Application of expert system and LSTM in extracting index of synaptic plasticity

Abstract

The indexes of synaptic plasticity, including long-term potentiation (LTP) and long-term depression (LTD), can usually be measured by evaluating the slope and/or magnitude of field excitatory postsynaptic potentials (fEPSPs). So far, the process depends on manually labeling the linear portion of fEPSPs one by one, which is not only a subjective procedure but also a time-consuming job. In the present study, a novel approach has been developed in order to objectively and effectively evaluate the index of synaptic plasticity. Firstly, we introduced an expert system applying symbolic rules to discard the contaminated waveform in an interpretable way, and further generate supervisory signals for subsequent seq 2seq model based on neural networks. For the propose of enhancing the system generalization ability to deal with the contaminated data of fEPSPs, we employed long short-term memory (LSTM) networks. Finally, the comparison was performed between the automatically labeling system and manually labeling system. These results show that the expert system achieves an accuracy of 96.22% on Type-I labels, and the LSTM supervised by the expert system obtains an accuracy of 96.73% on Type-II labels. Compared to the manually labeling system, the hybrids system is able to measure the index of synaptic plasticity more objectively and efficiently. The new system can reach the level of the human expert ability, and accurately produce the index of synaptic plasticity in a fast way.

Introduction

Synaptic plasticity has been broadly understood to offer the cellular basis for most models of learning, memory and development in neural circuits (Abbott and Nelson 2000; Liu et al. 2018; Wu et al. 2020), while spike-timing dependent plasticity was considered as one of the important factor to modulate LTP and LTD (Kim and Lim 2019, 2020). Based on the mechanisms of activity-dependent alterations of excitatory synaptic transmission, Synaptic plasticity points to an essential role in cognition with several classical functional indexes including long-term potentiation (LTP), long-term depression (LTD) and so on (Huang et al. 2012; Shang et al. 2019; Zhang et al. 2017, 2011). Field excitatory post-synaptic potentials (fEPSPs) are excitatory post-synaptic potential events. By labeling the linear region of fEPSPs and measuring its slope and/or amplitude, we can obtain the index of LTP or LTD accordingly.

It is well known that the slopes/amplitudes of fEPSPs were potentially affected by various factors including electrodes, recording sites and operating practices (Di Maio et al. 2018). Therefore, identifying and removing contaminated fEPSPs, which have an abnormal overall trend (Fig. 2), becomes important for properly measuring LTP or LTD. Usually, there are two steps to extract synaptic plasticity indexes from fEPSPs, which are (1) determining and discarding the fEPSPs contaminated by multi-factor and (2) measuring the linear portion which was used to extract synaptic strength indicators for those uncontaminated fEPSPs. In addition, there are several linear portions in a fEPSP wave and all of them are distributed on the rising and falling limbs. As a result, a part of researchers (Tamura et al. 2011) tend to extract the slope from the rising limb (‘a’ and ‘c’ in Fig. 1b), while the others (Kramar et al. 2004) from the falling one (‘b’ in Fig. 1b). Conventionally, neuroscientists are involved in labeling the linear region of fEPSPs by dragging two triangular cursors on a specific software, such as the pClamp™, to sandwich the region, and a linear portion ‘a’ can be seen in Fig. 1b. Due to the linear region of fEPSPs may appear at different times even in same experimental animal (Tamura et al. 2011), individuals are forced to labeling thousands of fEPSP waveform, separately. Therefore, this process is often a time-consuming, subjective and inefficient repetitive work.

Fig. 2.

The block diagram of an automatic labeling system. For training the system, fEPSP data are entered into the expert system to extract key facts and generate two types of labels, Type-I Label and Type-II Label. The former signifies if fEPSPs are contaminated and the latter represents uncontaminated fEPSPs which are entered into the neural network LSTM. The ES is only responsible for judging if a waveform is contaminated. The measurement of synaptic plasticity index is performed according to the linear portion marked by the LSTM

Fig. 1.

Experiment protocol and exemplar trace of fEPSP. a The experiment protocol of fEPSP recordings. b Representative trace of fEPSP waveform in the hippocampus DG region. The leftmost deep narrow valley of the waveform is caused by stimulus artifact. The colored lines mark the linear portion of the waveform. Neuroscientists extract the slope of specific area marked by red line by dragging the triangular cursor to sandwich it. c The main part of the waveform, used for training a sequential labeling model, is composed by 2 part, one includes artifact and another includes a M-shaped trace. The vertices are marked by colored ‘*’, and all linear portion is marked by sequence label representing as different color bar in there

There are a number of ways to automaticly labeling data in artificial intelligence research, including expert system (Pfefferkorn et al. 1985; Zhang and Maeda 2000) and supervised learning (Kotsiantis 2007). Among them, a supervised learning approach needs a large number of labeled samples and a supervised learning model can tag the unlabeled sample. Since there are few labeled samples, the increase of labeled sample numbers may lead to excessive time costs. Therefore, it is hard to train a supervised learning model in the study. It is well known that expert systems do not depend on labeled sample rather than rules that can be formulated by expert experience (Broner and Comstock 1997). An expert system is a computer system that imitates the decision-making ability of human experts (Peter 1998). Recently, expert systems were widely used in the field of biomedicine, such as adaptive decomposition of EEG spectrum (Herrmann et al. 2001) and breast cancer detection (Karabatak and Ince 2009). Compared with artificial neural network, expert systems tend to have poor generalization capability (Kroger 1989). Therefore, how to combine an expert system with neural networks is very important in artificial intelligence research (Broner and Comstock 1997; Kroger 1989; Šíma 1995). As we known that traditional methods include the use of neural networks to extract key facts from raw data that support consequently expert system (Becraft et al. 1991). In this case, the neural network acts as a “feature extractor” for the expert system. A previous study (Karabatak and Ince 2009) applied an expert system to extract features from the raw data and then trained the subsequent neural network for classification. In the study, the expert system served as a data cleaner and provided supervisory signals for subsequent supervised learning methods.

Neural networks exhibit greater generalization capabilities than that of expert systems. Usually, neural network training is performed on a high volume of data. In recent years, LSTM networks have been widely used in time series classification and natural language processing (Vargas et al. 2014; Venkitaramani et al. 2007). It was proposed by (Hochreiter and Schmidhuber 1997) as a type of RNN to solve the problem of gradient disappearance of vanilla-RNN. LSTM can capture long-term dependencies existing in time-series, therefore, it is widely used in time series classification and prediction (Gammulle et al. 2017; Salinas et al. 2017). The task of sequence labeling is to output a categorical label to each member of the sequence, which inspired us labeling fEPSPs using LSTM neural network. Since the automatic labeling of fEPSPs can be converted into a sequence-to-sequence (seq 2seq) labeling, it can be well solved by LSTM.

The first part gives a brief introduction of the study background and identifies the issues that need to be addressed. The second part introduces the approaches that are used in the study including the description of the expert system and LSTM, and offers the parameter details of the model. The third part shows the experimental results. Finally, the relationship between the present study and other investigations is carefully discussed.

Methods

Figure 2 shows the block diagram of the automatic labeling system. For training the model, the fEPSP data are entered into the expert system to extract the key fact and generate supervisory signals for LSTM. Firstly, the key vertices are extracted to support the rule-based inference and then the decision, which fEPSP should be discarded, has been made by the expert system (Type-I Label). Secondly, supervisory signals are generated by the expert system for the saved fEPSP (Type-II Label). Finally, a neural network with supervised learning is used to improve the generalization ability of the model. The measurement of LTP/LTD is performed according to the linear portion marked by the LSTM.

Database

The animal was anesthetized with 30% urethane and positioned on a stereotaxic frame. According to the mouse brain atlas, we gently implant a stimulation electrode in PP and another recording electrode in DG. Before giving theta-burst stimulation (TBS), the baseline of PP-DG pathway was determined by collecting 30 fEPSP waveforms in 30 min. After TBS, the long-term potential (LTP) was measured by recording 60 fEPSP waveforms in 60 min. Similarly, 60 fEPSP waveforms about depotentiation (DEP) was obtained after giving low-frequency stimuli (LFS). Experiment protocol of fEPSP recording was shown as Fig. 1a. In the experiment, the sampling rate was set to 20 kHz, and each fEPSP waveform had 1280 points (64 ms). All efforts have been made to minimize the number of animals used and their suffering. Experiments in the present study were carried out according to protocols approved by the Ethical Commission at Nankai University and abiding by the practices outlined in the NIH Guide for the Care and Use of Laboratory Animals. The experimental details can be obtained from our previous reports (Fu et al. 2017; Shang et al. 2017, 2016; Xiang et al. 2019)

There are 3 fEPSPs databases which have been used in this study. The first database consistes of data obtained from both the CON group and the DISC1 group (DISC1 knockdown), in which there are 5 mice in each group. The second database consistes of data obtained from both the WT group and the 5xFAD group, in which there are 5 mice in each group. The third database consistes of data obtained from both the Control group and the CUMS group, in which there are 5 mice in each group. Based on the first database, we developed an expert system for providing supervised signals of LSTM. Moreover, the second and the third databases were used for testing the performance of the system. The first database was established with collecting 1500 fEPSP waveforms from the hippocampus DG region and there were 150 fEPSP waveforms for each mouse including the three periods before and after TBS. The second database was established with collecting 900 fEPSP waveforms from the hippocampus DG region and there were 90 fEPSP waveforms for each mouse including the baseline and LTP periods. The third database is as same as the second one.

Expert system

In this part, we showed the process about extracting the vertex of fEPSP waveforms by the expert system (ES), and several key experience thresholds for determining those vertices. The core components of ES including the knowledge base and the inference engine were shown in Fig. 2.

(1) Facts extraction

Usually, there is an M-shaped trace in the fEPSP waveform obtained from the hippocampus DG region. The trace consists of two positive peaks and one negative peak, as shown in Fig. 1b. In order to describe conveniently, we named each peak as shown in Fig. 1c. The position and amplitude of these vertices are substantial facts for judging if the waveform is uncontaminated and also important for index extraction. We identified the location of the stimulus artifact and peak ‘c’ (Fig. 3a), and then determined the position of the remaining point by constantly searching the local maximum and minimum to the right. Since the time at which the stimulus artifact occurs is relatively stable, it is not hard to determine the position of the peak ‘d’ by searching for the maximum value within the time range specified (index between 120 and 180 in the waveform). After locating the peak ‘d’, we are able to perform search from ‘d’ to ‘e’ in the falling curve. There is usually a small peak and valley in the descending stage between ‘d’ and ‘e’. Then the threshold for searching is determined, which is not less than 10 points in order to prevent the expert system from locating the wrong valley. As soon as determining the negative peak ‘e’, the performance of search continues in the rising curve from ‘e’ to ‘a’ which is the sub-maximum point (Fig. 1c). This portion of fEPSPs is usually full of noise, so it is empirically set that the step length for searching is not able to be lower than 15. Considering that the position of negative peak ‘c’ is the minimum value within the range of 150 points from the peak ‘a’ to the right-side of the curve, the valley ‘c’ can be confirmed by determining the minimum value of the region. Finally, the position of the peak ‘b’ can be obtained by probing the maximum value in the second rising limb.

Fig. 3.

Fact extraction and rules construction. a A workflow for fact extraction. Capital letters represent peaks. Artifact and peak ‘c’ are extracted independently. pS is a peak search method and starts to search the vertex form left side of arrow until it finds a local maximum/minimum value. In normal, it should end at the point to the right of the arrow. b Determining a proper step parameter for peak search method. The left-side illustrates that searching from peak ‘e’ to peak ‘a’ needs to assign a step size to overcome some small disturbance. This step should be equal to the frequency of such disturbance. The right side represents that the disturbance mainly concentrates in the band that is larger than 4 kHz. c Simplifying process for rules of relationship between peak amplitudes. Red cross represents impossible rules based on the current fact detection methods. Blue cross represents unreasonable rules because they cannot be treated as a criterion for judging if a fEPSP is contaminated. The rule to the right of the arrow can be replaced by the rule to the left of the arrow

Step size is an important parameter in peak searching approach. To better understand why we need to set this parameter and how to set the parameter, an example is given (Fig. 3b-left). The performance, in which we are able to find the point ‘A’ from the point ‘E’ by searching for the maximum value to the right, is unreliable in noisy environments. A small fluctuation between point ‘A’ and ‘E’ possibly fall into a local maximum point. In order to sort it out, we can set a step size for the search instead of comparing the point one by one. Obviously, the appropriate value of step size should be equal to the number of points included in one cycle of the minimum noise waveform. Figure 3b-right shows that the minimum noise frequency is 4 kHz. Since our sample rate is 20 kHz, the step size has been set to 5.

(2) Rules and Inference

After obtaining the position of vertices, the rules were coded for determining the facts through the vertices in an IF–THEN style. There are thirteen symbol rules in the WHILE loop for searching evidences that support the inference. We divided these 13 rules into 4 categories and introduced them as follows.

1. Peak adjustment (3 items)

   Peak adjustment rule checks if the extracted facts are logical. This include a position check of peak ‘a’, which should be the maximum between the valley ‘e’ and the valley ‘c’. If not, the position of the peak ‘a’ was changed to the maximum value.

   The similar process was performed on examining the peak ‘e’ location and the peak ‘c’ location. All ruses was shown as below:

   • Rule 1	IF $\text{max}\left(x\left(E_{\mathit{idx}}:C_{\mathit{idx}}\right)\right)>x\left(A_{\mathit{idx}}\right)$ THEN $x\left(A_{\mathit{idx}}\right)=arg\text{max}\left(x\left(E_{\mathit{idx}}:C_{\mathit{idx}}\right)\right)$ WHILE CONTINUE
   • Rule 2	IF $\text{min}\left(x\left(D_{\mathit{idx}}:A_{\mathit{idx}}\right)\right)<x\left(E_{\mathit{idx}}\right)$ THEN $x\left(E_{\mathit{idx}}\right)=arg\text{min}(x\left(D_{\mathit{idx}}:A_{\mathit{idx}}\right)$. WHILE CONTINUE
   • Rule 3	IF $\text{max}\left(x\left(A_{\mathit{idx}}:B_{\mathit{idx}}\right)\right)<x\left(C_{\mathit{idx}}\right)$ THEN $x\left(C_{\mathit{idx}}\right)=arg\text{max}\left(x\left(A_{\mathit{idx}}:B_{\mathit{idx}}\right)\right)$ WHILE CONTINUE

2. Relationship between peak amplitudes (5 items)

   Originally, there are the total 25 rules by which the relationship between Peaks cold be established (Fig. 3c). ‘Amp(A) < Amp(D)’ shows that the waveform is a contaminated waveform if the amplitude of peak ‘a’ is less than peak ‘d’. The rules were constructed according to the waveform in Fig. 1b. However, a part of rules need to be removed from the total 25 rules. For example, since peak ‘e’ is searched by idendifying the local minimum value in the right of peak ‘d’ (Fig. 1c), ‘Amp(E) > Amp(D)’ is a rule that is impossible to happen. We marked the impossible rules by red cross and removed it from the 25 rules. Then, we merged the rules that are logically equivalent. Furthermore, because some rules cannot be used as a basis for judging if fEPSP is contaminated, they are defined as unreasonable rules that are removed from the total rules as well. These rules are marked with a blue cross. Finally, the rest of rules were accordingly merged and described below.

   • Rule 4	IF $x\left(A_{\mathit{idx}}\right)<x\left(D_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 5	IF $x\left(B_{\mathit{idx}}\right)<x\left(E_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 6	IF $x\left(A_{\mathit{idx}}\right)<x\left(C_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 7	IF $x\left(B_{\mathit{idx}}\right)<x\left(A_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 8	IF $x\left(A_{\mathit{idx}}\right)<x\left(At{f}_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK

3. Slope/amplitude of limb (3 items)

   The Slope or amplitude of limb rule are described below.

   • Rule 9	IF $A_{\mathit{idx}}-E_{\mathit{idx}}<10$ AND $abs\left(x\left(A_{\mathit{idx}}\right)-x\left(E_{\mathit{idx}}\right)\right)<0.08$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 10	IF $abs\left(x\left(C_{\mathit{idx}}\right)-x\left(B_{\mathit{idx}}\right)\right)<0.2$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 11	IF $C_{\mathit{idx}}-A_{\mathit{idx}}<25$ AND $abs\left(x\left(C_{\mathit{idx}}\right)-x\left(A_{\mathit{idx}}\right)\right)<0.1$ THEN Is_Contaminated = True WHILE BREAK

4. Other rules (2 items)

   Other rules including the overlap of the peak ‘d’ and peak ‘a’, and the collapse of the whole fEPSP waveforms are as following.

   • Rule 12	IF $\text{min}\left(x\left(D_{\mathit{idx}}:500\right)\right)==x\left(E_{\mathit{idx}}\right)$ AND $x\left(B_{\mathit{idx}}\right)-x\left(C_{\mathit{idx}}\right)<$ $x\left(A_{\mathit{idx}}\right)-x\left(E_{\mathit{idx}}\right)$ THEN Is_Contaminated = True WHILE BREAK
   • Rule 13	IF $x\left(B_{\mathit{idx}}\right)<mean\left(x\left(500:\right)\right)$ THEN Is_Contaminated = True WHILE BREAK

Long short-term memory network

The long short-term memory (LSTM) network is a special recurrent neural network (RNN) proposed by Hochreiter and Schmidhuber in 1997 (Hochreiter and Schmidhuber 1997). It was designed to overcome the gradient disappearance (explosion) problem in RNN. Compared with RNN, LSTM adds both a memory unit and three gating units inside. By writing valuable features into the memory and strictly controlling the data flow through the gating unit, LSTM protects important features from been squeezed by data appearing in following time steps (Chung et al. 2014).

Figure 4a, b shows a schematic diagram of LSTM and RNN. The j-th output neuron at time t is formulated as below:

$$h_t^j=o_t^{j}tanh\left(c_t^j\right)$$ 1

o^j_t represent the value of j-th output gate at time t. c^j_t represent the j-th memory unit at time t. They are computed by:

$$o_t^j=\sigma {\left(W_o{x}_t+U_o{h}_{t-1}+V_o{c}_t\right)}^j$$ 2

$$c_t^j=f_t^j{c}_{t-1}^j+i_t^j{\tilde{c}}_t^j$$ 3

where x_t is external input and f^j_t represent the value of j-th forgetting gate at time t; i^j_t is the value of j-th input gate; ${\overline{c}}_t^j$ is the value of j-th new memory. They are computed by:

$${\tilde{c}}_t^j=tanh{\left(W_c{x}_t+U_c{h}_{t-1}\right)}^j$$ 4

$$f_t^j=\sigma {\left(W_f{x}_t+U_f{h}_{t-1}+V_f{c}_{t-1}\right)}^j$$ 5

$$i_t^j=\sigma {\left(W_i{x}_t+U_i{h}_{t-1}+V_i{c}_{t-1}\right)}^j$$ 6

Fig. 4.

The structure of LSTM and traditional RNN. a The difference between LSTM cell and RNN cell. The memory unit and gate-control $i,f,o$ make LSTM overcome the gradient vanishing problem. b Six layers structure of LSTM network in the study

Weight matrix W, U, V in formula (2)(4)(5) and (6) can be solved separately by either back propagation through time (BPTT) algorithm or stochastic gradient descent with momentum (SGDM) algorithm.

Algorithm design of hybrid system

In the preceding section, the following two tasks will be performed by the means of the expert system including (1) determining if the waveform is proper to be used; (2) generating supervisory signals for LSTM. We denote to the first task output, generated by the expert system, as Type-I label and symbolize the second task output as Type-II label as shown in Fig. 2. LSTM neural networks are trained to perform classification in a supervised manner, and therefore require the input of supervisory signals. Accordingly, the output of expert system is entered as the supervisory signals of the LSTM.

To accelerating the convergence speed and exclude the irrelevant feature, we only input main part of fEPSPs series into the network as shown in Fig. 1c. Since the stimulus artifact of a normal fEPSP begins after t = 100, the right boundary of M-shaped trace can reach before t = 500 (Fig. 1c). Therefore, we only take fEPSP curves from t = 100 to t = 500 to training network, in which each 400-points curves is segmented into 10-points samples with 50% overlapping. This is mainly because if there is no 50% overlap, the rising and falling branches of a vertex may be divided into different segments. Consequently, the network loses the ability to characterize the segment where the vertex is located. As mentioned before, we let the output of expert system as a label for the LSTM network. Finally, LSTM can be implemented in MATLAB. The structure and initialization parameters of the neural network are configured as shown in the Fig. 4c and Table 1.

Table 1.

Parameter of LSTM neural networks for type-II classification

	$\mathit{LSTMfortagging}$
Parameter	$LSTMmodel:sequenceoutput$ $algorithm:SGDM$ $MaxEpochs:30$ $InitialLearnRate:0.01$ $inputSize:10$ $numHiddenUnits:100$ $droprate:0.2$ $numClass:6$

The parameters of LSTM are shown in Table 1, and there are totally 6 layers in the networks (Fig. 4c). In the model, a various number of hidden neurons was examined to choose proper ones, which make LSTM obtaining the highest classification performance. The data were illustrated in Table 2. Since the minimum noise frequency was 4 kHz, we set input size as 10 sampling spots. It makes sure that the input segment can represent the total trend rather than tiny fluctuation. The rest of the parameter including droprate and learning rate has been determained under the suggestion of MATLAB documentation. Learning rate was set to 0.01 to prevent the final set of weights from arriving on a sub-optimal. And Learning algorithm was set as SGDM.

Table 2.

Relationship between LSTM hidden neuron number and classification accuracy

Hidden neurons	60 (%)	80 (%)	100 (%)	150 (%)
1st -fold	82.53	92.66	94.96	80.14
2nd -fold	83.56	90.58	96.22	84.59
3th -fold	82.88	90.94	95.80	84.25
Mean accuracy	82.99	91.39	95.66	82.99

Results

Label generation based on the ES

The data obtained from the expert system are shown as below. For the uncontaminated fEPSP waveforms in the hippocampus PP and DG regions (Fig. 5a, c), two types of labels were identified. It is noted that there are different stimulation artifacts in Fig. 5a, c respectively. The stimulation artifact in Fig. 5a is a negative peak and it is a positive peak in Fig. 5c. Figure 5b showed that the expert system can’t find the important peak spots in the contaminated waveforms due to the unregularity of them. In this part, the expert system marked 487 waveforms as contaminated ones and 1003 waveforms as uncontaminated waveforms. In order to determine contaminated/uncontaminated waveforms, the output accuracy of the expert system was idendified through comparing the labels between the expert system and the manual system (Table 2). These results show that the expert system achieves an accuracy of 96.22% on Type-I samples. Furthermore, it was found that the value of false negative ratio is bigger than that of false positive ratio (Table 3). In order to ensure the accuracy of the label which were entered into the neural network LSTM, we increased the false negative rate of the expert system output (Table 3).

Fig. 5.

Results of expert system output a A typical exemplar shows uncontaminated fEPSPs and their vertex labeled by the expert system. b Another typical exemplar represents contaminated fEPSPs and their vertex labeled by the expert system. 6 type of vertices were mixed with each other. The blue cycles ‘°’ represented the artifact after stimuli. Peak D & B were marked by green ‘*’ and valley E & C were marked by blue ‘*’, finally, Peak A was labeled by red ‘*’. c A typical exemplar illustrates uncontaminated fEPSPs with positive artifacts. d The performance in the LSTM-based fEPSP labeling system. e The Normalized fEPSP amplitudes in both the WT group and the 5xFAD group. ‘*’ indicates a significant difference between these two groups (p < 0.05). Each small pot represents a mouse. (f) The Normalized fEPSP slopes for both the CON group and the CUMS group. ‘*’ indicates a significant difference between these two groups (p < 0.05). Each small pot represents a mouse

Table 3.

Confusion matrix

	Predict value
	True	False	Summary
Real value
Positive	210	11	221
Negative	1	76	77
Summary	211	87	298

Linear portion labeling on the LSTM

Classification accuracy of linear portion labeling

The samples labeled as ‘True’ in type-I classification were kept performing seq 2seq LSTM network for sequence tagging. The network parameters are described in the Table 1. The accuracy of three-fold cross validation was shown in Table 2. Figure 5d give an exemplar uncontaminated fEPSP to illustrate the performance of the labeling system. It can be seen that the linear portion across mice can be marked.

Effect during actual use

In order to verify the actual use of the system, we examined it using two different databases. A previous study showed that the value of LTP in the CUMS group was lower than that in the CON mice (Shang et al. 2017; Yu et al. 2016). Moreover, it is well known that the value of LTP in the 5xFAD group is lower than that in the WT group (Vargas et al. 2014; Venkitaramani et al. 2007). Therefore, the difference of LTP between either the CUMS group and the CON group or the 5xFAD group and the WT group, which was obtained from the hybrid system, was in accordance with that obtained from manually labeling system (Fig. 5e, f).

Figure 5e shows the fEPSP amplitude of either WT mice or 5xFAD mice, obtained from the both systems. Similor to the data measured by manually labeling, there is a significant difference of the fEPSP amplitude between the WT group and the 5xFAD group (Fig. 5e). It is worth noting that the hybrid system can reach a smaller intra-group variance than that of the manually labeling system (SEM = 3.23 and 3.28 for WT and 5xFAD animals in the hybrid system; SEM = 3.77 and 5.53 for WT and 5xFAD mice in the manually labeling system). In addition, Fig. 6f shows the fEPSP slope of either in the CON group or in the CUMS group, obtained from the both systems. Similor to the data measured by manually labeling, there is also a statistical difference of the fEPSP slope between the CON group and the CUMS group (Fig. 5f). Similarly, the hybrid system can reach a smaller intra-group variance than that of the manually labeling system (SEM = 6.31 and 4.18 in the CON group and the CUMS group in the hybrid system, SEM = 7.85 and 5.53 in the CON group and the CUMS group in the manually labeling system; Fig. 5f).

Fig. 6.

Comparation of expert system and neural network hybrid-system. a The structure of a hybrid system, in which the feature of fEPSP waveform is extracted by an expert system and then a supervised learning system is employed. b The structure of a hybrid system, in which the high-level feature is extracted by an unsupervised learning system and the category label is generated by expert system. c The structure of a hybrid system, in which the supervisory signal is generated by the neural network and then a supervised learning system is employed

Discussion and Conclusion

Objective and rapid extraction of the slope and amplitude of fEPSP is of great significance for measuring neuronal synaptic plasticity. The study presents a novel system consisting of two parts including the expert system and the neural network. Firstly, we found that the expert system performed efficiently in extracting the shape of uncontaminated fEPSPs compared with doing by hand in an interpretable manner but shortage with generalization. Moreover, the neural network LSTM, supervised by ES, increased the generalization capabilities of the whole system and boosted its performance. The flexible combination of both the expert system and the neural network exists in a variety of applications.

Data coupling with hybrid system

A previous study (Karabatak and Ince 2009) showed that the hybrid system, composed of AR + NN, achieved satisfactory result for breast cancer discrimination (Fig. 6a). Firstly, it applied an expert system in extracting important features from signals. These features are then used for training the perceptron network. In the above study, the expert system essentially acts as a feature filter and is therefore designed to solve the problem of dimensional curses. A previous study (Becraft et al. 1991) reported that an operator assistance system was generated for large-scale chemical process plants (Fig. 6b). At first, a neural network was used as a first-stage filter to extract fault features in the system. As a result, some fault features were then entered into the expert system for troubleshooting. Clearly, it could be a practical system, in which a neural network was initially employed for extracting robust features and then an expert system was used for judging faults based on the characteristics. From the point of view of the software engineering, there is a coupling between the expert system and the neural network (Broner and Comstock 1997; Karabatak and Ince 2009). In other words, the output of one module will act as the input of the another module.

Control coupling with hybrid system

In the practice of software engineering, another type of coupling between different modules is a control coupling. Patricia et al. proposed a hybrid model, in which the expert system was designed to select superparameters for a univariate time series prediction model based on neural network (Melin et al. 2012). It allows no data interaction between the expert system and neural network modules. The processing for the univariate time series is done by neural network alone, while the expert system is only responsible for selecting the hyperparameter of neural network for achieving better performance. A similar study showed that the creation, structure and training process of the neural network were completed under the guidance of the expert system (Nikolov and Bogdanov 2010).

In the present study, we initially employed an expert system for more objectively extracting the both slope and amplitude of fEPSPs, in which the waveforms of the fEPSPs were properly labeled. Firstly, we designed a knowledge-based system to labeling the different stages of fEPSP. These markers are also used as supervisory signals for neural networks. Since both the neural network and the expert system shared the same dataset, there was not only a control coupling but also a common coupling between the expert system and the neural network. The schematic diagram of our study is shown as Fig. 6c. It could be seen that the expert system was initially used to label a large number of fEPSPs in order to better supervise the neural network LSTM. After that, LSTM is used as a supervised learning approach. Since the neural network is based on a non-threshold method, it could significantly diminish the disruption of noise on the system, thereby improve the overall capability of the system. In the study, the main reason for the use of the expert system to generate supervisory signal is that the neural network is hard to acquire supervisory signals at the beginning. We understand that it is the most significant difference between our study and previous studies.

In summary, our results showed that the contaminated fEPSPs could be effectively distinguished from the all of fEPSPs by the expert system. Furthermore, the shortage of the expert system, caused by generalization ability, could be compensated by training the neural network LSTM supervised by the expert system.

Authors’ contributions

TZ, JZ: Conceived and designed the experiment, SZ, YCS, XX: Performed the experiments, SZ, ZY: Analyzed the data, SZ, TZ: Wrote the manuscript.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author (TZ) upon reasonable request.

Compliance with ethical standard

Conflict of interest

The authors declare that they have no conflict of interests.