Spindle-shaped waves of oscillations emerge in EEG scalp recordings during human and rodent non-REM sleep.  The association of these 10-16 Hz oscillations with events during prior wakefulness suggests a role in memory consolidation.  Human and rodent depth electrodes in the brain record strong spindles throughout the cortex and hippocampus, with possible origins in the thalamus.  However, the source and targets of the spindle oscillations from the hippocampus are unclear. Here, we employed an in vitro reconstruction of four subregions of the hippocampal formation with separate microfluidic tunnels for single axon communication between subregions assembled on top of a microelectrode array.  We recorded spontaneous 400-1000-ms-long spindle waves at 10-16 Hz in single axons passing between subregions, as well as from individual neurons in those subregions.  The highest amplitudes and most frequent occurrences suggest origins in the feedback axons from CA3 to DG. Spindle dissociation from spiking activity and recording in single axons from isolated hippocampal neurons suggests that spindle mechanisms are independent of action potentials and that consolidation of declarative-cognitive memories in the hippocampus may be separate from the more easily accessible consolidation of memories related to thalamic motor function.

Data was collected from the experiments of Vakilna (2021) using four PDMS compartments of primary neurons connected by microfluidic tunnels over a 120-electrode array (MEA) (Figure 1A).  Briefly, recordings were made using a Multichannel Systems MEA120 1100× amplifier (Multichannel Systems, Reutlingen, Germany). This study focused only on five minutes of spontaneous activity recordings, which were collected with the MC_Rack software at a sampling rate of 25 kHz at 37°C for 5 min, in humidified 5% CO₂, 9% O₂, as in culture. The recording begins after three weeks of culture in NbActiv4 media (Thermo-Fisher, Carlsbad, CA). Arrays with poor growth or less than 80% active electrodes were rejected (Vakilna 2021).  The four subregions contained neurons micro-dissected from the subregions of the hippocampal formation of postnatal day 4 Sprague-Dawley rats, namely the entorhinal cortex (EC), dentate gyrus (DG), CA3, and CA1+subiculum in order.  Fifty-one microfluidic tunnels of 3 x 10 x 400 µm allowed axons to connect to adjacent compartments.  Five of these tunnels per subregion were aligned over two 30 µm diameter electrodes to measure the direction of action potential propagation.  Recordings from nine arrays are reported.

Spike and Spindle detection

All the computations were performed using custom scripts written in MATLAB® 2022a. Spikes were detected using the method discussed in Vakilna et al. (2021) and the updated automation improvements (Lassers et al. 2023). A fourth-order elliptic bandpass filter from 300 Hz to 3000 Hz was applied to 5 min of raw data sampled at 25 kHz, which was used to cluster spikes in Wave_Clus (Chaure et al., 2018). In order to adequately capture large spikes in tunnels, two thresholds were used for the detection of the negative peak: 5 to 50 SD and 50.1 to 500 SD. Previous analyses of these tunnel devices showed ~63% of tunnels contained only one axon (Narula et al. 2017). Therefore, spikes could be combined into one cluster, and axons could reliably be found after normalized matching indexing (NMI) between the two tunnel electrodes to find conduction time delays, and by extension, the direction of transmission. Peaks in the NMI histogram distribution of conduction times can then be used to identify axons.

Spindle detection followed the filter methods of Silversmith 2020, Kim et al, and Sela et al. The time series data recorded from the tunnels was downsampled from 25 kHz to 1 kHz after applying a Kaiser window finite impulse response (FIR) anti-aliasing filter (50^th order) in Matlab. Spindles were detected by applying an 8-16 Hz bandpass filter (6th-order high-pass and 8th-order low-pass Butterworth infinite impulse response filter). Afterward, the envelope of the signal was extracted by computing the magnitude of the analytic signal from the Hilbert transform. A time window was classified as a spindle if the upper envelope exceeded a threshold of 2.5 standard deviations above the mean once and the lower envelope crossed a threshold of 1.5 standard deviations below the mean for more than 300 ms. Multiple detections of the same spindle were eliminated by combining the spindles that were less than 500 ms apart (Silversmith et al., 2020; Schimicek et al., 1994). The area under the curve, spindle duration, and inter-spindle interval were calculated for each spindle. The amplitude of each spindle was then calculated by dividing the area by the length. Phase locking of spikes with each spindle cycle was estimated by computing the polar histogram of phases and the phase angles of the spindle wave when the spike occurred. Time-frequency decomposition of the spindle was generated by computing a continuous wavelet transform using an analytic Morse wavelet ( = 3, = 60)  (Lilly, J. M., and S. C. Olhede, 2012).

Statistics

Each spindle-length quantity was fitted with a log-normal distribution, and the parameters were estimated by fittinga  Gaussian distribution to the log-transformed histogram. Means and standard deviations were computed for each histogram. Recordings with only one spindle oscillation in 5 min or spindles less than 0.4 sin length were not included. To measure oscillatory relationships of waveforms from two recording sites, the non-spindle regions of the recording were set to zero. Then, Pearson correlations (r) were calculated using the xcorr function of MATLAB, normalized, a nd unscaled for the entire five-minute recording. The significance ofthe difference in means was estimated by computing one-way analysis of variance (ANOVA), followed by Tukey’s HSD test.