CN-122000067-A - Nerve function prognosis prediction method and device after spinal cord injury operation

CN122000067ACN 122000067 ACN122000067 ACN 122000067ACN-122000067-A

Abstract

The invention discloses a nerve function prognosis prediction method and device after spinal cord injury operation, comprising the steps of inputting alignment data into a trained rehabilitation stage self-adaptive transducer model, outputting multi-modal fusion feature vectors, wherein the rehabilitation stage self-adaptive transducer model comprises a rehabilitation stage identification unit, a modal attention weight generator and a multi-modal attention mechanism module, the rehabilitation stage identification unit is used for identifying rehabilitation stage labels and then outputting rehabilitation stage feature vectors, the modal attention weight generator is used for dynamically generating attention weight matrixes of all modes according to the rehabilitation stage feature vectors, and the multi-modal attention mechanism module is used for fusing the multi-modal data according to the attention weight matrixes and outputting the multi-modal fusion feature vectors. The invention effectively improves the accuracy of prognosis prediction, provides reliable support for clinically making personalized treatment and rehabilitation schemes, and has good clinical application prospect.

Inventors

Gan Xinling
GAO HUI
JIANG ZEKUN
JIANG KUNYUAN

Assignees

四川大学华西医院

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (8)

1. A method for prognosis of neurological function after spinal cord injury surgery, comprising: Acquiring multi-mode data of each rehabilitation stage of spinal cord injury of a patient, wherein the multi-mode data is provided with a rehabilitation stage label, and the rehabilitation stage label comprises early stage, middle stage and later stage; preprocessing the multi-mode data to obtain preprocessed data; Carrying out alignment treatment on the preprocessed data to obtain alignment data; Inputting the aligned data into a trained rehabilitation stage self-adaptive transducer model to output a multi-mode fusion feature vector, wherein the rehabilitation stage self-adaptive transducer model comprises a rehabilitation stage identification unit, a modal attention weight generator and a multi-mode attention mechanism module, the rehabilitation stage identification unit is used for outputting the rehabilitation stage feature vector after identifying a rehabilitation stage label, the modal attention weight generator is used for dynamically generating an attention weight matrix of each mode according to the rehabilitation stage feature vector, and the multi-mode attention mechanism module is used for fusing the multi-mode data according to the attention weight matrix to output the multi-mode fusion feature vector; Carrying out average pooling and maximum pooling treatment on the multi-modal fusion feature vector, and outputting the pooled feature vector; And inputting the pooled feature vectors into a full connection layer and softmax activation function, and outputting the probability or grade of each predicted target in the corresponding rehabilitation stage.
2. The method of claim 1, wherein obtaining multi-modal data of a patient during a rehabilitation phase of spinal cord injury comprises: Acquiring early rehabilitation multi-modal data, wherein the early rehabilitation multi-modal data comprises clinical structural data, respiratory function monitoring data, complication risk data and basic vital sign data; acquiring multi-modal data in the mid-rehabilitation period, wherein the multi-modal data in the mid-rehabilitation period comprises clinical structural data, electromyographic signal data, gait data and MRI image review data; The method comprises the steps of acquiring later-stage multi-mode data of rehabilitation, wherein the later-stage multi-mode data of rehabilitation comprise clinical structural data, daily activity capacity data, auxiliary appliance use data and social adaptation data.
3. A method of prognostic prediction of neurological function after spinal cord injury according to claim 1 or 2, wherein the pretreatment comprises normalization, denoising and feature extraction.
4. The method for predicting prognosis of neurological function after spinal cord injury according to claim 1 or 2, wherein the aligning the preprocessed data to obtain aligned data comprises: converting the output format of the time sequence modal data in the multi-modal data into the same tensor format; And expanding the time dimension of the non-time sequence modal data in the multi-modal data to be the same as the time dimension of the same-stage time sequence modal data through repeated filling strategies.
5. The method for prognosis of neurological function after spinal cord injury according to claim 1 or 2, wherein the step of outputting the rehabilitation phase feature vector after identifying the rehabilitation phase label comprises the steps of: acquiring a single-heat code of a recovery stage label; Inputting the single thermal code into a fully-connected network, and processing by an activation function ReLU; A stage feature vector of dimension 256 is output.
6. The method for predicting prognosis of neurological function after spinal cord injury according to claim 2, wherein dynamically generating the attention weight matrix of each modality according to the feature vector of the rehabilitation stage comprises: If the convalescence stage feature vector determines that the convalescence stage label is early, generating an early attention weight matrix, wherein the attention weight of clinical data in the early attention weight matrix is 0.4, the attention weight of respiratory function data is 0.3, the attention weight of complication risk data is 0.2, and the attention weight of basic vital signs is 0.1; If the recovery stage label is determined to be a middle stage through the recovery stage feature vector, a middle stage attention weight matrix is generated, wherein the attention weight of clinical data in the middle stage attention weight matrix is 0.4, the attention weight of myoelectricity-gait synchronization features is 0.3, and the attention weight of MRI image data is 0.3; If the convalescence stage feature vector determines that the convalescence stage label is the later stage, a later stage attention weight matrix is generated, wherein the attention weight of clinical data in the later stage attention weight matrix is 0.1, the attention weight of daily activity capacity data is 0.4, the attention weight of auxiliary appliance use data is 0.3, and the attention weight of social adaptation data is 0.2.
7. The method for prognosis of nerve function after spinal cord injury according to claim 1, wherein the multi-modal attention mechanism module comprises: The first module is used for fusing early-stage clinical structural data and respiratory function data of rehabilitation according to an early-stage attention weight matrix, fusing the mid-stage clinical structural data and MRI image data of rehabilitation according to a mid-stage attention weight matrix, and fusing the later-stage clinical structural data and daily activity data of rehabilitation according to a later-stage attention weight matrix; The second module is used for fusing early respiratory function data and complication risk data of rehabilitation according to an early attention weight matrix, fusing electromyographic signal data and gait data of the middle rehabilitation according to a middle attention weight matrix, and fusing auxiliary appliance use data and social adaptation data of the later rehabilitation according to a later attention weight matrix; the third module is used for fusing the output results of the first module and the second module based on a cross attention mechanism to obtain a fusion result; And the fourth module is used for carrying out global weight distribution on all fusion results based on a multi-head global attention mechanism to obtain a multi-mode fusion feature vector.
8. An apparatus for use in a method for prognosis of neurological function after spinal cord injury surgery according to any one of claims 1 to 7, comprising: the acquisition module is used for acquiring multi-mode data of each rehabilitation stage of spinal cord injury of a patient, wherein the multi-mode data is provided with a rehabilitation stage label, and the rehabilitation stage label comprises early stage, middle stage and later stage; the preprocessing module is used for preprocessing the multi-mode data to obtain preprocessed data; The alignment module is used for carrying out alignment processing on the preprocessed data to obtain alignment data; The system comprises a fusion module, a multi-mode fusion module and a multi-mode fusion module, wherein the fusion module is used for inputting alignment data into a trained rehabilitation stage self-adaptive transducer model to output multi-mode fusion feature vectors, the rehabilitation stage self-adaptive transducer model comprises a rehabilitation stage identification unit, a modal attention weight generator and the multi-mode attention mechanism module, the rehabilitation stage identification unit is used for outputting the rehabilitation stage feature vectors after identifying rehabilitation stage labels, the modal attention weight generator is used for dynamically generating attention weight matrixes of all modes according to the rehabilitation stage feature vectors, and the multi-mode attention mechanism module is used for fusing the multi-mode data according to the attention weight matrixes and outputting the multi-mode fusion feature vectors; The pooling module is used for carrying out average pooling and maximum pooling treatment on the multi-mode fusion feature vectors and outputting the pooled feature vectors; And the output module is used for inputting the pooled feature vectors into the full-connection layer and the softmax activation function and outputting the probability or the grade of each predicted target in the corresponding rehabilitation stage.

Description

Nerve function prognosis prediction method and device after spinal cord injury operation Technical Field The invention relates to the technical field of medical artificial intelligence and spinal cord injury rehabilitation, in particular to a method and a device for predicting prognosis of nerve function after spinal cord injury operation. Background Cervical Spinal Cord Injury (CSCI) is a severe central nervous system injury caused by external violence or spinal cord compression, with a high incidence and a significant prognosis in the clinic. At present, the prognosis prediction of the nerve function after the CSCI operation mainly depends on a machine learning model driven by single-mode data, but the prediction performance is limited by the singleness of the data dimension, and the requirement of clinical accurate diagnosis and treatment is difficult to meet. The model is constructed by extracting clinical indexes such as age, sex, complications of patients, AIS classification when admitted, ASIA Motion Score (AMS) and the like and adopting algorithms such as gradient lifting tree (GBDT), extreme gradient lifting (XGBoost) and the like. However, because the key information reflecting the local structural damage in the spinal cord MRI image is not included, the prediction accuracy is limited, the existing accuracy is only 77.6-81.1%, the AUC value is between 0.857 and 0.867, and the rehabilitation potential of patients with different damage degrees can not be accurately distinguished. Another type of mainstream scheme is a prediction model based on MRI image data only, the model takes MRI images such as cervical vertebra T2 weighted sagittal position and the like as input, and the structural characteristics related to spinal cord injury are extracted through a Convolutional Neural Network (CNN), but the model has poor suitability to individual differences of patients due to the fact that clinical information such as general conditions and dynamic functional manifestations of the patients cannot be integrated, and actual rehabilitation potential of the patients is difficult to comprehensively reflect. In addition, a small number of existing multi-mode prediction schemes are spliced only through simple features, but the problem of heterogeneity of multi-mode data is not solved, active learning and alignment of cross-mode associated information are lacked, fusion efficiency is low, prediction performance is limited, contribution weights of various modes and features to prediction results cannot be quantized, interpretation is poor, and wide acceptance of clinicians is difficult to obtain. Therefore, the invention develops a method and a device for predicting the prognosis of the nerve function after spinal cord injury operation to solve the problems. Disclosure of Invention The invention provides a nerve function prognosis prediction method and device after spinal cord injury operation, which are used for solving the problems of low accuracy, poor suitability and poor interpretability of the traditional CMCI (physical and surgical control interface) postoperative nerve function prognosis prediction method. The invention realizes the above purpose through the following technical scheme: the invention discloses a nerve function prognosis prediction method after spinal cord injury operation, which comprises the following steps: Acquiring multi-mode data of each rehabilitation stage of spinal cord injury of a patient, wherein the multi-mode data is provided with a rehabilitation stage label, and the rehabilitation stage label comprises early stage, middle stage and later stage; preprocessing the multi-mode data to obtain preprocessed data; Carrying out alignment treatment on the preprocessed data to obtain alignment data; Inputting the aligned data into a trained rehabilitation stage self-adaptive transducer model to output a multi-mode fusion feature vector, wherein the rehabilitation stage self-adaptive transducer model comprises a rehabilitation stage identification unit, a modal attention weight generator and a multi-mode attention mechanism module, the rehabilitation stage identification unit is used for outputting the rehabilitation stage feature vector after identifying a rehabilitation stage label, the modal attention weight generator is used for dynamically generating an attention weight matrix of each mode according to the rehabilitation stage feature vector, and the multi-mode attention mechanism module is used for fusing the multi-mode data according to the attention weight matrix to output the multi-mode fusion feature vector; Carrying out average pooling and maximum pooling treatment on the multi-modal fusion feature vector, and outputting the pooled feature vector; And inputting the pooled feature vectors into a full connection layer and softmax activation function, and outputting the probability or grade of each predicted target in the corresponding rehabilitation stage. Further, acquiring multi