Research and optimisation of a deep learning model for positive thinking meditation based on bio-signal processing

Before the beginning of the experiment, the demographic information of all the subjects was collected. The main information contained the age of the subjects, whether they had mature experience in positive thinking meditation, etc. 100 subjects without any experience in positive thinking meditation were recruited through screening, with 50 males and 50 females [21]. The psychological state of the subjects was normal, and the average age of the subjects in the experimental group was 24.6 years old. All subjects did not suffer from mental diseases and were in good health.

The brain area concerned by this positive thinking meditation experiment is mainly located in the bilateral prefrontal lobe, so the design of the light pole plate covers the area where the prefrontal lobe is located. This experiment needs to design the light pole plate by itself, and the structure of its design is shown in Fig. 1, which includes 8 light sources (red numbers 1-8, wavelengths of 720 nm and 870 nm, respectively), which are indicated by the 8 transmitting probes. 16 receiving probes (blue numbers 1-16). Maintaining a fixed distance of 5 cm between the individual probes, this setup will result in the formation of detection channels between each transmitting probe and each receiving probe, forming a total of 28 channels (green straight lines). The electrode was placed in the prefrontal region of the subject’s head, and the entire electrode was attached to the subject’s head using a black hair band.

Figure 1.

Light plate structure design

This experiment was conducted in the laboratory of the Institute of Biophysics, and it was arranged that different subjects sat in the middle of a couch during the positive meditation experiment, where the main subject #1 sat behind the couch and recorded the NIRS data, and another main subject #2 sat in front of a table to collect the EEG data [22].

Simultaneous recording of EEG and f NIRS can increase the amount of data and may be particularly useful for some brain-computer interface applications as it helps to capture changes in the activity of biological signals such as EEG and haemodynamics corresponding to a single task. Just as methods for recording brain activity continue to evolve, so must techniques for decoding multiple data streams. Figure 2 shows a sequence of oxygenated hemoglobin concentration over time for a channel during an experiment. In the lower part of Figure A-H are the eight transmitting sources, and 1-16 are the detecting receiving sources, thus constituting 28 channels. The sequence of oxyhaemoglobin concentration over time for the black and red channels is shown, where the horizontal axis represents the time (s), and the vertical axis represents the change in oxyhaemoglobin concentration for two of the channels after the change in light intensity is converted to a change in oxygen concentration by a modified Lambert-Beer law hysteresis.

Figure 2.

The sequence of changes in concentration of oxygen and hemoglobin

The acquired signal is first band-pass filtered from 1-100 Hz to remove motion artefacts from the signal, and then 50 Hz trap filtering is performed to remove industrial frequency interference from the power supply. Raw EEG data collected from the scalp usually consists of a mixture of useful signals and irrelevant artefacts. Theoretically, these two are considered to be independent of each other, so the artefacts can be effectively separated using an independent component analysis (ICA) method.

The time-domain waveform of the original EEG signal is shown in Fig. 3, where (a) is the signal time-domain waveform, with the horizontal axis indicating time and the vertical axis indicating the relative amplitude change of the signal. (b) is the signal frequency domain graph, it can be clearly seen from the original EEG frequency domain signal that there is an obvious 50 Hz industrial frequency interference in the signal.

Figure 3.

Original brain electrical signal time domain and frequency domain waveform

The processed EEG waveform is shown in Fig. 4, where (a) is the signal time-domain waveform, with the horizontal axis indicating time and the vertical axis indicating the relative amplitude change of the signal. (b) is the signal frequency domain graph, and it can be seen from the processed EEG signal graph that the 50 Hz industrial frequency interference has been significantly eliminated.

Figure 4.

The time domain and frequency domain waveform after processing

For f NIRS data with fixed task states [23], with insufficient a priori assumptions, independent component analysis (ICA) is considered to decompose the meaningful neural activity components, while ICA methods can also strip out various types of noise components to achieve purification of f NIRS data [24]. ICA, as a data-driven method, attempts to use the ICA method to separate for noise removal. The current fNIRS-based ICA denoising method is basically temporal ICA, which is based on the temporal structure of the ICA method to make the original data under the base vector projection as free of linear correlation on time delay as possible, and the calculation formula can be written as: (1) C=E{ X(t)⋅X(t−τ)T }

The problem then boils down to coming to find a set of bases, the covariance matrix of the time delay is diagonal, and the objective function can be defined as: (2) argmin(W)=∑τ off (W^CW^T) (3) off(M)=∑i≠jmij2

The objective function J is minimised by an optimisation algorithm to obtain the corresponding W.

It is assumed that the temporal activity of noise and the neural temporal activity of interest for the meditation experiment are independent of each other and that the observed recorded blood oxygen signal consists of a linear summation of the changes in blood oxygen concentration due to these noise and neural activities paired with different combination coefficients: (4) { x1(t)⋅⋅xn(t) }=a1⋅s1(t)+a2⋅s2(t)+⋯+am⋅sm(t)

Where x_i(t) is the time series of the observed blood oxygen signal of the ind measurement channel, s_k(t) is the time series of the kth independent component, corresponding to the noise or neural activity component, which are independent of each other, and: (5) ak={ a1k⋅⋅ank },k∈{ 1,2,⋯,m }

Where a_ik is the weight of the k nd independent component on the i rd measurement channel, corresponding to the strength of noise or neural activity, and a set of a_i & s_i(t) is a source, e.g., a noise source or a neural activity source. The original data is decomposed using ICA to judge the selection of the noise components, and the process of removing the noise is to directly remove the noise term on the right side of the equation and then resume the non-noise term to get the denoised fNIRS data.

Fig. 5 shows a graph of the oxygenated haemoglobin concentration of a subject over time, calculated by the Lambert-Beer law. (a), (c) and (e) are graphs of the oxygenated haemoglobin concentration of the channel monitored by H-probe No. 9, the light source of the right brain of the subject’s positive thinking meditation, respectively. (b), (d), and (f) are graphs of the changes in oxyhaemoglobin concentration in the channel monitored by the subject’s positive mindfulness meditation left brain light source C-probe No. 7. a and b are data from the subject’s first positive mindfulness meditation, c and d are data from the subject’s third positive mindfulness meditation, and e and f are data from the subject’s eighth positive mindfulness meditation. At the time of the first positive thought meditation experiment, subjects experienced a substantial change in blood oxygen concentration in the aware breathing phase, several more pronounced fluctuations in the limb awareness phase, and strong changes in blood oxygen concentration from baseline in several other major phases compared to the pre-resting phase. As the number of positive meditation training sessions increased, the blood oxygen concentration gradually leveled off. This, to some extent, justifies the identification of the phases of positive thinking meditation through the signals of blood oxygen concentration changes.

Figure 5.

Changes in the concentration of oxygen and hemoglobin

After the processing operations described above and the processing of each meditation stage, the EEG and NIR data from the 8 meditation sessions of 100 subjects were divided according to each person per each stage to produce a dataset.

CNN is a feed-forward neural network inspired by the biological sensory field mechanism, with three structural properties: local connections, weight contributions, and convergence. CNN mainly consists of four parts: convolutional layer, pooling layer, activation function, and fully connected layer.

The problem of gradient explosion or gradient vanishing occurs when the data sequence is long, resulting in the model’s inability to capture long-distance dependencies. Therefore, in this paper, the long short-term memory network is used to solve this problem. The LSTM takes the internal state c_t-1 of the previous moment and the external state of the previous moment as inputs and introduces three gates, i.e., corresponding to the forgetting gate f_t, the input gate i_t, and the output gate o_t, and also introduces the tanh function and the Sigmoid function as activation functions, and introduces the operations of vector-most multiplication, vector-sum, and vector-splicing, and finally takes the current internal state c_i and the current external state h_t as outputs and sends them to the next loop unit. The main idea of LSTM is to update the hidden state, i.e., the external state h_i (i.e., short-term memory), by means of the forgetting gate f_i and the input gate i_i, and to provide the output gate o_i to transfer the information from the internal state c_i (which holds information for a longer period of time than the short-term memory h_i, but not as long as the long-term memory, and is also known as the long- and short-term memory) to the external state h_t.

Among them, the forgetting gate f_t maps the feature information into the interval [0, 1] by means of a Sigmoid function, so as to realize selective “forgetting” of the information in the internal state c_t-1 of the previous moment. Similarly, the input gate i_t and the output gate o_t also adopt the sigmoid function to realize the saving or outputting of the information in the current candidate state c˜t and the current internal state c_t, respectively.

The implementation of the forgetting gate f_t, input gate i_t and output gate o_t is shown in the following equations: (11) ft=σ(Wf⋅[ht−1,xt]+bf) (12) it=σ(Wi⋅[ht−1,xt]+bi) (13) ot=σ(Wo⋅[ht−1,xt]+bo)

Secondly, the candidate state c˜i at the current moment is generated by the hidden state h_t-1 at the previous moment and the input x_i at the current moment through the tanh function: (14) c˜t=tanh(WCo[ht−1,xt]+bC)

Finally, the hidden state h_t at the current moment is then similarly generated by the internal state c_t at the current moment through the tanh activation function and used as input to the next layer of LSTM units: (15) ht=ot*tanh(ct)

In this section, the migration learning and MobileNetV2 network model, the migration learning and ResNet50 network model, and the migration learning and Xception network model are integrated through a weighted average integration learning method to form a migration learning based integrated network for the positive thought meditation classification model. The model, i.e., Positive Thought Meditation Classification Model based on Migration Learning and Integrated Networks, is integrated to have a stronger generalisation ability than a single classification model.

The individual classification model with the best predictive performance is obtained by optimizing the training model. The predicted value of the positivity meditation classification model is obtained through the weighted average method, and then the error between the true value and the predicted value is calculated using the objective function, and then the weights are adaptively updated through the gradient descent and back-propagation algorithms to obtain the final positivity meditation predicted value. The specific process is as follows: first, pre-train the three classification models separately to achieve the best prediction performance. Then, the pre-trained models are used to obtain different predicted values, assuming that the predicted values obtained by these three classification models are V_A,i, V_B,i, V_C,i The initial predicted value of the integrated model is obtained by weighted average method, as shown in the following equation: (22) VO,i=WA,i*VA,i+WB,i*VB,i+WC,i+VC,i

Where, W_A,i, W_B,i and W_C,i denote the weights of the output values of the three models, W_A,i + W_B,i + W_C,i = 1.

The error between the true value and the predicted value is calculated using the cross-entropy loss function: (23) minWA,i,WB,i,WC,i=∑i=1N−[yilog(VO,i)+(1−yi)log(1−VO,i)]N

where y_i represents the true value of sample i and V_O,i is the predicted value of the integrated model.

Next, the backpropagation algorithm, using the gradient descent algorithm, adaptively updates the weight bar of the output values of each single classification model until the error between the output and predicted values of the integrated positive mindfulness meditation classification model is stabilised.