Detection of circadian rhythm disturbance of carbachol-induced beta wave dynamics in rat hippocampal slices using neural networks

Rats exhibit a circadian rhythm of locomotor activity. Recently, the relation between circadian rhythm and memory has been elucidated. Brain waves are related to memory processes and are modulated by circadian rhythm. Beta wave-like oscillations within 13–30 Hz can be induced by using carbachol, a cholinergic agonist, in rat hippocampal slices. In this study, we prepared hippocampal slices from rats maintained under light/dark (LD) or dark/dark (DD) conditions. We previously reported that carbachol-induced beta oscillations (CIBO) appeared at all times of the day, and oscillatory parameters, such as duration and inter-burst intervals, demonstrated significant diurnal changes. We also studied whether the time of day can be predicted using CIBO parameters and whether circadian disturbances can be detected with these parameters. Using CIBO parameters, a higher accuracy rate was obtained for time-of-day prediction than that when using the amount of locomotor activity during the day. The time of day was not correctly predicted when the parameters were applied under the DD condition, using the CIBO-parameter-trained neural network under LD conditions. Using the oscillatory parameters of the brain waves induced in the neuronal circuit will estimate the internal subjective time of the day and detect circadian disturbances such as DD condition. (199 words).


Introduction
Organisms have a one-day 24-h cycle or biological rhythm called a "circadian rhythm" that influences physiological processes [1].This rhythm is modulated by light signals, such as illumination [2].The light signal is transmitted to the suprachiasmatic nucleus (SCN) through the retina.The SCN regulates the activity of other brain areas and daily animal activities, including locomotor activity [3].From invertebrates to higher organisms, circadian rhythms regulate learning, memory, and cognitive processes [4].Rats exhibit high performance in hippocampal-dependent working memory tasks during the active dark phase [5,6].Long-term potentiation, which is a synaptic process critical for memory formation, is facilitated in the rat hippocampus during the dark phase [7].
Changes in circadian rhythm affect memory performance.For example, rapid changes in the light/dark (LD) cycle affect fear conditioning memory in mice [8].Constant dark or dark/dark (DD) conditions were shown to modulate memory performance during a hippocampal-dependent Morris water-maze task [9].Circadian rhythm disturbances can be seen in the early stages of Alzheimer's disease (AD) and later affect learning and memory [10].Thus, by detecting circadian disturbances earlier, memory deterioration can be prevented.Circadian rhythm disorders were predicted using actigraphy [11], which measures the sleep/wake cycle.Hence, detecting circadian disturbances before sleep disturbances is challenging.
Hippocampal brain waves are involved in learning and memory processes [12].Circadian rhythms regulate brain waves [13,14]; however, the mechanism underlying this regulation remains unclear.Whether the change in waves can be detected circadian disturbances is yet to be clarified.Hippocampal brain wavelike oscillations such as theta, beta, and gamma-like oscillations can be induced in rat hippocampal slices by applying the cholinergic agonist carbachol (CCh) [15][16][17].The slice model has a simple neuronal network and is thus suitable for investigating the neuronal mechanisms underlying circadian modulation of brain waves.
In this study, we measured CCh-induced beta oscillations (CIBO) at different times of the day under LD/DD conditions.The objective of this study was to investigate the possibility of predicting the time of day by circadian changes in hippocampal brain wave dynamics and whether the oscillatory parameters of brain waves can be used to detect circadian disturbances.We previously reported circadian changes in the CIBO frequency [18].Our paper presented at the SICE2021 conference [19] mainly demonstrated that circadian changes occur in other oscillatory parameters of CIBO and that inhibitory neurones are not involved in these changes.The paper also presented preliminary data for time zone prediction using a neural network (NN) with CIBO parameters.
In this study, we obtained the accuracy rate of prediction obtained by the NN using these parameters.In addition, we compared these results with those obtained using the locomotor activity in rats.Rats are nocturnal, so the circadian rhythm of locomotor activity can be observed.To allow the trained network to detect circadian rhythm disturbances, we first trained the NN using oscillatory parameters under LD conditions.In the test phase, we calculated the inaccuracy rate using the parameters under the DD condition as input and confirmed whether the NN could detect disturbances.

Materials and methods
The induction of CIBO in rat hippocampal slices and analysis of CIBO parameters are described in Ref. [19].CIBO occurred as a burst firing with a stable interval, as shown in Figure 1A, and we calculated the oscillatory parameters from the bursts.Four parameters were measured: duration, inter-burst interval (IBI), frequency, and amplitude.We defined the duration as the time between the start and end of each burst.The IBI is the time interval between the two bursts.The frequency was calculated using fast Fourier transformation (FFT) when the frequency was stable at approximately 1.5 s after burst onset.The amplitude was the peak-to-peak in the same time window as the frequency analysis.Rats in the LD group were placed in a cage with a 12:12 h illumination cycle.In the results, time indicates the Zeitgeber time (ZT).ZT from 0 to 12 was in the light phase and from 12 to 24 was in the dark phase.(Figure 1B).We assumed that ZT continued after the slices were obtained from the rat brains.The CIBO parameters were measured at different ZT values in the LD group.Rats in the DD group were housed in cages under constant dart conditions.The time after the onset of constant dark conditions is indicated as circadian time (CT).The time from CT0 to CT12 was the subjective light phase and that from CT12 to CT24 was in the subjective dark phase.CIBO parameters were measured using different CT scans in the DD group.We measured the parameters from 107 slices prepared from 65 and 42 rats placed under LD and DD conditions, respectively.ZT and CT had six time zones, each with a 4-h period (h bins).These parameters were assigned to six time zones.
Six time zones were predicted by an NN using CIBO parameters.The input vector comprises fourdimensional CIBO parameters, that is, duration, IBI, frequency, and amplitude.The three-layered network comprises 4 units in the input layer, 100 units in the intermediate layer, and 6 units in the output layer (Figure 2A).The number of units in the intermediate layers was set to 100 because the NN had the highest accuracy rate changing the number (Figure 2B).The output units comprised six units that refer to six time zones.In the NN, the activation function per unit was the rectified linear unit (ReLU) function, the loss function was the categorical cross-entropy, and the optimization function was the Adams function.The accuracy rate for the test data differed among the different numbers of epochs during training (Figure 2C).In addition, when we trained the NN for over 10,000 epochs, the accuracy rate was lower than that of the NN trained for 10,000 epochs.Therefore, we set the number of epochs to 10,000.We also trained the NN with the amount of locomotor activity at different times of the day in a cage, before the rats were dissected.Locomotor activity was measured at 10-min intervals using an infrared sensor (Figure 1C) and recorded using Com-pACT AMS software (MUROMACHI Co., Japan).The amount of locomotor activity was summed over an hour and used to train the NN.To make the dimensions of the activity the same as those of the CIBO parameters, we used six 4-h bin locomotor activities as the input vectors.The activity was measured in 96 and 26 rats in the LD and DD groups, respectively.
Three to seven cross-validations were conducted.The validation was repeated 10 times, and the mean accuracy rate per validation was obtained.Data are expressed as the mean ± SEM.
The cycle of CIBO parameters from LD rats was shifted in slices from DD rats, whereas the cycle of locomotion activities was not shifted [19].Thus, CIBO parameters may be more sensitive to circadian disturbances than that for locomotor activity.To clarify whether the NN can detect circadian disturbances under DD conditions, we first trained the network with the CIBO parameters in the slices from the LD rat group, added the parameters in the slices from the DD rat group into the NN as test data, and finally calculated the inaccuracy rate.To evaluate the effectiveness of circadian shift detection under DD conditions, we calculated the inaccuracy rate.The inaccuracy rate was calculated by subtracting the accuracy rate from the unit.This rate was reflected as a circadian shift in the DD group.
To validate the performance of the NN in classifying the CIBO parameters depending on time zones, we attempted to predict the time zones using the same dataset with a conventional classifier, the support vector machine (SVM).A Gaussian kernel function was used for SVM.The SVM function in the scikit-learn package of Python 3.6 was used.

Prediction of the time zone using NN
We trained an NN with six time zones, using four CIBO parameters.The accuracy rate was 95.9% ± 0.8% during training and 78% ± 1.0% during testing (Figure 3A).Using the locomotor activity, the accuracy rate during the testing was significantly lower than when using the CIBO parameters (39.3% ± 0.8%, Wilcoxon rank-sum test; p = 1.1 × 10 −5 ) (Figure 3B).The accuracy rate obtained using the SVM with CIBO parameters was 61.6% ± 1.1% during testing and significantly lower than that obtained using the NN (Wilcoxon rank-sum test; p = 1.1 × 10 −5 ; Figure 3C).Hence, NN classifiers using CIBO parameters provide better prediction of the time zone.
The CIBO amplitude was not modulated by the circadian rhythm in the previous physiological experiment [18]; therefore, using all four parameters may not be optimal for prediction.Thus, the amplitude component in an input vector was replaced with the data from the Gaussian distribution with the same mean and variances as the amplitude data, and the NN was trained in the same manner (Figure 4A).We obtained an accuracy of 51.3% ± 2.9% in the testing, which was significantly lower than that of the original input (Wilcoxon rank-sum test; p = 1.1 × 10 −5 ).We attempted to obtain the accuracy rates when components other than the amplitude in an input vector were replaced with random data.The accuracy rates were decreased in the case where all replaced components were used (randomized duration: 59.7% ± 2.1%, randomized IBI: 51.1% ± 2.7%, randomized frequency: 52.1% ± 1.9%), and they were significantly lower than that using the original input vector (Wilcoxon ranksum test; all parameters: p = 1.1 × 10 −5 , Figure 4B).Thus, all four CIBO parameter components in the input vector contributed equally to the time-zone prediction.

Detection of circadian rhythm disturbance by NN
The NN was trained in the time zone using all CIBO parameters obtained under LD conditions, as described in Section 3.1.We then provided the CIBO parameters obtained under DD conditions to the trained NN as test data and calculated the inaccuracy rate.The inaccuracy rate was 91.3% ± 0.6%.The rate was significantly higher than that under the LD conditions (22.3% ± 1.0%, Wilcoxon rank-sum test; p = 1.1 × 10 −5 ) (Figure 5A).The CIBO parameters under the DD condition were incorrectly classified as the +12-h time zone from the correct one (Figure 5B).In fact, the peak time of the circadian change in CIBO frequency seemed to be shifted to the +12-h time zone in physiological experiments [19].Thus, the results are consistent.
We also calculated the inaccuracy rate using locomotor activity in the same manner as the CIBO parameters.The NN was trained in the time zone of locomotor activity under LD conditions.We then applied activity under DD conditions to the trained NN.The inaccuracy rate using the locomotor activity was 80.0% ± 0.6%, which was significantly higher than that using the activity under the LD conditions (60.6% ± 1.6%, Wilcoxon rank-sum test; p = 1.1 × 10 −5 ) (Figure 5C).Locomotor activities under the DD condition were mostly classified in the +4-h time zone from the correct one (Figure 5D).Locomotor activity also predicts circadian disturbance.The inaccuracy rate of CIBO parameters was significantly higher than that of locomotor activity (Figures 5A,C).Thus, the CIBO parameters can detect the shift in the circadian rhythm, that is, circadian disturbance, under DD conditions more accurately than locomotor activity using NN.

Detection of time zone using hippocampal brain waves
We previously reported that CIBO parameters other than amplitude have circadian changes [19].In this study, all four CIBO parameters, that is, the duration, IBI, frequency, and amplitude, were input into a threelayered NN, and the NN was trained to predict the time zone when CIBO was recorded.The trained NN achieved a prediction with an accuracy rate of approximately 80%.The accuracy rate was significantly higher than that of the NN trained using locomotor activity.These results suggest that the time of day can be predicted using the oscillatory parameters of the hippocampal brain waves.Among the CIBO parameters, only the amplitude did not show a significant circadian change in physiological experiments [19].When the amplitude component was replaced with random data in the input vectors, the accuracy rate decreased significantly as well as those of the other parameters (Figure 4).Thus, all CIBO parameters in the input vectors will be needed equally to predict the time zone.Even if all parameters were used, the accuracy rate was 78% ± 1.0% during testing; hence, the accuracy rate should be increased in future studies.

Detection of circadian disturbance using hippocampal brain waves
We investigated whether hippocampal brain waves are useful in detecting circadian disturbances.In a slice experiment, the DD condition shifted the peak time of the circadian change in CIBO parameters [18].Using all oscillatory parameters of the CIBO, the change can be detected at a rate of approximately 90% (Figure 5A).The accuracy was significantly higher than that of the NN trained using the amount of locomotor activity (Figure 5C).The time zone predicted by the CIBO parameters under the DD condition was mostly classified as 12 h after the correct one (Figure 5B), whereas that predicted by the amount of locomotor activity was only classified 4 h later than the correct one (Figure 5D).Hippocampal EEG may change at intervals shorter than the amount of locomotor activity.The change in hippocampal EEG may therefore be able to detect disturbances in the circadian rhythm rather than changes in locomotor activity.CIBO can be induced in rat hippocampal slices [20], which have simple neuronal circuits.The CIBO is the field potential induced by a population of neurones.The EEG is also induced by a population of neurones.When recording hippocampal brain waves at the head surface, brain waves are possibly deformed.Humans have a much more complex neuronal circuit than rodents do.Moreover, our results were obtained from nocturnal rats; however, humans are not nocturnal.Human brain waves also change with the time of day [12,13].Using brain wave features, one can estimate the time of day of human circadian rhythms and detect circadian disturbances.The early detection of circadian rhythm disturbances is important for preventing memory deterioration in patients with AD.We herein suggest that oscillatory features of hippocampal brain waves may be useful in detecting circadian disturbances.We believe that by measuring human brain waves in different time zones, the time of day can be predicted by an NN with brain wave features.An NN can be trained using the oscillatory features of brain waves in a healthy individual.When a person experiences a circadian disturbance, the trained NN detects the disturbances with high accuracy.Furthermore, if hippocampal brain waves are indicative of disturbances in the circadian rhythm of a person, the application of melatonin can be a useful treatment [21] before a disturbance in the individual's memory occurs.

Figure 1 .
Figure 1.Experiment setup of locomotor activity and CIBO recording.(A) A sample waveform of CIBO.Scales are 5 s and 1 mV.(B) LD and DD illumination conditions.ZT and CT are defined as described in the figure.This figure was taken from [19].(C) Illumination chamber for rats.

Figure 2 .
Figure 2. Experiment setup of NN. (A) Three-layered NN architecture for the prediction of time zones using CIBO parameters.This figure was taken from [19].(B) Box plots depicting the accuracy rates of NN dependent on the unit number in the intermediate layer.(C) Box plots depicting accuracy rates of NN dependent on the epoch numbers in training.

Figure 3 .
Figure 3.Time zone predictions using an NN.(A) Accuracy rate of NN during training (blue line) and testing (orange line) at each epoch.(B) Box plots depicting accuracy rates of NN using CIBO parameters and locomotor activity as inputs.(C) Box plots depicting accuracy rates of NN and SVM.Wilcoxon rank sum test; * * * p < 0.001.

Figure 4 .
Figure 4. Time zone NN predictors using four CIBO parameters including the replaced amplitude with the randomized one.(A) Accuracy rate of the NN during training (blue line) and testing (orange line) at each epoch.(B) Box plots depicting the accuracy rate of NN using the original input vectors at the leftmost and using the input vectors including one of the components replaced with the randomized one.Wilcoxon rank sum test; p < 0.001 ( * * * ).

Figure 5 .
Figure 5. Detection of circadian disturbance by NN trained by CIBO parameters or locomotor activity under LD conditions.(A) Box plots depicting the inaccuracy rates of the NN trained by CIBO parameters.CIBO parameters under LD condition (left) or under DD condition (right) were given to the NN as inputs.A higher inaccuracy rate was obtained.(B) The classification rates of each time zone by the NN when CIBO parameters under DD conditions were given to the NN as inputs.Box plots depicting the inaccuracy rates of NN trained by locomotor activities under LD conditions.(C) The activities under LD condition (left) or under DD condition (right) were given to the NN as inputs.A higher inaccuracy rate was obtained.(D) The classification rates of each time zone by the NN when the locomotor activities under DD conditions were given to the NN, which was trained by the activities under LD conditions.Wilcoxon rank sum test; significance level of * * * p < 0.001.