Machine learning-based postoperative index analysis method for hematopoietic stem cell transplant patients

Abstract

The application relates to the technical field of data processing, and particularly discloses a postoperative index analysis method for hematopoietic stem cell transplant patients based on machine learning, which is used for solving the problem that when hematopoietic stem cell transplant patients have high-dimensional coupling of subjective scales and objective indexes at T1, T2 and T3 multiple time points, and there are missing values, batch effects and individual differences, it is difficult to realize accurate evaluation of anxiety and depression when carrying out progressive muscle relaxation training combined with music therapy, and the method comprises the following steps: collecting postoperative indexes, building a matrix and preprocessing, splitting into A and B according to columns and calculating a difference, extracting U1, U2, V1 and V2 through four-dimensional branch lifting, and obtaining U, V and X by optimization, principal component projection is carried out on X to obtain the first K dimensions and threshold pruning to obtain W, and W and group time point interaction items are input into GLMM fitting to predict anxiety and depression; through four-branch dimension lifting fusion, principal component projection and threshold pruning, compact features are formed, and GLMM is combined to model, so that the accuracy of anxiety and depression evaluation and trend prediction is improved, and the interference of data missing and batch effect is reduced.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for analyzing postoperative indicators in hematopoietic stem cell transplant patients based on machine learning. Background Technology

[0002] Patients undergoing hematopoietic stem cell transplantation often experience fluctuations in anxiety and depression during the admission, bone marrow suppression, and bone marrow reconstitution phases. Progressive muscle relaxation training and music therapy have been used as interventions in nursing care, with implementation recorded using intervention and group identifiers. Simultaneously, GAD-7, PHQ-9, systolic blood pressure, diastolic blood pressure, heart rate, and IL-1β, IFN-γ, TNF-α, IL-6, and IL-8 were repeatedly collected on the day of admission (T1), during the bone marrow suppression phase (T2), and during the bone marrow reconstitution phase (T3). This generated a longitudinal vector sequence with patients as rows and time points as columns, along with matrices showing differences, rates of change, correlation functions, and covariance characteristics between adjacent time points. However, issues remain regarding the strong correlation of multiple indicators and the nonlinear coupling between subjective and objective factors, skewed distribution of cytokines and batch drift, and the coexistence of deletions and outliers. The time intervals are not constant due to the influence of the treatment course. The differences in the caliber of different testing platforms and the significant heterogeneity of individual baselines can lead to collinearity, dimensionality expansion, and deterioration of the matrix condition number when traditional time-point tests or directly stacking the original covariates in generalized linear mixed models. This results in instability in the design matrix inversion and iterative solution of the simultaneous equation system in maximum likelihood or least squares estimation. Interaction terms need to be manually enumerated and it is difficult to capture key dimensions that are stable across time points. Furthermore, it is difficult to align the computational workload after screening with the original scale, vital signs, and cytokines item by item to support interpretable individualized trajectory analysis and reusable data processing flow. Moreover, when the sample size is limited, it is easy to introduce multiple comparison burden and overfitting, making it difficult for trend judgment and risk stratification results to be consistent across different batches and different populations. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a machine learning-based method for analyzing postoperative indicators in hematopoietic stem cell transplant patients. By forming compact features through four-branch dimensionality fusion, principal component projection and threshold pruning and jointly modeling with GLMM, the accuracy of anxiety and depression assessment and trend prediction is improved, and the interference of missing data and batch effect is reduced.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] Machine learning-based methods for analyzing postoperative indicators in hematopoietic stem cell transplant patients include:

[0006] For hematopoietic stem cell transplant patients who underwent progressive muscle relaxation training combined with music therapy, postoperative comprehensive assessment indicators at T1, T2, and T3 were collected to form matrix M, and missing markers, abnormal removal, standardization, and batch correction were performed.

[0007] Matrix M is split into matrices A and B by column segmentation, and the first and second difference components are calculated for each index. Matrix A, matrix B and their first and second difference components are used as input to the machine learning model. The machine learning model includes the first and second dimension-up branches for matrix A, and the third and fourth dimension-up branches for matrix B. The first, second, third and fourth dimension-up branches output U1, U2, V1 and V2 respectively. U is obtained by selecting one of U1 and U2 according to the preset selection rules, and V is obtained by selecting one of V1 and V2. U and V are concatenated to obtain the feature matrix X. Principal component projection is performed on X, and the first K dimensions are taken and the dimensions with absolute values ​​less than the threshold τ are deleted to obtain the pruned feature vector W.

[0008] The pruned feature vector W, along with the group, time point, and time point interaction term, is used as the input to a generalized linear mixture model. The patient identifier is used as the random intercept, and the model is fitted with anxiety assessment and depression assessment as dependent variables, respectively. The output model parameters and prediction results are then presented.

[0009] It should be noted that, based on the traditional approach of directly using postoperative multi-timepoint scales, vital signs, and cytokines as covariates in statistical models or time-series tests, this invention adds a new machine learning feature reconstruction chain for longitudinal multimodal data at T1, T2, and T3. First, matrix M is divided into A and B columns according to indicator type, and a first and second difference component are constructed. Then, through two branches in each of the two groups, the psychological and vital signs, and cytokine levels and differences are structurally upgraded. Following selection rules, one branch is taken from each group and concatenated to form the feature matrix X. Finally, principal component projection and threshold pruning are used. After obtaining a low-redundancy pruned feature vector W, it is input into a generalized linear mixture model for inference and prediction of group, time point, and interaction terms. This solves the problems of unstable estimation, difficulty in expressing interaction relationships, and easy overfitting of small samples when the indicators have high dimensionality and strong collinearity, coexistence of distribution skewness and batch drift, and significant individual differences and time correlation during the transplantation period. With a controllable dimension W, redundant noise is suppressed while retaining subjective and objective information and time change characteristics, making the mixture model fit more stable, trend recognition and inter-group difference detection more sensitive, and obtaining reusable individualized trajectory analysis output.

[0010] As a further aspect of this invention, the postoperative comprehensive assessment indicators include psychological assessment, vital signs, and inflammatory cytokine levels. The psychological assessment includes the scores of the Generalized Anxiety Disorder-7 (GAD-7) and the Patient Health Questionnaire-9 (PHQ-9). The GAD-7 score is obtained by summing the scores of seven items—tension and anxiety, uncontrollable worry, excessive worry, difficulty relaxing, restlessness, irritability, and ominous premonitions—based on preset frequencies within the past two weeks. The PHQ-9 score is obtained by summing the scores of nine items—decreased interest, depressed mood, sleep disturbances, fatigue, abnormal appetite, self-blame or low self-esteem, decreased attention, psychomotor retardation or agitation, and self-harming thoughts—based on preset frequencies within the past two weeks. Vital signs include systolic blood pressure, diastolic blood pressure, and heart rate. Inflammatory cytokine levels include the measured values ​​of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8.

[0011] As a further aspect of the present invention, the first difference component is the value of the same indicator at time T2 minus the value at time T1, and the second difference component is the value of the same indicator at time T3 minus the value at time T2; the inputs of the first and second dimensionality-upgraded branches include psychological assessment quantities and vital sign quantities, respectively; the inputs of the third and fourth dimensionality-upgraded branches include the level quantity and difference component of inflammatory cytokine quantities, respectively; and the inputs of all dimensionality-upgraded branches also include the training content vector R of progressive muscle relaxation training and the playback content vector S of music therapy; K is the principal component dimension, and τ is the pruning threshold.

[0012] As a further aspect of the present invention, the first differential component includes the first differential component of GAD-7, the first differential component of PHQ-9, the first differential component of systolic blood pressure, the first differential component of diastolic blood pressure, the first differential component of heart rate, the first differential component of IL-1β, the first differential component of IFN-γ, the first differential component of TNF-α, the first differential component of IL-6, and the first differential component of IL-8.

[0013] As a further aspect of the present invention, the second differential component includes the second differential component of GAD-7, the second differential component of PHQ-9, the second differential component of systolic blood pressure, the second differential component of diastolic blood pressure, the second differential component of heart rate, the second differential component of IL-1β, the second differential component of IFN-γ, the second differential component of TNF-α, the second differential component of IL-6, and the second differential component of IL-8.

[0014] As a further aspect of the present invention, the dimension-uppering operators of the first, second, third, and fourth dimension-uppering branches are all composed of three fully connected linear transformations and nonlinear activation functions connected in series, namely, a first fully connected layer, a first activation layer, a second fully connected layer, a second activation layer, a third fully connected layer, and a third activation layer; wherein, the input of the first fully connected layer includes a training content vector R of progressive muscle relaxation training; the input of the second fully connected layer includes a playback content vector S of music therapy; the third fully connected layer performs adaptive adjustment on the output of the second fully connected layer, the adaptive adjustment being: generating a gating coefficient vector G based on the patient identifier, time point identifier, and the first and second difference components, and performing a linear transformation on the output vector of the second fully connected layer dimension by dimension by the gating coefficient vector G to obtain the corresponding dimension-uppering vectors U1, U2, V1, or V2.

[0015] It should be noted that, compared to existing technologies that primarily use time-point t-tests and rank-sum tests for anxiety and depression analysis in hematopoietic stem cell transplant patients, or directly input original covariates such as GAD-7, PHQ-9, blood pressure, heart rate, and cytokines, along with a few artificial interaction terms into generalized linear mixed models, and treat progressive muscle relaxation training and music therapy merely as group labels / intervention status without detailing the intervention prescription, this invention constructs a matrix of postoperative comprehensive assessment indicators for hematopoietic stem cell transplant patients undergoing progressive muscle relaxation training and music therapy, based on psychological assessment, vital signs, and inflammatory cytokine levels, and establishes T1-T2 and T2-T3 difference components. Different modal inputs are then distributed to four dimensionality-upgrading branches for structured dimensionality enhancement. The training content vector R and the playback content vector S are injected into the dimensionality-upgrading operator as computable prescription features. In the third fully connected layer, a gating coefficient vector G generated by patient identification, time point identification and difference components is introduced to adaptively modulate the feature expression. Then, low-redundancy features are obtained by principal component projection and threshold pruning before being input into the hybrid model. This solves the problems of unstable estimation caused by direct input of multiple indicators with strong collinearity, skewness and batch drift, individual differences and time correlation, the inability to input intervention dosage and content differences into the model, and the difficulty in uniformly characterizing subjective and objective coupling and time series changes. While retaining prescription information and time change information, it suppresses redundant noise, improves the fitting stability and trend recognition consistency of the hybrid model, and enhances the usability of intervention-related trajectory differences and individualized prediction.

[0016] As a further aspect of the present invention, the training content vector R is generated from the muscle group number sequence, the duration of tension maintenance for a single muscle group, the duration of relaxation for a single muscle group, respiratory rhythm parameters, the total duration of a single training session, and the number of training sessions. During generation, the muscle group number sequence is first encoded into a muscle group embedding vector. The tension duty cycle is then obtained by comparing the duration of tension in a single muscle group with the duration of relaxation in a single muscle group. With segmental period Respiratory rate is obtained from respiratory rhythm parameters. Compared to inhalation and exhalation The training dose is obtained from the total duration of a single training session and the number of training sessions. and will , , , , , The training content vector R is obtained by concatenating and normalizing the data in a preset order.

[0017] Compared to existing technologies that record progressive muscle relaxation training using only coarse-grained statistics such as whether it was implemented and the number of times / duration, or rely on textual descriptions in nursing records that are difficult to incorporate into the model as calculable variables, this invention parameterizes the training prescription into a training content vector R: encoding the muscle group number sequence to obtain the muscle group embedding vector. Furthermore, the tension duty cycle is derived from the duration of tension and relaxation. With segmental period Extracting respiratory rate from respiratory rhythm parameters Compared to inhalation and exhalation The training dose is calculated from the total duration of a single training session and the number of training sessions. Then , , , , , By splicing and normalizing the data in a fixed order to form a vectorized input with a unified caliber, the problem of discrete elements, inconsistent dimensions, and difficulty in aligning them to T1 / T2 / T3 time points and jointly modeling them with psychological assessment, vital signs, and inflammatory cytokine levels is solved. This allows training content, rhythm, and dosage to be stably input into machine learning and hybrid models with structured features, improving the comparability and modeling usability of data differences in training protocols for different patients, and reducing feature drift and noise amplification caused by differences in recording caliber.

[0018] As a further aspect of the present invention, the playback content vector S is generated from the track number, beat parameter BPM, playback volume, frequency band parameter, single playback duration, and number of playbacks. During generation, the track number is first encoded into a track embedding vector. The rhythm intensity is then obtained from the beat parameter BPM and the playback volume. The spectrum vector is obtained from the frequency band parameters. The music dosage is obtained from the duration of a single playback and the number of playbacks. and will , , , The playback content vector S is obtained by concatenating and normalizing the data in a preset order.

[0019] Compared to existing technologies that often use vague records of music listening, playback duration, or frequency, and treat differences in music tracks as entirely equivalent to acoustic properties, failing to reflect the differentiated effects of music stimulation on emotional arousal and attentional engagement, this invention addresses the perceptible stimulus characteristics of music therapy by providing a structured expression of playback prescriptions: mapping track numbers to track embedding vectors. 's' represents the differences in track category and style, and the beat parameter BPM is combined with the playback volume to form rhythm intensity. To characterize the combined load of rhythmic drive and loudness stimulus, frequency band parameters are extracted into spectral vectors. To characterize the energy distribution characteristics of low, medium and high frequencies, and to calculate the music dose based on the duration of a single playback and the number of playbacks. To characterize the intensity of exposure, ultimately , , , By concatenating and normalizing the playback content vector S in a fixed order, this method solves the problem in music therapy where different stimuli attributes during playback lead to incomparable intervention variables and the inability to establish calculable correlations with psychological assessments, vital signs, and inflammatory cytokine levels at T1 / T2 / T3 time points. This allows the style, rhythm, loudness, and spectral differences of music stimuli to be quantified and reused in the input model, thereby improving the precision of prescription difference characterization and the consistency of cross-patient data alignment.

[0020] As a further aspect of the present invention, the time point interaction term is an interaction term between group and time point. The group is represented by a binary indicator variable X, where X=0 represents the control group and X=1 represents the experimental group. The time point is represented by indicator variables T2 and T3, where T2=1 indicates that the current observation is in the bone marrow suppression period, otherwise it is 0, and T3=1 indicates that the current observation is in the bone marrow reconstitution period, otherwise it is 0. T1 is used as the reference time point. The time point interaction term includes two items, X×T2 and G×T3, and is used as a fixed-effect input to the generalized linear mixture model.

[0021] It should be noted that by constructing interaction terms X×T2 and X×T3 by multiplying the binary indicator variable X of the group with the indicator variables T2 of the bone marrow suppression period and T3 of the bone marrow reconstruction period, and then using X, T2, and T3 together as fixed-effect inputs, a generalized linear mixture model with patient identification as the random intercept is constructed. This solves the problem that existing models, when only group main effects or time main effects are set, treat the intervention effect as constant, and cannot characterize the stage shifts and changes of the experimental group relative to the control group at T2 and T3. This leads to the averaged trend of indicators at each stage and estimation bias caused by interference from individual baseline heterogeneity. Thus, under a unified standard with T1 as the reference, the inter-group difference parameters of T2 and T3 are output separately, making the detection of stage differences more sensitive, the interpretation of trends clearer, and the prediction more stable.

[0022] As a further aspect of the present invention, the preset selection rules and the gating coefficient vector G are determined in the following manner: Time stability indices J1 and J2 are calculated for the first upgraded branch output U1 and the second upgraded branch output U2, respectively. These time stability indices are constituted by the weighted sum of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the training dose D in the training content vector R. When J1 is not less than J2, U=U1 is selected; otherwise, U=U2 is selected. Time stability indices J3 and J4 are calculated for the third upgraded branch output V1 and the fourth upgraded branch output V2, respectively. These time stability indices are constituted by the weighted sum of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the music dose in the playback content vector S. Determine; when J3 is not less than J4, select V=V1; otherwise, select V=V2; the gating coefficient vector G is generated by the difference driving vector and the prescription driving vector. The difference driving vector is obtained by concatenating the first and second difference components of matrix A and matrix B in a preset order. The prescription driving vector C is obtained by concatenating the training content vector R and the playback content vector S in a preset order. The gating coefficient vector G is obtained by inputting the linear combination of the difference driving vector and the prescription driving vector into a nonlinear activation function. The dimension of G is consistent with the dimension of the output vector of the second fully connected layer and is used for dimension-wise multiplication gating.

[0023] It should be noted that, compared to existing technologies that typically employ fixed branch splicing, manual experience-based path selection, or offline selection solely based on loss minimization for multi-branch feature extraction, and whose gating often relies on a single input or is generated only from the patient's static baseline, making it difficult to reflect the temporal stability and intervention prescription differences during the transplantation period (T1, T2, T3), this invention uses time stability index to drive branch selection. It calculates pairwise cosine similarity weighted sums of the output vectors of U1, U2 and V1, V2 at T1, T2, and T3 to form J1 to J4, and uses the training dose D and music dose Q as weights to determine the branch outputs with higher consistency across time points as U and V. Simultaneously, it uses differential driving vectors and prescriptions... The driving vector C generates a gating coefficient vector G, which is simultaneously constrained by the magnitude of index changes (first difference component, second difference component) and intervention prescription parameters (R, S). The output of the second fully connected layer is gated and modulated using a dimension-aligned, dimension-wise multiplication method. This solves the problems of unstable branch output and non-reproducible selection in multimodal, multi-time point data under strong noise, collinearity, and individual differences, as well as the dilution of inter-group and stage differences due to the inability of intervention dosage and content differences to enter feature expression. This makes branch selection computable, cross-time point expression more consistent, and gating more adaptive to prescriptions and changes, thereby improving the robustness of the pruned feature W and the consistency of subsequent hybrid model parameter estimation and prediction.

[0024] Compared with the prior art, the technical effects of the present invention are as follows:

[0025] This invention constructs a matrix M, which consists of comprehensive postoperative evaluation indicators for hematopoietic stem cell transplant patients at T1, T2, and T3 stages, by column segmentation and difference. Candidate features are generated using an upgraded representation of two branches for each of the two groups. Branches are then selected across groups according to a selection rule, followed by principal component projection and threshold pruning to obtain a low-redundancy pruned feature vector W. W, along with group, time point, and interaction terms, is then used as input to a hybrid model. This achieves a stable and compressed expression of subjective and objective indicators and their temporal changes, even under conditions of small sample size, strong collinearity, skewness, batch drift, and significant individual baseline heterogeneity. It reduces the interference of redundant noise on parameter estimation and improves the identifiability and fitting stability of group and stage difference trends, resulting in more consistent model parameters and individualized trajectory prediction results. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention;

[0027] Figure 2 This is a model network topology diagram of the present invention. Detailed Implementation

[0028] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described technical solutions are only a part of this invention, not all of it. Based on the content of this invention, all other technical solutions obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0029] like Figure 1 As shown, the present invention provides a machine learning-based method for analyzing postoperative indicators in hematopoietic stem cell transplant patients, comprising:

[0030] For hematopoietic stem cell transplant patients who underwent progressive muscle relaxation training combined with music therapy, postoperative comprehensive assessment indicators at T1, T2, and T3 were collected to form matrix M, and missing markers, abnormal removal, standardization, and batch correction were performed.

[0031] Matrix M is split into matrices A and B by column segmentation, and the first and second difference components are calculated for each index. Matrix A, matrix B and their first and second difference components are used as input to the machine learning model. The machine learning model includes the first and second dimension-up branches for matrix A, and the third and fourth dimension-up branches for matrix B. The first, second, third and fourth dimension-up branches output U1, U2, V1 and V2 respectively. U is obtained by selecting one of U1 and U2 according to the preset selection rules, and V is obtained by selecting one of V1 and V2. U and V are concatenated to obtain the feature matrix X. Principal component projection is performed on X, and the first K dimensions are taken and the dimensions with absolute values ​​less than the threshold τ are deleted to obtain the pruned feature vector W.

[0032] The pruned feature vector W, along with the group, time point, and time point interaction term, is used as the input to a generalized linear mixture model. The patient identifier is used as the random intercept, and the model is fitted with anxiety assessment and depression assessment as dependent variables, respectively. The output model parameters and prediction results are then presented.

[0033] In practice, firstly, a patient ID and time point markers (T1, T2, T3) are established for each hematopoietic stem cell transplant patient undergoing progressive muscle relaxation training combined with music therapy. Postoperative comprehensive assessment indicators are collected on the day of admission (T1), during the bone marrow suppression period (T2), and during the bone marrow reconstitution period (T3), and written into a data table. These indicators include GAD-7, PHQ-9, systolic blood pressure, diastolic blood pressure, heart rate, and the levels of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8. A matrix M is then formed using patient-time point as the row index and indicator fields as the column index. Subsequently, M is processed according to preset rules, including missing marker and missing value handling, outlier removal or truncation, standardization by field dimensions, and batch correction for cytokine detection batches. Finally, the data is segmented by column to obtain... Matrix A (psychological assessment and vital signs) and matrix B (inflammatory cytokine levels) are used, and the first difference component Δ12 and the second difference component Δ23 are calculated for each field. A, B, and Δ12 and Δ23 are input into a machine learning model. The model sets two dimensionality-increasing branches for A, outputting U1 and U2, and sets two dimensionality-increasing branches for B, outputting V1 and V2. U and V are determined according to preset selection rules and concatenated to form a feature matrix X. Then, principal component projection is performed on X to obtain the first K dimensions, and dimensions with |Z| less than the threshold τ are deleted to form a pruned feature vector W. Finally, W, along with the group, time point, and their interaction terms, are input into a generalized linear mixture model. The patient ID is set as the random intercept, and the model is fitted with GAD-7 and PHQ-9 as dependent variables, respectively, and the fixed-effects parameters, random-effects parameters, and corresponding predicted values ​​are output.

[0034] It should be noted that the postoperative comprehensive assessment indicators include psychological assessment, vital signs, and inflammatory cytokine levels. The psychological assessment includes the scores of the 7-item Generalized Anxiety Disorder (GAD-7) and the 9-item Patient Health Questionnaire (PHQ-9). The GAD-7 score is obtained by summing the scores of seven items—tension and anxiety, uncontrollable worry, excessive worry, difficulty relaxing, restlessness, irritability, and ominous premonitions—based on preset frequencies within the past two weeks. The PHQ-9 score is obtained by summing the scores of nine items—loss of interest, depressed mood, sleep disturbances, fatigue, abnormal appetite, self-blame or low self-esteem, decreased attention, psychomotor retardation or agitation, and self-harming ideas—based on preset frequencies within the past two weeks. Vital signs include systolic blood pressure, diastolic blood pressure, and heart rate. Inflammatory cytokine levels include the measured values ​​of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8.

[0035] In practice, on the day of admission (T1), during the bone marrow suppression period (T2), and during the bone marrow reconstitution period (T3), nursing staff collected psychological assessment data within the same observation window. Using a pre-set four-level frequency scale, seven items—tension and anxiety, uncontrollable worry, excessive worry, difficulty relaxing, restlessness, irritability, and premonition of misfortune—were scored and summed to obtain the GAD-7 score. Simultaneously, nine items—loss of interest, depressed mood, sleep disturbances, fatigue, abnormal appetite, self-blame or low self-esteem, decreased attention, psychomotor retardation or agitation, and self-harming thoughts—were scored and summed to obtain the PHQ-9 score. At the same time, systolic and diastolic blood pressure were measured at rest using bedside monitoring or a standard cuff blood pressure monitor, and heart rate was recorded as vital signs. At each time point, 2 ml of venous blood was collected, and the concentrations of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8 were detected using a flow cytometry microsphere capture chip as inflammatory cytokine levels. Finally, these indicators were recorded in a data table according to patient and time point identifiers.

[0036] It should be noted that the first difference component is the value of the same indicator at time T2 minus the value at time T1, and the second difference component is the value of the same indicator at time T3 minus the value at time T2; the inputs of the first and second dimensionality-up branches include psychological assessment and vital sign quantities, respectively; the inputs of the third and fourth dimensionality-up branches include the level and difference components of inflammatory cytokine quantities, respectively; and the inputs of all dimensionality-up branches also include the training content vector R of progressive muscle relaxation training and the playback content vector S of music therapy; K is the principal component dimension, and τ is the pruning threshold.

[0037] In specific implementation, after the comprehensive postoperative assessment indicators at three time points (T1, T2, and T3) are entered into the database, the invention calculates the first difference component Δ12 = indicator (T2) − indicator (T1) and the second difference component Δ23 = indicator (T3) − indicator (T2) for each indicator, forming a difference field. Subsequently, the psychological assessment quantity and its Δ12 and Δ23, along with the training content vector R and the playback content vector S, are input into the first dimension-upgrading branch. The vital signs quantity and its Δ12 and Δ23, along with R and S, are input into the second dimension-upgrading branch. The inflammatory cytokine level quantity, along with R and S, is input into the third dimension-upgrading branch. The inflammatory cytokine difference components Δ12 and Δ23, along with R and S, are input into the fourth dimension-upgrading branch to obtain U1, U2, V1, and V2, respectively. In the machine learning model, principal component projection is performed on the subsequent spliced ​​features according to the preset principal component dimension K, and the projection vector is filtered out dimension by dimension using a threshold τ to form pruned features for subsequent modeling.

[0038] In practice, the preset principal component dimension K and threshold τ are set according to data structure constraints, modeling stability constraints, and task caliber constraints: K is preferentially determined based on the eigenvalue spectrum of the training set covariance matrix, so that the cumulative explained variance of the first K principal components reaches a preset proportion (e.g., 0.85 to 0.95) and the corresponding eigenvalues ​​are not less than a preset lower limit, while simultaneously satisfying the sample size constraint so that K does not exceed a preset proportion of the effective sample size to avoid overparameterization of the mixture model; when prediction is the objective, K is determined by minimizing the log-likelihood or AIC / BIC under cross-validation; the threshold τ is determined based on the noise floor and scale of the projection vector Z in the training set, and can be taken as the quantile threshold of the absolute value distribution of each dimension of Z or as a multiple of the scale estimate (e.g., absolute deviation of the median), and the frequency of the retained dimension under different sampling is tested through self-sampling to ensure that it reaches the preset stability, thereby controlling the number of pruned dimensions within a range that can be stably fitted.

[0039] It should be noted that the first differential component includes the first differential component of GAD-7, the first differential component of PHQ-9, the first differential component of systolic blood pressure, the first differential component of diastolic blood pressure, the first differential component of heart rate, the first differential component of IL-1β, the first differential component of IFN-γ, the first differential component of TNF-α, the first differential component of IL-6, and the first differential component of IL-8.

[0040] It should be noted that the second difference includes the second difference of GAD-7, the second difference of PHQ-9, the second difference of systolic blood pressure, the second difference of diastolic blood pressure, the second difference of heart rate, the second difference of IL-1β, the second difference of IFN-γ, the second difference of TNF-α, the second difference of IL-6, and the second difference of IL-8.

[0041] In specific implementation, after collecting and storing the GAD-7 score, PHQ-9 score, systolic blood pressure, diastolic blood pressure, heart rate, and the detection values ​​of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8 at three time points (T1, T2, and T3), this invention generates difference fields item by item according to patient identification and indicator fields: First, the first difference component is calculated for each indicator, specifically: GAD-7 first difference component = GAD-7(T2)−GAD-7(T1), PHQ-9 first difference component = PHQ-9(T2)−PHQ-9(T1), systolic blood pressure first difference component = systolic blood pressure (T2)−systolic blood pressure (T1), diastolic blood pressure first difference component = diastolic blood pressure (T2)−diastolic blood pressure (T1), heart rate first difference component = heart rate (T2)−heart rate (T1), and IL-1β first difference component = I L-1β(T2)−IL-1β(T1), IFN-γ first difference component = IFN-γ(T2)−IFN-γ(T1), TNF-α first difference component = TNF-α(T2)−TNF-α(T1), IL-6 first difference component = IL-6(T2)−IL-6(T1), IL-8 first difference component = IL-8(T2)−IL-8(T1); then calculate the second difference component for the same index, specifically calculate the second difference component of GAD-7 = GAD-7(T3)−GAD-7(T2) until the second difference component of IL-8 = IL-8(T3)−IL-8(T2); write the above first difference component and second difference component into the difference matrix according to the unified field naming rule and align it with the original index matrix by row index, for subsequent dimensionality increase branch input and gating generation.

[0042] It should be noted that the dimension-upgrading operators of the first, second, third, and fourth dimension-upgrading branches are all composed of three fully connected linear transformations and nonlinear activation functions cascaded together, namely, the first fully connected layer, the first activation layer, the second fully connected layer, the second activation layer, the third fully connected layer, and the third activation layer, respectively. The input of the first fully connected layer includes the training content vector R of progressive muscle relaxation training; the input of the second fully connected layer includes the playback content vector S of music therapy; the third fully connected layer performs adaptive adjustment on the output of the second fully connected layer. This adaptive adjustment involves generating a gating coefficient vector G based on the patient identifier, time point identifier, and the first and second difference components, and then performing a linear transformation on the output vector of the second fully connected layer after multiplying it dimension by the gating coefficient vector G to obtain the corresponding dimension-upgrading vectors U1, U2, V1, or V2.

[0043] In practice, the machine learning model treats each dimensionality-upgrading branch as a computationally compliant pipeline with the same three fully connected layers, two activations, and a gated third layer. The only difference is that the four branches receive different underlying metrics (psychological / physical / cytokine levels / cytokine differences). The implementation steps are as follows:

[0044] 1) Prepare branch input vectors

[0045] For each patient at each time point t∈{T1,T2,T3}, first form the basic input vector x(t) for that branch:

[0046] First dimensional branch: x(t) = psychological assessment quantity and its Δ12 and Δ23 (concatenated in field order);

[0047] Second dimensional branch: x(t) = vital signs and their Δ12 and Δ23;

[0048] Third dimension-upgrading branch: x(t) = cytokine level;

[0049] Fourth dimension-upgrading branch: x(t) = cytokine difference components Δ12, Δ23;

[0050] The training content vector R and the playback content vector S are then incorporated into the input of subsequent layers at preset positions.

[0051] 2) Inject R (corresponding to relaxation training content) into the first fully connected layer.

[0052] Execute on each branch at time t:

[0053] The first layer input h0(t) = [x(t), R] is formed by concatenation.

[0054] Output of the first fully connected layer: a1(t) = W1·h0(t) + b1;

[0055] First activation layer: z1(t)=φ1(a1(t)).

[0056] Here, R enters the first layer in vector form, allowing the network to see the structure and dosage information of the training prescription when increasing the dimensionality.

[0057] 3) Second-layer fully connected injection of S (corresponding to music prescription)

[0058] The second layer input is formed as h1(t) = [z1(t), S].

[0059] The output of the second fully connected layer is: a2(t) = W2·h1(t) + b2;

[0060] Second activation layer: z2(t)=φ2(a2(t)).

[0061] The reason for incorporating S into the second layer is to allow the rhythm, spectrum, and dosage of the music's stimulus characteristics to participate in the combination in the intermediate representation layer.

[0062] 4) Generate the gating coefficient vector G (the source of adaptive modulation)

[0063] For each patient and at each time point t, construct the gated input g0(t):

[0064] g0(t)=[ID_embed,time_embed,Δ12,Δ23], where ID_embed is the patient identifier embedding, time_embed is the time point embedding; Δ12 and Δ23 are the first and second difference vectors of each indicator of the patient (concatenated in a preset order).

[0065] Then calculate the gating coefficient:

[0066] G(t) = σ(Wg·g0(t) + bg), where σ is either Sigmoid or Hard-Sigmoid, ensuring that G(t) falls on the interval (0,1) or [0,1] dimension by dimension.

[0067] And constrain the dimension of G(t) to be consistent with the dimension of z2(t) (if it is insufficient, it is filled by linear mapping; if it is excessive, it is compressed by projection).

[0068] 5) Perform a gating linear transformation on the third fully connected layer (to obtain U1, U2, V1, V2).

[0069] First, perform element-wise gating: z2'(t) = z2(t)⊙G(t) (⊙ represents element-wise multiplication);

[0070] The output of the third fully connected layer is: a3(t) = W3·z2'(t) + b3;

[0071] Third activation layer: U(t)=φ3(a3(t)).

[0072] U(t) is the upgraded output of this branch at time t; the four branches are denoted as U1(t), U2(t), V1(t), and V2(t).

[0073] 6) Key points for practical implementation during training and inference

[0074] The four branches have the same structure but independent parameters (the branches W1, W2, W3 and Wg can share or not share; usually, it is more stable for U1 and U2 to share a set of gating networks, and it is more stable for V1 and V2 to share a set of gating networks).

[0075] The gating inputs Δ12 and Δ23 enable G to adaptively shrink and amplify the intermediate representation as the index changes, time_embed makes it sensitive to the stage differences of T2 and T3, and ID_embed ensures individualized modulation.

[0076] During the reasoning phase, U1, U2, V1, and V2 can be obtained by performing forward calculations for each patient at each time point according to the above pipeline, which can then be used for subsequent selection, splicing, PCA, and pruning steps.

[0077] It should be noted that the training content vector R is generated from the muscle group number sequence, the duration of tension for a single muscle group, the duration of relaxation for a single muscle group, respiratory rhythm parameters, the total duration of a single training session, and the number of training sessions. During generation, the muscle group number sequence is first encoded into a muscle group embedding vector. The tension duty cycle is then obtained by comparing the duration of tension in a single muscle group with the duration of relaxation in a single muscle group. With segmental period Respiratory rate is obtained from respiratory rhythm parameters. Compared to inhalation and exhalation The training dose is obtained from the total duration of a single training session and the number of training sessions. and will , , , , , The training content vector R is obtained by concatenating and normalizing the data in a preset order.

[0078] In practice, the training content vector R is automatically generated from the training prescription record table and execution record table, and archived according to patient identification and training date (or observation window corresponding to T1, T2, and T3 time points). The implementation process is as follows:

[0079] 1) Collect training prescriptions and execution parameters

[0080] Before each progressive muscle relaxation training session begins, the following information is entered and confirmed by the system or nursing staff: muscle group number sequence (muscle group codes arranged in the training order), duration of tension and relaxation for each muscle group, respiratory rhythm parameters (including at least respiratory rate and inspiratory-expiratory rhythm ratio), total duration of a single training session, and number of training sessions for the day; missing items are filled in or marked as missing according to preset rules.

[0081] 2) Generate muscle group embedding vectors

[0082] A muscle group encoding dictionary is established, mapping the muscle group number sequence to an index sequence. For each muscle group index, the corresponding embedding vector is obtained by looking up the table and concatenating them according to the training order. Then, a fixed-length muscle group embedding vector is generated by pooling. Pooling is a stepwise averaging or summing operation of the sequence embedding vectors, and a unified dimension is achieved by a preset linear mapping when the dimension of the embedding vectors is insufficient or too long.

[0083] 3) Calculate the duty cycle With segmental period

[0084] Duration of tension on each muscle group With relaxation time Calculate the tension duty cycle of this muscle group. With segmental period When multiple muscle groups are present, for and The global average can be obtained by averaging the training results in either the training order or by weighting the duration of each muscle group. With the global .

[0085] 4) Extract respiratory rate Compared to inhalation and exhalation

[0086] Read or calculate respiratory rate from respiratory rhythm parameters. (Number of breaths per minute or number of breaths per second), and calculate the inspiratory-to-expiratory ratio. =Inhalation duration / Exhalation duration; When the respiratory rhythm is given in beats or guided speech rhythms, convert the beats into respiratory cycles and calculate accordingly. and .

[0087] 5) Calculate training dose

[0088] Based on the total duration of a single training session With training times Calculate training dose When there are multiple training periods, the dose parameters at the corresponding time points are obtained by summing or averaging the values ​​of D according to the observation windows of T1, T2, and T3.

[0089] 6) Concatenate and normalize to obtain the training content vector R

[0090] According to the preset field order , , , , , Concatenate into a vector ;right Normalization is performed dimension by dimension. The normalization method is to linearly scale the minimum-maximum range obtained from the training set statistics or to standardize the mean and standard deviation. The normalized result is recorded as the training content vector R. R, along with the patient identifier and time window identifier, is written into the feature library for subsequent dimension-up branches.

[0091] It should be noted that the playback content vector S is generated from the track number, beat parameter (BPM), playback volume, frequency band parameter, single playback duration, and number of playbacks. During generation, the track number is first encoded into a track embedding vector. The rhythm intensity is then obtained from the beat parameter BPM and the playback volume. The spectrum vector is obtained from the frequency band parameters. The music dosage is obtained from the duration of a single playback and the number of playbacks. and will , , , The playback content vector S is obtained by concatenating and normalizing the data in a preset order.

[0092] In practice, the playback content vector S is automatically generated from the music therapy playlist combined with the playback log, and summarized according to patient identification and corresponding observation window (time window matching T1, T2, and T3). The implementation process is as follows:

[0093] 1) Collect and play prescriptions and play logs

[0094] Before or during each music therapy session, record the track number (library ID or hash ID), beat parameter (BPM) (from library metadata or offline analysis results), playback volume (device output level or volume level conversion value), frequency band parameters (track spectral energy distribution or pre-labeled frequency band characteristics), and duration of each playback session. With the number of plays If a session contains multiple tracks, the tracks are arranged in the order they are played and the corresponding parameters for each track are saved.

[0095] 2) Generate track embedding vectors

[0096] A song library number dictionary is established, mapping song numbers to indices. For each song index, a lookup table is used to obtain the corresponding embedding vector. Then, the embedding vectors of songs within the same session are weighted and summed or mean-squared according to playback duration to obtain a fixed-length song embedding vector. When the track sequences have inconsistent lengths, the above-mentioned weighted pooling method ensures... The dimension is fixed.

[0097] 3) Calculate rhythm intensity

[0098] Read the beat parameter BPM and playback volume for each track or session. The volume levels are mapped to continuous values ​​according to preset conversion rules, and then the rhythm intensity is calculated. If there are multiple tracks in the session, the playback duration will be calculated based on the length of each track. The rhythm intensity of the session is obtained by performing a weighted average.

[0099] 4) Extract the spectrum vector

[0100] The energy percentage or energy integral value of each frequency band is obtained from the frequency band parameters, and a fixed-length vector is formed according to the preset frequency band division. The preset frequency band division includes at least three frequency bands: low frequency, mid frequency, and high frequency, resulting in corresponding spectrum vectors. If the frequency band parameters are given in a finer granularity, they are converged to a fixed dimension according to a preset mapping matrix. .

[0101] 5) Calculate music dosage

[0102] Based on the duration of a single playback With the number of plays Calculate music dosage When multiple playback sessions exist within the same view window, for The total music dose for the observation window can be obtained by accumulating the data from each session, or by averaging the data according to preset rules to obtain the dose parameters corresponding to each time point.

[0103] 6) Concatenate and normalize to obtain the playback content vector S

[0104] According to the preset field order , , , Concatenate into a vector ;right Normalization is performed dimension by dimension. The normalization method is to linearly scale the minimum-maximum range obtained from the training set statistics or to standardize it according to the mean and standard deviation. The normalized result is recorded as the playback content vector S. S, along with the patient identifier and time window identifier, is written into the feature library for subsequent input of the dimension-up branch and gating generation.

[0105] It should be noted that the time point interaction term is the interaction term between group and time point. The group is represented by a binary indicator variable X, where X=0 represents the control group and X=1 represents the experimental group. The time point is represented by indicator variables T2 and T3, where T2=1 indicates that the current observation is in the bone marrow suppression period, otherwise it is 0, and T3=1 indicates that the current observation is in the bone marrow reconstitution period, otherwise it is 0. T1 is used as the reference time point. The time point interaction term includes two terms, X×T2 and G×T3, and is used as a fixed-effects input to the generalized linear mixed model.

[0106] In specific implementation, this invention first writes a group indicator variable X for each observation record, assigning a value of 0 to the control group and a value of 1 to the experimental group; then, it generates time point indicator variables T2 and T3 based on the time point to which the record belongs. For records during the bone marrow suppression period, T2 is set to 1 and T3 to 0; for records during the bone marrow reconstitution period, T3 is set to 1 and T2 to 0; and for records on the day of admission (T1), T2 is set to 0 and T3 to 0 as a reference. Subsequently, interactive features X×T2 and X×T3 are generated by multiplying each item. X, T2, T3, X×T2, X×T3, and the pruned feature vector W are used together as fixed effects input to the generalized linear mixture model, and the patient identifier is used as the random intercept to complete the fitting.

[0107] It should be noted that the preset selection rules and the gating coefficient vector G are determined by the following generation method: Time stability indices J1 and J2 are calculated for the outputs U1 and U2 of the first and second dimensional branches, respectively. These indices are composed of the weighted sum of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the training dose D in the training content vector R. When J1 is not less than J2, U=U1 is selected; otherwise, U=U2 is selected. Time stability indices J3 and J4 are calculated for the outputs V1 and V2 of the third and fourth dimensional branches, respectively. These indices are composed of the weighted sum of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the music dose in the playback content vector S. Determine; when J3 is not less than J4, select V=V1; otherwise, select V=V2; the gating coefficient vector G is generated by the difference driving vector and the prescription driving vector. The difference driving vector is obtained by concatenating the first and second difference components of matrix A and matrix B in a preset order. The prescription driving vector C is obtained by concatenating the training content vector R and the playback content vector S in a preset order. The gating coefficient vector G is obtained by inputting the linear combination of the difference driving vector and the prescription driving vector into a nonlinear activation function. The dimension of G is consistent with the dimension of the output vector of the second fully connected layer and is used for dimension-wise multiplication gating.

[0108] In practical implementation, a machine learning model using a four-branch multilayer perceptron (MLP) with gated subnetworks can be adopted: After forward computation of four sets of upgraded outputs U1(Ti), U2(Ti), V1(Ti), and V2(Ti) for the same patient at T1, T2, and T3 respectively, the time stability index is first calculated. For example, for U1, J1 is defined as J1 = w12·cos(U1(T1), U1(T2)) + w13·cos(U1(T1), U1(T3)). J2 is obtained by applying cosine similarity to U1(T2), U1(T3)) + w23·cos(U1(T2), U1(T3)). The weights w12, w13, and w23 are obtained by mapping the training dose D. For example, D is normalized to α∈[0,1] according to the training set range, and w12=1−α, w13=α / 2, w23=α / 2 are set to reflect that a larger training dose emphasizes consistency across later time points. If J1≥J2, then U=U1 is selected; otherwise, U=U2 is selected. 2; J3 and J4 are calculated for V1 and V2 in the same way, but the weights are normalized to β by the music dose Q and w12=1−β, w13=β / 2, w23=β / 2 are set. If J3≥J4, V=V1 is selected, otherwise V=V2 is selected; then a gated input is constructed. The differential driving vector Δ is concatenated by the first and second differential components of matrix A and matrix B in the order of preset fields. The prescription driving vector C is concatenated by the training content vector R and the playback content vector S. The gated sub-network calculates G=σ(WΔ·Δ+WC·C+b), and σ is Sigmoid to ensure that G falls between 0 and 1 in each dimension and its dimension is consistent with the output z2 of the second fully connected layer. Finally, for each branch, the dimension-wise gate z2′=z2⊙G is performed on the output of the second activation layer and fed into the third fully connected layer to obtain the corresponding U1, U2, V1 or V2. Thus, the branch selection is determined by cross-time point consistency + dose weight and the feature expression is adaptively modulated by differential change + prescription information.

[0109] In practice, upon admission, the relaxation training combined with music therapy intervention program is initiated. Fixed times are set for mornings and nights, twice daily for 20-30 minutes each time, three times a week, continuing until the bone marrow reconstitution period. Before each training session, the nursing staff completes the relaxation preparation process, ensuring a quiet and comfortable environment. Patients are instructed to empty their bladder and bowels, wear loose clothing, and assume a supine or sitting position, gently closing their eyes and placing their hands on their abdomen or thighs. Before training begins, patients take 3-5 deep breaths and focus their attention on the target muscle group. During the training phase, guided by audio, the entire body's muscles are divided into 16 groups. Alternating tension and relaxation movements are performed sequentially: arms, head and face, neck, shoulders, chest, abdomen, back, buttocks, lower limbs, and feet. Tension is held for 10-15 seconds, and relaxation for 15-20 seconds, repeated 2-3 times for each muscle group. Patients are prompted to identify the feeling of tension during contraction and to experience the state of relaxation during relaxation. Simultaneously, selected five-element musical pieces in the angular mode are played. The audio-guided program consists of tracks from the "Traditional Chinese Five Elements Music (Zheng Mode)" in the Jiao mode and the Tianyun Five Elements Music in the Jiao mode. It can be played in rotation from a list of tracks such as "Eighteen Songs of a Nomad Flute," "Green Grass and Trees," "Spring Breeze," "Jiangnan Good," "Jiangnan Silk and Bamboo Music," "Rainbow Skirt Song," "Spring in the Red River," "Green Leaves Welcoming the Wind," "Spring Waltz," "One Grain of Soil, Ten Thousand Loads Harvest," "Zhuangzi's Dream of a Butterfly," "Liezi Riding the Wind," and "Walking the Street." The specific tracks are adaptively adjusted according to the patient's preferences and actual situation. Playback volume and track switching are based on the patient's subjective comfort. Track numbers, BPM, volume, frequency band parameters, single playback duration, and number of playbacks are recorded to generate the playback content vector S. Similarly, muscle group sequences, tension / relaxation duration, respiratory rhythm, total single training duration, and number of training sessions are recorded to generate the training content vector R. This R is then aligned with the psychological assessment, vital signs, and inflammatory cytokine levels collected at time points T1, T2, and T3 before being input into the machine learning and hybrid model analysis process of this invention.

[0110] like Figure 2As shown, the model network topology of the method proposed in this invention takes hematopoietic stem cell transplant patients undergoing progressive muscle relaxation training combined with music therapy as the target, and consists of five levels: intervention information injection, feature dimensionality increase modeling, optimal fusion, dimensionality reduction pruning, and statistical modeling. First, the postoperative comprehensive evaluation indicators collected from the patient at three time points (T1, T2, and T3) are aggregated into a matrix M. After missing data labeling, anomaly removal, standardization, and batch correction, it is divided into matrices A and B by columns, and the first and second order difference vectors of each indicator are calculated simultaneously. Subsequently, four parallel dimensionality increase branches are set in the machine learning model. The first and second dimensionality increase branches receive matrix A and its difference components, and the third and fourth dimensionality increase branches receive matrix B and its difference components. Each branch completes nonlinear dimensionality increase mapping under a unified three-layer fully connected, dual-activation, and gating structure, outputting U1, U2, V1, and V2 respectively. The intervention information is explicitly modulated by injecting the training content vector R and the playback content vector S. Subsequently, U is generated from U1 and U2 according to preset rules, and V is generated from V1 and V2. U and V are then concatenated to form a feature matrix X. Principal component projection is performed on X, and the first K dimensions are selected. Dimensions with absolute values ​​less than a threshold τ are pruned to obtain a compact and more discriminative feature vector W. Finally, the pruned feature vector W, along with the group, time point, and their interaction terms, is input into a generalized linear mixture model. Using patient identification as the random intercept, the model fits and predicts anxiety and depression assessments, thereby achieving joint modeling of subjective and objective indicators, individualized analysis, and interpretable output.

[0111] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0112] In conclusion, the above description is merely a specific implementation of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for analyzing postoperative indicators in hematopoietic stem cell transplant patients, characterized in that, include: For hematopoietic stem cell transplant patients who underwent progressive muscle relaxation training combined with music therapy, postoperative comprehensive assessment indicators at T1, T2, and T3 were collected to form matrix M, and missing markers, abnormal removal, standardization, and batch correction were performed. Matrix M is split into matrices A and B by column segmentation, and the first and second difference components are calculated for each index. Matrix A, matrix B and their first and second difference components are used as input to the machine learning model. The machine learning model includes the first and second dimension-up branches for matrix A, and the third and fourth dimension-up branches for matrix B. The first, second, third and fourth dimension-up branches output U1, U2, V1 and V2 respectively. U is obtained by selecting one of U1 and U2 according to the preset selection rules, and V is obtained by selecting one of V1 and V2. U and V are concatenated to obtain the feature matrix X. Principal component projection is performed on X, and the first K dimensions are taken and the dimensions with absolute values ​​less than the threshold τ are deleted to obtain the pruned feature vector W. The pruned feature vector W, along with the group, time point, and time point interaction term, is used as the input to a generalized linear mixture model. The patient identifier is used as the random intercept, and the model is fitted with anxiety assessment and depression assessment as dependent variables, respectively. The output model parameters and prediction results are then presented.

2. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 1, characterized in that, Postoperative comprehensive assessment indicators include psychological assessment, vital signs, and inflammatory cytokine levels. Psychological assessment includes the scores of the Generalized Anxiety Disorder-7 (GAD-7) and the Patient Health Questionnaire-9 (PHQ-9). The GAD-7 score is obtained by summing the scores of seven items—tension and anxiety, uncontrollable worry, excessive worry, difficulty relaxing, restlessness, irritability, and ominous premonitions—based on preset frequencies within the past two weeks. The PHQ-9 score is obtained by summing the scores of nine items—loss of interest, depressed mood, sleep disturbances, fatigue, abnormal appetite, self-blame or low self-esteem, decreased attention, psychomotor retardation or agitation, and self-harming ideas—based on preset frequencies within the past two weeks. Vital signs include systolic blood pressure, diastolic blood pressure, and heart rate. Inflammatory cytokine levels include the measured values ​​of IL-1β, IFN-γ, TNF-α, IL-6, and IL-8.

3. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 2, characterized in that, The first difference component is the value of the same indicator at time T2 minus the value at time T1, and the second difference component is the value of the same indicator at time T3 minus the value at time T2. The inputs of the first and second dimensionality-upgraded branches include psychological assessment and vital sign quantities, respectively. The inputs of the third and fourth dimensionality-upgraded branches include the level and difference components of inflammatory cytokine quantities, respectively. The inputs of all dimensionality-upgraded branches also include the training content vector R of progressive muscle relaxation training and the playback content vector S of music therapy. K is the principal component dimension, and τ is the pruning threshold.

4. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 3, characterized in that, The first differential components include the first differential component of GAD-7, the first differential component of PHQ-9, the first differential component of systolic blood pressure, the first differential component of diastolic blood pressure, the first differential component of heart rate, the first differential component of IL-1β, the first differential component of IFN-γ, the first differential component of TNF-α, the first differential component of IL-6, and the first differential component of IL-8.

5. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 3, characterized in that, The second difference components include the second difference components of GAD-7, PHQ-9, systolic blood pressure, diastolic blood pressure, heart rate, IL-1β, IFN-γ, TNF-α, IL-6, and IL-8.

6. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 3, characterized in that, The dimension-uppering operators of the first, second, third, and fourth dimension-uppering branches are all composed of three fully connected linear transformations and nonlinear activation functions cascaded together, namely, the first fully connected layer, the first activation layer, the second fully connected layer, the second activation layer, the third fully connected layer, and the third activation layer. The input of the first fully connected layer includes the training content vector R of progressive muscle relaxation training; the input of the second fully connected layer includes the playback content vector S of music therapy; the third fully connected layer performs adaptive adjustment on the output of the second fully connected layer. This adaptive adjustment involves generating a gating coefficient vector G based on the patient identifier, time point identifier, and the first and second difference components, and then performing a linear transformation on the output vector of the second fully connected layer after multiplying it dimension by the gating coefficient vector G to obtain the corresponding dimension-uppering vectors U1, U2, V1, or V2.

7. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 6, characterized in that, The training content vector R is generated from the muscle group number sequence, the duration of tension for a single muscle group, the duration of relaxation for a single muscle group, respiratory rhythm parameters, the total duration of a single training session, and the number of training sessions. During generation, the muscle group number sequence is first encoded into a muscle group embedding vector. The tension duty cycle is then obtained by comparing the duration of tension in a single muscle group with the duration of relaxation in a single muscle group. With segmental period Respiratory rate is obtained from respiratory rhythm parameters. Compared to inhalation and exhalation The training dose is obtained from the total duration of a single training session and the number of training sessions. and will , , , , , The training content vector R is obtained by concatenating and normalizing the data in a preset order.

8. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 6, characterized in that, The playback content vector S is generated from the track number, beat parameter (BPM), playback volume, frequency band parameter, single playback duration, and number of playbacks. During generation, the track number is first encoded into a track embedding vector. The rhythm intensity is then obtained from the beat parameter BPM and the playback volume. The spectrum vector is obtained from the frequency band parameters. The music dosage is obtained from the duration of a single playback and the number of playbacks. and will , , , The playback content vector S is obtained by concatenating and normalizing the data in a preset order.

9. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 1, characterized in that, The time point interaction term is the interaction term between group and time point. The group is represented by a binary indicator variable X, where X=0 represents the control group and X=1 represents the experimental group. The time point is represented by indicator variables T2 and T3, where T2=1 indicates that the current observation is in the bone marrow suppression period, otherwise it is 0, and T3=1 indicates that the current observation is in the bone marrow reconstitution period, otherwise it is 0. T1 is used as the reference time point. The time point interaction term includes two terms, X×T2 and G×T3, and is used as a fixed-effects input to the generalized linear mixed model.

10. The method for analyzing postoperative indicators of hematopoietic stem cell transplant patients based on machine learning according to claim 6, characterized in that, The preset selection rules and gating coefficient vector G are determined by the following generation method: Time stability indices J1 and J2 are calculated for the outputs U1 and U2 of the first and second dimensional branches, respectively. These indices are weighted sums of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the training dose D in the training content vector R. U=U1 is selected when J1 is not less than J2; otherwise, U=U2 is selected. Time stability indices J3 and J4 are calculated for the outputs V1 and V2 of the third and fourth dimensional branches, respectively. These indices are weighted sums of the pairwise cosine similarities of the output vectors of the corresponding branches at T1, T2, and T3, with the weights determined by the music dose D in the playback content vector S. Determine; when J3 is not less than J4, select V=V1; otherwise, select V=V2; the gating coefficient vector G is generated by the difference driving vector and the prescription driving vector. The difference driving vector is obtained by concatenating the first and second difference components of matrix A and matrix B in a preset order. The prescription driving vector C is obtained by concatenating the training content vector R and the playback content vector S in a preset order. The gating coefficient vector G is obtained by inputting the linear combination of the difference driving vector and the prescription driving vector into a nonlinear activation function. The dimension of G is consistent with the dimension of the output vector of the second fully connected layer and is used for dimension-wise multiplication gating.

Images