Stroke aphasia acupuncture rehabilitation decision-making method and system based on multi-modal data fusion

Abstract

This invention relates to the field of multimodal data fusion technology, and in particular provides a method and system for decision-making regarding acupuncture rehabilitation for post-stroke aphasia based on multimodal data fusion. The method includes collecting clinical scale data, functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and longitudinal relaxation time structural imaging data from post-stroke aphasia patients; preprocessing the image data to generate a standardized multimodal image feature set; extracting radiomics features based on the multimodal image feature set; identifying a set of key predictive features strongly correlated with acupuncture efficacy; inputting the selected key predictive features into various machine learning algorithms to train and construct an acupuncture efficacy prediction model; and forming an intelligent support model that can assist clinical decision-making, used to predict the patient's acupuncture rehabilitation effect and guide individualized treatment plans. This invention achieves quantitative prediction and personalized guidance for the acupuncture rehabilitation effect of post-stroke aphasia.

Description

Technical Field

[0001] This invention relates to the field of multimodal data fusion technology, and in particular to a decision-making method and system for acupuncture rehabilitation of post-stroke aphasia based on multimodal data fusion. Background Technology

[0002] Stroke primarily affects the brain, and post-stroke aphasia often results from imbalances in the Yin and Yang of the internal organs, disrupted Qi and blood flow, and disturbance of the brain's spirit. This leads to a loss of Qi and ultimately, tongue stiffness and speechlessness. Acupuncture aims to restore consciousness and open the orifices by harmonizing Yin and Yang and clearing the meridians. Currently, clinical diagnosis and treatment rely mainly on detailed medical history, progressive acupuncture procedures, and real-time patient feedback to assess acupuncture efficacy. Clinical research assesses efficacy differences primarily through scale evaluations, physical and chemical examinations, and neuroimaging assessments to continuously monitor and record patient differences after acupuncture treatment. Currently, there is no precise method for screening acupuncture efficacy differences and identifying the dominant population in PSA patients. Clinicians relying on personal experience to predict the efficacy of acupuncture intervention for post-stroke aphasia lack the visualization to guide clinical practice and widespread application. Therefore, it is necessary to establish a PSA acupuncture dominant population screening method that integrates clinical data assessment and key neuroimaging feature identification to address the impact of individual patient heterogeneity on the diagnosis, treatment, and prognosis of post-stroke aphasia.

[0003] The ultimate goal of clinical treatment for PSA is to maximize an individual's daily communication ability. The first three months after a stroke are the golden window for disability rehabilitation, while rehabilitation after six months is significantly less effective. Therefore, seizing the optimal window for language function rehabilitation and finding effective solutions to improve PSA outcomes is the first priority. Currently, the high cost of speech therapy is a major obstacle for most families to bear long-term. Acupuncture, a simple and inexpensive treatment, can benefit stroke patients with aphasia while significantly reducing the financial burden on families. Evaluating the effects of acupuncture is a crucial means of assessing its effectiveness; scale assessments can effectively identify the efficacy of acupuncture in improving PSA patients. Scale assessments provide tools for quantifying and standardizing clinical effects, quantifying disease progression, tracking patient rehabilitation, and evaluating efficacy differences, and the assessment system is constantly expanding. Therefore, extracting the trajectory characteristics of scale scores in acupuncture studies in the PSA population, and combining them with other clinical data and neuroimaging differential characterizations for joint analysis, can provide insights into disease treatment effects and prognostic outcomes, support clinical decision-making, and optimize treatment pathways. The individual heterogeneity of the damage-reorganization mechanisms in the brain language areas of PSA patients, such as lesion location / volume and the status of brain language function network reorganization and recovery, may be an important factor influencing the effect of acupuncture intervention on PSA. Neuroimaging technology is an objective tool for uncovering the individual differences in acupuncture efficacy and predicting efficacy. It can visually reveal the mechanisms and imaging characteristics of differences in acupuncture efficacy, clarify the reorganization of structure and function in various brain regions and the changes that occur with the progression of the disease during acupuncture intervention, and may provide a new entry point for in-depth research on the rehabilitation mechanisms of post-stroke aphasia. With the further development and application of neuroimaging technology, the prediction and individualized analysis of the effects of acupuncture in clinical treatment will become more accurate.

[0004] Functional magnetic resonance imaging (fMRI), with its non-invasive, radiation-free, non-traumatic, and precisely localizable characteristics, provides a visual approach for exploring the neuroimaging representations of the differential effects of acupuncture. fMRI can reveal the impact of changes in low-frequency fluctuation (ALFF), regional homogeneity (ReHo), and functional connectivity (FC) on language function. Furthermore, the normal performance of language function is influenced by the information input from the subcortical white matter reticular formation and the integrity of information exchange between subcortical structures. Diffusion tensor imaging (DTI) studies often use FA values ​​to reflect white matter integrity, considering it one of the most sensitive indicators of white matter damage. The magnitude of FA is closely related to myelin sheath integrity, fiber density, and parallelism, providing information about nerve fiber damage. Neuroimaging representations such as white matter fiber integrity are correlated with multiple language scores and have significant predictive value. Therefore, fMRI imaging technology can help uncover the specific correlation between the differentiated effects of acupuncture treatment and key neuroimaging predictive features such as structural damage and functional reorganization in the brain's language areas. Combined with individualized clinical information, it can also reveal the disease progression process, providing significant value for timely intervention and efficacy prediction. During the occurrence and recovery of PSA, acupuncture, while promoting language function improvement, is accompanied by changes in whole-brain functional network characteristics, decreased stability of dynamic network indicators, and increased variability. This includes stimulating dynamic functional reorganization, altering gray matter structural characteristics, and stimulating functional recovery of language pathway-related nodes in the subcortical region. Therefore, with the assistance of fMRI visualization, conducting precise research on efficacy differences and prognosis based on key predictive factors such as neuroimaging, clinical scale effects, and individualized information is a forward direction for integrating medical resources and promoting efficient diagnosis and treatment, and it also represents a crucial missing link in the current application of acupuncture in the treatment of PSA.

[0005] In the context of artificial intelligence and big data, machine learning is a crucial tool for analyzing and visualizing data. The 2024 Nobel Prize in Physics was awarded to Professors John J. Hopfield and Geoffrey E. Hinton for their groundbreaking work on machine learning based on artificial neural networks. Neural network theory and machine learning tools have significantly advanced research in neuroscience. Machine learning's advantages and potential in data integration are key AI technologies for governing and integrating acupuncture clinical data in the context of big data. This enables the computational accessibility of multi-dimensional clinical data and knowledge across various levels of acupuncture, supporting intelligent decision-making in acupuncture and ultimately providing crucial support for improving decision-making and computational capabilities. Deep learning, as a subset of machine learning, has significant advantages in tasks such as image, speech, and natural language processing. Using machine learning, radiomics can transform traditional medical images into high-throughput image features that can be mined. This allows for the quantitative description of spatial and temporal heterogeneity in images, revealing image features invisible to the naked eye. Effectively converting medical images into a high-dimensional, identifiable feature space, and performing statistical analysis on the generated feature space, allows for the establishment of models with diagnostic, prognostic, or predictive value, providing valuable information support for medical decision-making.

[0006] However, the application of artificial intelligence and machine learning in the field of PSA is currently lacking, and a large amount of data remains to be explored. Radiomics, in its data analysis capabilities, provides a crucial foundation for analyzing factors related to acupuncture effects and predicting the clinical efficacy of acupuncture intervention in PSA in the era of big data. Furthermore, machine learning research that correlates key representations such as clinical effects, personalized information, and neuroimaging can ensure the digital adaptability and efficiency of acupuncture clinical big data, providing stable, efficient, and high-quality decision support for diverse acupuncture application scenarios, and driving original knowledge transformation, theoretical iteration, and innovative applications in acupuncture in the digital age. Summary of the Invention

[0007] To achieve the above objectives, the present invention adopts the following technical solution: One aspect of the present invention provides a decision-making method for acupuncture rehabilitation of post-stroke aphasia based on multimodal data fusion, comprising the following steps: Clinical scale data, functional magnetic resonance imaging, diffusion tensor imaging, and longitudinal relaxation time structure image data of patients with post-stroke aphasia were collected. The image data were preprocessed by format conversion, head motion correction, time layer correction, spatial standardization, denoising, and covariate regression to generate a standardized multimodal image feature set. Based on multimodal image feature sets, radiomics features such as gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions are extracted. Combined with clinical data, principal component analysis and LASSO regression are used to perform feature dimensionality reduction and screening to identify key predictive feature sets that are strongly correlated with acupuncture efficacy. The selected key predictive features are input into various machine learning algorithms such as random forest, support vector machine, and logistic regression to train and build an acupuncture efficacy prediction model. The model performance is evaluated through cross-validation, ROC curve analysis, accuracy and sensitivity, forming an intelligent support model that can assist clinical decision-making, predict the acupuncture rehabilitation effect of patients and guide individualized treatment plans.

[0008] Another aspect of the present invention provides a decision-making system for acupuncture rehabilitation of post-stroke aphasia based on multimodal data fusion, comprising: The multimodal data acquisition and preprocessing module is used to acquire clinical scale data, functional magnetic resonance imaging, diffusion tensor imaging, and longitudinal relaxation time structure image data of patients with post-stroke aphasia; it performs preprocessing on the image data such as format conversion, head motion correction, time layer correction, spatial standardization, denoising, and covariate regression to generate a standardized multimodal image feature set. The radiomics feature extraction and screening module is used to extract radiomics features such as gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions based on multimodal image feature sets. Combined with clinical data, principal component analysis and LASSO regression are used to perform feature dimensionality reduction and screening to identify key predictive feature sets that are strongly correlated with acupuncture efficacy. The intelligent decision support model building module is used to input the selected key predictive features into various machine learning algorithms such as random forest, support vector machine and logistic regression to train and build an acupuncture efficacy prediction model. The model performance is evaluated through cross-validation, ROC curve analysis, accuracy and sensitivity and other indicators to form an intelligent support model that can assist clinical decision-making, predict the patient's acupuncture rehabilitation effect and guide individualized treatment plan.

[0009] This invention, based on the principles of identifying advantageous populations, recognizing specific characteristics, and constructing decision-making models, utilizes artificial intelligence technology to extract, integrate, and analyze information from images and structured data to predict the efficacy of acupuncture and assist in clinical treatment decisions for PSA patients. It proposes a multidimensional coupling relationship between neuroimaging representations such as the damage patterns in the brain's language areas and the reorganization status of the brain's language function network in PSA patients and the efficacy of acupuncture. Multimodal fMRI combined with machine learning can accurately analyze the brain injury-reorganization association mechanism and uncover its specific characteristics. By combining individual information of PSA patients, an efficacy prediction model is constructed, hypothesizing precise clinical decision support. Multimodal imaging technology and machine learning methods are applied to identify advantageous populations for PSA acupuncture efficacy and to uncover neuroimaging representations related to differentiated acupuncture efficacy in PSA patients. An innovative clinical intelligent decision support model for acupuncture intervention in PSA populations is constructed, integrating radiomics features and multidimensional clinical evaluation indicators to assess the treatment benefit ratio, comprehensively deepen the clinical efficacy of acupuncture intervention for PSA, guide patients' optimal golden recovery period treatment strategies, focus on scientific frontiers to promote precision medicine, and provide original theoretical support. Attached Figure Description

[0010] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the acupuncture rehabilitation decision-making method for post-stroke aphasia based on multimodal data fusion provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion provided in Embodiment 1 of the present invention. Figure 3 This is a process diagram of generating a standardized multimodal image feature set provided in Embodiment 3 of the present invention; Figure 4 This is a flowchart illustrating the process of identifying key predictive feature sets strongly correlated with acupuncture efficacy, as provided in Embodiment 7 of the present invention. Figure 5 This is a process diagram of forming an intelligent support model that can assist clinical decision-making, as provided in Embodiment 9 of the present invention; Figure 6 This is a block diagram of the acupuncture rehabilitation decision system for post-stroke aphasia based on multimodal data fusion provided in Embodiment 13 of the present invention; Figure 7 A block diagram of the electronic device provided by the present invention; Figure 8 The present invention provides an alpha error curve for each of the 10 folds and a schematic diagram of the optimal alpha; Figure 9 This is a schematic diagram of the characteristic coefficients as a function of alpha, provided by the present invention. Figure 10 A schematic diagram of the seven selected features and their coefficients provided by this invention; Figure 11 A schematic diagram of the ROC curve of a random forest provided by this invention; Figure 12 A schematic diagram of the ROC curve of the support vector machine provided for this invention; Figure 13 A schematic diagram of the logistic regression ROC curve provided by this invention; Reference numerals: 1. Multimodal data acquisition and preprocessing module; 2. Radiomics feature extraction and screening module; 3. Intelligent decision support model construction module; 4. Central processing unit / microprocessor / main control chip; 5. Storage medium; 6. Data bus; 7. Input / output bus / external bus / device bus; 8. Display; 9. Input / output device; 10. Computer-readable instructions; 11. Non-transitory computer-readable storage medium. Detailed Implementation

[0011] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0012] Hereinafter, the terms "first," "second," etc., are used for descriptive convenience only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "multiple" means two or more.

[0013] In this invention, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, a connection can be a fixed mechanical connection, a detachable mechanical connection, or an integral part; or, a connection can be a direct connection or an indirect connection through an intermediate medium. Furthermore, unless otherwise explicitly specified and limited, the term "coupling" should be interpreted broadly. For example, coupling can be a direct electrical connection, such as physical contact and electrical conduction between two components; it can also be understood as an electrical connection between different components in a circuit structure through physical lines capable of transmitting electrical signals, such as copper foil or wires on a printed circuit board (PCB), to transmit electrical signals; or, coupling can be an indirect electrical connection between two components through an intermediate medium; or, coupling can be an electrical connection between two components in a non-contact manner, such as an electrical connection between two components using capacitive coupling to transmit electrical signals.

[0014] In this embodiment of the invention, directional terms such as up, down, left, and right may be defined relative to the orientation of the components shown in the accompanying drawings. It should be understood that these directional terms can be relative concepts, used for relative description and clarification, and can change accordingly depending on the orientation of the components in the accompanying drawings.

[0015] Post-stroke aphasia (PSA) is a severe brain dysfunction following stroke. PSA patients have a 9.5% higher risk of in-hospital mortality and a 1.5-year mortality rate that is doubled compared to non-aphasia patients, significantly increasing the global burden of stroke-related death and disability, the difficulty of clinical rehabilitation, and the economic burden on healthcare. Currently, modern medical rehabilitation treatments are costly and not suitable for all patients, and there are no specific drug recommendations for drug therapy. In 2022, the *BMJ* listed PSA as a disease for which acupuncture is advantageous. A meta-analysis including 28 randomized controlled trials (RCTs) involving 1747 patients showed that acupuncture can effectively improve functional communication and language abilities in PSA patients. Previous multicenter randomized controlled trials (RCTs) by our research group have demonstrated that the "Awakening the Brain and Opening the Orifices" acupuncture method, established by Academician Shi Xuemin, is effective in improving spontaneous language, repetition ability, comprehension, naming, and reading scores. It has obtained the highest-level quality clinical evidence for acupuncture intervention in PSA and has been widely applied. However, clinical practice and research show that the acupuncture effect varies among some PSA patients. It should be noted that in addition to differences in acupuncture methods, individual patient heterogeneity is an unavoidable factor affecting efficacy. Clinical practice urgently needs to break through the empirical treatment paradigm and establish an intelligent efficacy prediction system based on objective characteristics. This system should analyze the population characteristics associated with clinical outcomes to increase benefits through precise intervention. Patients with advantageous outcomes should receive "Awakening the Brain and Opening the Orifices" acupuncture intervention as soon as possible, while those with disadvantaged outcomes should promptly combine it with other intervention methods or adjust individual factors affecting the acupuncture effect to ensure optimal rehabilitation. Therefore, individualized and precise identification of the language function outcome of acupuncture intervention in PSA is crucial for assisting clinical rehabilitation program decisions, seizing the golden period for PSA patients' recovery, and is also a major challenge currently faced in clinical practice. Based on the research group's previous clinical and neuroimaging studies on acupuncture intervention in PSA patients, this invention further establishes an acupuncture intervention PSA radiomics database. The aim is to screen patients with different PSA acupuncture efficacy, analyze the common characteristics of the dominant population, and, based on the analysis of the dual-pathway rehabilitation correlation mechanism of PSA brain language area damage-reorganization and the mining of related neuroimaging specific representations and prediction of acupuncture efficacy, establish a clinical intelligent decision support model to assist in early clinical decision-making in PSA patients, guide patients in the best golden period of rehabilitation treatment strategies, effectively reduce the family rehabilitation burden, and focus on the scientific frontier to promote precision medicine.

[0016] Example 1: As Figure 1 As shown, this embodiment of the invention provides a decision-making method for acupuncture rehabilitation of post-stroke aphasia based on multimodal data fusion, comprising the following steps: Step S100: Collect clinical scale data, functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and longitudinal relaxation time T1 structural image data of patients with post-stroke aphasia; perform preprocessing on the image data, including format conversion, head motion correction, time layer correction, spatial standardization, denoising, and covariate regression, to generate a standardized multimodal image feature set; Step S200: Based on the multimodal image feature set, extract radiomics features such as gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions; combined with clinical data, use principal component analysis and LASSO regression to perform feature dimensionality reduction and screening, and identify key predictive feature sets that are strongly correlated with acupuncture efficacy. Step S300: Input the selected key prediction features into various machine learning algorithms such as random forest, support vector machine and logistic regression to train and build an acupuncture efficacy prediction model; evaluate the model performance through cross-validation, ROC curve analysis, accuracy and sensitivity and other indicators to form an intelligent support model that can assist clinical decision-making, and be used to predict the patient's acupuncture rehabilitation effect and guide individualized treatment plan.

[0017] The clinical scale data comprises quantitative clinical indicators collected through standardized neuropsychological and language function assessment tools, primarily including the Western Aphasia Test System (WATS) score, the Boston Aphasia Severity Scale (Boston Aphasia Severity Scale), and the Pinprick Sensation Scale. As the gold standard for efficacy evaluation, it provides quantitative labels for the degree of language function impairment and rehabilitation effects, used for defining outcome variables in supervised learning and for comparing true values ​​during model training. Functional magnetic resonance imaging (fMRI) is a non-invasive brain functional imaging technique based on the blood oxygen level-dependent BOLD effect; it acquires low-frequency oscillation signals (0.01-0.1 Hz) in resting-state brain regions, reflecting the synchronous changes in spontaneous neuronal activity; these signals are used to calculate regional coherence (ReHo), low-frequency amplitude (ALFF), and the whole-brain functional connectivity (FC) matrix, quantifying the dynamic reorganization state of the brain's language function network. Diffusion tensor imaging (DTI) and longitudinal relaxation time (T1) structural images were used. DTI is a magnetic resonance imaging technique based on the anisotropic diffusion of water molecules in brain white matter fibers. T1 structural images are high-resolution three-dimensional anatomical images used to differentiate between gray matter, white matter, and cerebrospinal fluid. DTI mainly extracts parameters such as fractional anisotropy (FA) and mean diffusion rate (MD) to reflect the integrity of white matter fiber bundles. T1 images are used for brain region volume measurement and cortical thickness analysis, providing a structural morphological basis. DTI is used to assess the degree of damage to language-related white matter pathways, such as the arcuate fasciculus and superior longitudinal fasciculus. T1 images are used for lesion localization, gray matter volume calculation, and image spatial normalization registration. The multimodal image feature set preprocesses fMRI, DTI, and T1 data, aligning and integrating them in a unified spatial coordinate system, such as the MNI standard space, to form a high-dimensional data structure. It contains multi-dimensional quantitative information at both the brain region-level (ROI) and voxel-level, with each sample corresponding to a set of multimodal feature vectors covering structure, function, and fiber connectivity. Serving as a unified data source for key predictive feature extraction, it comprehensively characterizes the patient's brain state from different physical perspectives. Radiomics features are quantitative image features extracted from medical images through high-throughput processing, transforming visual images into a mineable data feature space. This includes first-order statistical features (gray-level mean, variance, etc.), second-order texture features (gray-level co-occurrence matrix), morphological features (lesion volume, shape), and higher-order functional connectivity features (graph theory indices, network efficiency, etc.). It reveals microscopic structural changes and functional reorganization patterns invisible to the naked eye, providing potential biomarkers reflecting differences in acupuncture efficacy. The key predictive feature set strongly correlated with acupuncture efficacy was selected through principal component analysis (PCA) dimensionality reduction and LASSO regression feature selection. This subset of features was the most strongly correlated with the efficacy outcome variable, the rate of change of WAB score, and had the lowest redundancy. A few high-contribution features with non-zero feature coefficients and stable performance in cross-validation were selected, such as 7 features selected from 145 features in the preliminary experiment. This reduced model complexity, prevented overfitting, and identified neuroimaging markers that played a decisive role in the differences in acupuncture response.The intelligent support model is a classification or regression model built on supervised learning algorithms, capable of predicting the probability of individual acupuncture efficacy based on input features. It is a deployable decision-making tool formed after training, validation, and performance evaluation using algorithms such as random forest and support vector machine. It realizes the transformation from experience-based treatment to data-driven decision-making, and uses computer technology to assist clinicians in identifying advantageous populations and formulating individualized acupuncture treatment plans during the golden recovery period.

[0018] In the above embodiments, this embodiment integrates clinical scales, functional magnetic resonance imaging, diffusion tensor imaging, and structural imaging data to achieve a complete process from data acquisition to decision generation. During the data preprocessing stage, operations such as format conversion, head movement correction, temporal correction, spatial standardization, denoising, and covariate regression improve the quality and comparability of multimodal imaging data, laying the foundation for key predictive feature extraction. By extracting radiomics features such as gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions, combined with clinical data, the complex changes in brain structure and function in post-stroke aphasia patients are captured. Principal component analysis and LASSO regression are used for feature dimensionality reduction and screening, effectively identifying key predictive features strongly correlated with acupuncture efficacy, reducing data redundancy, and enhancing model interpretability and generalization ability. The selected key predictive features are input into machine learning algorithms such as random forest, support vector machine, and logistic regression to construct an acupuncture efficacy prediction model. Through cross-validation, ROC curve analysis, and evaluation of accuracy and sensitivity, the model demonstrates reliable predictive performance and stability. The resulting intelligent support model can assist clinical decision-making, predict the acupuncture recovery effect of patients, and provide objective basis for the formulation of individualized treatment plans.

[0019] In summary, this embodiment achieves quantitative prediction and personalized guidance of the rehabilitation effect of acupuncture for post-stroke aphasia by multimodal data fusion, feature selection and machine learning modeling, thereby improving the accuracy and scientific nature of rehabilitation treatment.

[0020] Example 2: As Figure 3 As shown, based on Example 1, the process of generating a standardized multimodal image feature set in step S100 of this embodiment of the invention specifically includes the following steps: Step S101: Convert the acquired functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and longitudinal relaxation time T1 structural image data into a unified storage format; perform head motion correction on the time series of fMRI images, eliminate inter-slice misalignment caused by slight head displacement during scanning through rigid registration, and simultaneously complete time-slice correction to compensate for differences in acquisition time of each slice; output corrected image data with reduced motion artifacts and aligned time points. Step S102: The corrected images are mapped to a standard three-dimensional spatial coordinate system through a nonlinear registration algorithm, so that images of different individuals and different modalities have spatial correspondence at the same anatomical position; spatial smoothing filtering is used to remove high-frequency random noise in the images, and standardized image data with spatial alignment and noise suppression are output. Step S103: Extract physiological interference sources such as brain white matter signal and cerebrospinal fluid signal from the standardized images as covariates, and eliminate their influence on the BOLD signal through regression analysis to eliminate non-neurophysiological noise; perform voxel-level alignment of data from three modalities, namely functional magnetic resonance imaging, diffusion tensor imaging and T1 structural imaging, in a unified spatial coordinate system, and integrate them into a data structure corresponding to a set of high-dimensional feature vectors for each sample. The data structure is a standardized multimodal image feature set.

[0021] In the above embodiments, this embodiment achieves consistent storage of data from different sources by uniformly converting functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and T1 structural images, providing a basic data interface for processing. Head motion correction and inter-slice time correction are applied to the time series of fMRI images to eliminate spatial misalignment and temporal asynchrony caused by head displacement and acquisition timing differences, reducing motion artifacts and improving the spatial alignment accuracy of the time series data. Nonlinear registration maps multimodal images to a standard spatial coordinate system, establishing anatomical positional correspondences between different individuals and modalities. Spatial smoothing filtering suppresses high-frequency random noise, improving the spatial comparability and signal-to-noise ratio of the image data. Brain white matter and cerebrospinal fluid signals are extracted as covariates for regression analysis to eliminate the influence of physiological noise on blood oxygenation level-dependent signals and reduce interference from non-neural sources. Voxel-level alignment and integration of multimodal data are completed under a unified spatial coordinate system, forming a structured feature set corresponding to a high-dimensional feature vector for each sample, providing a spatially standardized, noise-suppressed, and multimodal fusion-based quantitative image feature foundation for analysis.

[0022] Example 3: Based on Example 2, the process in step S102 of this embodiment of the invention, which establishes a spatial correspondence between images of different individuals and different modalities at the same anatomical location, specifically includes the following steps: Step S1021: The corrected longitudinal relaxation time T1 structural image data is globally linearly aligned with the anatomical template corresponding to the pre-constructed standard three-dimensional spatial coordinate system; by calculating the overall translation, rotation and scaling parameters, the T1 image is initially aligned with the template space on a macro scale, eliminating the overall size and orientation differences between individuals, and outputting the initially registered T1 image. Step S1022: Using the pre-registered T1 image as an intermediate reference coordinate system, the corrected functional magnetic resonance imaging and diffusion tensor imaging data are aligned with the spatial position of the functional and diffusion images with the T1 image by adjusting the position and orientation, so that all modal data are unified under the same individual spatial coordinate system, and the functional and diffusion images aligned with the T1 space are output. Step S1023: Based on the initially registered T1 images, calculate the local deformation field between them and the standard spatial anatomical template, and perform fine nonlinear adjustments on the T1 images to ensure that fine anatomical structures such as cerebral sulci and gyri are completely matched with the template; at the same time, apply the local deformation field to the aligned functional and diffusion tensor images, and finally obtain standardized image data in the standard three-dimensional spatial coordinate system in which the three modalities of functional magnetic resonance imaging, diffusion tensor imaging and T1 structural images all have anatomical correspondence.

[0023] In the above embodiments, this embodiment achieves preliminary spatial alignment at a macroscopic scale by globally linearly aligning the T1 structural image with a standard anatomical template, eliminating differences in overall size, translation, and rotation between individuals, and establishing a basic coordinate system for fine registration. Using the registered T1 image as an intermediate reference, functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) data are aligned with it, unifying multimodal data into a single volumetric spatial coordinate system and ensuring consistent spatial relationships between different modal images at the individual level. By calculating the local nonlinear deformation field between the T1 image and the standard template, fine anatomical adjustments are made to the T1 image to match details such as cerebral sulci and gyri with the template. This deformation field is then simultaneously applied to functional and diffusion images, ultimately achieving voxel-level anatomical correspondence of multimodal data in a standard spatial coordinate system, providing a precise spatial basis for quantitative comparison and analysis across individuals and modalities.

[0024] Example 4: Based on Example 3, the process of achieving preliminary positional correspondence between the T1 image and the template space at a macroscopic scale in step S1021 of this embodiment of the invention specifically includes the following steps: Step S10211: In the corrected longitudinal relaxation time T1 structural image data, the brain parenchyma region is divided according to the grayscale difference between brain tissue and surrounding non-brain tissue; the spatial coordinates of all voxels in the brain parenchyma region are statistically analyzed, and the weighted average is used to obtain the spatial centroid of the whole brain region; the T1 image is translated as a whole so that the spatial centroid coincides with the center point of the anatomical template corresponding to the pre-constructed standard three-dimensional spatial coordinate system, thus completing the preliminary alignment of the overall image position, and outputting the T1 image after position adjustment. Step S10212: In the T1 image after position adjustment, the spatial trajectories of the interhemispheric fissure and lateral fissure of the brain are located by utilizing the grayscale changes of the sulci and gyri in the image, and the principal axis direction vectors of each are fitted; at the same time, the reference direction of the corresponding fissure is obtained from the standard template; the angular deviation between the fissure direction of the T1 image and the reference direction of the template is calculated, and the image is rotated in three dimensions to eliminate the deviation, so that the two directions are aligned, and the T1 image after direction adjustment is output. Step S10213: In the T1 image after orientation adjustment, scan the outer boundary of the brain tissue region along the three orthogonal directions of front-back, left-right, and up-down to determine the maximum span value in each direction; compare the span value in the same direction with that of the standard template to obtain the scaling ratio in the three directions; perform isotropic or anisotropic scaling on the T1 image to make its overall size consistent with the template, and finally output the T1 image after preliminary registration in position, orientation and scale.

[0025] In the above embodiments, this embodiment measures the span of brain tissue along three orthogonal directions and compares it with the template. After determining the scaling ratio of each direction, it performs scale transformation to match the overall size of the individual brain image with the template. Finally, it completes the comprehensive adjustment of position, direction and scale, and achieves the initial alignment of the individual T1 image and the standard template in macroscopic anatomical space, providing a stable initial spatial correspondence for nonlinear fine registration.

[0026] Example 5: Based on Example 4, the process of calculating the angular deviation between the T1 image crack direction and the template reference direction in step S10212 provided in this embodiment of the invention specifically includes the following steps: Step S102121: In the T1 image after position adjustment, based on the continuous trajectory of the interhemispheric fissure and lateral fissure in the image grayscale, fit the main direction lines of these two fissures in three-dimensional space to obtain the main direction lines of the T1 image fissures. Step S102122: Extract the reference main orientation lines of the interhemispheric fissure and lateral fissure in the template space from the anatomical template corresponding to the pre-constructed standard three-dimensional spatial coordinate system to obtain the template reference fissure main orientation lines; Step S102123: Place the main orientation line of the crack in the T1 image and the main orientation line of the crack reference in the template in the same three-dimensional spatial coordinate system. Through rotation simulation around the front-back direction, left-right direction and up-down direction, make the two main orientation lines of the T1 image completely coincide with the two reference main orientation lines of the template in spatial direction. Record the rotation angles around the three directions required to make the two orientations consistent, thus obtaining the angular deviation between the crack direction of the T1 image and the reference direction of the template.

[0027] In the above embodiments, this embodiment accurately extracts the geometric orientation features of key anatomical structures of the individual brain in space by fitting the three-dimensional main orientation lines of the interhemispheric fissure and lateral fissure in T1 images based on grayscale continuity trajectory; by extracting the reference main orientation lines of the corresponding fissures from the standard anatomical template, a standardized spatial reference benchmark is provided for the orientation correction of individual images; by placing the individual fissure orientation lines and the template reference lines in the same coordinate system and performing three-dimensional rotation simulation, the rotation angles around the front-back, left-right, and up-down directions required to make their spatial orientations coincide are quantitatively calculated, thereby obtaining the precise angular deviation of the individual T1 image relative to the standard template in the anatomical direction, providing direct parameter basis for image rotation correction, and ensuring that the main anatomical structures of the individual brain are accurately aligned with the template in spatial orientation.

[0028] Example 6: Based on Example 2, the process of eliminating the influence of regression analysis on the BOLD signal in step S103 of this embodiment of the invention specifically includes the following steps: Step S1031: In the standardized image data, the white matter region and the cerebrospinal fluid region are delineated according to the brain tissue segmentation results. The time series of all voxels in these two regions are extracted respectively. The signal mean of all voxels in the region is calculated at each time point to obtain the average time series of white matter and the average time series of cerebrospinal fluid, which are used as characteristic waveforms of physiological interference. Step S1032: For each voxel in the BOLD data, its time series is synchronously compared with the obtained average time series of white matter and average time series of cerebrospinal fluid to determine the part of the voxel signal that is consistent with the changing trend of the two average waveforms, and these covariant parts are removed from the original time series of the voxel to obtain the pure BOLD time series of the voxel after removing physiological noise. Step S1033: Reorganize the BOLD time series after removing physiological noise from all voxels according to their original spatial positions to form BOLD image data with physiological noise removed. Align this data with the diffusion tensor imaging data and T1 structural image data in the same standard spatial coordinate system at the voxel level. Finally, integrate them into a standardized multimodal image feature set corresponding to each sample, which contains functional, diffusion and structural information.

[0029] In the above embodiments, this embodiment extracts the average time series of the white matter and cerebrospinal fluid regions based on the brain tissue segmentation results to obtain characteristic waveforms reflecting physiological noise, such as respiration and heartbeat, providing noise estimation references for signal correction. By synchronously comparing and regressing the BOLD time series of each voxel with the noise characteristic waveforms, components covariant with physiological noise in the signal are removed, thereby effectively suppressing the influence of physiological interference on blood oxygen level-dependent signals at the voxel level and improving the signal-to-noise ratio and specificity of functional magnetic resonance imaging data. By recombining the noise-removed voxel time series according to their original spatial positions and aligning them with diffusion tensor imaging and T1 structural imaging data of the same sample at the voxel level in standard space, a spatially registered and signal-clean multimodal image feature set is finally formed, providing a high-quality data foundation for subsequent cross-modal correlation analysis and functional network modeling.

[0030] Example 7: Figure 4 As shown, based on Example 1, the process of identifying the key predictive feature set strongly correlated with acupuncture efficacy in step S200 of this embodiment of the invention specifically includes the following steps: Step S201: Extract the radiomics feature vector of each sample from the multimodal image feature set, and combine the feature vectors of all samples into a feature matrix; perform a linear orthogonal transformation on the feature matrix to generate a new set of comprehensive features, each of which is a linear combination of the original features and is linearly independent of each other; calculate the proportion of the variance of each comprehensive feature to the total variance of all comprehensive features, and accumulate them from high to low proportions. When the accumulated value first exceeds the preset variation contribution threshold, retain the first few comprehensive features as the dimensionality-reduced image feature set. Step S202: Combine the obtained dimensionality-reduced image features with the continuous and categorical indicators in the clinical data to construct a joint feature vector for each sample; use the rate of change in Western aphasia test scores before and after acupuncture treatment as the efficacy outcome variable to establish a regression model; introduce a constraint mechanism that can simultaneously achieve parameter estimation and feature selection when solving the regression model, which forces the coefficients of the features to shrink to zero while minimizing the prediction error, and only the coefficients of the features that contribute significantly to the efficacy prediction remain non-zero; extract the features with non-zero coefficients as the initial set of key prediction features; Step S203: Resample the original samples multiple times with replacement to generate multiple sample subsets; perform the feature selection process independently for each sample subset and record whether each feature is retained in each selection; after all sample subsets are selected, calculate the proportion of the number of times each feature is retained to the total number of resampling times; set a stability threshold and retain only the features whose retention frequency exceeds the stability threshold as the key prediction feature set.

[0031] In the above embodiments, this embodiment maps high-dimensional radiomics features into linearly independent comprehensive features through linear orthogonal transformation, and retains the main variation information based on the variance contribution threshold, thereby effectively compressing the feature dimension and reducing the computational complexity and multicollinearity interference of modeling. By merging the dimensionality-reduced image features with clinical indicators, a regression model with the change rate of acupuncture efficacy as the outcome is established. With the help of a constraint mechanism that can simultaneously perform parameter estimation and feature selection, features that make significant contributions to efficacy prediction are automatically screened to form an initial feature set, thus initially realizing the identification of core variables related to efficacy from multimodal data. Through a stability selection method based on resampling, the selection frequency of each feature under multiple subset sampling is evaluated, and features with high repeatability are retained according to a preset threshold. Finally, a key predictive feature set that is statistically stable and strongly correlated with acupuncture efficacy is obtained, improving the robustness and interpretability of the feature selection results and laying the foundation for building a reliable efficacy prediction model.

[0032] Example 8: Based on Example 7, the process of establishing a regression model in step S202 of this embodiment of the invention specifically includes the following steps: Step S2021: Using each feature in the joint feature vector as the independent variable and the efficacy outcome variable as the dependent variable, construct a linear regression model framework with coefficient compression constraints; while minimizing the sum of squared prediction errors of all samples, the linear regression model framework applies a compression constraint proportional to the total absolute value of the coefficients to the regression coefficients of each independent variable, and the compression force is controlled by a non-negative adjustment parameter. Step S2022: Set a set of adjustment parameter values ​​arranged from largest to smallest. For each adjustment parameter value, solve the regression coefficients that satisfy the constraints by updating each feature in turn, and obtain a set of coefficient values ​​under the adjustment parameters. As the adjustment parameters change from largest to smallest, record the numerical change path of each feature coefficient under different adjustment parameters. Step S2023: Divide all samples into several parts randomly and evaluate the predictive performance of the regression model under different adjustment parameters by taking turns to validate: For each adjustment parameter, take one part as the validation set and the rest as the training set. Solve the coefficients on the training set and calculate the error between the predicted value and the actual value on the validation set. Repeat all samples to obtain the average prediction error under the adjustment parameter. Step S2024: Compare the average prediction errors corresponding to all adjustment parameters, and select the adjustment parameter with the smallest average prediction error as the optimal adjustment parameter; extract the features whose coefficients are not zero under the optimal adjustment parameter to form the initial key prediction feature set.

[0033] In the above embodiments, this embodiment constructs a linear regression model framework with coefficient compression constraints. While minimizing the prediction error, it imposes constraints on the feature coefficients based on the total absolute value, achieving simultaneous model fitting and feature selection. This automatically suppresses features that contribute weakly or redundantly to efficacy prediction while retaining features with strong predictive power. By setting a set of decreasing adjustment parameters and using a solution method that updates features one by one, the continuous change path of each feature coefficient under different constraint strengths is obtained. This systematically reveals the dynamic behavior of the importance of features in the model as the constraint strength changes, providing complete coefficient evolution information for parameter selection. By randomly dividing the samples into several subsets and performing round-robin validation, the generalization prediction performance of the model under different adjustment parameters is evaluated. Based on the objective standard of average prediction error, the adjustment parameters that make the model perform optimally on new samples are selected from the candidate parameters, avoiding overfitting of the training data. By selecting the adjustment parameter with the smallest average prediction error as the final parameter and extracting features with non-zero coefficients under this parameter, a set of preliminary key prediction features that statistically contribute significantly to the prediction of efficacy outcomes and shows stable generalization ability is obtained, providing a reliable basis for stability verification.

[0034] Example 9: As Figure 5 As shown, based on Example 1, the process of forming an intelligent support model that can assist clinical decision-making in step S300 of this embodiment of the invention specifically includes the following steps: Step S301: Randomly divide all samples in the key prediction feature set into a model building set and a model validation set according to a preset ratio. The model building set is used for model training, and the model validation set is used to evaluate the generalization ability of the model. Step S302: On the model construction set, three machine learning algorithms based on different principles are used to construct preliminary prediction models: the decision tree ensemble algorithm constructs multiple decision trees and combines their voting results to form a model; the margin maximization classification algorithm separates samples with different therapeutic effects by finding the optimal classification hyperplane; and the probability estimation-based linear regression algorithm outputs classification probabilities by fitting the linear relationship between features and therapeutic outcomes. For each algorithm, the internal parameters are adjusted through multiple rounds of resampling within the model construction set to minimize the prediction error of the model on the construction set, resulting in three preliminary candidate models. First, for decision tree ensemble algorithms, multiple sample subsets are extracted with replacement from the model building set, a decision tree is grown for each subset, and the prediction results of all decision trees are combined by majority voting to form an ensemble model. By trying different numbers of trees and growth depths, the parameter combination that minimizes the prediction error of the building set is selected.

[0035] Secondly, for the margin-maximizing classification algorithm, sample features are mapped to a high-dimensional space to find a classification surface that can separate samples of different therapeutic effects and maximize the margin between the two classes. By trying different mapping methods and margin relaxation, the model complexity is adjusted to minimize the classification error rate on the constructed set.

[0036] Furthermore, for linear regression algorithms based on probability estimation, a linear relationship between features and therapeutic outcomes is established, and a probability transformation function maps the linear output to the probability of the therapeutic category. By trying different combinations of features and constraint strengths, the deviation between the probability estimates on the constructed set and the actual outcomes is minimized. After adjusting the parameters as described above, three preliminary candidate models that performed well on the model building set were obtained. Step S303: Apply the three candidate models to the model validation set respectively, predict the acupuncture efficacy of each sample in the validation set, and obtain the prediction results; compare the prediction results of each model with the actual efficacy outcome of the sample, calculate the prediction accuracy and sensitivity, and draw a curve reflecting the relationship between sensitivity and specificity by changing the judgment threshold and calculate the area under the curve; comprehensively compare the prediction performance of the three models, select the candidate model with the best overall performance on the validation set as the final acupuncture efficacy intelligent support model, which will be used to guide the individualized clinical treatment plan; Features from the model validation set are input into each candidate model, which outputs a predicted efficacy category. The accuracy is calculated by comparing the predicted category with the actual efficacy category, and the sensitivity is calculated by calculating the proportion of correctly predicted categories among the actual valid samples. Simultaneously, for models that output continuous probabilities, multiple sets of sensitivity and specificity are obtained by setting different thresholds, and curves are plotted and the area under the curve is calculated. The three indicators of the three models are compared, and the model with the best overall performance is selected as the final intelligent support model. This model will be used to predict the acupuncture rehabilitation effect of new patients and to assist in decision-making regarding treatment plans.

[0037] The core parameters and their candidate ranges for each algorithm are determined as follows: For decision tree ensemble algorithms, the core parameters are the number of decision trees and the growth depth of each tree; the candidate range for the number of trees is set according to the sample size of the model construction set, ranging from 50 to 500 trees, with candidate value sequences generated at intervals of 50 trees; the candidate range for growth depth is from 2 to 10 layers, with candidate value sequences generated at intervals of 1 layer; For margin maximization-based classification algorithms, the core parameters are the feature mapping method, namely the kernel function type and the margin relaxation degree. Three candidate kernel function types are preset based on data characteristics: linear mapping, polynomial mapping, and radial basis mapping; the candidate values ​​for margin relaxation degree take multiple orders of magnitude, including 0.001, 0.01, 0.1, 1, 10, and 100; For linear regression algorithms based on probability estimation, the core parameter is the constraint strength, i.e., the compression degree controlling the feature weights; the candidate values ​​for constraint strength also take multiple orders of magnitude, including 0.001, 0.01, 0.1, 1, 10, and 100, while simultaneously considering whether to use all feature combinations with constraint strength.

[0038] A grid traversal strategy is used to generate all candidate parameter combinations. The core parameter candidate values ​​of each algorithm are fully combined to form a parameter set. For example, the decision tree ensemble algorithm has 10 candidate tree numbers and 9 candidate depths, for a total of 90 sets; the margin maximization algorithm has 3 mapping methods and 6 relaxation levels, for a total of 18 sets; the probabilistic linear regression algorithm has 6 constraint strengths, for a total of 6 sets; each set of parameters represents a candidate configuration to be evaluated.

[0039] Within the model construction set, the generalization performance of each parameter group is evaluated through multiple rounds of resampling. The model construction set is evenly divided into five parts, with each part used alternately as internal validation data and the remaining four parts as internal training data. For each set of candidate parameters, the model is trained on the internal training data in each round, and the prediction error is calculated on the internal validation data. For classification tasks, the proportion of misclassified samples to the total number of validation samples is used. After five rounds, the average of the five error proportions for each parameter group is taken as the average validation error for that parameter group. The parameter group with the smallest average validation error is selected as the optimal configuration for the algorithm. The average validation errors of all candidate parameter groups for each of the three algorithms are compared, and the parameter group with the smallest error is selected. If multiple parameter groups have the same and smallest error, the group with smaller parameter values, such as fewer trees or weaker constraint strength, is selected as the optimal configuration. The model is retrained on the complete model construction set using the optimal parameters. The selected optimal parameters are then used for the corresponding algorithm, and training is performed once on the complete model construction set to obtain the final preliminary candidate model for the algorithm. At this point, the model parameters perfectly fit the information in the construction set, and the tuning process is reproducible.

[0040] In the above embodiments, the process of forming an intelligent support model that can assist clinical decision-making is achieved through the integration of feature selection, model construction, and validation evaluation, realizing the systematic development and optimization of the acupuncture efficacy prediction model. The specific technical features, when used in combination, achieve the following technical effects: Step S301 ensures the independence of model training and validation data by randomly dividing the sample set, avoiding a decline in generalization ability due to overfitting of the training data, and laying the foundation for reliable evaluation of subsequent model performance. Step S302 uses multiple machine learning algorithms with different principles to build models in parallel, covering different modeling approaches such as decision tree ensemble, margin maximization classification, and probabilistic linear regression. Internal parameter tuning ensures that each algorithm reaches its optimal fit on the training set, thereby providing multiple candidate models with different structural characteristics, increasing the diversity and robustness of model selection. Step S303 uses an independent validation set to evaluate the performance of each candidate model, comprehensively comparing the model prediction effects through multiple indicators such as accuracy, sensitivity, and area under the curve. This process effectively identifies the model with the best generalization ability, avoiding performance misjudgments caused by overfitting of the training set, and ensuring that the finally selected model can be stably applied to the prediction of new samples.

[0041] In summary, this embodiment constructs an intelligent support model with high predictive accuracy and clinical applicability by combining training-validation separation, multi-algorithm comparison, and comprehensive performance evaluation. This model can output efficacy prediction results based on patient characteristics, providing quantitative and interpretable decision-making basis for clinical development of individualized acupuncture treatment plans, thereby improving the scientific rigor and targeted nature of treatment strategies.

[0042] Example 10: Based on Example 9, the process of obtaining three preliminary candidate models in step S302 of this embodiment of the invention specifically includes the following steps: Step S3021: Based on the characteristics of the model construction set, determine the value sequence of the core parameters for the three algorithms, and pair the values ​​of each parameter to generate a set of parameters to be evaluated for each type of algorithm. For decision tree ensemble algorithms, the number of trees and the growth depth of trees are set according to the sample size of the model building set. The values ​​of the two sequences are then combined pairwise to obtain the parameter set to be evaluated for the algorithm. For margin maximization-based classification algorithms, three mapping methods are preset as candidates based on data characteristics, and multiple orders of magnitude of relaxation levels are set. The mapping methods and relaxation levels are then combined pairwise to obtain the parameter set to be evaluated for the algorithm. For linear regression algorithms based on probability estimation, multiple orders of magnitude of constraint strength are set, and whether all features are enabled is combined with the constraint strength to obtain the parameter set to be evaluated for the algorithm. Step S3022: Divide the model building set into multiple subsets of equal size, and take turns using one subset as internal validation data and the remaining subsets as internal training data. For each set of parameters to be evaluated for each type of algorithm, build the model on the internal training data in each round and calculate the prediction error on the internal validation data. Then take the average of the errors in all rounds as the average validation error of the set of parameters. For each set of parameters to be evaluated, multiple rounds of internal training and validation are performed sequentially: In each round, the internal training data of the current round is used to build a model based on the set of parameters, and then the model is applied to the internal validation data of the current round. The proportion of samples where the model prediction results are inconsistent with the actual efficacy category is counted as the error of that round. After all rounds are completed, the arithmetic mean of the round errors is calculated to obtain the average validation error of the parameters. Step S3023: Compare the average validation error of all parameter groups to be evaluated for each type of algorithm, and select the parameter group with the smallest error as the optimal configuration for that type of algorithm; retrain the model using the optimal configuration on the complete model building set to obtain the preliminary candidate model of the algorithm; For each type of algorithm, select the group with the smallest average verification error from all the parameter groups to be evaluated for the corresponding algorithm. If there are multiple groups with the same and smallest error, the group with smaller parameter values ​​shall be selected first. Use the selected optimal parameters for the algorithm and train it once on the complete model building set to obtain the preliminary candidate model of the algorithm. Perform the above operation on the three types of algorithms respectively to finally obtain three preliminary candidate models.

[0043] In the above embodiments, this embodiment realizes a systematic parameter tuning and model evaluation process; by covering a wide parameter space, adopting a stable cross-validation evaluation method, and selecting the optimal configuration and retraining based on the principle of minimizing error, the final candidate model has been fully optimized in parameter settings, has good generalization ability and prediction accuracy, and lays a reliable foundation for model integration or selection.

[0044] Example 11: Based on Example 10, the process of building a model on the internal training data and calculating the prediction error on the internal validation data in step S3022 of this embodiment of the invention specifically includes the following steps: Step S30221: Combine the current round of internal training data with the parameter group to be evaluated to generate a set of efficacy judgment rules under the parameter group to be evaluated; The feature values ​​and actual efficacy categories of all samples are extracted from the training data in the current round. According to the algorithm structure specified by the parameter group to be evaluated, such as the number of trees and growth depth in the decision tree ensemble algorithm, the mapping method and relaxation degree in the interval maximization algorithm, and the constraint strength in the probabilistic linear regression algorithm, the correspondence between features and efficacy categories is fitted to form a set of rules that can convert input features into efficacy category judgments. Step S30222: Input the features of the current round's internal verification data into the decision rule set to generate the prediction result sequence of the current round's verification samples; The feature values ​​of each sample in the current round of internal validation data are sequentially fed into the obtained set of judgment rules. The set of rules outputs the corresponding efficacy category judgment based on the input features. The efficacy category judgments are arranged in the order of the samples to obtain the prediction result sequence of the current round of internal validation samples. Step S30223: Compare the predicted result sequence with the actual efficacy category of the internal validation data, and calculate the proportion of inconsistent samples as the prediction error for the current round; The obtained prediction result sequence is compared with the actual efficacy category recorded in the current round of internal validation data one by one. The number of samples that are inconsistent between the two is counted. The number of samples is divided by the total number of samples in the current round of internal validation to obtain the prediction error value on the current round of internal validation data.

[0045] In the above embodiments, this embodiment completes the entire process from parameter configuration to model construction, and then to prediction output and error calculation; by generating a set of parameter-dependent rules based on training data, performing prediction and calculating errors on validation data, it realizes the direct evaluation of the generalization ability of a specific parameter set; it ensures the empirical and repeatable nature of parameter evaluation and provides data-driven performance indicators for parameter selection.

[0046] Example 12: Based on Example 11, the process of fitting the correspondence between features and efficacy categories in step S30221 of this embodiment of the invention specifically includes the following steps: Step S302211: Based on the rule structure specified by the parameter group to be evaluated, match the feature values ​​in the internal training data with the efficacy category to generate the original data matrix used to construct the judgment rules; The feature values ​​of all samples are extracted from the training data within the current round. According to the algorithm structure specified by the parameter group to be evaluated, such as the number of trees and growth depth in the decision tree ensemble algorithm, the mapping method and relaxation degree in the interval maximization algorithm, and the constraint strength in the probabilistic linear regression algorithm, the feature values ​​are arranged into a matrix form that is compatible with the rule structure. The corresponding efficacy category is added as an additional column to form the original data matrix. Step S302212: On the original data matrix, according to the rules for generating the parameter group to be evaluated, calculate the correlation weight between the feature values ​​and the efficacy category one by one to determine the role of each feature in the judgment. For the decision tree ensemble algorithm, the splitting features and thresholds are selected by calculating the degree to which each feature improves the purity of the efficacy category under different values, and the node splitting rules of each decision tree are generated. For the margin maximization algorithm, the normal vector and intercept of the classification hyperplane are determined by solving the support vectors of each category sample in the feature space. For the probabilistic linear regression algorithm, the feature weights are iteratively adjusted to minimize the deviation between the probability value after linear combination and the actual efficacy category, and the weighting coefficients of each feature are obtained. Step S302213: Integrate the obtained feature weights or splitting rules according to the structure specified by the parameter group to be evaluated to form a complete set of efficacy judgment rules; For the decision tree ensemble algorithm, the node rules of each decision tree are organized according to the tree structure, and the outputs of all trees are combined into an overall judgment rule through majority voting. For the margin maximization algorithm, the hyperplane normal vector and intercept are fixed to form a rule with the linear combination of features as the discrimination. For the probabilistic linear regression algorithm, the feature weighting coefficients are combined with the probability transformation function to form a rule with the probability threshold as the discrimination. A set of rules that can convert input features into efficacy category judgments is obtained.

[0047] In the above embodiments, this embodiment realizes a complete construction process from structured data to a parameterized rule set. Through data matrix preparation, feature weight calculation, and rule structure integration, the final generated decision rule set not only reflects the model form and complexity specified by the parameter configuration, but also includes feature and category association knowledge learned from the training data; the rule set has a clear mathematical form and executability, providing a foundation for prediction applications and ensuring a direct correspondence between model performance and parameter settings.

[0048] Example 13: As Figure 6 As shown, based on Examples 1-12, the post-stroke aphasia acupuncture rehabilitation decision-making system based on multimodal data fusion provided in this embodiment of the invention includes: The multimodal data acquisition and preprocessing module 1 is used to acquire clinical scale data, functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and longitudinal relaxation time T1 structural image data of patients with post-stroke aphasia; and to perform preprocessing on the image data such as format conversion, head motion correction, time layer correction, spatial standardization, denoising, and covariate regression to generate a standardized multimodal image feature set. The radiomics feature extraction and screening module 2 is used to extract radiomics features such as gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions based on multimodal image feature sets; combined with clinical data, principal component analysis and LASSO regression are used to perform feature dimensionality reduction and screening to identify key predictive feature sets that are strongly correlated with acupuncture efficacy. The intelligent decision support model construction module 3 is used to input the selected key prediction features into various machine learning algorithms such as random forest, support vector machine and logistic regression to train and build an acupuncture efficacy prediction model; the model performance is evaluated through cross-validation, ROC curve analysis, accuracy and sensitivity and other indicators to form an intelligent support model that can assist clinical decision-making, and is used to predict the patient's acupuncture rehabilitation effect and guide individualized treatment plan.

[0049] In the above embodiments, this embodiment integrates a multimodal data acquisition and preprocessing module, a radiomics feature extraction and screening module, and a smart decision support model construction module to achieve end-to-end technical integration from raw data to clinical decision-making. The multimodal data acquisition and preprocessing module ensures the consistency and comparability of multi-source heterogeneous data in spatial and temporal dimensions, providing high-quality input for key predictive feature extraction and model construction. The radiomics feature extraction and screening module extracts radiomics features reflecting brain structure, white matter fiber integrity, functional connectivity, and neural activity levels, and performs comprehensive analysis in conjunction with clinical data; through principal component analysis and LASSO regression, feature dimensionality reduction and screening are achieved, effectively identifying key biomarker features closely related to acupuncture rehabilitation efficacy, reducing redundant information, and improving the representativeness and interpretability of the feature set. The smart decision support model construction module inputs the screened key predictive features into machine learning algorithms such as random forest, support vector machine, and logistic regression to construct an acupuncture efficacy prediction model. The model's performance was evaluated and optimized through cross-validation, ROC curve analysis, and metrics such as accuracy and sensitivity, resulting in a decision support model with high predictive stability and generalization ability. This model can predict acupuncture rehabilitation effects based on multimodal patient data and provide quantitative evidence for the development of individualized treatment plans. In summary, this embodiment achieves standardized processing of multimodal data, precise screening of key biomarkers, and reliable construction of a efficacy prediction model, ultimately forming an intelligent support system that can assist clinical decision-making, improving the accuracy, individualization, and scientific decision-making capabilities of acupuncture rehabilitation treatment for post-stroke aphasia.

[0050] Figure 7A block diagram of an exemplary electronic device suitable for implementing embodiments of the present invention is shown. The electronic device may include a central processing unit / microprocessor / main control chip 4; and a storage medium 5 coupled to the central processing unit / microprocessor / main control chip 4, wherein computer-executable instructions are stored for performing steps of various methods of embodiments of the present invention when executed by a processor. The central processing unit / microprocessor / main control chip 4 may include, but is not limited to, one or more processors or microprocessors. The storage medium 5 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, and computer storage media (e.g., hard disk, floppy disk, solid-state drive, removable disk, CDROM, DVDROM, Blu-ray disc, etc.). In addition, the electronic device may include (but is not limited to) a data bus 6, an input / output bus / external bus / device bus 7, a display 8, and input / output devices 9 (e.g., keyboard, mouse, speaker, etc.). The central processing unit / microprocessor / main control chip 4 may communicate with external devices (8, 9, etc.) via the input / output bus / external bus / device bus 7 via a wired or wireless network (not shown). Storage medium 5 may also store at least one computer-executable instruction for performing the steps of various functions and / or methods in the embodiments described herein when run by the central processing unit / microprocessor / main control chip 4. In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product, wherein one or more computer-executable instructions are run by a processor to perform the steps of various functions and / or methods in the embodiments described herein.

[0051] This invention has preliminarily completed the analysis of relevant data of 41 PSA patients before and after 4 weeks of acupuncture treatment, calculated relevant parameters, and established a preliminary acupuncture efficacy prediction model. It also preliminarily explored the application value of the efficacy prediction model based on pre-treatment fMRI in evaluating acupuncture efficacy, providing experience for the establishment of multimodal radiomics models and the analysis of complex correlational data. Feature extraction: Leave-one-out permutation test was used, and LASSO feature selection was performed separately in each leave-one-out permutation. Dimensionality reduction selection was performed on the imaging features of the 41 samples. 10-fold cross-validation was used to determine the optimal alpha value as 0.41153977197036584, with a corresponding -log(alpha) of 0.38558818739557. The coefficients of each feature were identified based on the alpha value, and features with non-zero coefficients were selected to screen out the features most strongly correlated with treatment (see...). Figure 8 and Figure 9 Ultimately, seven features with the highest correlation to acupuncture efficacy were selected from 145 features and used to construct the predictive model. The selected features and their corresponding regression coefficients are shown below. Figure 10As shown in the figure. A preliminary model for predicting the efficacy of acupuncture was established: Support Vector Machine (SVM), Logistic Regression, and Random Forest were used to model the seven selected features. The ROC curve analysis results for each classifier are shown in the figure. Figures 11-13 As shown. After detailed performance comparison, the random forest model showed the best performance on the dataset, with an AUC of 0.78 (95% CI: 0.64–0.92).

[0052] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A decision-making method for acupuncture rehabilitation of post-stroke aphasia based on multimodal data fusion, characterized in that, Includes the following steps: Based on the multimodal image feature set obtained from the preprocessing of acquired image data, radiomics features of gray matter volume, white matter fiber integrity, functional connectivity strength, and low-frequency amplitude in brain regions are extracted; combined with clinical data, feature dimensionality reduction and screening are performed to identify key predictive feature sets that are strongly correlated with acupuncture efficacy. The selected key predictive features are input into various machine learning algorithms to train and build an acupuncture efficacy prediction model. The model performance is evaluated through cross-validation, ROC curve analysis, and accuracy and sensitivity indicators to form an intelligent support model that can assist clinical decision-making.

2. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 1, characterized in that, The process of identifying a set of key predictive features strongly correlated with the efficacy of acupuncture includes the following steps: Extract the radiomics feature vector of each sample from the multimodal image feature set, and combine the feature vectors of all samples into a feature matrix; A linear orthogonal transformation is performed on the feature matrix to generate a new set of comprehensive features. Each comprehensive feature is a linear combination of the original features and they are linearly independent of each other. Calculate the proportion of the variance of each comprehensive feature to the total variance of all comprehensive features, and accumulate them from high to low proportions. When the accumulated value first exceeds the preset variation contribution threshold, retain the first few comprehensive features as the image feature set after dimensionality reduction. The obtained dimensionality-reduced image features are combined with continuous and categorical indicators in clinical data to construct a joint feature vector for each sample; the rate of change in Western aphasia test scores before and after acupuncture treatment is used as the efficacy outcome variable to establish a regression model; when solving the regression model, a constraint mechanism that can simultaneously achieve parameter estimation and feature selection is introduced, which forces the coefficients of the features to shrink to zero while minimizing the prediction error, and only the feature coefficients that contribute to the efficacy prediction remain non-zero. Extract features with non-zero feature coefficients as the initial set of key prediction features; The original sample is resampled multiple times with replacement to generate multiple sample subsets; a feature selection process is performed independently for each sample subset, and it is recorded whether each feature is retained in each selection; after all sample subsets are selected, the proportion of the number of times each feature is retained out of the total number of resampling times is calculated. Set a stability threshold and retain only features whose retention frequency exceeds the stability threshold as the key prediction feature set.

3. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 1, characterized in that, The process of developing an intelligent support model that can assist clinical decision-making includes the following steps: All samples in the key prediction feature set are randomly divided into a model building set and a model validation set according to a preset ratio. The model building set is used for model training, and the model validation set is used to evaluate the generalization ability of the model. In the model building set, three different machine learning algorithms were used to build preliminary prediction models: the decision tree ensemble algorithm formed a model by constructing multiple decision trees and combining their voting results; the margin maximization classification algorithm separated samples with different therapeutic effects by finding the optimal classification hyperplane. Linear regression algorithms based on probability estimation output classification probabilities by fitting a linear relationship between features and therapeutic outcomes; For each algorithm, the internal parameters of the algorithm are adjusted through multiple rounds of resampling within the model construction set to minimize the prediction error of the model on the construction set, resulting in three preliminary candidate models. Three candidate models were applied to the model validation set to predict the acupuncture efficacy of each sample in the validation set and obtain the prediction results. The prediction results of each model were compared with the actual efficacy outcomes of the samples, and the prediction accuracy and sensitivity were calculated. A curve reflecting the relationship between sensitivity and specificity was plotted by changing the judgment threshold and the area under the curve was calculated. By comprehensively comparing the predictive performance of the three models, the candidate model with the best overall performance on the validation set was selected as the final intelligent support model for acupuncture efficacy, which is used to guide individualized clinical treatment plans.

4. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 3, characterized in that, The process of obtaining three preliminary candidate models includes the following steps: Based on the characteristics of the model construction set, the value sequences of the core parameters are determined for the three algorithms, and the values ​​of each parameter are fully paired to generate the set of parameters to be evaluated for each type of algorithm. The model building set is divided into multiple subsets of equal size. One subset is used as internal validation data and the remaining subsets are used as internal training data in turn. For each set of parameters to be evaluated for each type of algorithm, the model is built on the internal training data in each round and the prediction error is calculated on the internal validation data. Then the average of the errors in all rounds is taken as the average validation error of the parameters. Compare the average verification error of all parameter groups to be evaluated for each type of algorithm, and select the parameter group with the smallest error as the optimal configuration of the algorithm; The model is retrained using the optimal configuration on the complete model building set to obtain the initial candidate model of the algorithm.

5. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 4, characterized in that, The process of building a model on internal training data in each round and calculating the prediction error on internal validation data includes the following steps: The current round of internal training data is combined with the parameter set to be evaluated to generate a set of efficacy judgment rules under the parameter set to be evaluated. Input the features of the current round's internal validation data into the decision rule set to generate a sequence of predicted results for the current round's validation samples; The predicted sequence is compared with the actual efficacy categories in the internal validation data, and the proportion of inconsistent samples is used as the prediction error for the current round.

6. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 5, characterized in that, The process of fitting the correspondence between features and efficacy categories in generating the set of efficacy judgment rules for the parameter group to be evaluated includes the following steps: Based on the rule structure specified by the parameter group to be evaluated, the feature values ​​in the internal training data are matched with the efficacy category to generate the original data matrix used to construct the judgment rules; Based on the original data matrix, according to the rules for generating the parameter group to be evaluated, the correlation weight between the feature values ​​and the efficacy category is calculated one by one to determine the role of each feature in the judgment. The obtained feature weights or splitting rules are integrated according to the structure specified by the parameter group to be evaluated to form a complete set of efficacy judgment rules.

7. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 1, characterized in that, Clinical scale data, functional magnetic resonance imaging, diffusion tensor imaging, and longitudinal relaxation time structure data of patients with post-stroke aphasia were collected. The image data were preprocessed by format conversion, head motion correction, temporal correction, spatial standardization, denoising, and covariate regression to generate a standardized multimodal image feature set.

8. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 7, characterized in that, The process of generating a standardized multimodal image feature set includes the following steps: The acquired functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and longitudinal relaxation time T1 structural image data are converted into a unified storage format. Head motion correction is performed on the time series of fMRI images. Rigid registration is used to eliminate inter-slice misalignment caused by slight head displacement during scanning, and time-slice correction is performed simultaneously to compensate for differences in acquisition time of each slice. The corrected image data with reduced motion artifacts and aligned time points are output. The corrected images are mapped to a standard three-dimensional spatial coordinate system through a nonlinear registration algorithm, so that images of different individuals and different modalities have spatial correspondence at the same anatomical position; spatial smoothing filtering is used to remove high-frequency random noise in the images, and outputs standardized image data that is spatially aligned and noise-suppressed. Physiological interference sources of brain white matter and cerebrospinal fluid signals were extracted from standardized images as covariates. Their influence on the BOLD signal was eliminated through regression analysis, thus eliminating non-neurophysiological noise. Data from three modalities—functional magnetic resonance imaging, diffusion tensor imaging, and T1 structural imaging—were voxel-level aligned in a unified spatial coordinate system and integrated into a data structure that corresponds to a set of high-dimensional feature vectors for each sample. The data structure is a standardized multimodal image feature set.

9. The decision-making method for post-stroke aphasia rehabilitation based on multimodal data fusion as described in claim 8, characterized in that, The process of establishing spatial correspondence between images of different individuals and modalities at the same anatomical location includes the following steps: The corrected longitudinal relaxation time T1 structural image data is globally linearly aligned with the anatomical template corresponding to the pre-constructed standard three-dimensional spatial coordinate system. By calculating the overall translation, rotation and scaling parameters, the T1 image is initially aligned with the template space on a macro scale, eliminating the differences in overall size and orientation between individual images, and outputting the initially registered T1 image. The corrected functional magnetic resonance imaging and diffusion tensor imaging data are used with the pre-registered T1 image as an intermediate reference coordinate system. By adjusting the position and orientation, the spatial positions of the functional and diffusion images are aligned with the T1 image, so that all modal data are unified under the same individual spatial coordinate system, and the functional and diffusion images aligned with the T1 space are output. Based on the initially registered T1 images, the local deformation field between them and the standard spatial anatomical template is calculated, and the T1 images are finely nonlinearly adjusted to ensure that the fine anatomical structure of the brain sulci and gyri is completely matched with the template. At the same time, the local deformation field is synchronously applied to the aligned functional and diffusion tensor images, and finally standardized image data with anatomical correspondence in the standard three-dimensional spatial coordinate system are obtained for the three modalities of functional magnetic resonance imaging, diffusion tensor imaging and T1 structural images.

10. A decision-making system for post-stroke aphasia acupuncture rehabilitation based on multimodal data fusion, used to implement the decision-making method for post-stroke aphasia acupuncture rehabilitation based on multimodal data fusion as described in any one of claims 1 to 9, characterized in that, include: The multimodal data acquisition and preprocessing module is used to acquire clinical scale data, functional magnetic resonance imaging, diffusion tensor imaging, and longitudinal relaxation time structure image data of patients with post-stroke aphasia; it performs format conversion, head motion correction, time layer correction, spatial standardization, denoising, and covariate regression preprocessing on the image data to generate a standardized multimodal image feature set. The radiomics feature extraction and screening module is used to extract radiomics features of brain region gray matter volume, white matter fiber integrity, functional connectivity strength and low frequency amplitude based on multimodal image feature sets; combined with clinical data, principal component analysis and LASSO regression are used to perform feature dimensionality reduction and screening to identify key predictive feature sets that are strongly correlated with acupuncture efficacy. The intelligent decision support model building module is used to input the selected key predictive features into various machine learning algorithms such as random forest, support vector machine and logistic regression to train and build an acupuncture efficacy prediction model. The model performance is evaluated through cross-validation, ROC curve analysis and accuracy and sensitivity indicators to form an intelligent support model that can assist clinical decision-making, predict the patient's acupuncture rehabilitation effect and guide individualized treatment plans.

Images