978). In rodents, large-scale recordings revealed the occurrence of accelerated ‘replays’ of the spatial information encoded in the population activity during an experience (e.g. traversal of an arm in a maze), both during sleeping and when in a resting state (Davidson et al., 2009; Diba and Buzsáki, 2007; Foster and Wilson, 2006; Karlsson and Frank, 2009; Lee and Wilson, 2002; Wilson and McNaughton, 1994). Structured firing of experience-related activity was also reported in entorhinal (O’Neill et al., 2017; Ólafsdóttir et al., 2016), neocortical (Hoffman and McNaughton, 2002; Ji and Wilson, 2007; Peyrache et al., 2009; Qin et al., 1997; Ribeiro et al., 2004; Rothschild et al., 2017) and subcortical (Girardeau et al., 2017; Gomperts et al., 2015; Lansink et al., 2008; Pennartz et al., 2004; Valdés et al., 2015) circuits.

Correlative analyses of the multi-neuronal spike content of hippocampal replay suggest that distinct replay sequences might be used for sharing information with the neocortex, specifically for system-level memory consolidation (Axmacher et al., 2009; Dupret et al., 2010; Lewis and Durrant, 2011) and for the creation and update of an internal model for covert evaluation of possible outcomes during memory-guided decision-making (Ólafsdóttir et al., 2018; Ólafsdóttir et al., 2017; Penny et al., 2013; Pezzulo et al., 2014; Pezzulo et al., 2017;Pezzulo et al., 2014; Pfeiffer and Foster, 2013;Wu et al., 2017). Closed-loop manipulations of replay-related activity are to date limited to the use of real-time detection of fast sharp-wave ripple (SWR) oscillations (120 – 250 Hz) (Ego-Stengel and Wilson, 2010; Girardeau et al., 2009; Girardeau et al., 2014; Jadhav et al., 2012; Kovács et al., 2016; Maingret et al., 2016;Nokia et al., 2010; Nokia et al., 2012; Novitskaya et al., 2016; Roux et al., 2017; Talakoub et al., 2016; van de Ven et al., 2016) that reflect merely the bursting state of the neuronal population but not the actual replay content. Any causal link between replay content and behavior or physiology thus remains untested.

To provide a direct causal link between a specific transient neural activity pattern and its hypothesized contribution to cognition, correlational and computational studies are not sufficient and a closed-loop manipulation is necessary (Hady, 2016; Potter et al., 2014; Zrenner et al., 2016). A critical element for ‘closing the loop’ around a neural pattern of interest is the ability to detect such pattern in real-time (Grosenick et al., 2015), but a complete methodology for detecting (re)expressed spike patterns within tens of milliseconds has not yet been demonstrated.

The rapid detection of multi-neuronal patterns in a closed-loop setting is challenging because offline identification algorithms generally require all data to be available or are otherwise not suitable for an online scenario. Moreover, the algorithms often have intensive computational demands that pose challenges for real-time execution at tens of millisecond latency. Such requirements are not present in brain–computer interfaces (BCIs) and other neuroprostheses for motor control as they generally operate sufficiently with latencies of 120 – 200 ms (Ibáñez et al., 2014; Velliste et al., 2008; Xu et al., 2014), one order of magnitude greater than is required for the rapid detection of transient cortical multi-neuronal sequences.

To bridge this gap, we devised a system for online spike pattern detection with minimal computational time delays. The system builds on our previous work describing an efficient offline implementation of spike-sorting-less neural decoding (Kloosterman et al., 2014; Sodkomkham et al., 2016) and a general software framework (Falcon) for complex online neural processing in closed-loop settings (Ciliberti and Kloosterman, 2017). Here, we added parallelized neural decoding to Falcon and built on top a concrete implementation of a full real-time spike pattern identification and classification system.

Using this novel closed-loop system, we performed real-time population decoding of spatial information from CA1 assemblies at the stringent 10 ms temporal resolution and detected distinct classes of replay at a low 50 ms average latency. Real-time identification of population bursts with replay had more than 70% average sensitivity and specificity, whereas content classification accuracy was on average over 95%. We demonstrated this technology in a freely behaving rat during both resting and awake states using in vivo large-scale recordings. We also illustrated that detection algorithm parameters can be adjusted, depending on the requirements of the experiment, to trade off sensitivity and specificity or latency and accuracy. We further showed how detection performance is affected by the number of recording channels.

Our ready-to-use brain–computer interface will enable, for the first time, selective manipulation of specific hippocampal and extra-hippocampal ensemble firing sequences at a millisecond timescale and will thus pave the way towards a deeper understanding of their contribution to cognitive processes.

We built a fully functional closed-loop system for real-time detection and identification of replay patterns. Our system addressed three general major technical challenges. The first challenge derives from the desire to respond to an identified replay as soon as possible after the event has started (and well before it is over). Thus, the determination of replay content should be completed with minimal data, which inevitably introduces a trade-off between algorithmic detection latency and detection accuracy. Second, any approach for online detection of hippocampal replay content must be suitable for data arriving as a continuous live stream; in such a scenario, the exact start (and end) times of replay events are unknown and cannot be predicted. The constraints of the online scenario require the implementation to make use of past-observed information only. As a third challenge, the real-time computations needed to identify replay content in the continuous high-volume and high-rate stream should only minimally increase the overall detection latency, which should be dominated by its algorithmic contribution. The need for fast computations introduces another possible latency–accuracy trade-off, in which approximate algorithms may be required to meet real-time deadlines.

Built on top of the Falcon open source framework for closed-loop neuroscience (Ciliberti and Kloosterman, 2017), the system reads and processes a continuous stream of multi-channel neural signals, analyzes the signals to search for potential replay events and emits a digital output pulse upon the detection of a replay event containing target content (Figure 1a, Figure 1—figure supplement 1a). We defined the target as a replay trajectory of one of the tracks in a multi-arm maze (e.g. a Y-maze). To detect replay content, we applied neural population decoding followed by the characterization of the spatial-temporal structure in the decoded locations (Figure 1b), similar to the approach that is often used for offline replay analysis (Davidson et al., 2009; Kloosterman, 2011). The incoming data stream is continuously decoded in 10 ms bins and, on the basis of the most recent 30 ms of decoded spiking activity, an assessment is made of whether or not a population burst with replay of one of the maze arms is present.

Figure 1 with 4 supplements see all

Download asset Open asset

In the online replay classification algorithm, burst and content detections are combined (Figure 1c, Figure 1—figure supplement 1b, see 'Materials and methods'). The burst detection component responds to the crossing of a user-defined threshold (${\theta }_{mua}$) in the causal moving average of multi-unit activity (MUA). The content detection component searches for sharp high-fidelity position estimates that are consistently located in the same maze arm in the last three 10-ms time bins. The sharpness of the estimates is computed as the causal moving average of the locally integrated probability around the maximum-a-posteriori (MAP) estimate and needs to cross a user-defined threshold (${\theta }_{sharp}$). A replay event is identified if the burst, sharpness and consistency conditions are all met simultaneously.

The algorithms for neural decoding and replay content detection were adapted to fit the specific constraints of the online use case. In our approach, the most demanding computation is the online evaluation of the encoding model that directly relates spike waveform features to the animal’s position without the need for prior spike sorting (Kloosterman et al., 2014). Modest compression of the encoding model reduced the computational costs, resulting in an online decode time below 1 ms per spike, without affecting the system's ability to decode the animal’s position during run epochs (Figure 1c). The effect of compression on the integrity of posterior probability distributions in candidate replay population bursts is small for low to moderate compression levels, in particular for posteriors that have a MAP estimate with high probability (Figure 1d).

Overall, the implementation of the algorithms, including the spike-by-spike construction of posterior probability distributions and parallel processing (see 'Materials and methods'), made sure that the determination of replay content and the triggering of a corresponding digital output pulse occurred within 2.2 ms of the time at which the neural data were acquired in over 95% of the cases (Figure 1e; median added latency: 1.1 ms). Stress tests of the system with artificially generated spikes showed that the added latency remains low for high per-tetrode spike rates and for a large number of tetrodes (up to 32) on a 32-core workstation (Figure 1—figure supplement 2a), thus demonstrating the scalability of our solution. Additional tests on a 4-core workstation revealed that added latency has higher variability, especially in the case of 32 tetrodes and high spike rates (Figure 1—figure supplement 2a), but that performance is similar to that of the 32-core workstation for up to 24 tetrodes and per-tetrode spike rates that match our datasets (Figure 1—figure supplement 2b).

Real-time detection of replay content during resting and awake states

To test our closed-loop framework of online detection of replay content, we recorded hippocampal ensemble activity using tetrodes in a rat that repeatedly explored the arms of a 3-arm radial maze. In each of the three separate recording sessions, a first exposure to the maze (RUN1) was used to prepare the encoding model for online neural decoding. Cross-validation showed good decoding performance in RUN1 with a median error below 10 cm (Figure 1—figure supplement 3a), which required only a few traversals of the maze (Figure 1—figure supplement 4). Real-time closed-loop replay content detection was activated in a subsequent rest epoch (REST) and in a second exposure to the radial maze (RUN2). The encoding model built using data from RUN1 was also valid for decoding the animal’s position in RUN2 (Figure 1—figure supplement 3b), indicating that recordings were stable and spatial representations had not remapped (Anderson and Jeffery, 2003; Latuske et al., 2017). To evaluate the online detection performance without distortions of the signal, no actual feedback perturbation of brain activity was applied to the brain.

Reference hippocampal population bursts were defined offline on the basis of the detrended MUA signal and classified as either associated or not associated with replay (Table 1). The offline reference data set contained a total of 3506 population bursts that were recorded in REST (n = 1321) and RUN2 (n = 2185) across all three sessions. The median duration of reference bursts was 83.0 ms (inter-quartile range (iqr) [65.3, 119.8]; REST: median 89.0 ms, iqr [70.0, 126.0]; RUN2: median 79.0 ms, iqr [63.0, 116.0]). In RUN2, the majority of reference bursts occurred when the animal paused at the distal reward platforms.

Table 1

Dataset	Epoch	# bursts	Burst rate [Hz]	# replay bursts	# bursts w/joint content	Mean burst duration [ms]
1	REST	297	0.40	42	15	107.2
1	RUN2	663	0.56	345	85	94.4
2	REST	537	0.43	68b) F	34	101.2
2	RUN2	808	0.64	315	77	93.6
3	REST	487	0.38	134	46	102.5
3	RUN2	714	0.59	334	62	95.6

Online, closed-loop replay identification resulted in a total of 2123 positive hits, a large majority of which (87.3%, 1854/2123) were associated with reference bursts. Very few online detections occurred immediately before and within 20 ms after reference burst onset, or after burst offset (Figure 2a). A small fraction of hits were non-burst false positives (FP_{non_burst}), predominantly in the third test session (Table 2). There was a slightly higher fraction of positive hits inside reference bursts in REST (91.9%, 694/755) as compared to RUN2 (84.8%, 1160/1368). The higher rate of FP_{non_burst} in RUN2 is possibly due to the more prominent spatially modulated firing present during active exploration behavior, which led to spurious crossings of the ${\theta }_{mua}$ threshold with posterior densities representing the actual, rather than a replayed, position of the animal. Indeed, 80.6% of FP_{non_burst} in RUN2 occurred when the rat was moving (speed > 5 cm/s), whereas the majority of online detections inside reference bursts (73.8%) occurred when the rat was immobile (speed < 5 cm/s). Overall, the majority of online detections corresponded to population bursts.

If positive online hits that were associated with reference bursts correctly indicated replay content, then it is expected that, as compared to the ignored bursts, the detected bursts scored higher on metrics used for offline replay identification. Indeed, reference population bursts that were identified online as replay events had higher posterior bias (bias_max and bias_max score) for one of the maze arms than did reference bursts that were ignored by the closed-loop system (Figure 2b). Likewise, online detected events scored higher on metrics for linear spatial-temporal structure (line fit score, Pearson correlation and sequence score; Figure 2c,d). These results are consistent with the closed-loop system preferentially detecting replay-containing reference bursts.

Figure 2

Download asset Open asset

For quantification purposes, and absent ground-truth replay labels, we labeled offline reference bursts with arm-biased posteriors (bias_max score >3) and trajectory structure (line fit score >0.1) as containing replay content for one of the maze arms (Table 1). Across the three sessions, 35.3% of reference bursts were associated with replay (1238/3506; REST — 18.5%, 244/1321; RUN2 — 45.5%, 994/2185), of which 25.8% were classified as putative joint replay events that span more than one maze arm (Wu and Foster, 2014).

In both REST and RUN2, single and joint replay events were correctly detected and classified in real-time (Figure 3a,b). The system also correctly ignored bursts that did not contain consistent representations of any of the three maze arms (Figure 3c). False-positive detections were mainly caused by local matches to the identification criteria in the 30 ms window without significant reference replay present across the whole population burst (see examples in Figure 3—figure supplement 1a). False-negative detections were mainly caused by failure of noisy MUA, sharpness and consistency signals to cross threshold values concurrently (see examples in Figure 3—figure supplement 1b). Occasionally, false-positive/-negative detections were the result of a mismatch between the offline and online decoded position estimates, possibly resulting from compression of the encoding model or differences in spike detection.

Figure 3 with 1 supplement see all

Download asset Open asset

To quantify the performance, we treated the online detection of replay content in reference bursts as a binary classification problem and determined the number of true/false positives/negatives in the corresponding confusion matrix (Figure 4a). Overall, a high fraction of reference bursts with replay content were correctly flagged as such online (median sensitivity 0.78, range: [0.60, 0.95]) and a high fraction of non-replay reference bursts were correctly ignored by the online system (median specificity 0.71, range: [0.44, 0.80]) (Figure 4b, top; Table 3). The online detection favored a low number of false negatives at the expense of a higher number of false positives (median false omission rate (FOR) 0.11, false discovery rate (FDR) 0.54), in particular during REST (FOR 0.02, FDR 0.68) as compared to RUN2 (FOR 0.28, FDR 0.36) (Figure 4b, middle; Table 3).

Figure 4

Download asset Open asset

Table 3

Dataset	Epoch	Sensitivity	Specificity	False omission rate	False discovery rate	Informedness	Markedness	Correlation
1	REST	0.95	0.74	0.01	0.65	0.69	0.34	0.49
1	RUN2	0.60	0.68	0.35	0.36	0.29	0.29	0.29
2	REST	0.86	0.75	0.02	0.72	0.61	0.26	0.40
2	RUN2	0.69	0.80	0.18	0.34	0.49	0.48	0.48
3	REST	0.95	0.44	0.03	0.68	0.40	0.30	0.34
3	RUN2	0.71	0.57	0.28	0.44	0.28	0.28	0.28

To further characterize the online detection of reference bursts with replay content, we computed unbiased performance measures informedness, markedness and Matthews correlation coefficient (Table 3), which are not sensitive to population prevalence and label bias (Powers, 2011). All three measures were well above zero, the performance score of a random classifier (Figure 4b bottom; Table 3), indicating good online identification of population bursts with replay content. For all datasets, the correlation between online hits and offline replay content was highly significant (${\chi }^2$>50, p<$2\times {10}^{-12}$).

For the reference bursts that were correctly flagged by the online system as events with replay, the actual replay content (i.e. which of the three maze arms was represented in the replay event) was classified with high accuracy (Table 2; median content accuracy 0.95, range [0.83, 1.00]; REST 0.95, RUN2 0.95).

Table 2

Dataset	Epoch	Out of burst rate [min−1]	Content accuracy	Median absolute latency [ms]	Median relative latency [%]
1	REST	0.24	0.95	51.77	50.85
1	RUN2	1.83	0.99	53.77	67.54
2	REST	0.33	0.96	49.77	54.10
2	RUN2	1.05	1.00	57.39	63.81
3	REST	2.37	0.83	44.62	50.49
3	RUN2	7.46	0.87	49.95	62.53

Taken together, these results demonstrate the capability of our closed-loop system to identify and classify hippocampal replay content with high accuracy under realistic experimental conditions during both resting and active animal behavior.

Next, we analyzed the latency of online detection, which is a critical metric for experiments that require on-time closed-loop feedback. Across all reference bursts that were detected online, the median closed-loop detection latency (Figure 4a; Table 2) relative to the offline-defined burst onset was 50.7 ms (90% CI [31.4, 116.2]) and the median relative detection latency, as percentage of the reference burst duration, was 59.5% (90% CI: [21.9, 89.0]) (Figure 4c,d; Table 2). The latency for online detection of reference bursts that contain replay content was similar (median detection latency 51.0 ms, 90% CI [30.3, 119.6]; median relative latency 53.6%, 90% CI [19.5, 87.4]). For long reference replay bursts (duration >100 ms), the median detection latency was slightly higher at 58.3 ms (90% CI [30.0, 127.2]), but the median relative latency of 41.9% (90% CI [18.1, 84.7]) was lower (Figure 4c,d). These results demonstrate that online replay detection occurred well before burst offset time, thus making our hardware–software system suitable for experiments requiring replay content-specific closed-loop manipulation.

Effect of parameters on replay content detection

In the experimental tests, the parameters for online replay content identification and classification were chosen on the basis of pilot experiments and offline simulations carried out beforehand. However, such parameters may be tuned to meet the specific requirements of an experiment. For example, it may be desirable to detect the target replay content as completely as possible (i.e. high sensitivity) at the expense of a higher number of false positives (reduced specificity). Conversely, experimenters may want to favor high specificity at the expense of a higher number of missed target replay events.

To characterize the influence of parameters on online detection performance, experimental data were played back offline while applying the online detection algorithm, using a range of values for the MUA and posterior sharpness thresholds ${\theta }_{mua}$ and${\theta }_{sharp}$. As expected, increasing values of ${\theta }_{mua}$ and ${\theta }_{sharp}$ parameters reduced the false-positive and false-discovery rates, at the expense of a decrease in sensitivity and an increase in false-omission rate during both REST (Figure 5—figure supplement 1) and RUN2 (Figure 5—figure supplement 2).

We next asked how close the chosen parameters in the live test were to the pair of optimal parameters that maximize the Matthews correlation coefficient (Figure 5a), which balances all four elements of the confusion matrix (true positives (TP), true negatives (TN), in-burst false positives (FP_burst), and false negatives (FN)). For each optimal pair of parameters, we separately assessed the corresponding rate of non-burst detections (FP_{non_burst}) (Figure 5b), which is not accounted for in the maximization of the Matthews correlation coefficient.

Figure 5 with 3 supplements see all

Download asset Open asset

We found that we used sub-optimal parameters in REST epochs, suggesting that a better detection could have been achieved by increasing ${\theta }_{mua}$ and ${\theta }_{sharp}$ thresholds (Figure 5 a,b Table 4). On the other hand, in the RUN2 epochs, the maximum attainable correlations were only marginally and not significantly higher than those obtained in the live tests (Table 4; Figure 5a, bottom). Moreover, the highest correlation for RUN2 was obtained at the same ${\theta }_{sharp}$ used in the live tests, but with lower ${\theta }_{mua}$ thresholds (Figure 5a, bottom; Table 4), thus allowing for a dramatically lower FP_{non_burst} rate than the parameters that maximize the Matthews correlation coefficient (Figure 5b, bottom; Table 4).

Both content accuracy and detection latency were largely unaffected by changes in ${\theta }_{mua}$ or ${\theta }_{sharp}$ thresholds (Figure 5—figure supplement 1e, f and Figure 5—figure supplement 2e,f). Content accuracy is mainly dependent on the online decoding performance and is influenced by compression threshold (see Figure 1d) and the number of recorded neurons (or proxies thereof, such as the number of tetrodes). Detection latency is mainly dependent on the size of the integration window used to determine replay content, which was fixed to 30 ms (three 10 ms bins) in the live tests. The simulations show that detection latency varied with the size of the integration window (Figure 5—figure supplement 3). As expected, the use of smaller bin size and/or a lower number of N_bins decreases relative detection latency, at the expense of reduced detection accuracy (i.e. lower Matthews correlation coefficient). Thus, experimenters may also tune the integration window size to accommodate faster or more accurate online detections.

Online detection of replay content requires sufficiently high sampling of ensemble activity

Accurate real-time replay detection requires replay content to be predictable from the initial portion of the bursting event. Therefore, higher signal-to-noise and neatly structured decoded replay trajectories are easier to detect online as they are less susceptible to short timescale signal variability. Since the signal-to-noise ratio of replay is strongly influenced by the level of sampling of the spiking neuronal population (although biological factors might play a role too [Roumis and Frank, 2015; Tang et al., 2017]), spiking data recorded from poorly sampled neuronal populations are likely to lead to an increase in online replay detection errors.

To determine how spike data availability affects online replay identification, we progressively excluded tetrodes in one dataset and performed offline simulations of online replay detection in REST and RUN2 (Figure 6 and Figure 6—figure supplement 1). Consistent with a higher signal-to-noise ratio, a higher tetrode number decreased the variability of the MUA rate estimate within 10 ms bins and increased the separability of the MUA rate distributions for non-burst periods, bursts without replay and bursts with replay (Figure 6—figure supplement 2a,c). The sharpness of the posterior probability distributions increased with higher tetrode numbers, whereas the sharpness distributions for non-burst periods and bursts with replay showed a clear separation (Figure 6—figure supplement 2b,d).

Figure 6 with 2 supplements see all

Download asset Open asset

To quantify the replay detection performance, a fixed replay reference was used (computed on the original dataset), so that the results were not affected by degradation of the reference. For each simulation, optimal ${\theta }_{mua}$ and ${\theta }_{sharp}$ thresholds were computed separately (Figure 6—figure supplement 1b, d). As expected, replay identification and classification performance improved when more tetrodes were included, as evidenced by higher correlation and content accuracy measures (Figure 6a). Relative detection latency decreased marginally with the number of tetrodes; whereas FP_{non_burst} rate remained low for REST and was variable for RUN2 (probably because ${\theta }_{mua}$ and ${\theta }_{sharp}$ thresholds were not optimized for low out-of-burst detections).

Matthews correlation coefficient does not appear to saturate at the maximum number of tetrodes in the test dataset (14), suggesting that further improvement is possible by increasing the number of sampled cells. Interestingly, Matthews correlation shows little variation for low number of tetrodes and improves most strongly from 10 tetrodes and up. This non-linear dependency is even more clearly present for content accuracy. It is conceivable that, given heterogeneity in spiking data across tetrodes, there is a critical subset of tetrodes that provide most of the information for accurate determination of replay content. The higher Matthews correlation was mainly due to an increased sensitivity and a reduction in false discovery and omission rates (Figure 6—figure supplement 1a,c).

Next, we asked to what extent run decoding performance, evaluated through cross-validation prior to closed-loop manipulation, is predictive of good performance of subsequent (online) replay content identification (Maboudi et al., 2018; van der Meer et al., 2017). As expected, cross-validated run decoding improved with increasing number of tetrodes, as in-arm decoding error decreased and the fraction of arm-correct position estimates increased (Figure 6b). Run-decoding performance improved rapidly up to four tetrodes and more gradually for higher tetrode counts. As a result, replay detection performance and run decoding performance show a low and highly non-linear correlation (Figure 6c), which suggests that a low run-decoding error (e.g. median error below 10 cm) is a necessary although not sufficient condition for accurate online replay detection.

We demonstrated a solution for real-time detection of hippocampal replay content in freely moving rodents that perform navigation tasks. Our solution provides rapid closed-loop feedback that is specifically triggered by replay of an experimenter-defined content, such as replay of a trajectory through one segment of a maze explored by the animal. This technical advance makes it possible to perform replay content-specific perturbations to address directly the causal link between specific replay events and memory consolidation/retrieval processes.

Our system goes beyond recent efforts to decode or respond to hippocampal activity online. Previous work described systems that provide real-time feedback stimulation in response to spiking activity (Nguyen et al., 2014; Patel et al., 2017). These systems perform online decoding of a rat’s physical location from a limited number of hippocampal place cells (Guger et al., 2011) or implement online spike-sorting with simultaneous detection of ripple oscillations (Sethi and Kemere, 2015). However, none of these prior studies addressed the combination of short timescale neural decoding and low-latency detection of replayed spike sequences. de Lavilléon et al. (2015) showed that coupling spikes of a single hippocampal place cell during rest to electrical stimulation of the medial forebrain bundle induced a subsequent preference of the animal for the cell’s place field. Still, the association with replay was only speculative as no multi-neuronal pattern was being detected. Finally, Deng et al. (2016) developed an algorithm for rapid hippocampal replay content classification in the awake state using state-space models. Despite the low algorithmic latency and the solid mathematical framework of their strategy, the computational latency was still prohibitive and the algorithm's application to live data streams in an experiment was neither demonstrated nor fully conceptualized.

Our replay detection technology has similarities and differences with the large body of research in brain–computer interfaces (BCIs) and neuroprosthetics (Lebedev and Nicolelis, 2006). In common with some prototypes of invasive BCIs used for controlling external devices (Fraser et al., 2009; Li and Li, 2017), we applied online population decoding to unsorted large-scale single-unit recordings. The use of high-level signals that relate to cognitive processes such action planning or memory rather than to pure motor function is not new in neuroprosthetics (see cognitive neural prosthetics (Andersen et al., 2010; Hampson et al., 2018). In contrast to other BCI and neuroprosthetics applications, the development of a ‘replay BCI’ faced two unique challenges. First, while it is possible to test the accuracy of BCIs by measuring the difference between desired and actual output, the content of hippocampal replay does not have a ‘ground truth’ and requires assumptions and statistical analyses to quantify the goodness of the output. The second and most difficult challenge relates to the strict latency specifications in generating an online feedback. BCIs can operate with a processing latency of less than 50 ms when using slowly sampled (below 4800 Hz) EEG signals on few (8–32) channels (Fischer et al., 2014; Wilson et al., 2010). BCIs using rapidly sampled signals for spike detection on larger electrode arrays can reach latencies as low as 120 ms (Velliste et al., 2008); lower latencies are however not necessary as the natural control of motor output has a natural delay of 150 – 200 ms (Velliste et al., 2008; Xu et al., 2014). Here, we surpassed these limits by using more than 50 channels sampled at 32 kHz to generate estimates of brain state (animals’ position) every 10 ms and to classify hippocampal replay patterns with ~50 ms average latency.

As with any system that aims to provide on-time feedback, a trade-off was made between the earliest possible prediction of replay content in a population burst and the most accurate possible prediction. This trade-off is reflected in the numbers of incorrect detections and omissions, when compared to a reference replay content classification. Ultimately, the experimental goals determine how the best trade-off is made. For a balanced trade-off between true and false positives/negatives, we opted for the Matthews correlation coefficient as a measure of performance. For some experiments, however, it may be desirable to be more content-specific and a corresponding increase in false negatives (lower sensitivity) is deemed acceptable. On the other hand, it may be more important to detect as many events with target replay content as possible and an increased rate of false positives (lower specificity) is taken for granted. The two main parameters in our replay content detection algorithm, ${\theta }_{mua}$ and ${\theta }_{sharp}$, provide an intuitive means to tune the detection performance and to obtain the desired trade-off. While the exact parameter values depend on the data and optimal values are not known prior to the experiment, similar datasets collected beforehand (without disruptive closed-loop manipulations) are a guide to the actual parameter selection.

The trade-off between accuracy and latency is affected by the performance of neural decoding. A high cell yield in the recordings will improve decoding performance and hence is expected to lead to better online replay detection. The no-spike-sort decoding approach (Kloosterman et al., 2014), which is at the core of our solution, helps by utilizing the information from all available spikes, whether or not they are part of well-isolated cell clusters. Compression of the encoding model and software parallelization in Falcon provided the necessary computational speed-up, which in our system results in the availability of the animal’s replayed position within 2 ms of data arrival, with acceptable decoding accuracy loss.

A second factor that contributes to the latency-accuracy trade-off is the amount of recent data considered for classifying replay content online. In the content classification algorithm, a shorter window of integration (i.e. fewer bins and/or smaller bin sizes) results in lower detection latencies, and also in an increased focus on local, noisy features of the data that may not be consistent with replay content across the complete event. We settled on a window of 30 ms, which resulted in good replay content detection within short latencies. In future work, further decrease of false-positive and negative detection could be achieved by an extension of our algorithm that adaptively increases the time window to accumulate additional evidence until a certain level of confidence that the event under consideration does or does not represent the target replay content has been reached.

Further improvement of online replay detection is possible by increasing the sampling of spiking data (e.g. by increasing number of recording channels), as the performance in our tests with up to 14 tetrodes did not yet reach a plateau. The increased spiking data will improve both the MUA burst detection and replay content determination components of the online algorithm. This may also allow a further reduction of the integration window, as this is the main determinant of the detection latency. The system as presented will be able to handle the additional computational burden without increasing the computational latency significantly, as it scales to at least 128 recording channels (32 tetrodes) and probably more when run on a 32-core workstation. Scaling to several hundreds or thousands of recording channels in future implementations may require offloading the computationally intensive neural decoding steps to the graphics processing unit (GPU).

The detection of hippocampal replay events has to take into account that ground truth replay labels are unavailable. To evaluate the performance of online replay detection, we showed that detected events score higher on measures that correlate with replay content and that detections largely agree with an offline-defined reference. The quality of the offline reference – that is, how well it captures the desired target replay content – depends on the parameters of the offline detector and on sufficient sampling of spiking data. The offline detector has the advantage of access to all data and of not being bound by limited compute time, whereas the the online detector includes compromises to perform fast and early real-time detection. Hence, one would generally have a higher confidence in the offline reference and would accordingly tune the parameters of the online detector using either prior knowledge or training data. Still, depending on parameters, the offline reference may vary from being too inclusive (e.g. all population bursts are marked as replay) to overly exclusive (e.g. no population bursts are marked as replay) and reliance on an inappropriate reference without careful verification will make experiments that manipulate online detected replay events difficult to interpret. Another cautionary note should be made about using decoding performance during run to predict subsequent online replay detection performance: in our tests, run decoding was less sensitive to data degradation and although low run decoding error may be required for good decoding of replay content, it is not a sufficient condition.

Detection latency of online replay detection was on average ~50 ms. In our previous work, using the same hardware–software system, we demonstrated online burst detection (without replay content identification) at ~40 ms latency (Ciliberti and Kloosterman, 2017). Here, by combining burst and replay content detection in the algorithm, on average we added only 10 ms to the time that is needed for a basic population burst detection in the same closed-loop system.

The rate of false-positive detections outside candidate replay population bursts were generally low in rest, but more prevalent when the rat was actively exploring the maze, probably due to a higher mean population firing rate during motion than during rest. In offline analyses, information about the animal’s behavior (e.g. running speed) or replay-associated electrophysiological markers (e.g. high-frequency ripple oscillation power) are often used to reduce this category of false positives (Davidson et al., 2009). The integration of this additional information in a Falcon processing pipeline would be relative straightforward, but this strategy may increase detection latency and introduce new sources of errors.

The incorporation of our real-time replay detection solution in an experiment requires that an encoding model is built from spiking data recorded during an initial exploration of the maze. At present, the encoding model is constructed manually offline and uploaded into the Falcon software during a brief pause of 5 – 15 min between the end of the initial maze exploration and the start of the real-time replay detection. Future enhancements will enable the construction and validation of the encoding model online rather than offline: besides reducing experimental down-time, this will provide immediate feedback on the decoding performance and will enable content-specific closed-loop manipulation of awake replay events as soon as the quality of the encoding model is sufficient.

In our demonstration, the target replay content for online detection was the representation of one of three maze arms. The system can be trivially extended to environments that are composed of more than three arms (e.g. 8-arm radial mazes or set of linear tracks) and to target joint replay content that covers multiple maze arms in succession. For more strict detection of replay trajectories, a local spatial derivative criterion may be added to the replay classification algorithm. Together with an extension of the encoding model with running direction (Davidson et al., 2009), such a trajectory criterion would form the basis of a detector for forward and reverse replay events (Diba and Buzsáki, 2007; Foster and Wilson, 2006). By altering the encoding model, the system could also be used to target other types of replay content, including reactivation of immobility-associated place cells (Yu et al., 2017), non-spatial task-related replay (Takahashi, 2015), sequence reactivation of time cells (Pastalkova et al., 2008) or spiking sequences in nested theta/gamma oscillations during active behavior (Zheng et al., 2016). Our approach is not limited to hippocampal replay bursts and could be applied to cortical (Ji and Wilson, 2007; Peyrache et al., 2009) or subcortical (Lansink et al., 2009) replay as long as it is possible to define a decodable and detectable content.

In conclusion, we successfully bridged the gap between the basic formulation of online replay identification algorithms and an actual experiment-ready system that addresses the challenges associated with the detection of replay content in live neural data streams. The approach presented here will pave the way for a new generation of closed-loop experiments aimed at elucidating the causal role of hippocampal and non-hippocampal internally generated neural activity in memory processes and cognition.

Datasets are available on the Collaborative Research in Computational Neuroscience website (http://crcns.org). The code for online decoding is integrated into the Falcon software platform and the latest version can be retrieved from https://bitbucket.org/kloostermannerflab/falcon-server (copy archived at https://github.com/elifesciences-publications/falcon-server). The code for offline analysis can be obtained from https://bitbucket.org/kloostermannerflab/ciliberti_elife2018_realtimereplay (copy archived at https://github.com/elifesciences-publications/Ciliberti_Elife2018_RealTimeReplay).

1. 1. Davide Ciliberti
   2. Frédéric Michon
   3. Fabian Kloosterman

   (2018) Extracellular recordings from rat hippocampal area CA1 during rest and exploration of a 3-arm maze along with results of real-time and offline detection of hippocampal replay content

   Publicly available at Collaborative Research in Computational Neuroscience - Data sharing.

   http://dx.doi.org/10.6080/K0PN93T4

1. 1. Ambrose RE
   2. Pfeiffer BE
   3. Foster DJ

   (2016) Reverse replay of hippocampal place cells is uniquely modulated by changing reward

   Neuron 91:1124–1136.

   https://doi.org/10.1016/j.neuron.2016.07.047

   • PubMed
   • Google Scholar
2. 1. Andersen RA
   2. Hwang EJ
   3. Mulliken GH

   (2010) Cognitive neural prosthetics

   Annual Review of Psychology 61:169–190.

   https://doi.org/10.1146/annurev.psych.093008.100503

   • PubMed
   • Google Scholar
3. 1. Anderson MI
   2. Jeffery KJ

   (2003) Heterogeneous modulation of place cell firing by changes in context

   The Journal of Neuroscience 23:8827–8835.

   https://doi.org/10.1523/JNEUROSCI.23-26-08827.2003

   • Google Scholar
4. 1. Axmacher N
   2. Draguhn A
   3. Elger CE
   4. Fell J

   (2009) Memory processes during sleep: beyond the standard consolidation theory

   Cellular and Molecular Life Sciences 66:2285–2297.

   https://doi.org/10.1007/s00018-009-0019-1

   • PubMed
   • Google Scholar
5. 1. Cheng S
   2. Frank LM

   (2008) New experiences enhance coordinated neural activity in the Hippocampus

   Neuron 57:303–313.

   https://doi.org/10.1016/j.neuron.2007.11.035

   • PubMed
   • Google Scholar
6. 1. Ciliberti D
   2. Kloosterman F

   (2017) Falcon: a highly flexible open-source software for closed-loop neuroscience

   Journal of Neural Engineering 14:045004.

   https://doi.org/10.1088/1741-2552/aa7526

   • PubMed
   • Google Scholar
7. 1. Davidson TJ
   2. Kloosterman F
   3. Wilson MA

   (2009) Hippocampal replay of extended experience

   Neuron 63:497–507.

   https://doi.org/10.1016/j.neuron.2009.07.027

   • PubMed
   • Google Scholar
8. 1. de Lavilléon G
   2. Lacroix MM
   3. Rondi-Reig L
   4. Benchenane K

   (2015) Explicit memory creation during sleep demonstrates a causal role of place cells in navigation

   Nature Neuroscience 18:493–495.

   https://doi.org/10.1038/nn.3970

   • PubMed
   • Google Scholar
9. 1. Deng X
   2. Liu DF
   3. Karlsson MP
   4. Frank LM
   5. Eden UT

   (2016) Rapid classification of hippocampal replay content for real-time applications

   Journal of Neurophysiology 116:2221–2235.

   https://doi.org/10.1152/jn.00151.2016

   • PubMed
   • Google Scholar
10. 1. Diba K
    2. Buzsáki G

    (2007) Forward and reverse hippocampal place-cell sequences during ripples

    Nature Neuroscience 10:1241–1242.

    https://doi.org/10.1038/nn1961

    • PubMed
    • Google Scholar
11. 1. Dupret D
    2. O'Neill J
    3. Pleydell-Bouverie B
    4. Csicsvari J

    (2010) The reorganization and reactivation of hippocampal maps predict spatial memory performance

    Nature Neuroscience 13:995–1002.

    https://doi.org/10.1038/nn.2599

    • PubMed
    • Google Scholar
12. 1. Ego-Stengel V
    2. Wilson MA

    (2010) Disruption of ripple-associated hippocampal activity during rest impairs spatial learning in the rat

    Hippocampus 20:1–10.

    https://doi.org/10.1002/hipo.20707

    • PubMed
    • Google Scholar
13. 1. Fischer J
    2. Milekovic T
    3. Schneider G
    4. Mehring C

    (2014) Low-latency multi-threaded processing of neuronal signals for brain-computer interfaces

    Frontiers in Neuroengineering 7:1.

    https://doi.org/10.3389/fneng.2014.00001

    • PubMed
    • Google Scholar
14. 1. Foster DJ
    2. Wilson MA

    (2006) Reverse replay of behavioural sequences in hippocampal place cells during the awake state

    Nature 440:680–683.

    https://doi.org/10.1038/nature04587

    • PubMed
    • Google Scholar
15. 1. Fraser GW
    2. Chase SM
    3. Whitford A
    4. Schwartz AB

    (2009) Control of a brain-computer interface without spike sorting

    Journal of Neural Engineering 6:055004.

    https://doi.org/10.1088/1741-2560/6/5/055004

    • PubMed
    • Google Scholar
16. 1. Girardeau G
    2. Benchenane K
    3. Wiener SI
    4. Buzsáki G
    5. Zugaro MB

    (2009) Selective suppression of hippocampal ripples impairs spatial memory

    Nature Neuroscience 12:1222–1223.

    https://doi.org/10.1038/nn.2384

    • PubMed
    • Google Scholar
17. 1. Girardeau G
    2. Cei A
    3. Zugaro M

    (2014) Learning-induced plasticity regulates hippocampal sharp wave-ripple drive

    Journal of Neuroscience 34:5176–5183.

    https://doi.org/10.1523/JNEUROSCI.4288-13.2014

    • PubMed
    • Google Scholar
18. 1. Girardeau G
    2. Inema I
    3. Buzsáki G

    (2017) Reactivations of emotional memory in the hippocampus-amygdala system during sleep

    Nature Neuroscience 20:1634–1642.

    https://doi.org/10.1038/nn.4637

    • PubMed
    • Google Scholar
19. 1. Gomperts SN
    2. Kloosterman F
    3. Wilson MA

    (2015) VTA neurons coordinate with the hippocampal reactivation of spatial experience

    eLife 4:e05360.

    https://doi.org/10.7554/eLife.05360

    • PubMed
    • Google Scholar
20. 1. Grosenick L
    2. Marshel JH
    3. Deisseroth K

    (2015) Closed-loop and activity-guided optogenetic control

    Neuron 86:106–139.

    https://doi.org/10.1016/j.neuron.2015.03.034

    • PubMed
    • Google Scholar
21. 1. Grosmark AD
    2. Buzsáki G

    (2016) Diversity in neural firing dynamics supports both rigid and learned hippocampal sequences

    Science 351:1440–1443.

    https://doi.org/10.1126/science.aad1935

    • PubMed
    • Google Scholar
22. 1. Guger C
    2. Gener T
    3. Pennartz CM
    4. Brotons-Mas JR
    5. Edlinger G
    6. Bermúdez I Badia S
    7. Verschure P
    8. Schaffelhofer S
    9. Sanchez-Vives MV

    (2011) Real-time position reconstruction with hippocampal place cells

    Frontiers in Neuroscience 5:85.

    https://doi.org/10.3389/fnins.2011.00085

    • PubMed
    • Google Scholar
23. Book

    1. Hady AE

    (2016)

    Closed loop neuroscience

    Elsevier.

    • Google Scholar
24. 1. Hampson RE
    2. Song D
    3. Robinson BS
    4. Fetterhoff D
    5. Dakos AS
    6. Roeder BM
    7. She X
    8. Wicks RT
    9. Witcher MR
    10. Couture DE
    11. Laxton AW
    12. Munger-Clary H
    13. Popli G
    14. Sollman MJ
    15. Whitlow CT
    16. Marmarelis VZ
    17. Berger TW
    18. Deadwyler SA

    (2018) Developing a hippocampal neural prosthetic to facilitate human memory encoding and recall

    Journal of Neural Engineering 15:036014.

    https://doi.org/10.1088/1741-2552/aaaed7

    • Google Scholar
25. 1. Harris KD

    (2005) Neural signatures of cell assembly organization

    Nature Reviews Neuroscience 6:399–407.

    https://doi.org/10.1038/nrn1669

    • PubMed
    • Google Scholar
26. 1. Hoffman KL
    2. McNaughton BL

    (2002) Coordinated reactivation of distributed memory traces in primate neocortex

    Science 297:2070–2073.

    https://doi.org/10.1126/science.1073538

    • PubMed
    • Google Scholar
27. 1. Ibáñez J
    2. Serrano JI
    3. del Castillo MD
    4. Monge-Pereira E
    5. Molina-Rueda F
    6. Alguacil-Diego I
    7. Pons JL

    (2014) Detection of the onset of upper-limb movements based on the combined analysis of changes in the sensorimotor rhythms and slow cortical potentials

    Journal of Neural Engineering 11:056009.

    https://doi.org/10.1088/1741-2560/11/5/056009

    • PubMed
    • Google Scholar
28. 1. Jadhav SP
    2. Kemere C
    3. German PW
    4. Frank LM

    (2012) Awake hippocampal sharp-wave ripples support spatial memory

    Science 336:1454–1458.

    https://doi.org/10.1126/science.1217230

    • PubMed
    • Google Scholar
29. 1. Ji D
    2. Wilson MA

    (2007) Coordinated memory replay in the visual cortex and Hippocampus during sleep

    Nature Neuroscience 10:100–107.

    https://doi.org/10.1038/nn1825

    • PubMed
    • Google Scholar
30. 1. Karlsson MP
    2. Frank LM

    (2009) Awake replay of remote experiences in the Hippocampus

    Nature Neuroscience 12:913–918.

    https://doi.org/10.1038/nn.2344

    • PubMed
    • Google Scholar
31. 1. Kloosterman F
    2. Davidson TJ
    3. Gomperts SN
    4. Layton SP
    5. Hale G
    6. Nguyen DP
    7. Wilson MA

    (2009) Micro-drive array for chronic <em>in vivo</em> recording: drive fabrication

    Journal of Visualized Experiments 26:.

    https://doi.org/10.3791/1094

    • Google Scholar
32. Book

    1. Kloosterman F

    (2011) Analysis of hippocampal memory replay using neural population decoding

    In: Fellin T, Halassa M, editors. Neuronal Network Analysis. Springer. pp. 259–282.

    https://doi.org/10.1007/7657_2011_8

    • Google Scholar
33. 1. Kloosterman F
    2. Layton SP
    3. Chen Z
    4. Wilson MA

    (2014) Bayesian decoding using unsorted spikes in the rat Hippocampus

    Journal of Neurophysiology 111:217–227.

    https://doi.org/10.1152/jn.01046.2012

    • PubMed
    • Google Scholar
34. 1. Kovács KA
    2. O'Neill J
    3. Schoenenberger P
    4. Penttonen M
    5. Ranguel Guerrero DK
    6. Csicsvari J

    (2016) Optogenetically blocking sharp wave ripple events in sleep does not interfere with the formation of stable spatial representation in the CA1 area of the Hippocampus

    PLOS ONE 11:e0164675.

    https://doi.org/10.1371/journal.pone.0164675

    • PubMed
    • Google Scholar
35. 1. Lansink CS
    2. Goltstein PM
    3. Lankelma JV
    4. Joosten RN
    5. McNaughton BL
    6. Pennartz CM

    (2008) Preferential reactivation of motivationally relevant information in the ventral striatum

    Journal of Neuroscience 28:6372–6382.

    https://doi.org/10.1523/JNEUROSCI.1054-08.2008

    • PubMed
    • Google Scholar
36. 1. Lansink CS
    2. Goltstein PM
    3. Lankelma JV
    4. McNaughton BL
    5. Pennartz CM

    (2009) Hippocampus leads ventral striatum in replay of place-reward information

    PLOS Biology 7:e1000173.

    https://doi.org/10.1371/journal.pbio.1000173

    • PubMed
    • Google Scholar
37. 1. Latuske P
    2. Kornienko O
    3. Kohler L
    4. Allen K

    (2017) Hippocampal remapping and its entorhinal origin

    Frontiers in Behavioral Neuroscience 11:253.

    https://doi.org/10.3389/fnbeh.2017.00253

    • PubMed
    • Google Scholar
38. 1. Lebedev MA
    2. Nicolelis MA

    (2006) Brain-machine interfaces: past, present and future

    Trends in Neurosciences 29:536–546.

    https://doi.org/10.1016/j.tins.2006.07.004

    • PubMed
    • Google Scholar
39. 1. Lee AK
    2. Wilson MA

    (2002) Memory of sequential experience in the Hippocampus during slow wave sleep

    Neuron 36:1183–1194.

    https://doi.org/10.1016/S0896-6273(02)01096-6

    • PubMed
    • Google Scholar
40. 1. Lewis PA
    2. Durrant SJ

    (2011) Overlapping memory replay during sleep builds cognitive schemata

    Trends in Cognitive Sciences 15:343–351.

    https://doi.org/10.1016/j.tics.2011.06.004

    • PubMed
    • Google Scholar
41. 1. Li J
    2. Li Z

    (2017) Sums of spike waveform features for motor decoding

    Frontiers in Neuroscience 11:406.

    https://doi.org/10.3389/fnins.2017.00406

    • PubMed
    • Google Scholar
42. 1. Luczak A
    2. Bartho P
    3. Marguet SL
    4. Buzsaki G
    5. Harris KD

    (2007) Sequential structure of neocortical spontaneous activity in vivo

    Proceedings of the National Academy of Sciences 104:347–352.

    https://doi.org/10.1073/pnas.0605643104

    • Google Scholar
43. 1. Luczak A
    2. Barthó P
    3. Harris KD

    (2009) Spontaneous events outline the realm of possible sensory responses in neocortical populations

    Neuron 62:413–425.

    https://doi.org/10.1016/j.neuron.2009.03.014

    • PubMed
    • Google Scholar
44. Book

    1. Luczak A

    (2015)

    Packets of sequential seural activity in sensory cortex

    In: Tatsuno M, editors. Analysis and Modeling of Coordinated Multi-Neuronal Activity. New York: Springer. pp. 163–182.

    • Google Scholar
45. 1. Maboudi K
    2. Ackermann E
    3. de Jong LW
    4. Pfeiffer BE
    5. Foster D
    6. Diba K
    7. Kemere C

    (2018) Uncovering temporal structure in hippocampal output patterns

    eLife 7:e34467.

    https://doi.org/10.7554/eLife.34467

    • PubMed
    • Google Scholar
46. 1. Maingret N
    2. Girardeau G
    3. Todorova R
    4. Goutierre M
    5. Zugaro M

    (2016) Hippocampo-cortical coupling mediates memory consolidation during sleep

    Nature Neuroscience 19:959–964.

    https://doi.org/10.1038/nn.4304

    • PubMed
    • Google Scholar
47. 1. Nguyen TKT
    2. Navratilova Z
    3. Cabral H
    4. Wang L
    5. Gielen G
    6. Battaglia FP
    7. Bartic C

    (2014) Closed-loop optical neural stimulation based on a 32-channel low-noise recording system with online spike sorting

    Journal of Neural Engineering 11:046005.

    https://doi.org/10.1088/1741-2560/11/4/046005

    • Google Scholar
48. 1. Nokia MS
    2. Penttonen M
    3. Wikgren J

    (2010) Hippocampal ripple-contingent training accelerates trace eyeblink conditioning and retards extinction in rabbits

    Journal of Neuroscience 30:11486–11492.

    https://doi.org/10.1523/JNEUROSCI.2165-10.2010

    • PubMed
    • Google Scholar
49. 1. Nokia MS
    2. Mikkonen JE
    3. Penttonen M
    4. Wikgren J

    (2012) Disrupting neural activity related to awake-state sharp wave-ripple complexes prevents hippocampal learning

    Frontiers in Behavioral Neuroscience 6:84.

    https://doi.org/10.3389/fnbeh.2012.00084

    • PubMed
    • Google Scholar
50. 1. Novitskaya Y
    2. Sara SJ
    3. Logothetis NK
    4. Eschenko O

    (2016) Ripple-triggered stimulation of the locus coeruleus during post-learning sleep disrupts ripple/spindle coupling and impairs memory consolidation

    Learning & Memory 23:238–248.

    https://doi.org/10.1101/lm.040923.115

    • PubMed
    • Google Scholar
51. 1. Ólafsdóttir HF
    2. Carpenter F
    3. Barry C

    (2016) Coordinated grid and place cell replay during rest

    Nature Neuroscience 19:792–794.

    https://doi.org/10.1038/nn.4291

    • PubMed
    • Google Scholar
52. 1. Ólafsdóttir HF
    2. Carpenter F
    3. Barry C

    (2017) Task demands predict a dynamic switch in the content of awake hippocampal replay

    Neuron 96:925–935.

    https://doi.org/10.1016/j.neuron.2017.09.035

    • PubMed
    • Google Scholar
53. 1. Ólafsdóttir HF
    2. Bush D
    3. Barry C

    (2018) The role of hippocampal replay in memory and Planning

    Current Biology 28:R37–R50.

    https://doi.org/10.1016/j.cub.2017.10.073

    • PubMed
    • Google Scholar
54. Book

    1. O’Keefe J
    2. Nadel L

    (1978)

    The Hippocampus as a Cognitive Map

    Oxford: Clarendon Press.

    • Google Scholar
55. 1. O’Neill J
    2. Boccara CN
    3. Stella F
    4. Schoenenberger P
    5. Csicsvari J

    (2017) Superficial layers of the medial entorhinal cortex replay independently of the Hippocampus

    Science 355:184–188.

    https://doi.org/10.1126/science.aag2787

    • PubMed
    • Google Scholar
56. 1. Pasquale V
    2. Martinoia S
    3. Chiappalone M

    (2017) Stimulation triggers endogenous activity patterns in cultured cortical networks

    Scientific Reports 7:9080.

    https://doi.org/10.1038/s41598-017-08369-0

    • PubMed
    • Google Scholar
57. 1. Pastalkova E
    2. Itskov V
    3. Amarasingham A
    4. Buzsáki G

    (2008) Internally generated cell assembly sequences in the rat Hippocampus

    Science 321:1322–1327.

    https://doi.org/10.1126/science.1159775

    • PubMed
    • Google Scholar
58. 1. Patel YA
    2. George A
    3. Dorval AD
    4. White JA
    5. Christini DJ
    6. Butera RJ

    (2017) Hard real-time closed-loop electrophysiology with the Real-Time eXperiment interface (RTXI)

    PLOS Computational Biology 13:e1005430.

    https://doi.org/10.1371/journal.pcbi.1005430

    • PubMed
    • Google Scholar
59. 1. Pennartz CM
    2. Lee E
    3. Verheul J
    4. Lipa P
    5. Barnes CA
    6. McNaughton BL

    (2004) The ventral striatum in off-line processing: ensemble reactivation during sleep and modulation by hippocampal ripples

    Journal of Neuroscience 24:6446–6456.

    https://doi.org/10.1523/JNEUROSCI.0575-04.2004

    • PubMed
    • Google Scholar
60. 1. Penny WD
    2. Zeidman P
    3. Burgess N

    (2013) Forward and backward inference in spatial cognition

    PLOS Computational Biology 9:e1003383.

    https://doi.org/10.1371/journal.pcbi.1003383

    • PubMed
    • Google Scholar
61. 1. Peyrache A
    2. Khamassi M
    3. Benchenane K
    4. Wiener SI
    5. Battaglia FP

    (2009) Replay of rule-learning related neural patterns in the prefrontal cortex during sleep

    Nature Neuroscience 12:919–926.

    https://doi.org/10.1038/nn.2337

    • PubMed
    • Google Scholar
62. 1. Pezzulo G
    2. van der Meer MA
    3. Lansink CS
    4. Pennartz CM

    (2014) Internally generated sequences in learning and executing goal-directed behavior

    Trends in Cognitive Sciences 18:647–657.

    https://doi.org/10.1016/j.tics.2014.06.011

    • PubMed
    • Google Scholar
63. 1. Pezzulo G
    2. Kemere C
    3. van der Meer MAA

    (2017) Internally generated hippocampal sequences as a vantage point to probe future-oriented cognition

    Annals of the New York Academy of Sciences 1396:144–165.

    https://doi.org/10.1111/nyas.13329

    • PubMed
    • Google Scholar
64. 1. Pfeiffer BE
    2. Foster DJ

    (2013) Hippocampal place-cell sequences depict future paths to remembered goals

    Nature 497:74–79.

    https://doi.org/10.1038/nature12112

    • PubMed
    • Google Scholar
65. 1. Potter SM
    2. El Hady A
    3. Fetz EE

    (2014) Closed-loop neuroscience and neuroengineering

    Frontiers in Neural Circuits 8:115.

    https://doi.org/10.3389/fncir.2014.00115

    • PubMed
    • Google Scholar
66. 1. Powers D

    (2011)

    Evaluation: from precision, recall and F-Measure to Roc, informedness, markedness & correlation

    Journal of Machine Learning Technologies 2:37–63.

    • Google Scholar
67. 1. Qin YL
    2. McNaughton BL
    3. Skaggs WE
    4. Barnes CA

    (1997) Memory reprocessing in corticocortical and hippocampocortical neuronal ensembles. philosophical transactions of the royal society of London

    Series B, Biological Sciences 352:1525–1533.

    https://doi.org/10.1098/rstb.1997.0139

    • Google Scholar
68. 1. Ribeiro S
    2. Gervasoni D
    3. Soares ES
    4. Zhou Y
    5. Lin SC
    6. Pantoja J
    7. Lavine M
    8. Nicolelis MA

    (2004) Long-lasting novelty-induced neuronal reverberation during slow-wave sleep in multiple forebrain Areas

    PLOS Biology 2:e24.

    https://doi.org/10.1371/journal.pbio.0020024

    • PubMed
    • Google Scholar
69. 1. Riehle A
    2. Grün S
    3. Diesmann M
    4. Aertsen A

    (1997) Spike synchronization and rate modulation differentially involved in motor cortical function

    Science 278:1950–1953.

    https://doi.org/10.1126/science.278.5345.1950

    • PubMed
    • Google Scholar
70. 1. Rothschild G
    2. Eban E
    3. Frank LM

    (2017) A cortical-hippocampal-cortical loop of information processing during memory consolidation

    Nature Neuroscience 20:251–259.

    https://doi.org/10.1038/nn.4457

    • PubMed
    • Google Scholar
71. 1. Roumis DK
    2. Frank LM

    (2015) Hippocampal sharp-wave ripples in waking and sleeping states

    Current Opinion in Neurobiology 35:6–12.

    https://doi.org/10.1016/j.conb.2015.05.001

    • Google Scholar
72. 1. Roux L
    2. Hu B
    3. Eichler R
    4. Stark E
    5. Buzsáki G

    (2017) Sharp wave ripples during learning stabilize the hippocampal spatial map

    Nature Neuroscience 20:845–853.

    https://doi.org/10.1038/nn.4543

    • PubMed
    • Google Scholar
73. Conference

    1. Sethi A
    2. Kemere CT

    (2015) Mulitchannel real time spike sorting for decoding ripple sequences

    7th Annual International IEEE/EMBS Conference on Neural Engineering . pp. 956–959.

    https://doi.org/10.1109/NER.2015.7146784

    • Google Scholar
74. 1. Singer AC
    2. Frank LM

    (2009) Rewarded outcomes enhance reactivation of experience in the Hippocampus

    Neuron 64:910–921.

    https://doi.org/10.1016/j.neuron.2009.11.016

    • PubMed
    • Google Scholar
75. 1. Sodkomkham D
    2. Ciliberti D
    3. Wilson MA
    4. Fukui K-ichi
    5. Moriyama K
    6. Numao M
    7. Kloosterman F

    (2016) Kernel density compression for real-time bayesian encoding/decoding of unsorted hippocampal spikes

    Knowledge-Based Systems 94:1–12.

    https://doi.org/10.1016/j.knosys.2015.09.013

    • Google Scholar
76. 1. Takahashi S

    (2015) Episodic-like memory trace in awake replay of hippocampal place cell activity sequences

    eLife 4:e08105.

    https://doi.org/10.7554/eLife.08105

    • Google Scholar
77. 1. Talakoub O
    2. Gomez Palacio Schjetnan A
    3. Valiante TA
    4. Popovic MR
    5. Hoffman KL

    (2016) Closed-Loop interruption of hippocampal ripples through fornix stimulation in the Non-Human primate

    Brain Stimulation 9:911–918.

    https://doi.org/10.1016/j.brs.2016.07.010

    • PubMed
    • Google Scholar
78. 1. Tang W
    2. Shin JD
    3. Frank LM
    4. Jadhav SP

    (2017) Hippocampal-Prefrontal reactivation during learning is stronger in awake compared with sleep states

    The Journal of Neuroscience 37:11789–11805.

    https://doi.org/10.1523/JNEUROSCI.2291-17.2017

    • PubMed
    • Google Scholar
79. 1. Valdés JL
    2. McNaughton BL
    3. Fellous JM

    (2015) Offline reactivation of experience-dependent neuronal firing patterns in the rat ventral tegmental area

    Journal of Neurophysiology 114:1183–1195.

    https://doi.org/10.1152/jn.00758.2014

    • PubMed
    • Google Scholar
80. 1. van de Ven GM
    2. Trouche S
    3. McNamara CG
    4. Allen K
    5. Dupret D

    (2016) Hippocampal offline reactivation consolidates recently formed cell assembly patterns during sharp Wave-Ripples

    Neuron 92:968–974.

    https://doi.org/10.1016/j.neuron.2016.10.020

    • PubMed
    • Google Scholar
81. 1. van der Meer MAA
    2. Carey AA
    3. Tanaka Y

    (2017) Optimizing for generalization in the decoding of internally generated activity in the Hippocampus

    Hippocampus 27:580–595.

    https://doi.org/10.1002/hipo.22714

    • PubMed
    • Google Scholar
82. 1. Velliste M
    2. Perel S
    3. Spalding MC
    4. Whitford AS
    5. Schwartz AB

    (2008) Cortical control of a prosthetic arm for self-feeding

    Nature 453:1098–1101.

    https://doi.org/10.1038/nature06996

    • PubMed
    • Google Scholar
83. 1. Wilson JA
    2. Mellinger J
    3. Schalk G
    4. Williams J

    (2010) A procedure for measuring latencies in brain-computer interfaces

    IEEE Transactions on Biomedical Engineering 57:1785–1797.

    https://doi.org/10.1109/TBME.2010.2047259

    • PubMed
    • Google Scholar
84. 1. Wilson MA
    2. McNaughton BL

    (1994) Reactivation of hippocampal ensemble memories during sleep

    Science 265:676–679.

    https://doi.org/10.1126/science.8036517

    • PubMed
    • Google Scholar
85. 1. Wu CT
    2. Haggerty D
    3. Kemere C
    4. Ji D

    (2017) Hippocampal awake replay in fear memory retrieval

    Nature Neuroscience 20:571–580.

    https://doi.org/10.1038/nn.4507

    • PubMed
    • Google Scholar
86. 1. Wu X
    2. Foster DJ

    (2014) Hippocampal replay captures the unique topological structure of a novel environment

    Journal of Neuroscience 34:6459–6469.

    https://doi.org/10.1523/JNEUROSCI.3414-13.2014

    • PubMed
    • Google Scholar
87. 1. Xu R
    2. Jiang N
    3. Lin C
    4. Mrachacz-Kersting N
    5. Dremstrup K
    6. Farina D

    (2014) Enhanced low-latency detection of motor intention from eeg for closed-loop brain-computer interface applications

    IEEE Transactions on Bio-Medical Engineering 61:288–296.

    https://doi.org/10.1109/TBME.2013.2294203

    • PubMed
    • Google Scholar
88. 1. Yu JY
    2. Kay K
    3. Liu DF
    4. Grossrubatscher I
    5. Loback A
    6. Sosa M
    7. Chung JE
    8. Karlsson MP
    9. Larkin MC
    10. Frank LM

    (2017) Distinct hippocampal-cortical memory representations for experiences associated with movement versus immobility

    eLife 6:e27621.

    https://doi.org/10.7554/eLife.27621

    • PubMed
    • Google Scholar
89. 1. Zheng C
    2. Bieri KW
    3. Hsiao YT
    4. Colgin LL

    (2016) Spatial sequence coding differs during slow and fast gamma rhythms in the Hippocampus

    Neuron 89:398–408.

    https://doi.org/10.1016/j.neuron.2015.12.005

    • PubMed
    • Google Scholar
90. 1. Zrenner C
    2. Belardinelli P
    3. Müller-Dahlhaus F
    4. Ziemann U

    (2016) Closed-Loop neuroscience and Non-Invasive brain stimulation: a tale of two loops

    Frontiers in Cellular Neuroscience 10:92.

    https://doi.org/10.3389/fncel.2016.00092

    • PubMed
    • Google Scholar

Author details

1. Davide Ciliberti

   1. Neuro-Electronics Research Flanders, Leuven, Belgium
   2. Brain and Cognition, KU Leuven, Leuven, Belgium
   3. VIB, Leuven, Belgium

   Contribution

   Conceptualization, Data curation, Software, Formal analysis, Validation, Visualization, Methodology, Writing—original draft, Writing—review and editing

   For correspondence

   davide.ciliberti@nerf.be

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1229-642X

2. Frédéric Michon

   1. Neuro-Electronics Research Flanders, Leuven, Belgium
   2. Brain and Cognition, KU Leuven, Leuven, Belgium
   3. VIB, Leuven, Belgium

   Contribution

   Investigation, Visualization, Writing—review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1289-2133

3. Fabian Kloosterman

   1. Neuro-Electronics Research Flanders, Leuven, Belgium
   2. Brain and Cognition, KU Leuven, Leuven, Belgium
   3. VIB, Leuven, Belgium
   4. imec, Leuven, Belgium

   Contribution

   Conceptualization, Software, Supervision, Funding acquisition, Visualization, Project administration, Writing—review and editing

   For correspondence

   fabian.kloosterman@nerf.be

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-6680-9660

• 2,683

  views

• 356

  downloads

• 27

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.