ANet: Autoencoder-Based Local Field Potential Feature Extractor for Evaluating An Antidepressant Effect in Mice after Administering Kratom Leaf Extracts

We collected KT leave from the natural resources in Surat Thani province, Thailand. Supplementary materials explain more details about the preparation of KT extract and the quantification of KT major alkaloids, mitragynine (MT), that accumulated in each form of KT extracts.

The committee approved all operations involving animals employed in the scientific study from the Prince of Songkla University’s Institute of Animal Care and Use, which followed the criteria of the International Committee on Laboratory Animal Science (ICLAS) [project license number: MHESI 6800.11/845 and reference number: 57/2019]. We recruited Male Swiss albino ICR mice from the Nomura Siam International Company, Bangkok, Thailand. They were appropriately at rest for one week before the experiment began to minimize the stress. Moreover, mice stayed in separated stainless steel cages (17 x 28.5 x 17 cm) under the standard condition room (12/12 h light/dark cycle, 22°C, and 55.1$\%$ relative humidity). Water and commercial food pellets were accessed freely. We conducted the studies between eight a.m. and four p.m.

Male ICR mice (four months of age) had been through an intraperitoneal injection of a mixed solution of 16 mg/kg xylazine hydrochloride (Xylavet, Sigma-Aldrich International GmbH, Switzerland) and 50 mg/kg zoletil (Tiletamine – zolazepam, Vibac Ah, Inc., USA) to be deeply anesthetized. The head of the animals was then fixed with stereotaxic apparatus before the scalp on the dorsal head in the middle line was exposed. According to the mouse brain atlas [13], electrodes were stereotaxically implanted on the left hemisphere of the brain, from the bregma to the medial prefrontal cortex (mPFC) (AP; +2.5 mm, ML; 0.2 mm; DV; 1.5 mm), hippocampus (HP) (AP; -2.5 mm, ML; 2.0 mm; DV; 1.5 mm) and the nucleus accumbens (NAc) (AP; +1.3 mm, ML; 1.0 mm; DV; 4.2 mm). Over the cerebellum (AP; -6.0 mm, ML; 0.0 mm; DV; 1.5 mm), a ground electrode was inserted. Dental acrylic was applied to hold and secure all placed electrodes. The antibiotic ampicillin (100 mg/kg) (General Drug House Co., Ltd., Bangkok, Thailand) and carprofen (10 mg/kg) (Best Equipment Center Co., Ltd., Thailand) were given intramuscularly once a day for three days [14] to prevent infection and relieve pain. It took at least two weeks to fully recuperate from surgery before they could begin the experiment.

As soon as animals recovered from surgery, they went to the experimental period adapted from the previous investigations [15, 5], as shown in supplementary materials. In brief, after animals habituated for three consecutive days, they went to the testing phase. During this phase, we placed mice individually in the recording chamber (25 cm x 18 cm x 25 cm) to familiarize the experimental conditions for one hour. After that, the LFP signals were harvested in the recording session for 30 minutes after the injection of the treated substances (number of mice = 7 per group) for one hour. The concentrations of KT extract in each sample existed in Table I according to the high-performance liquid chromatography analysis. KT extract calculated doses are from the whole amount of sample containing MT, fixed at 10 mg/kg, a dose that effectively cured the animal model of depression [16]. The AD flu at 10 mg/kg was selected as the standard drug since it showed positive results in previous findings [3].

Details of instrumental setup for LFP signals recording are with the previous works [14]. In summary, the Dual Bio Amp (AD Instruments, Castle Hill. NSW, Australia) and the PowerLab 16/35 system (AD Instruments, Castle Hill. NSW, Australia) with 16-bit A/D at a sampling frequency for two kHz are amplification and digitization for LFP signals, respectively. Also, we filtered the power line noise artifact at 50 Hz. LabChart 7.3.7 Pro software was a tool for recording LFP signals.

This experiment’s purpose was to compare LFP responses induced by different substances traditionally. The substances used are control, flu, KT syrup, KT alkaloids, or KT aqueous. First, we downsampled raw LFP to 200 Hz using the MNE python package (version 0.23.4) [17]. Then we computed the spectral power of the prepared LFP for all channels by applying the Morlet wavelets transform and divided into six frequency bands: Delta, Theta, Alpha, Beta, Gamma I, and Gamma II according to the previous work [14]. Finally, we conducted the statistical analysis using a non-parametric Kruskal-Wallis test with Bonferroni multiple comparisons to test group differences across the frequency bands at each brain region.

Comparing statistical results to the deep learning approaches, the recorded LFP signals at each brain region were further split into four seconds/segment, followed by spectral power of the frequency band extraction as the input to ANet and the multi-task AE in the following two experiments as shown in Figure 2 and Figure 3.

This experiment aimed to evaluate the similarity of LFP features responded to each KT extract relative to the reference drug (AD flu). We first designed the structure of the auto-encoder or ANet as the following two components.

This study extended ANet by including the supervised learning component to the AE, which turned out to be the multi-task AE. We aimed to perform the feasibility of extracting discriminative features from multiple LFP responses simultaneously.

ANet and Supervised Learning: We added the supervised learning component by having the latent vector layer of ANet as the input, which made the proposed architecture form the multi-task AE, as shown in Figure 1. At the end of the supervised task, the softmax function classified the latent vector via layer activation in the FC layer. These allowed our model to learn the reconstruction along with the classification.

During the experiment, we explored the possibility of the multi-task AE or the model recognizing the multi-class LFP features for classifying LFP responses. We believed that this experiment could benefit in developing computational tools for future studies of the substance that affected mice’s brain activities. The experimental protocol followed Figure 3. Firstly, we split the inputs for training, validating, and testing sets. Then we optimized the model according to the following Subsections III-D and III-E. Eventually, the following bullets and equation explain quantitative and qualitative assessments of the model capability in LFP feature extraction.

• •

  $t$-SNE projection is for the qualitative assessment of the visualize reduced dimensional data [18]. In this experiment, we defined the latent vector as the input of t-SNE projection algorithm to present the probability distribution of unlearned and learned features, representing the LFP datasets before and after feature extraction by the proposed model, respectively.

• •

  Maximum Mean Discrepancy (MMD) distance is for the quantitative assessment of the distance between the means of the embedding features of two probability distributions in t-SNE projection. They evaluated whether data in set X and Y generated a probability distribution equally. Here, the ideal concept is that if the MMD distance of LFP features between substance X and Y is low. Then, we might imply that those two LFP inputs seem to originate from a very close origination. The standard calculation of MMD is the statistical measure of the samples from the different sets, expressed in Equation (1).

Where $\varphi$ represents the map of features of $X$ or $Y\rightarrow\kappa$, $\kappa$ denotes a reproducing kernel Hilbert space. $P_{X}$ and $P_{Y}$ represent the probability distribution of LFP in sets $X$ and $Y$, respectively. In addition, the multi-task AE performance also exhibited averaged percentage accuracy, F1-score, and confusion matrix.

In both Experiment II and III, we implemented our model using Keras API (TensorFlow version 2.2.0 as backend) under NVIDIA Tesla v100 GPU setup with 32G memory. With the optimization by Adam optimizer, the learning rate was between 10^-4 and 10^-5. If there were no improvement in validation loss for five consecutive epochs, the learning rate gradually alleviated with a decay rate of 0.5. The batch size was 128 samples. We applied the function of early stopping to stop the training iteration when the validation loss had not reduced for 20 continuous epochs.

According to the described experiments earlier, we optimized mean square error (MSE) loss alone for Experiment II or the study of ANet. In contrast, we optimized both MSE and cross-entropy (CE) losses for Experiment III, which studied the multi-task AE. MSE and CE losses correspond to AE and supervised learning components, respectively, as described below:

• •

  Mean square error (MSE) loss was adopted to monitor and minimize errors between the input signal and reconstructed data. It was expressed in Equation (2).

$C$ represents a number of channels while $x$ and $\hat{x}$ are input and reconstructed data of the channel.

• •

  Cross-entropy (CE) loss was used in supervised learning to assess the error between the actual label and classification probability, as shown in Equation (3).

Where $y_{j}$ and $\hat{y}_{j}$ denote true and predicted labels of data in class, and $c$ is the number of classes.

The summation of the losses, as mentioned earlier, finally evaluated the total loss: $L_{MSE}$ and $L_{CE}$ demonstrated in Equation (2) and Equation (3), respectively, as shown below.

$N$ is the total number of input signals and the weight of loss assigned for 1.0.

Experiment II validated ANet for anomaly detection with the subject-independent scheme. The stratified 5-folds cross-validation was adopted to separate data for testing, and training sets, both outer and inner loops, using Scikit-Learn (version 1.1.1). Only the spectral power of LFP that responded to the AD flu was assigned as training and validation sets, while unseen data from KT extracts were used as a testing set, as illustrated in Figure 2.

The multiple LFP responses classification required a larger scale of samples in the network training, so Experiment III validated the multi-task AE with the subject-dependent scheme. The spectral powers of multi-LFP classes responded to each substance: control, AD flu, KT syrup, KT alkaloids, and KT aqueous gathered from seven mice per class, as shown in Figure 3.

The conventional hand-crafted feature of LFP signals is spectral power. This experiment used statistical testing across the quantitative spectral power analysis to assess the significant difference in LFP responses among the group of mice treated with different substances. As reported in Table II, the results found that LFP responses to KT syrup and KT alkaloids have no significant difference from AD flu in most brain regions and oscillation bands. In summary, the results convinced both KT extracts could be potential candidates for alternative drugs given the similar effects on the brain to AD flu. However, this conventional approach demonstrated the weakness in differentiating LFP responses from KT syrup and KT alkaloids. Thus, we addressed the weakness by proposing ANet and the multi-task AE approaches in identifying the higher similarity of LFP responses to KT syrup and AD flu compared to those to KT alkaloids.

Experiment II demonstrated the performance of the proposed anomaly detector or ANet in assessing the similarity of the unseen LFP compared to LFP response to AD flu was the reference class or standard antidepressant substance in producing the drugs. The findings revealed that the number of LFP that responded to KT syrup appeared to distribute mainly in the normal zone indicating the high similarity to the standard drug or AD flu Figure 4(a), followed by KT alkaloids Figure 4(b), and KT aqueous Figure 4(c) extract, respectively. Quantitative data found that the normal detecting rate or defined as the similarity rate for this study, was 85.62 ± 0.29%, 83.64 ± 0.32%, and 40.69 ± 0.38% in detecting LFP responses or testing sets from KT syrup, KT alkaloids, and KT aqueous extract, respectively, referred to LFP response of AD flu or the training sets.

To investigate an insightful optimization process of the multi-task AE, we assessed the alterations of training and validation losses in the multi-LFP class recognition during the training process. Based on the least iteration stopped by early stopping, Figure 5(a), (b), and (c) demonstrated training and validation loss while training for 70 epochs for MSE loss, CE loss, and total loss, respectively. The findings demonstrated that MSE loss reached a relatively stable level after data training for ten epochs. Meanwhile, high fluctuation patterns were in the CE and the total losses during the 20^th-40^th epochs. However, validation loss converged to the training loss, suggesting that the model training had not been overfitted, confirming this study’s well-trained model. All measures of the performance were applied together with the early stopping function. In summary, the proposed multi-task AE reached accuracy, and the F1-score of the model was 79.78 ± 0.39% and 79.53 ± 0.00%, respectively.

Furthermore, we also illustrated a confusion matrix of the predicted outputs. While the matrix revealed that almost all LFP samples that responded to the treated substances were primarily correctly classified, it was obvious that there were some misclassifications between the AD flu and KT syrup, as shown in Figure 5(d). Thus, LFP samples from both classes might have similar spectral power features originating from sharing mechanisms exerted on the CNS and the brain. The finding was consistent with Subsections IV-A and IV-B.

The 2-dimension of the t-SNE projection visualized the probability distribution of reduced dimensional data at latent space. It was found that t-SNE of baseline, processed from raw spectral power, diffused and mixed among classes of substances evenly to form one cluster, as shown in Figure 6(a). Meanwhile, the characteristics of t-SNE resulting from the learned features or represented latent vectors appeared to be formed for correct clusters. For example, consistency with the experiments mentioned earlier and results, among KT extracts, the KT syrup cluster shared the intersect clusters of the AD flu mostly, as depicted in Figure 6(b).

Although discriminative LFP features of KT extract distributed with the AD flu overlappingly in the t-SNE visualization, this result seems to provide only qualitative estimation. To get a concrete explanation, MMD distance, a value estimated from the t-SNE projection-based-latent vector evaluation, was thus analyzed quantitatively. The results exposed that LFP features of the AD flu and KT syrup were the first groups showing the shortest MMD distance to control mice and were followed by KT alkaloids and KT aqueous, respectively. Moreover, features in LFP responses from the AD flu: KT syrup and AD flu: KT alkaloids were the first two paired classes that produced the shortest MMD distance. In contrast, the longest MMD distance was detected in the LFP features of AD flu to KT aqueous, as shown in Figure 7. This finding also brought us to the conclusive similarity of LFP responses obtained from the group of mice treated with KT syrup and AD flu.