Individualized pain prediction method and system based on mixed effect

Abstract

The invention discloses an individualized pain prediction method and system based on a mixing effect. The method comprises the following steps: acquiring pain scores and corresponding central nervous system data when a plurality of healthy individuals receive pain stimulation; performing data processing on the obtained central nervous system data to obtain central nervous system physiological signals of pain stimulation received by corresponding individuals; combining the obtained physiological signals of the central nervous system and the corresponding pain scores to construct a mixed effect model; predicting the pain state of the individual by using the constructed mixed effect model; the system comprises a pain stimulation device, a central nervous detection system, a data acquisition module, a data analysis module, a prediction model construction module and a pain recognition module. The constructed mixed effect model gives consideration to the characteristics of an individualized model and a group model, not only can new individuals be migrated, but also individual differences are considered, the hierarchical characteristics of data are fully utilized, and the pain score prediction effectiveness can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of pain detection, and in particular to an individualized pain prediction method and system based on mixed effects. Background Art

[0002] Pain is one of the most common symptoms of disease, and chronic pain plagues hundreds of millions of patients, causing huge health and economic burdens on patients and society. Building an accurate pain prediction model is of great significance for objectively quantifying the degree of pain and developing personalized pain treatment plans. Machine learning has been widely used in pain prediction, but the existing individualized and group models for pain prediction have shortcomings such as low accuracy and poor applicability.

[0003] The current pain prediction models are mainly divided into two categories: individualized models and group models. Individualized models build different models for each individual, while group models build a single model applicable to the entire group. Although individualized models take into account individual differences, the model has poor transferability and cannot be transferred to new individuals. In addition, a large amount of data needs to be collected from each individual to build a model with better predictive effectiveness. Although group models have good transferability, it is difficult to consider the impact of individual differences on the model's prediction effect, and they also ignore the hierarchical data structure caused by the correlation of the same individual's data. These shortcomings are the important reasons for the low predictive effectiveness and poor applicability of existing pain prediction models.

[0004] The mixed effect model is a model that can organically integrate individual characteristics and group characteristics, taking into account the advantages of individual models and group models, making up for their shortcomings and improving the prediction effectiveness of the model. Therefore, the development of a pain prediction model based on mixed effects will have important clinical application value. Summary of the invention

[0005] In order to solve the above-mentioned technical problems and achieve accurate prediction of individual pain status, the present invention provides an individualized pain prediction method and system based on mixed effects, which uses central physiological signals such as EEG and functional magnetic resonance imaging, peripheral nervous system signals such as skin electricity and electromyography, and non-physiological signals such as gender, age, and psychological scale scores to predict pain, taking into account individual differences and group characteristics, and is easy to migrate to predict psychological states and clinical characteristics other than pain.

[0006] The present invention adopts the following technical solution:

[0007] In one aspect, the present invention provides an individualized pain prediction method based on mixed effects, the method comprising:

[0008] Step 1, collecting pain scores and corresponding central nervous system data of several healthy individuals when receiving pain stimulation;

[0009] Step 2, processing the obtained central nervous system data to obtain a central nervous system physiological signal corresponding to the individual receiving pain stimulation;

[0010] Step 3, combining the obtained central nervous system physiological signals and the corresponding pain scores to construct a mixed effects model;

[0011] Step 4: Use the constructed mixed effects model to predict individual pain status.

[0012] Furthermore, the data collected in step 1 also includes peripheral nervous system data generated when a healthy individual receives pain stimulation, and the data is processed in step 2.

[0013] Preferably, the mixed effect model constructed in step 3 is:

[0014] y ij =f(X ij )+b i Z i +e ij

[0015] Where: y ij is the continuous pain score of individual i in the jth trial;

[0016] X ij is the physiological signal matrix of individual i in the jth trial;

[0017] f( ) is a linear or nonlinear transformation function;

[0018] f(X ij ) constitutes the prediction y ij Fixed effects model of ;

[0019] Z i is the characteristic corresponding to the random effect, which is assumed to be normally distributed as N(0,D);

[0020] b i is the random effect coefficient of individual i;

[0021] e ij is the error term, which is assumed to be normally distributed as N(0,σ 2 I).

[0022] Furthermore, the expectation maximization algorithm is used to estimate the model parameters in the constructed mixed effect model. The specific method is:

[0023] Step 3.1, initialize random effect coefficients Error variance σ 2 =1, random effect covariance matrix D = 1;

[0024] Step 3.2, in the tth iteration, use the current estimate Calculate the residual of the expected target value

[0025]

[0026] Step 3.3, use fixed effects model fitting

[0027]

[0028] Step 3.4, update the random effect coefficients

[0029]

[0030] in

[0031] Step 3.5, update the error variance σ 2(t+1) :

[0032]

[0033] Step 3.6, update the random effect covariance matrix D (t+1) :

[0034]

[0035] Step 3.7, calculate the generalized log-likelihood value GLL after the current iteration (t+1) , when GLL (t+1) With GLL (t) When the absolute value of the difference is less than the set threshold τ, the iteration is stopped and the model parameters of the mixed effect model are obtained.

[0036] Preferably, the algorithm of generalized log-likelihood value adopted is:

[0037]

[0038] in: n i is the number of trials of the i-th subject, and det(·) is the determinant.

[0039] Preferably, the central nervous system data collected in step 2 is functional magnetic resonance imaging data, and the specific processing method is:

[0040] The obtained functional magnetic resonance imaging data were analyzed using SPM12, removing the first three brain volumes of each scan and using the middle layer for temporal slice correction;

[0041] The time-slice-corrected functional images were registered with the T1 structural images, then normalized to the MNI template and smoothed;

[0042] The smoothed data were analyzed using a general linear model, with the regressors including several pain stimulus conditions and six head motion parameter estimates convolved with the canonical hemodynamic response function;

[0043] The signal was high-pass filtered, and the first-order autoregressive model (AR(1)) was used to reduce the impact of temporal autoregression on the results, thus obtaining the processed functional magnetic resonance physiological signal.

[0044] On the other hand, the present invention also provides an individualized pain prediction system based on mixed effects, the system comprising: a pain stimulation device, which acts on the body of a healthy individual to produce pain stimulation and obtain a self-assessed pain score;

[0045] A central nervous system detection system for collecting functional magnetic resonance imaging data of healthy individuals when receiving painful stimuli;

[0046] A data collection module, used to collect pain scores and corresponding central nervous system data of healthy individuals when receiving pain stimulation;

[0047] A data analysis module processes the obtained central nervous system data to obtain a central nervous system physiological signal corresponding to the individual receiving pain stimulation;

[0048] A prediction model building module, which combines the obtained central nervous system physiological signals and the corresponding pain scores to establish a mixed effect model;

[0049] The pain recognition module uses the constructed mixed effect model to predict the pain state of the individual to be tested and outputs it.

[0050] Furthermore, the system also includes a peripheral nerve detection system for collecting peripheral nervous system data such as skin electricity and myoelectricity of healthy individuals.

[0051] The technical solution of the present invention has the following advantages:

[0052] A. The present invention provides an individualized pain prediction method and system based on mixed effects, in which the constructed mixed effect model takes into account the characteristics of both individualized models and group models. It can not only be migrated to new individuals, but also takes into account individual differences, fully utilizes the hierarchical characteristics of the data, and can effectively improve the prediction effectiveness of pain scores.

[0053] B. The mixed effect model constructed by the present invention based on mixed effect has been verified by experiments. In terms of mean absolute error, root mean square error and error percentage measurement indicators, the prediction power of the present invention is improved by 20% compared with the traditional fixed effect model. If the explanation rate of score variation R is 2 As a measure of the index, the predictive power of the present invention is improved by more than 95% compared with the traditional fixed effect model, which is significantly higher than the traditional fixed effect model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the specific embodiments of the present invention, the drawings required for use in the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A flow chart of the individualized pain prediction method based on mixed effects provided by the present invention;

[0056] Figure 2 A diagram showing the composition of the individualized pain prediction system provided by the present invention;

[0057] Figure 3 A diagram of the random slope distance and random slope model provided by the present invention;

[0058] Figure 4 It is a comparison chart of the predictive power between the mixed effect model constructed by the present invention based on mixed effect and the traditional fixed effect model. DETAILED DESCRIPTION

[0059] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, the present invention provides an individualized pain prediction method based on mixed effects, comprising the following steps:

[0061] [S01] Collect pain scores and corresponding central nervous system data, such as EEG and functional magnetic resonance imaging data, from several healthy individuals when they receive pain stimulation; of course, it also includes peripheral nervous system data generated when receiving pain stimulation, such as skin electricity signals and electromyography signals.

[0062] [S02] Processing the obtained central nervous system data to obtain central nervous system physiological signals corresponding to the individual receiving pain stimulation.

[0063] The central nervous system data is preferably functional magnetic resonance imaging data, and the specific processing method is: using SPM12 to analyze the obtained functional magnetic resonance imaging data. To ensure the stability of the signal, the present invention removes the first three brain volumes of each scan. The functional magnetic resonance imaging data image uses the middle layer for time layer correction. After time layer correction, the functional magnetic resonance imaging data image is aligned with the T1 structural image and then standardized to the MNI template (resampled voxel size: 3×3×3mm 3 ), and finally smoothing (Gaussian kernel, such as half-height width of 6mm).

[0064] The smoothed data were analyzed using a general linear model. The regression factors included several pain stimulation conditions (e.g., 10 stimulations × 2 stimulation intensities, a total of 20 pain stimulation conditions) and six head motion parameter estimates after convolution with the typical hemodynamic response function. In addition, the signal was high-pass filtered (cutoff period: 128s), and the first-order autoregressive model (AR(1)) was used to reduce the influence of time series autoregression on the results. Finally, the processed central nervous system physiological signals were obtained.

[0065]

S03

[0066] (1) Model setting

[0067] The mixed effects model based on mixed effects is as follows:

[0068] y ij =f(X ij )+b i Z i +e ij

[0069] where y ij is the continuous pain score of individual i in the jth trial, X ij is the physiological signal matrix of individual i in the jth trial, f( ) is a linear or nonlinear transformation function, f(X ij ) constitutes the prediction y ij The fixed effect part, Z i is the feature corresponding to the random effect, b i is the random effect coefficient of individual i, e ij is the error term. b i and e ij All are assumed to follow normal distribution.

[0070] The model of the present invention is a method of introducing mixed effects into a machine learning regression model. ij It can represent central nervous system physiological signals such as EEG and MRI, peripheral nervous system physiological signals such as skin electricity and myoelectricity, as well as non-physiological signals such as gender, age, and psychological scale scores. ij The conversion can play a dual role: first, the original feature X ij The first is to perform transformations such as standardization and dimensionality reduction, and the second is to perform nonlinear transformations on features. Depending on the f( ), the mixed effect model can be constructed using common algorithms such as linear regression, principal component regression, random forest, and support vector regression.

[0071] Z i represents the random effect. Depending on the difference of Z, this model can include the random intercept model and the random slope model. In general, Z i Including several features, the model also considers f(X ij ) predict y ij When different individuals have different intercepts and slopes (e.g. Figure 3 The intercepts α1, α2, α3 and slopes β1, β2, β3 of the different individual fitting lines represented by different colors are all different), which is a random slope model. If z is only a unit vector, the model only considers f(X ij ) predict y ij When the intercepts of different individuals are different, it is a random intercept model.

[0072] When the constructed mixed effect model predicts new individual data, if the new data is collected from individuals included in the training set, the model's prediction will take into account the random effect of the individual, and the predictive power will be improved; if the new data comes from new individuals not included in the training set, the model's prediction will be based on the fixed effect part, and the predictive power will be equivalent to the traditional model. Therefore, the predictive power of this mixed effect-based model will theoretically be no less than that of the model trained using traditional machine learning methods.

[0073] (2) Model parameter estimation

[0074] The model parameters are estimated using the Expectation-Maximization algorithm (EM algorithm). i =f(X i )+b i Z i +e i In the above example, we assume that b i The distribution is N(0,D), e i The distribution is N(0,σ2 I). The specific steps of the expectation maximization algorithm to estimate parameters are as follows:

[0075]

S031

[0076] The random effect coefficients are initialized to zero: Error variance σ 2 =1, random effects covariance matrix D=I.

[0077] The EM algorithm maximizes the log-likelihood function of the observed data by alternating the following steps:

[0078]

S032

[0079]

[0080]

S033

[0081]

[0082] Next, update the random effect parameters and model parameters σ 2(t+1) and D (t+1) ):

[0083]

S034

[0084]

[0085] in

[0086]

S035

[0087]

[0088]

S036

[0089]

[0090]

S033

[0091] Calculate the generalized log-likelihood (GLL) after the current iteration:

[0092]

[0093] in

[0094] For the pre-set threshold τ, if |GLL (t+1) -GLL (t) |<τ, the iteration stops; if the condition is not met, the iteration continues.

[0095]

S04

[0096] In addition, if Figure 2 As shown, the present invention also provides an individualized pain prediction system based on mixed effects, including: a pain stimulation device, a central nervous system detection system, a data acquisition module, a data analysis module, a prediction model construction module and a pain recognition module; the pain stimulation device acts on the body of a healthy individual to produce pain stimulation and obtain a self-assessed pain score; the central nervous system detection system is used to collect central nervous system data of healthy individuals when receiving pain stimulation; the data acquisition module is used to collect the pain score and corresponding central nervous system data of healthy individuals when receiving pain stimulation; the data analysis module performs data processing on the obtained central nervous system data to obtain the central nervous system physiological signal of the corresponding individual receiving pain stimulation; the prediction model construction module combines the obtained central nervous system physiological signal and the corresponding pain score to establish a mixed effect model; the pain recognition module uses the constructed mixed effect model to predict the pain state of the individual to be tested, and outputs it for the user to view. Of course, the system further includes a peripheral nerve detection system for collecting peripheral nervous system data such as skin electricity and myoelectricity of healthy individuals.

[0097] Example

[0098] In order to test the effect of the present invention on the prediction of pain scores, the pain scores and corresponding functional magnetic resonance imaging data of 100 healthy volunteers (37 males and 63 females, age [M±SD] 22.5±6.2) when receiving pain stimulation were collected. All subjects had no acute / chronic pain, neurological diseases, cerebrovascular diseases, coronary heart disease and mental disorders. The volunteers signed an informed consent form before the experiment began.

[0099] Pain stimulation was achieved using a short radiant heat pulse generated by a Nd:YAP laser (spot diameter ~7mm, duration 4ms). The stimulation intensity was 3.0J and 3.5J, and volunteers were required to use a digital rating scale to rate these two stimuli from 0 to 10. Among them, 0 points represented no feeling, 10 points represented unbearable pain, and the higher the score, the stronger the feeling.

[0100] In the experiment, each subject received 10 laser pain stimuli of 3.0J and 3.5J respectively. At the beginning of the experiment, a black screen with a white "+" was first displayed on the computer screen for 6 seconds, and then the laser pain stimulation was applied. After the pain stimulation was presented for 10 seconds, a 5-second scoring interface appeared on the screen. The volunteers used the keys to move the cursor on the screen to report the perceived assessment stimulus intensity (0-10 points), and then the black screen with a white "+" appeared on the screen again, and the next laser stimulation was applied 1s or 2s later.

[0101] While the volunteers were performing the pain stimulus rating task, their brain activity was collected using functional magnetic resonance imaging (fMRI). Magnetic resonance data were collected using a 3.0T GE magnetic resonance imaging instrument.

[0102] The T1 structural image scan used the SPGR sequence with the following scanning parameters: echo time (TE) = 2.992 ms, repletion time (TR) = 6.896 ms, inversion time (TI) = 450 ms, sagittal slices = 176, thickness = 1 mm, gap = 0 mm, acquisition matrix = 256 × 256; in-plane resolution = 1 × 1 mm.

[0103] The T2 structural image scan used the Cube T2 sequence with the following scanning parameters: TE = Maximum, TR = 2500 ms, thickness = 1 mm, gap = 0 mm, acquisition matrix = 256 × 256; in-plane resolution = 1 × 1 mm.

[0104] The task state scan used the Gradient Echo sequence, and the scanning parameters were: TR = 2000ms, TE = 29ms, axial slices = 43, thickness = 3.0mm, gap = 0mm, acquisition matrix = 64×64, flip angle = 90°, field of view = 192×192mm 2 , in-plane resolution=3×3mm.

[0105] MRI data were analyzed using SPM12. To ensure signal stability, the first three brain volumes of each scan were removed. Brain images were time-slice corrected using the middle layer. After time-slice correction, functional images were registered with T1 and then normalized to the MNI template (resampled voxel size: 3×3×3mm 3 ), and finally smoothed (Gaussian kernel, half-width at half maximum of 6 mm).

[0106] The regressors in the individual-level analysis included 20 pain stimulus conditions (10 stimuli × 2 intensities) and six head motion parameter estimates after convolution with the canonical hemodynamic response function. In addition, the signal was high-pass filtered (cutoff period: 128 s), and a first-order autoregressive model (AR(1)) was used to reduce the influence of temporal autoregression on the results.

[0107] After completing the individual-level analysis, the t-value statistical graph corresponding to each stimulus was obtained and used as the feature for building the model, while the pain score of each stimulus was used as the target for prediction.

[0108] In order to compare the pain prediction mixed effect model proposed in the present invention with the conventional fixed effect model, a mixed effect model and a fixed effect model were constructed respectively. To evaluate the effectiveness of the model, the data of all 100 volunteers were randomly divided into two halves, one half was used as the discovery set (Discovery Set) for building the model, and the other half was used as a separate test set (Test Set).

[0109] The fixed effect model uses Lasso regression (LASSO). First, the t-value map corresponding to the pain stimulus is standardized. Then, the feature dimension is reduced using principal component analysis, but all principal components are retained. Finally, the Lasso regression is constructed with the principal components as features. In order to determine the optimal adjustment parameter λ of the Lasso regression, 100 samples in [10 -2 , 10 2 ] and used five-fold cross validation to adjust the parameter. The prediction effectiveness of the model was also measured using five-fold cross validation, and the following common model prediction evaluation indicators were used to evaluate the prediction power: mean absolute error (MAE), root mean square error (RMSE), R 2 In addition, the predictive power of the model was tested on a separate reserved test set.

[0110] The construction of the mixed effect model is as described in the previous model setting section. In order to improve the comparability of the prediction power of the two models, the fixed effect f(X ij ) uses the same lasso regression and parameter λ as the fixed effect model, and the random effect b i Z i It only contains random intercepts, that is, Z i is a vector of all 1s.

[0111] like Figure 2As shown, the mixed effect model has a significantly higher prediction power for pain scores of a single stimulus than the traditional fixed effect model. In the discovery set and test set, MAE dropped from 1.39 and 1.36 of the fixed effect model to 1.09 and 1.03 of the mixed effect model, RMSE dropped from 1.76 and 1.73 to 1.39 and 1.34, and the error percentage dropped from 25% and 25% to 20% and 19%. It can be seen that if measured by these indicators, the mixed effect model provided by the present invention has a prediction power that is about 20% higher than that of the fixed effect model. If the explanation rate of score variation R is used, the mixed effect model provided by the present invention is 20%. 2 As a measure of the indicators, the predictive power of the mixed effect increased from 0.25 and 0.27 to 0.55 and 0.56 compared with the fixed effect, with an improvement of more than 95%.

[0112] Anything not described in the present invention is applicable to the prior art.

[0113] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. However, the obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A personalized pain prediction method based on mixed effects, characterized in that: The method comprises: Step 1, collecting pain scores and corresponding central nervous system data of several healthy individuals when receiving pain stimulation; Step 2, processing the obtained central nervous system data to obtain a central nervous system physiological signal corresponding to the individual receiving pain stimulation; Step 3, combining the obtained central nervous system physiological signals and the corresponding pain scores to construct a mixed effects model; Step 4: Use the constructed mixed effects model to predict individual pain status.

2. The individualized pain prediction method based on mixed effects according to claim 1, characterized in that: The data collected in step 1 also includes peripheral nervous system data generated when a healthy individual receives pain stimulation, and the data is processed in step 2.

3. The individualized pain prediction method based on mixed effects according to claim 1, characterized in that: The mixed effects model constructed in step 3 is: y ij =f(X ij )+b i Z i +e ij Where: y ij is the continuous pain score of individual i in the jth trial; X ij is the physiological signal matrix of individual i in the jth trial; f() is a linear or nonlinear transformation function; f(X ij ) constitutes the prediction y ij Fixed effects model of ; Z i is the characteristic corresponding to the random effect, which is assumed to be normally distributed as N(0,D); b i is the random effect coefficient of individual i; e ij is the error term, which is assumed to be normally distributed as N(0,σ 2 I).

4. The individualized pain prediction method based on mixed effects according to claim 3, characterized in that: The expectation maximization algorithm is used to estimate the model parameters in the constructed mixed effect model. The specific method is: Step 3.1, initialize random effect coefficients Error variance σ 2 =1, random effects covariance matrix D = I; Step 3.2, in the tth iteration, use the current estimate Calculate the residual of the expected target value Step 3.3, use fixed effects model fitting Step 3.4, update the random effect coefficients in Step 3.5, update the error variance σ 2(t+1) : Step 3.6, update the random effect covariance matrix D (t+1) : Step 3.7, calculate the generalized log-likelihood value GLL after the current iteration (t+1) , when GLL (t+1) With GLL (t) When the absolute value of the difference is less than the set threshold τ, the iteration is stopped and the model parameters of the mixed effect model are obtained.

5. The individualized pain prediction method based on mixed effects according to claim 4, characterized in that: The algorithm used for the generalized log-likelihood is: in: n i is the number of trials of the i-th subject, and det(·) is the determinant.

6. The individualized pain prediction method based on mixed effects according to claim 1, characterized in that: The central nervous system data collected in step 2 is functional magnetic resonance imaging data, and the specific processing method is: The obtained functional magnetic resonance imaging data were analyzed using SPM12, removing the first three brain volumes of each scan and using the middle layer for temporal slice correction; The time-slice-corrected functional images were registered with the T1 structural images, then normalized to the MNI template and smoothed; The smoothed data were analyzed using a general linear model, with the regressors including several pain stimulus conditions and six head motion parameter estimates convolved with the canonical hemodynamic response function; The signal was high-pass filtered, and the first-order autoregressive model (AR(1)) was used to reduce the impact of temporal autoregression on the results, thus obtaining the processed functional magnetic resonance physiological signal.

7. An individualized pain prediction system based on mixed effects, characterized in that: The system comprises: a pain stimulation device, which acts on the body of a healthy individual to produce pain stimulation and obtain a self-assessed pain score; Central nervous system detection system, used to collect central nervous system data of healthy individuals when receiving pain stimulation; A data collection module, used to collect pain scores and corresponding central nervous system data of healthy individuals when receiving pain stimulation; A data analysis module processes the obtained central nervous system data to obtain a central nervous system physiological signal corresponding to the individual receiving pain stimulation; A prediction model building module, which combines the obtained central nervous system physiological signals and the corresponding pain scores to build a mixed effect model; The pain recognition module uses the constructed mixed effect model to predict the pain state of the individual to be tested and outputs it.

8. The mixed effect-based individualized pain prediction system according to claim 7, characterized in that: The system also includes a peripheral nerve detection system for collecting skin electricity and myoelectric peripheral nervous system data of a healthy individual.

Images