A deep learning-based method for evaluating the effectiveness of executive interventions

Abstract

The present invention relates to the field of artificial intelligence and data processing technology, and in particular to a method for evaluating the effect of an executive intervention based on deep learning, which is as follows: recording the heart rate signals of autistic children during intervention task training through a wearable heart rate monitoring device; preprocessing the collected heart rate signals; constructing a data set and dividing it into a training set, a validation set, and a test set; performing dynamic time segmentation and adaptive noise suppression on each sample data in the data set; constructing an executive intervention effect evaluation model based on deep learning, inputting the denoised training set into the model for training, and verifying the performance indicators of the model through the validation set; inputting the data in the test set into the trained model to obtain the final evaluation results. The present invention can objectively estimate the effect of the intervention and assist manual judgment by analyzing and processing the heart rate signals reflecting psychological load, cognitive engagement, and emotional fluctuations.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data processing technology, and in particular to an executive intervention effect evaluation method based on deep learning. Background Art

[0002] Currently, executive-based training is gaining increasing attention in rehabilitation interventions for children with autism as a key intervention approach that directly impacts children's cognition, behavioral control, and task completion. However, the evaluation of executive-based interventions often relies on subjective methods such as teacher observation, questionnaire scoring, or video review. These methods are subject to delays, inconsistent standards, and difficulty quantifying individual differences, making it difficult to provide immediate, refined, and personalized feedback on the effectiveness of interventions. With the widespread adoption of wearable devices and the advancement of real-time acquisition technologies for physiological signals such as heart rate, heart rate variability, as an objective indicator reflecting children's mental workload, cognitive engagement, and emotional fluctuations, has been recognized as a potential channel for quantifying intervention effectiveness. However, heart rate signals exhibit complex characteristics such as non-stationarity, high noise levels, and significant individual variability. Conventional signal processing and modeling methods have significant limitations in modeling time series and extracting intervention responses. For example, fixed-window denoising methods cannot adapt to the phase transitions of intervention tasks, and conventional neural networks struggle to identify task-induced physiological responses, neglecting the temporal impact of intervention events. Furthermore, the lack of feature extraction mechanisms and evaluation models designed for intervention evaluation tasks limits the practical application of heart rate signals in intelligent assessment.

[0003] Therefore, the present invention proposes an executive intervention effect evaluation method based on deep learning to solve the above problems. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention develops an executive intervention effect evaluation method based on deep learning. By accurately retaining relevant features, enhancing modeling capabilities, capturing subtle fluctuations and performing in-depth comprehensive optimization of multiple models, the present invention can improve the accuracy of heart rate signal analysis and processing, thereby obtaining more accurate intervention effect evaluation results.

[0005] The technical solution to the technical problem of the present invention is a deep learning-based execution-based intervention effect evaluation method, which includes the following steps:

[0006] S1. Data collection: Wearable heart rate monitoring devices were used to record the heart rate signals of children with autism during intervention task training.

[0007] S2. Data preprocessing: preprocess the collected heart rate signal, remove abnormal data, and normalize and time-align the data;

[0008] S3. Dataset construction and partitioning: When constructing the dataset, the preprocessed data is divided into multiple samples according to the different intervention tasks. Each sample contains a complete heart rate sequence within the time period of an intervention task, as well as the label corresponding to the intervention task. The dataset is then divided into a training set, a validation set, and a test set.

[0009] S4, data denoising: dynamically segment each sample data in the data set into time segments, and then perform adaptive noise suppression on the data in each time segment;

[0010] S5. Build an executive intervention effect evaluation model based on deep learning. Input the denoised training set into the model for training to obtain the evaluation results of the intervention effect. At the same time, verify the performance indicators of the model through the validation set. When the performance indicators of the model reach the preset standards, the model training stops and the trained model is obtained. The training content of the model includes dynamic attention encoding of heart rate sequences, construction of multi-scale differential features, and deep spiking neural network modeling.

[0011] S6. Input the data in the test set into the trained model, output the classification probability of the intervention effect, and take the category with the highest probability as the final evaluation result.

[0012] S1 is as follows:

[0013] In the test of the effectiveness of executive intervention for children with autism, a standardized intervention task training process was designed to stimulate the physiological and psychological responses of children with autism through intervention tasks. Then, a wearable heart rate monitoring device was used to collect the children's heart rate signals in real time, obtaining continuous heart rate time series data.

[0014] Intervention tasks include cognitive training, social skills training, and behavior modification training.

[0015] S4 is as follows:

[0016] The pre-processed heart rate signal in the dataset is dynamically segmented, and the continuous heart rate signal is divided into multiple time segments of variable length according to the time span of each intervention stage in the intervention training task;

[0017] Then, within each time segment, a wavelet basis function that matches the temporal characteristics of the segment is selected, and the wavelet window width is dynamically adjusted according to the mean and variance statistical characteristics of the data within the segment. Then, a double-threshold screening is performed on the data within each segment to obtain the denoised heart rate signal for each segment.

[0018] Finally, the denoised heart rate signals of all time segments are spliced ​​into a complete heart rate signal time series in chronological order, and the heart rate value corresponding to the starting moment of each intervention task anchor point is extracted from it.

[0019] The dynamic attention encoding process of the heart rate sequence is as follows:

[0020] Dynamically weight the denoised heart rate signal time series of each sample based on the start time of the intervention task. Specifically, at each moment, the cosine similarity between the denoised heart rate signal time series and the heart rate values ​​at all task anchor points is calculated. The weight of the high similarity is amplified by an exponential function to obtain the attention weight of each anchor point at the current moment.

[0021] Then, the time difference of each task anchor point is input into the multi-layer perceptron to generate the time difference related modulation factor, and then each sample The complete heart rate sequence time series at each moment is multiplied element-wise by the time difference related modulation factor, and finally the attention weights of each anchor point are used for attention enhancement and then accumulated to obtain the heart rate encoding features that integrate the task timing information.

[0022] The process of constructing multi-scale differential features is as follows:

[0023] By dynamically adjusting the differential kernel width to capture heart rate changes at different time scales, the variance of the heart rate in a local window near the current moment is calculated. The differential kernel width is adaptively determined based on the variance to obtain a dynamic kernel width. This kernel width is then used to calculate the difference between the previous and next heart rate values. The difference result is then smoothed using an exponential function combined with the local variance to obtain the heart rate differential feature.

[0024] Then the denoised heart rate signal time series of each sample is concatenated with the heart rate differential features of different kernel widths according to the time points to form a multi-scale joint feature.

[0025] The process of deep spiking neural network modeling is as follows:

[0026] The structure of the deep pulse neural network is defined as a multi-layer impulse response fully connected neural network, including an input layer, a multi-layer pulse activation hidden layer and an output layer;

[0027] The multi-scale joint features are input to the input layer, and then the hidden layer uses an impulse parameterized activation function to perform deep abstraction and capture of nonlinear features through multiple layer stacking. Residual connections and dynamic gating mechanisms are used between layers, and finally the output layer is fully connected. The classification probability of the intervention effect is calculated at the first level, and the category with the highest probability is taken as the final evaluation result;

[0028] The training process of the deep spiking neural network is as follows: defining the activation function of the deep spiking neural network, performing forward propagation of the deep spiking neural network, calculating the joint loss function of the deep spiking neural network, and performing backpropagation of the deep spiking neural network.

[0029] The process of defining the activation function of a deep spiking neural network is as follows:

[0030] Define an activation function with an adjustable waveform shape, dynamically adjust the nonlinear strength of the activation function through learnable parameters, and superimpose a sine function to phase modulate the activation output, thereby obtaining a pulsed parameterized activation function. The calculation formula is as follows:

[0031] ,

[0032] in, is a pulsed parameterized activation function; is the neuron input of the deep spiking neural network. For the first layer of the deep spiking neural network, its input is the multi-scale joint feature, and for the next layer of the deep spiking neural network, its input is the output of the previous layer; are the learnable control parameters of deep spiking neural networks; is the nonlinear intensity adjustment factor, , To control the parameter curvature; is a sine function; is the frequency parameter of the sine function, set to 0.5; is the phase shift parameter of the sine function.

[0033] The forward propagation process of a deep spiking neural network is as follows:

[0034] In the forward propagation stage of the deep spiking neural network, a pulse forward propagation mechanism based on gated residual is used to capture the multi-stage impulse response and non-stationary features in the heart rate signal. Layer pulse activation unit, the calculation process during forward propagation is as follows:

[0035] ,

[0036] in, For the samples Moment in the Deep Spiking Neural Network Output features of the layer; For the samples Moment in the Deep Spiking Neural Network Output features of the layer; For Hadamard; Deep Spiking Neural Network Layers have learnable weights; Deep Spiking Neural Network Layers can learn biases; is a pulsed parameterized activation function; Deep Spiking Neural Network Learnable control parameters of the layer; is a dynamic gating function.

[0037] The joint loss function calculation process of the deep spiking neural network is as follows:

[0038] A joint loss function is defined that includes the differences between the intervention effectiveness classification task and the heart rate sequence prediction task, and is jointly optimized in a weighted manner. First, the focus loss function is used to weight the error-prone samples in the classification task. Then, the correlation coefficient between the predicted value and the true value is used in the regression loss to impose a higher penalty on the error of low-correlation samples. Finally, the independence between feature dimensions is maintained by defining a feature orthogonality constraint. The calculation formula of the joint loss function is as follows:

[0039] ,

[0040] in, is the joint loss function; A hyperparameter that controls the weight between classification and regression losses; is the classification loss; is the regression loss; is the weight hyperparameter of the feature orthogonality constraint loss; is the feature orthogonality constraint loss;

[0041] (1) The classification loss adopts the focus loss mechanism, which increases the weight of the misclassified samples and reduces the contribution of the easy-to-classify samples to the loss. The calculation of is as follows:

[0042] ,

[0043] in, is the number of classification categories; Indicates the prediction for The probability of each category; represents the adjustment factor for focus loss; For the The true label vector of each category;

[0044] (2) The regression loss is based on the mean square error and combines the correlation coefficient between the predicted heart rate value and the true value as a dynamic weight. The lower the correlation coefficient, the greater the error penalty. The regression loss The calculation of is as follows:

[0045] ,

[0046] in, is the total number of time steps in the heart rate time series; For the samples Predicted heart rate value at all times; For the samples Real heart rate value at all times; Calculate the Pearson correlation coefficient function;

[0047] The predicted heart rate value is the output of the heart rate value prediction model, which is essentially a recurrent neural network. The multi-scale joint features are cascaded with the features output by the last layer of the deep pulse neural network, and then input into the heart rate value prediction model. The heart rate value prediction model performs regression modeling on the input features and outputs the predicted heart rate value;

[0048] (3) The feature orthogonality constraint loss forces the encoded feature matrix to maintain orthogonality in different dimensions. The calculation of is as follows:

[0049] ,

[0050] in, It is the feature output by the last layer of the deep spiking neural network; for device; is the identity matrix; is the Frobenius norm.

[0051] The back propagation process of the deep spiking neural network is as follows:

[0052] During the back-propagation process of the deep spiking neural network, an adaptive pulse momentum gradient update method is used to update the weights of each layer in the deep spiking neural network by combining the conventional gradient descent method with the local pulse momentum mechanism.

[0053] Deep Spiking Neural Network The weight update rules of the layer network are as follows:

[0054] ,

[0055] ,

[0056] in, Deep Spiking Neural Network Layer Weight after round update; Deep Spiking Neural Network Layer Weight after round update; is the learning rate of the deep spiking neural network; is the momentum coefficient; is the gradient memory decay parameter; Deep Spiking Neural Network Layer The weight after round update.

[0057] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0058] The present invention divides time segments based on the triggering moment of the intervention task and adaptively selects the wavelet basis function and window width. This can accurately preserve the task-related heart rate fluctuation characteristics, avoiding the lack of flexibility of conventional denoising methods and processing strategies, and the difficulty in dealing with the non-stationary characteristics of the heart rate signal during the intervention phase.

[0059] The present invention adopts a similarity weighting mechanism between the task anchor point and the current heart rate, combined with a time difference modulation factor, which can effectively enhance the modeling ability of the physiological response to the intervention event, avoiding the situation where the model lacks a response mechanism to the task triggering event and has difficulty distinguishing spontaneous fluctuations from task-induced physiological changes.

[0060] The present invention dynamically adjusts the differential kernel width according to the intensity of local heart rate fluctuations and integrates the variance-weighted exponential smoothing method to enhance the ability to capture subtle fluctuations. This avoids the problem that conventional differential and fluctuation extraction methods are rough, lack adaptability, and cannot accurately extract microscopic heart rate features for intervention assessment.

[0061] The present invention constructs a sinusoidal modulation activation function based on the heart rate pulse response characteristics, and combines classification, trend prediction and orthogonal constraint triple joint loss to achieve comprehensive optimization of the deep model. It can also avoid the problem of limited model expression ability and training stability caused by the lack of activation functions and optimization strategies adapted to physiological data.

[0062] In summary, the present invention can provide timely, precise and personalized evaluation feedback by analyzing and processing heart rate signals that reflect psychological load, cognitive engagement and emotional fluctuations, and assist manual judgment by providing objective intervention effect evaluation results, thereby contributing to the rehabilitation intervention practice of autistic children. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0064] Figure 1 Schematic diagram of the method of the present invention.

[0065] Figure 2 This is a comparison chart of the classification accuracy of the denoising method in the present invention and the existing denoising method.

[0066] Figure 3 This is a comparison chart of the denoising effects of the denoising method in the present invention and the fixed window denoising method.

[0067] Figure 4 3 is a comparison chart of the effects of the timing modeling method in the present invention and the existing timing modeling method.

[0068] Figure 5 is a visual graph of the dynamic attention weight distribution.

[0069] Figure 6 This is a comparison chart of the overall performance of the deep learning-based executive intervention effect evaluation model in the present invention and the existing models. DETAILED DESCRIPTION

[0070] In order to clearly illustrate the technical features of this solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0071] Example 1

[0072] A deep learning-based execution-based intervention effect evaluation method includes the following steps:

[0073] S1. Data collection: Wearable heart rate monitoring devices were used to record the heart rate signals of children with autism during intervention task training.

[0074] S2. Data preprocessing: preprocess the collected heart rate signal, remove abnormal data, and normalize and time-align the data;

[0075] Abnormal data refers to data caused by equipment failure or interference with children's activities;

[0076] By normalizing the collected data, the heart rate can be adjusted to a uniform range, which makes it easier to sort out the heart rate differences between different individuals when processing the data later;

[0077] Time alignment of the data ensures that the start and end times of each intervention task correspond precisely to the timestamps of the heart rate data;

[0078] S3. Dataset Construction and Partitioning: When constructing the dataset, the preprocessed data is divided into multiple samples based on the different intervention tasks. Each sample contains a complete heart rate sequence within the time period of an intervention task, as well as the label corresponding to the intervention task. For example, the classification labels of the intervention effect are "effective" and "ineffective". The dataset is then divided into training, validation, and test sets, which are used for model training, hyperparameter adjustment, and final performance evaluation, respectively.

[0079] S4, data denoising: dynamically segment each sample data in the data set into time segments, and then perform adaptive noise suppression on the data in each time segment;

[0080] S5. Build an executive intervention effect evaluation model based on deep learning. Input the denoised training set into the model for training to obtain the evaluation results of the intervention effect. At the same time, verify the performance indicators of the model through the validation set. When the performance indicators of the model reach the preset standards, the model training stops and the trained model is obtained. The training content of the model includes dynamic attention encoding of heart rate sequences, construction of multi-scale differential features, and deep spiking neural network modeling.

[0081] S6. Input the data in the test set into the trained model, output the classification probability of the intervention effect, and take the category with the highest probability as the final evaluation result.

[0082] In a specific implementation manner, S1 is specifically as follows:

[0083] In the test of the effectiveness of executive intervention for children with autism, a standardized intervention task training process was designed to stimulate the physiological and psychological responses of children with autism through intervention tasks. Then, a wearable heart rate monitoring device was used to collect the children's heart rate signals in real time, obtaining continuous heart rate time series data.

[0084] Intervention tasks include cognitive training, social skills training, and behavior modification training, designed to stimulate through specific activities.

[0085] In a specific implementation manner, S4 is specifically as follows:

[0086] The pre-processed heart rate signal in the data set is dynamically segmented, and the continuous heart rate signal is divided into multiple time segments of variable length according to the time span of each intervention stage in the intervention training task. The heart rate value at the moment is expressed as ;

[0087] Then, within each time segment, a wavelet basis function that matches the temporal characteristics of the segment is selected, and the wavelet window width is dynamically adjusted according to the mean and variance statistical characteristics of the data within the segment. Then, a double-threshold screening is performed on the data within each segment to obtain the denoised heart rate signal for each segment.

[0088] The specific process of denoising is to calculate the mean and standard deviation of each local area in the clip, set the lower and upper thresholds, and only retain the heart rate fluctuations that exceed the threshold range. The rest are regarded as noise and set to zero, thereby retaining the high fluctuation characteristics related to the task. The calculation formula is as follows:

[0089] ,

[0090] in, For the The denoised heart rate sequence of the sth sample in the time segment reflects the heart rate data after local feature enhancement; The number of local areas divided into each time segment; For the The mean of the data in a time segment; For the The standard deviation of the time segment data; For the The wavelet basis function corresponding to each time segment is: Indicates the The time range of each dynamic time segment is determined based on the task event triggering moment or the heart rate change boundary, such as the task triggering moment or the heart rate change boundary, which can avoid the loss of local features caused by fixed windows; express Heart rate value at the moment; represents a dual-threshold selection function;

[0091] Adaptively adjust the wavelet window width according to the variance of the segment data For example, when the wavelet basis function is selected When using the Dobesi wavelet function, calculate the wavelet window width , the calculation formula is as follows:

[0092] ,

[0093] in, Indicates the The number of time points in a dynamic time segment, Indicates the The mean of the data in a time segment;

[0094] The dual-threshold selection function can divide the heart rate fluctuation into task-related features and irrelevant noise. Suppose the input of the dual-threshold selection function is , the calculation formula is as follows:

[0095] ,

[0096] in, is the low threshold, is the high threshold;

[0097] Finally, the heart rate signals after denoising all time segments are spliced ​​into a complete heart rate signal time series in chronological order. , and extract the heart rate value corresponding to the starting moment of each intervention task anchor point , Indicates the samples The complete heart rate sequence time series at each moment, Indicates the The intervention task Anchor point start time The corresponding heart rate value.

[0098] It should be noted that dynamic time segmentation is a special treatment for the non-stationary and sudden fluctuation characteristics of heart rate data. It can flexibly divide time segments according to the triggering time of task events or the boundary of heart rate changes, which is impossible to achieve with conventional fixed window methods. Adaptive noise suppression can adaptively remove noise according to the statistical characteristics of each time segment by dynamically adjusting the wavelet window width and dual threshold screening, which can better adapt to the complex characteristics of heart rate data and improve the quality of the signal. The wavelet basis function is selected The Dobesi wavelet function, combined with the calculation of the dual-threshold selection function, can realize non-fixed window wavelet denoising and solve the problem that conventional fixed windows are prone to losing dynamic features; the reasonable setting of low and high thresholds can effectively screen significant heart rate fluctuations related to the task and suppress irrelevant physiological noise, such as resting fluctuations or sensor errors. It is a special setting specifically made for executive intervention tasks.

[0099] In a specific implementation, the dynamic attention encoding process of the heart rate sequence is as follows:

[0100] Dynamically weight the denoised heart rate signal time series of each sample based on the start time of the intervention task. Specifically, at each moment, the cosine similarity between the denoised heart rate signal time series and the heart rate values ​​at all task anchor points is calculated. The weight of the high similarity is amplified by an exponential function to obtain the attention weight of each anchor point at the current moment.

[0101] The correlation between the current heart rate value and the task anchor value is measured by cosine similarity, and the anchor value of each task is calculated. The attention weight at a certain moment is calculated as follows:

[0102] ,

[0103] in, For the The anchor point of the task is The attention weight of the moment, is the total number of anchor points in the intervention task; is the attention coefficient, which is used to amplify the high similarity weight. ; is the cosine similarity function; is the attention weight mapping matrix, which is a learnable parameter matrix used to enhance the model's ability to capture complex dynamic relationships;

[0104] Then, the time difference of each task anchor point is input into the multi-layer perceptron to generate the time difference related modulation factor, and then each sample The complete heart rate sequence time series at each moment is multiplied element-wise by the time difference related modulation factor, and finally the attention weights of each anchor point are used for attention enhancement and then accumulated to obtain the heart rate encoding features that integrate the task timing information.

[0105] The calculation formula of the heart rate encoding feature is as follows:

[0106] ,

[0107] in, For the Samples in Heart rate encoding features at each moment; For Hadamard; is the mapping function of the multilayer perceptron; The time difference between the current moment and the task anchor point;

[0108] The heart rate encoding feature integrates the task anchor point attention and time difference modulation features, which can reflect the influence of the heart rate sequence changing with the time of the task stage and capture the dynamic physiological pattern of the task response of autistic children.

[0109] It should be noted that the dynamic attention mechanism is a special treatment for the problem that the heart rate response triggered by the intervention task is difficult to distinguish from spontaneous fluctuations in the heart rate data. Through dynamic weighting and time difference modulation, the model can better capture the impact of the intervention task on the heart rate, which is difficult to achieve with conventional LSTM or GRU models; the heart rate response triggered by the intervention task is usually concentrated near the anchor point, and dynamic attention encoding helps to capture the task timing correlation and distinguish spontaneous fluctuations; the time difference modulation factor can further enhance the model's ability to model time information, enabling the model to better understand the impact of the time sequence and duration of the intervention task on heart rate changes.

[0110] In a specific implementation, the process of constructing multi-scale differential features is as follows:

[0111] By dynamically adjusting the differential kernel width to capture heart rate changes at different time scales, the variance of the heart rate in a local window near the current moment is calculated. The differential kernel width is adaptively determined based on the variance to obtain a dynamic kernel width. This kernel width is then used to calculate the difference between the previous and next heart rate values. The difference result is then smoothed using an exponential function combined with the local variance to obtain the heart rate differential feature.

[0112] The calculation formula of the heart rate difference feature is as follows:

[0113] ,

[0114] in, For the current moment Centered at Heart rate differential characteristics; After denoising Time right interval Heart rate value; After denoising Time left interval Heart rate value; is the smoothing coefficient, set to 0.7; for local variance of the moment;

[0115] The dynamic kernel width is calculated from the local variance and is dynamically adjusted according to the local window variance to reflect the severity of local heart rate fluctuations. When the variance is large, the kernel width is reduced to capture microscopic changes, and vice versa, the kernel width is expanded to smooth noise. The calculation formula is as follows:

[0116] ,

[0117] in, For rounding operation; It is a logarithmic function, and the default base is 10;

[0118] The local variance can measure the stability or intensity of the local heart rate and directly reflect the microscopic changes in the heart rate signal. The calculation formula is as follows:

[0119] ,

[0120] ,

[0121] in, is the width of the local window, is the total length of the window; After denoising Time right interval Heart rate value; is the mean of the heart rate data in the local window;

[0122] Then the denoised heart rate signal time series of each sample is concatenated with the heart rate differential features of different kernel widths according to the time points to form a multi-scale joint feature.

[0123] The calculation formula of multi-scale joint features is as follows:

[0124] ,

[0125] in, for Multi-scale joint features of moments; is the fixed heart rate differential feature; The dynamic kernel width is Heart rate differential characteristics; The dynamic kernel width is Heart rate differential characteristics.

[0126] It should be noted that the dynamic kernel width heart rate differential feature can adaptively adjust the differential kernel width according to the local characteristics of the heart rate data, thereby better capturing the microscopic patterns of heart rate changes and subtle fluctuations in heart rate during the intervention process, thereby improving the model's assessment accuracy of the intervention effect; As an exponential smoothing term, it can suppress noise in low-variance areas and retain fluctuation details in high-variance areas. Smoothing can further enhance the robustness of features, enabling the model to better cope with noise interference in heart rate data. It is a necessary step to cooperate with dynamic kernel width to better capture the microscopic patterns of heart rate changes.

[0127] In a specific implementation, the process of deep pulse neural network modeling is as follows:

[0128] The structure of the deep pulse neural network is defined as a multi-layer impulse response fully connected neural network, including an input layer, a multi-layer pulse activation hidden layer and an output layer;

[0129] The multi-scale joint features are input to the input layer, and then the hidden layer uses an impulse parameterized activation function to perform deep abstraction and capture of nonlinear features through multiple layer stacking. The residual connection and dynamic gating mechanism are used between layers to suppress gradient disappearance and network overfitting. Finally, the output layer is fully connected. The classification probability of the intervention effect is calculated at the first level, and the category with the highest probability is taken as the final evaluation result;

[0130] The training process of the deep spiking neural network is as follows: defining the activation function of the deep spiking neural network, performing forward propagation of the deep spiking neural network, calculating the joint loss function of the deep spiking neural network, and performing backpropagation of the deep spiking neural network.

[0131] In a specific implementation, the activation function of a deep spiking neural network is defined as follows:

[0132] An activation function with an adjustable waveform shape is defined. The nonlinear strength of the activation function is dynamically adjusted through learnable parameters, so that the activation function remains linear in the low-amplitude region and quickly saturates in the high-amplitude region. At the same time, a sinusoidal function is superimposed to phase-modulate the activation output, so that the waveform of the activation function matches the pulsed rising / falling trend of the heart rate signal, thereby obtaining a pulsed parameterized activation function. The calculation formula is as follows:

[0133] ,

[0134] in, is a pulsed parameterized activation function; is the neuron input of the deep spiking neural network. For the first layer of the deep spiking neural network, its input is the multi-scale joint feature, and for the next layer of the deep spiking neural network, its input is the output of the previous layer; are the learnable control parameters of deep spiking neural networks; is the nonlinear intensity adjustment factor, , To control the parameter curvature, set it to 0.3, by The shape of the activation function can be controlled so that the low-amplitude region linearly maintains details and the high-amplitude region saturates to suppress noise; is a sine function; is the frequency parameter of the sine function, set to 0.5; The phase offset parameter of the sine function is set to 1.

[0135] It should be noted that the heart rate signal has an impulse response characteristic. The impulse parameterized activation function is a special treatment made for the impulse response characteristic of the heart rate data, especially the As a sinusoidal phase modulation strategy, it can better model the rising and falling trends of heart rate signals; based on the pulsed parameterized activation function, the use of dynamic gating mechanism and residual connection can effectively suppress gradient vanishing and network overfitting, and improve the training stability and generalization ability of the model.

[0136] In a specific embodiment, the forward propagation process of the deep spiking neural network is as follows:

[0137] In the forward propagation stage of the deep spiking neural network, a pulse forward propagation mechanism based on gated residual is used to capture the multi-stage impulse response and non-stationary features in the heart rate signal. Layer pulse activation unit, the calculation process during forward propagation is as follows:

[0138] ,

[0139] in, For the samples Moment in the Deep Spiking Neural Network Output features of the layer; For the samples Moment in the Deep Spiking Neural Network Output features of the layer; Deep Spiking Neural Network Layers have learnable weights; Deep Spiking Neural Network Layers can learn biases; is a pulsed parameterized activation function; Deep Spiking Neural Network Learnable control parameters of the layer; is a dynamic gating function.

[0140] The dynamic gating function can control the flow intensity of the pulse characteristics and adapt to the signal fluctuation characteristics of different intervention stages. The calculation formula is as follows:

[0141] ,

[0142] in, is the Sigmoid function; For the The weight of the layer gated layer; For the The bias of the layer gated layer; Deep Spiking Neural Network The differential features of the layer features can capture instantaneous changes and enhance the adaptability of the gate to fluctuations in the intervention stage. , express Moment in the Deep Spiking Neural Network The output features of the layer, express Moment in the Deep Spiking Neural Network The output characteristics of the layer; Characterize the concatenation of the previous layer features of a deep spiking neural network and its differential features, highlighting the impact of instantaneous changes on the gating mechanism.

[0143] In a specific embodiment, the joint loss function calculation process of the deep spiking neural network is as follows:

[0144] A joint loss function is defined that includes the differences between the intervention effectiveness classification task and the heart rate sequence prediction task, and is jointly optimized in a weighted manner. First, the focus loss function is used to weight and emphasize the error-prone samples in the classification task, which can improve the model's learning ability for minority classes. Then, the correlation coefficient between the predicted value and the true value is used in the regression loss. A higher penalty is imposed on the error of low-correlation samples, which can ensure that the model focuses on the trend of heart rate changes rather than single-point errors. Finally, by defining feature orthogonality constraints to maintain the independence between feature dimensions, the redundancy of the feature space during multi-task learning can be avoided. The calculation formula of the joint loss function is as follows:

[0145] ,

[0146] in, is the joint loss function; To control the weight between classification and regression loss, the hyperparameter is set to ; is the classification loss; is the regression loss; is the weight hyperparameter of the feature orthogonality constraint loss, set to ; is the feature orthogonality constraint loss;

[0147] (1) The classification loss adopts the focus loss mechanism, which increases the weight of the misclassified samples and reduces the contribution of the easy-to-classify samples to the loss. It can weight and emphasize the difficult-to-classify samples and enhance the model's ability to judge the effectiveness of the intervention. The calculation of is as follows:

[0148] ,

[0149] in, is the number of classification categories; Indicates the prediction for The probability of each category; Indicates the adjustment factor of focus loss, set to 2; For the The true label vector of each category;

[0150] (2) The regression loss is based on the mean square error and combines the correlation coefficient between the predicted heart rate value and the true value as a dynamic weight. The lower the correlation coefficient, the greater the error penalty. This can strengthen the model's grasp of the overall trend and highlight the true dynamic trend of the heart rate change during the intervention process. The regression loss The calculation of is as follows:

[0151] ,

[0152] in, is the total number of time steps in the heart rate time series; For the samples Predicted heart rate value at all times; For the samples Real heart rate value at all times; Calculate the Pearson correlation coefficient function;

[0153] The predicted heart rate value is the output of the heart rate value prediction model, which is essentially a recurrent neural network. The multi-scale joint features are cascaded with the features output by the last layer of the deep pulse neural network, and then input into the heart rate value prediction model. The heart rate value prediction model performs regression modeling on the input features and outputs the predicted heart rate value;

[0154] (3) The feature orthogonality constraint loss forces the encoded feature matrix to maintain orthogonality in different dimensions, which can avoid feature redundancy and ensure the validity and independence of features. The calculation of is as follows:

[0155] ,

[0156] in, It is the feature output by the last layer of the deep spiking neural network; for device; is the identity matrix; is the Frobenius norm.

[0157] It should be noted that the joint loss function comprehensively considers the differences between the intervention effectiveness classification task and the heart rate sequence prediction task. Through joint optimization in a weighted manner, it can better balance the needs of classification and regression tasks. The design of feature orthogonality constraints can avoid feature redundancy and improve the generalization ability of the model. Therefore, the joint optimization of the three losses can effectively solve the problems of model overfitting and trend neglect.

[0158] In a specific embodiment, the back propagation process of the deep spiking neural network is as follows:

[0159] During the back-propagation process of deep spiking neural networks, an adaptive pulse momentum gradient update method is used to combine the conventional gradient descent method with the local pulse momentum mechanism to update the weights of each layer in the deep spiking neural network. This can more effectively optimize the network parameters and avoid the gradient vanishing and explosion problems.

[0160] Deep Spiking Neural Network The weight update rules of the layer network are as follows:

[0161] ,

[0162] ,

[0163] in, Deep Spiking Neural Network Layer Weight after round update; Deep Spiking Neural Network Layer Weight after round update; is the learning rate of the deep spiking neural network, set to 0.001; is the momentum coefficient; is the gradient memory decay parameter, set to 0.7; Deep Spiking Neural Network Layer The weight after round update.

[0164] The momentum coefficient is dynamically calculated through the gradient volatility of the network, and the calculation formula is as follows:

[0165] ,

[0166] in, is the basic momentum parameter, set to 0.9; is the pulse sensitivity parameter, set to 0.5; represents the gradient norm.

[0167] It should be noted that, except for the weight parameters of the deep spiking neural network, the remaining learnable parameters are updated using the conventional gradient descent method; the adaptive pulse momentum gradient update method combines the traditional gradient descent method with the local pulse momentum mechanism, and does not directly add the momentum-related parameters to the parameter update increments of the conventional gradient descent method. Instead, it comprehensively considers the adaptive weighted average of the current round gradient and the historical gradient difference, and uses the gradient memory decay parameter for calculation. It can more effectively consider the historical iterative optimization situation to further optimize the network parameters and improve the training efficiency and stability of the model.

[0168] Example 2

[0169] like Figure 2 As shown in the figure, in order to verify the influence of dynamic time segmentation and adaptive noise suppression strategy on the accuracy of intervention effect evaluation, the performance of traditional sliding average filtering, fixed window wavelet denoising and the method of the present invention in classification tasks are compared; Figure 2 The box plot results in the figure show that the conventional sliding average filter loses local features due to the fixed window, and the classification results are more discrete; although the fixed window wavelet denoising improves a certain stability, it is not adaptable enough to sudden heart rate fluctuations, and the upper limit of the accuracy is still limited; in contrast, the method of the present invention can more accurately distinguish between noise and task-related features by dynamically dividing time segments and adaptively adjusting the wavelet basis function window width. The box position is significantly higher than other methods and the distribution range is more concentrated, indicating that it has stronger robustness and classification consistency on different samples, reflecting the effectiveness of the dynamic segmentation strategy for extracting complex heart rate signal features.

[0170] like Figure 3 As shown, in order to verify the advantage of the dynamic time segmentation denoising method in retaining task-related physiological characteristics, the denoising effect of the dynamic time segmentation denoising method in the present invention is compared with that of the fixed window denoising method. Figure 3As can be seen in the figure, the original heart rate signal (grey solid line) has significant fluctuations near the intervention task anchor points (20 seconds, 35 seconds, and 50 seconds marked by the green dotted lines). Although the fixed window denoising method (blue dotted line) can smooth the noise, it excessively flattens the task-related fluctuation details in the anchor point area. The dynamic time segmentation denoising method (red solid line) of the present invention automatically shrinks the analysis window around the task anchor point, which not only suppresses high-frequency noise but also fully preserves the pulse-like fluctuation characteristics. Figure 3 The experimental results shown in the figure show that on the signal curve of the anchor point area, the red curve shows an upward and downward trend consistent with the original signal, proving that the method of the present invention can accurately distinguish between noise and effective physiological responses based on the characteristics of the task stage, overcoming the problem of feature loss caused by denoising using conventional methods.

[0171] Example 3

[0172] like Figure 4 As shown in Figure 2, in order to verify the ability of the dynamic attention encoding mechanism to capture task-related heart rate characteristics, the receiver operating characteristic curves of the long short-term memory network, the gated recurrent unit and the dynamic attention encoding method proposed in this invention are compared. Figure 4 It can be seen that the true positive rate of the long short-term memory network and the gated recurrent unit increases slowly with the false positive rate because it is difficult to distinguish between spontaneous fluctuations and task response characteristics, and the area under the curve is relatively small; while the method of the present invention adopts the cosine similarity weight of the task anchor point and the time difference modulation factor, so that the model can focus on the key heart rate change nodes in the intervention stage. The curve shows a steeper upward trend and is overall close to the upper left corner of the coordinate system, indicating that the dynamic attention mechanism can significantly enhance the model's sensitivity to task triggering features and effectively improve the recognition ability of true positive samples.

[0173] like Figure 5 As shown in Figure 2, by visualizing the dynamic attention weight distribution, we analyze the impact of task anchors on heart rate feature encoding. Figure 5 The red curve in the middle represents the change of attention weight over time, and the green dotted line marks the starting moment of the intervention task. Figure 5 It can be observed that the weight rapidly climbs to a peak at the beginning of each task and then decays exponentially, forming a pulsed weight distribution centered on the anchor point. This pattern is highly consistent with the physiological response of the heart rate signal (light blue curve) after the task begins. When the weight increases, the corresponding heart rate value shows characteristic task-triggered fluctuations. By strengthening the characteristics of heart rate changes during the critical period of the task, the attention mechanism enables the model to automatically focus on the physiological response segments related to the intervention, while downplaying the influence of daily spontaneous fluctuations. Figure 5 The phenomenon that the three weight peaks are strictly aligned with the anchor points proves that this method can accurately capture the correlation between task timing and physiological responses.

[0174] Example 4

[0175] like Figure 6 As shown in the figure, the multi-dimensional advantages of the deep learning-based executive intervention effect evaluation model constructed in this invention are analyzed through comprehensive performance comparison experiments. The bar chart on the left represents the accuracy and inference speed, the line chart on the right represents the memory usage, and the feature dimension labeling below the model name reflects the network complexity; Figure 6 It can be seen that although traditional time series models (long short-term memory network LSTM, temporal convolutional network TCN, Transformer network) have a certain accuracy rate, they have the defects of high memory usage and slow inference speed; the pulse neural network in the executive intervention effect evaluation model of the present invention maintains the highest accuracy while having the lowest memory usage and the fastest inference speed, indicating that the adopted pulse activation function is adapted to the characteristics of physiological signals and reduces redundant parameter calculations by simulating the nonlinear characteristics of heart rate pulse response.

[0176] Although the above describes the specific implementation methods of the invention in conjunction with the accompanying drawings, it does not limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A deep learning-based executive intervention effect evaluation method, characterized by: The following steps are involved: S1. Data collection: Wearable heart rate monitoring devices were used to record the heart rate signals of children with autism during intervention task training. S2. Data preprocessing: preprocess the collected heart rate signal, remove abnormal data, and normalize and time-align the data; S3. Dataset construction and partitioning: When constructing the dataset, the preprocessed data is divided into multiple samples according to the different intervention tasks. Each sample contains a complete heart rate sequence within the time period of an intervention task, as well as the label corresponding to the intervention task. The dataset is then divided into a training set, a validation set, and a test set. S4, data denoising: dynamically segment each sample data in the data set into time segments, and then perform adaptive noise suppression on the data in each time segment; S4 is as follows: The pre-processed heart rate signal in the dataset is dynamically segmented, and the continuous heart rate signal is divided into multiple time segments of variable length according to the time span of each intervention stage in the intervention training task; Then, within each time segment, a wavelet basis function that matches the temporal characteristics of the segment is selected, and the wavelet window width is dynamically adjusted according to the mean and variance statistical characteristics of the data within the segment. Then, a double-threshold screening is performed on the data within each segment to obtain the denoised heart rate signal for each segment. Finally, the heart rate signals after denoising all time segments are spliced ​​into a complete heart rate signal time series in chronological order, and the heart rate value corresponding to the starting moment of each intervention task anchor point is extracted from it; S5. Build an executive intervention effect evaluation model based on deep learning. Input the denoised training set into the model for training to obtain the evaluation results of the intervention effect. At the same time, verify the performance indicators of the model through the validation set. When the performance indicators of the model reach the preset standards, the model training stops and the trained model is obtained. The training content of the model includes dynamic attention encoding of heart rate sequences, construction of multi-scale differential features, and deep spiking neural network modeling. The process of constructing multi-scale differential features is as follows: By dynamically adjusting the differential kernel width to capture heart rate changes at different time scales, the variance of the heart rate in a local window near the current moment is calculated. The differential kernel width is adaptively determined based on the variance to obtain a dynamic kernel width. This kernel width is then used to calculate the difference between the previous and next heart rate values. The difference result is then smoothed using an exponential function combined with the local variance to obtain the heart rate differential feature. Then, the denoised heart rate signal time series of each sample is concatenated with the heart rate differential features of different kernel widths at each time point to form a multi-scale joint feature. S6. Input the data in the test set into the trained model, output the classification probability of the intervention effect, and take the category with the highest probability as the final evaluation result.

2. The method for evaluating the effectiveness of an executive intervention based on deep learning according to claim 1, characterized in that: S1 is as follows: In the test of the effectiveness of executive intervention for children with autism, a standardized intervention task training process was designed to stimulate the physiological and psychological responses of children with autism through intervention tasks. Then, a wearable heart rate monitoring device was used to collect the children's heart rate signals in real time, obtaining continuous heart rate time series data. Intervention tasks include cognitive training, social skills training, and behavior modification training.

3. The method for evaluating the effectiveness of an executive intervention based on deep learning according to claim 2, wherein: The dynamic attention encoding process of the heart rate sequence is as follows: Dynamically weight the denoised heart rate signal time series of each sample based on the start time of the intervention task. Specifically, at each moment, the cosine similarity between the denoised heart rate signal time series and the heart rate values ​​at all task anchor points is calculated. The weight of the high similarity is amplified by an exponential function to obtain the attention weight of each anchor point at the current moment. Then, the time difference of each task anchor point is input into the multi-layer perceptron to generate the time difference related modulation factor, and then each sample The complete heart rate signal time series at each moment is multiplied element-wise by the time difference related modulation factor, and finally the attention is enhanced by the attention weights of each anchor point and then accumulated to obtain the heart rate encoding features of the fusion task timing information.

4. According to the deep learning-based executive intervention effect evaluation method of claim 3, the process of deep spiking neural network modeling is as follows: The structure of the deep pulse neural network is defined as a multi-layer impulse response fully connected neural network, including an input layer, a multi-layer pulse activation hidden layer and an output layer; The multi-scale joint features are input to the input layer, and then the hidden layer uses an impulse parameterized activation function to perform deep abstraction and capture of nonlinear features through multiple layer stacking. Residual connections and dynamic gating mechanisms are used between layers, and finally the output layer is connected through a fully connected layer. The classification probability of the intervention effect is calculated at the first level, and the category with the highest probability is taken as the final evaluation result; The training process of the deep spiking neural network is as follows: defining the activation function of the deep spiking neural network, performing forward propagation of the deep spiking neural network, calculating the joint loss function of the deep spiking neural network, and performing backpropagation of the deep spiking neural network.

5. The method for evaluating the effectiveness of an executive intervention based on deep learning according to claim 4, wherein: The process of defining the activation function of a deep spiking neural network is as follows: Define an activation function with an adjustable waveform shape, dynamically adjust the nonlinear strength of the activation function through learnable parameters, and superimpose a sine function to phase modulate the activation output, thereby obtaining a pulsed parameterized activation function. The calculation formula is as follows: ， in, is a pulsed parameterized activation function; is the neuron input of the deep spiking neural network. For the first layer of the deep spiking neural network, its input is the multi-scale joint feature, and for the next layer of the deep spiking neural network, its input is the output of the previous layer; are the learnable control parameters of deep spiking neural networks; is the nonlinear intensity adjustment factor, , To control the parameter curvature; is a sine function; is the frequency parameter of the sine function; is the phase shift parameter of the sine function.

6. The method for evaluating the effectiveness of an executive intervention based on deep learning according to claim 5, characterized in that: The forward propagation process of a deep spiking neural network is as follows: In the forward propagation stage of the deep spiking neural network, a pulse forward propagation mechanism based on gated residual is used to capture the multi-stage impulse response and non-stationary features in the heart rate signal. Layer pulse activation unit, the calculation process during forward propagation is as follows: ， in, For the samples Moment in the Deep Spiking Neural Network Output features of the layer; For the samples Moment in the Deep Spiking Neural Network Output features of the layer; For Hadamard; Deep Spiking Neural Network Layers have learnable weights; Deep Spiking Neural Network Layers can learn biases; is a pulsed parameterized activation function; Deep Spiking Neural Network Learnable control parameters of the layer; is a dynamic gating function.

7. The method for evaluating the effectiveness of an executive intervention based on deep learning according to claim 6, wherein: The joint loss function calculation process of the deep spiking neural network is as follows: A joint loss function is defined that includes the differences between the intervention effectiveness classification task and the heart rate sequence prediction task, and is jointly optimized in a weighted manner. First, the focus loss function is used to weight the error-prone samples in the classification task. Then, the correlation coefficient between the predicted value and the true value is used in the regression loss to impose a higher penalty on the error of low-correlation samples. Finally, the independence between feature dimensions is maintained by defining a feature orthogonality constraint. The calculation formula of the joint loss function is as follows: ， in, is the joint loss function; A hyperparameter that controls the weight between classification and regression losses; is the classification loss; is the regression loss; is the weight hyperparameter of the feature orthogonality constraint loss; is the feature orthogonality constraint loss; (1) The classification loss adopts the focus loss mechanism, which increases the weight of the misclassified samples and reduces the contribution of the easy-to-classify samples to the loss. The calculation of is as follows: ， in, is the number of classification categories; Indicates the prediction for The probability of each category; represents the adjustment factor for focus loss; For the The true label vector of each category; (2) The regression loss is based on the mean square error and combines the correlation coefficient between the predicted heart rate value and the true value as a dynamic weight. The lower the correlation coefficient, the greater the error penalty. The regression loss The calculation of is as follows: , in, is the total number of time steps in the heart rate time series; For the samples Predicted heart rate value at all times; For the samples Real heart rate value at all times; Calculate the Pearson correlation coefficient function; The predicted heart rate value is the output of the heart rate value prediction model, which is essentially a recurrent neural network. The multi-scale joint features are cascaded with the features output by the last layer of the deep pulse neural network, and then input into the heart rate value prediction model. The heart rate value prediction model performs regression modeling on the input features and outputs the predicted heart rate value; (3) The feature orthogonality constraint loss forces the encoded feature matrix to maintain orthogonality in different dimensions. The calculation of is as follows: ， in, It is the feature output by the last layer of the deep spiking neural network; for device; is the identity matrix; is the Frobenius norm.

8. The deep learning-based execution-based intervention effect evaluation method according to claim 7 is characterized in that: The back propagation process of the deep spiking neural network is as follows: During the back-propagation process of the deep spiking neural network, an adaptive pulse momentum gradient update method is used to update the weights of each layer in the deep spiking neural network by combining the conventional gradient descent method with the local pulse momentum mechanism. Deep Spiking Neural Network The weight update rules of the layer network are as follows: ， ， in, Deep Spiking Neural Network Layer Weight after round update; Deep Spiking Neural Network Layer Weight after round update; is the learning rate of the deep spiking neural network; is the momentum coefficient; is the gradient memory decay parameter; Deep Spiking Neural Network Layer The weight after round update.

Images