Multimodal fusion new-born brain injury degree ai evaluation and prognosis prediction system

Abstract

The present application relates to the technical field of medical data processing and artificial intelligence, in particular to a multi-modal fusion neonatal brain injury degree AI evaluation and prognosis prediction system. The present application obtains the stress fluctuation characteristics of the current trigger time based on the accumulation of the voltage value of the brain wave in the intervention time window sequence of the current trigger time. According to the data characteristics of the gray value of the voxel in the nuclear magnetic resonance brain three-dimensional matrix of the current trigger time, the normal equivalent value of each voxel is analyzed. The stress fluctuation characteristics of the current trigger time and the historical trigger time, the joint distribution characteristics of the normal equivalent value of each voxel and the adjacent voxel in the nuclear magnetic resonance brain three-dimensional matrix, and the accumulation of the normal equivalent value of the voxel in the nuclear magnetic resonance brain three-dimensional matrix of the current trigger time are fused to obtain the comprehensive risk index of the current trigger time, and the prognosis score of the current trigger time is determined.

Description

Technical Field

[0001] This invention relates to the fields of medical data processing and artificial intelligence technology, specifically to a multimodal fusion AI assessment and prognosis prediction system for the degree of neonatal brain injury. Background Technology

[0002] In the neonatal intensive care unit, for newborns with hypoxic-ischemic encephalopathy, it is usually necessary to simultaneously collect bedside continuous electroencephalogram (EEG) signals and static magnetic resonance imaging (MRI), and use artificial intelligence and multimodal data processing technology to assess the degree of brain damage and predict the prognostic trend.

[0003] However, in real clinical applications, routine nursing interventions such as suctioning and turning over that children receive regularly can cause stress-induced fluctuations in the brain waveform. Furthermore, the ability of the child's brain to calm these fluctuations in response to external stimuli depends not only on the total volume of normal brain tissue but also on the spatial connectivity of the normal brain tissue.

[0004] Current technologies for handling such scenarios typically employ a direct feature stitching method: extracting the proportion of abnormal lesion volume from static MRI images and the EEG amplitude features within discrete time windows, then directly inputting these isolated feature vectors into a fully connected network for evaluation. This approach fails to extract the spatially fragmented features of abnormal lesions and construct their physical constraints with dynamic stress fluctuations. It also cannot effectively isolate spurious fluctuations caused by routine nursing procedures, resulting in irregular jumps in the output brain injury assessment values ​​with external routine interventions, thus offering limited reference value for predicting disease progression in clinical practice. Summary of the Invention

[0005] To address the problem that existing technologies offer limited reference value for clinical assessment of the severity of brain injury in newborns with hypoxic-ischemic encephalopathy and for predicting prognostic trends, this invention aims to provide a multimodal fusion AI-based system for assessing the severity of neonatal brain injury and predicting prognosis. The specific technical solution adopted is as follows: This invention proposes a multimodal fusion AI-based assessment and prognostic prediction system for the severity of neonatal brain injury, the system comprising: The data acquisition module is used to acquire the post-intervention time window sequence and the MRI brain three-dimensional matrix of the target child at the current triggering time; the post-intervention time window sequence includes the EEG voltage value at each sampling time; the MRI brain three-dimensional matrix includes the gray value of each voxel; The basic feature analysis module is used to obtain the stress fluctuation characteristics at the current triggering moment based on the cumulative EEG voltage values ​​in the post-intervention time window sequence; and to analyze the normal equivalent value of each voxel based on the data characteristics of the gray values ​​of voxels in the MRI brain three-dimensional matrix. The comprehensive feature analysis module is used to integrate the stress fluctuation characteristics of the current triggering time and the historical triggering time, the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the MRI brain, and the cumulative situation of the normal equivalent values ​​of voxels in the three-dimensional matrix of the MRI brain at the current triggering time, to obtain the comprehensive risk index at the current triggering time. The prediction result output module is used to determine the prognostic score based on the numerical characteristics of the comprehensive risk index; the prognostic score is used to predict the degree of neonatal brain injury.

[0006] Furthermore, the stress fluctuation characteristics at the current triggering moment are obtained based on the cumulative EEG voltage values ​​in the post-intervention time window sequence, including: The stress fluctuation characteristics at the current triggering time are obtained by multiplying the absolute value of the difference between the EEG voltage value at each sampling moment in the post-intervention time window sequence and the preset EEG baseline value, and then multiplying the result with the time interval between any two adjacent sampling moments.

[0007] Further, the step of analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the three-dimensional matrix of the MRI brain includes: Based on the degree of deviation between the gray value of each voxel in the three-dimensional matrix of the MRI brain and the preset reference gray value, the lesion equivalent value of each voxel is determined. The normal equivalent value of each voxel is determined based on the negative correlation coefficient of the pathological equivalent value of each voxel.

[0008] Furthermore, the combined distribution characteristics of stress fluctuations at the current triggering time and historical triggering times, the normal equivalent values ​​of each voxel and its adjacent voxels in the 3D MRI brain matrix, and the cumulative normal equivalent values ​​of voxels in the 3D MRI brain matrix at the current triggering time are used to obtain a comprehensive risk index for the current triggering time, including: Based on the stress fluctuation characteristics of the current trigger time and the historical trigger time within the preset backtracking window, a normalization operation is performed on the stress fluctuation characteristics of the current trigger time to obtain the standard stress characteristics of the current trigger time. Based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the brain at the current trigger time and the historical trigger time within the preset retrospective window, the standard lesion spatial characteristics at the current trigger time are determined. Based on the cumulative normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain at the current triggering time, analyze the standard normal characteristics at the current triggering time. By integrating the standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics, a comprehensive risk index is obtained.

[0009] Further, the determination of the standard lesion spatial features at the current triggering time based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the MRI brain three-dimensional matrix at the current triggering time and the historical triggering time within the preset retrospective window includes: Multiply the normal equivalent value between each voxel in the 3D matrix of the brain at any trigger time within the preset backtracking window and each adjacent voxel to obtain the effective product between each voxel and each adjacent voxel. The ratio of the sum of all effective products at any given triggering time to the total number of effective products is used as the lesion spatial feature at any given triggering time. Based on the lesion spatial features at the current trigger time and the historical trigger time within the preset backtracking window, a normalization operation is performed on the lesion spatial features at the current trigger time to obtain the standard lesion spatial features at the current trigger time.

[0010] Furthermore, the analysis of standard normal characteristics at the current triggering time, based on the cumulative normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain from the MRI at the current triggering time, includes: The sum of the normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain at the current triggering time is taken as the equivalent normal feature. The ratio of the equivalent normal feature to the preset capacity benchmark is used as the standard normal feature at the current triggering time.

[0011] Furthermore, the fusion of the standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics yields a comprehensive risk index, including: The ratio of the square of the standard stress feature to the product of the standard lesion spatial feature and the standard normal feature is used as a comprehensive risk indicator.

[0012] Furthermore, determining the prognostic score based on the numerical characteristics of the comprehensive risk index includes: Based on the relative relationship between the comprehensive risk index and the preset full-scale risk benchmark, the prognostic score is obtained.

[0013] Furthermore, the prediction result output module is also used for: The data fitting algorithm is used to process the prognostic scores of the current trigger time and the historical trigger time within the preset backtracking window to obtain the score change rate and the basic score; Based on the score change rate and the base score, a linear calculation is performed for each future triggering time to obtain the prognostic score for each future triggering time.

[0014] Furthermore, after analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the three-dimensional matrix of the MRI brain, the method further includes: If the normal equivalent value of any voxel in the 3D matrix of the MRI brain is less than or equal to 0, then the normal equivalent value of any voxel is updated to a preset minimum positive number; if the normal equivalent value of any voxel in the 3D matrix of the MRI brain is greater than 0, then the normal equivalent value of any voxel is retained.

[0015] The present invention has the following beneficial effects: This invention, based on the cumulative EEG voltage values ​​of the target child within a post-intervention time window sequence at the current trigger moment, obtains the stress fluctuation characteristics at the current trigger moment. It effectively integrates the absolute amplitude shift of the EEG fluctuations with the duration of the abnormal state, achieving a reliable conversion of long-sequence AC fluctuation signals into a single stable scalar feature. This allows for a true and objective quantification of the overall overreaction load borne by the brain in response to external physical stimuli. Based on the data characteristics of the voxel grayscale values ​​in the three-dimensional matrix of the target child's MRI brain at the current trigger moment, the normal equivalent value of each voxel is analyzed, transforming visual image features into digital node attributes. This provides a foundational data basis for subsequent quantification of the spatial fragmentation distribution of abnormal regions.

[0016] By integrating the stress fluctuation characteristics of the current trigger moment and historical trigger moments, the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the 3D matrix of MRI brain, and the cumulative situation of the normal equivalent values ​​of voxels in the 3D matrix of MRI brain at the current trigger moment, a comprehensive risk index for the current trigger moment is obtained. The physical structure base in the spatial dimension applies constraints to the dynamic abnormal fluctuations in the temporal dimension, filtering out false EEG fluctuation noise amplified by the fragmentation of the spatial distribution of normal brain tissue, and improving the robustness of the risk assessment results output when facing high-frequency routine clinical nursing interference.

[0017] Based on the numerical characteristics of comprehensive risk indicators, prognostic scores are determined, transforming abstract absolute characteristic values ​​into relative risk scales. This achieves unified dimensions of multimodal data and alignment with individual assessment benchmarks, enhancing the reference value of prognostic scores and providing digital decision-making basis for medical staff to adjust nursing intervention strategies and deploy medical care plans in advance. Attached Figure Description

[0018] Figure 1 The diagram shows the structure of a multimodal fusion AI assessment and prognosis prediction system for neonatal brain injury severity, provided in one embodiment of the present invention. Figure 2 This is a module structure diagram of a comprehensive feature analysis module provided in one embodiment of the present invention. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the multimodal fusion AI assessment and prognosis prediction system for neonatal brain injury severity provided by this invention.

[0021] Please see Figure 1 The diagram illustrates a multimodal fusion AI assessment and prognosis prediction system for neonatal brain injury severity provided by an embodiment of the present invention. The system includes: a data acquisition module 101, a basic feature analysis module 102, a comprehensive feature analysis module 103, and a prediction result output module 104.

[0022] The data acquisition module 101 is used to acquire the post-intervention time window sequence and the MRI brain three-dimensional matrix of the target child at the current triggering time; the post-intervention time window sequence includes the EEG voltage value at each sampling time; the MRI brain three-dimensional matrix includes the gray value of each voxel.

[0023] This system receives nursing operation logs from the hospital information system, MRI brain files from the medical image archiving and communication system, and raw continuous EEG signals collected at a fixed frequency by the bedside EEG monitor of the target child. The raw continuous EEG signals collected by the bedside EEG monitor include the EEG voltage values ​​at each sampling time. The MRI brain file from the medical image archiving and communication system is a 3D MRI brain matrix generated from the brain scan of the target child. The three data dimensions of the 3D MRI brain matrix correspond to the sagittal, coronal, and vertical axes of the child's brain in physical space, respectively. The data units located by the 3D array index within the 3D MRI brain matrix are called voxels. The 3D array index of each voxel corresponds to the actual 3D physical location of the target child's brain. Each voxel records a gray value converted from the electromagnetic feedback intensity of the MRI signal, representing the MRI gray value at the corresponding 3D physical location of the target child's brain.

[0024] This system identifies instruction records in the nursing operation logs of the hospital information system that include suctioning or turning operations for the target child, and records such instructions with the exact timestamp of their generation in the hospital information system as the trigger time.

[0025] In this embodiment, to avoid data contamination caused by overlapping data time windows due to continuous and frequent nursing operations, the system activates an anti-contamination lockout mechanism upon identifying any trigger moment. Specifically, if a new trigger moment is identified within a preset time period after the initial trigger moment, the new trigger moment is forcibly ignored until the preset time period corresponding to the initial trigger moment has completely ended, at which point the identification of new trigger moments resumes.

[0026] The preset time length represents a complete coverage of the normal physiological time cycle of a newborn's nervous system after physical intervention, from the generation of an irritation response to its gradual decline and return to a calm and peaceful state.

[0027] Taking any trigger moment as the target trigger moment, and using the target trigger moment as an example, the system extracts the raw continuous EEG signal data at this time. A preset time length of raw continuous EEG signal is extracted from the future timeline. The EEG voltage values ​​within this preset time length are arranged in chronological order of sampling time to form the raw post-intervention sequence. To filter out baseline drift and high-frequency environmental electromagnetic noise, the system uses a digital bandpass filter to smooth the raw post-intervention sequence at the target trigger moment, preserving the raw continuous EEG signal within a preset frequency band. The preset frequency band represents the truncation of ultra-low frequency and high-frequency noise, and encompasses the most core EEG physiological frequency bands (such as Delta, Theta, Alpha, and Beta waves) that clinically reflect the central nervous system state.

[0028] Furthermore, after iterating through the filtered original post-intervention sequence, for any given sampling time, if the absolute value of the EEG voltage at that sampling time is greater than a preset amplitude threshold, then the EEG voltage at that sampling time is determined to be an electromyographic or displacement artifact caused by body movement, and the EEG voltage at that sampling time is discarded. A linear interpolation is then performed using the EEG voltage values ​​from two adjacent sampling times to obtain the EEG voltage value for that sampling time. The preset amplitude threshold represents the upper limit of the safe envelope of the neonatal resting physiological EEG potential. If the preset amplitude threshold is exceeded, it is likely due to non-physiological hardware extreme noise caused by the infant's violent struggle, large head displacement, or poor electrode contact, and therefore needs to be discarded. If the absolute value of the EEG voltage at that sampling time is less than or equal to the preset amplitude threshold, then the EEG voltage value at that sampling time is retained.

[0029] It should be understood that digital bandpass filtering and linear interpolation algorithms are well-known conventional data smoothing techniques in this field, and will not be elaborated further.

[0030] Specifically, if the sampling time is the first or last sampling time in the filtered original post-intervention sequence, since linear interpolation cannot be performed with only one adjacent sampling time, the EEG voltage values ​​obtained by the above processing of the adjacent sampling times are updated to the EEG voltage values ​​of the sampling time.

[0031] The original post-intervention sequence, after final filtering, smoothing, and data cleaning, will be used as the post-intervention time window sequence at the target trigger time.

[0032] Meanwhile, considering that the generation time of the MRI brain file in the medical image archiving and communication system is unlikely to coincide with the target trigger time, and that there may not be a corresponding MRI brain file at the target trigger time, if a corresponding MRI brain file exists at the target trigger time, the MRI brain 3D matrix in the MRI brain file corresponding to the target trigger time is used as the MRI brain 3D matrix at the target trigger time. If no corresponding MRI brain file exists at the target trigger time, the MRI brain file generated before the target trigger time and closest to the target trigger time in the medical image archiving and communication system is extracted, and the MRI brain 3D matrix in the MRI brain file is used as the MRI brain 3D matrix at the target trigger time. The closest MRI brain file refers to the historical MRI brain file with the smallest time difference between its generation time and the target trigger time.

[0033] In this embodiment of the invention, if the target child has just been admitted to the hospital and has not yet undergone an MRI scan, or if there is a communication interruption between the medical image archiving and communication system, resulting in the inability to extract the corresponding MRI brain file, it is determined that the current spatial feature data is missing, the current round of analysis is abandoned, a data missing prompt is issued, and the system continues to listen and wait for the next trigger moment.

[0034] In this embodiment of the invention, the nursing operation log of the hospital information system is searched to obtain the time interval between each trigger moment before the current absolute physical time and the current absolute physical time. The most recent trigger moment whose time interval is greater than or equal to a preset time length is taken as the current trigger moment. Essentially, this operation uses the current absolute physical time as a benchmark, and sequentially determines whether the time interval between each trigger moment and the current absolute physical time is greater than or equal to a preset time length, in reverse chronological order. The most recent trigger moment that meets the condition is taken as the current trigger moment.

[0035] In one specific implementation of this invention, the fixed frequency of the bedside EEG monitor is set to 250 Hz; the preset time length is set to 1800 seconds; the preset frequency band of the digital bandpass filter is set to 0.5 Hz to 30 Hz; and the preset amplitude threshold is set to 100 microvolts.

[0036] The basic feature analysis module 102 is used to obtain the stress fluctuation characteristics at the current triggering time based on the cumulative EEG voltage values ​​in the post-intervention time window sequence; and to analyze the normal equivalent value of each voxel based on the data characteristics of the gray values ​​of voxels in the three-dimensional matrix of the MRI brain.

[0037] Stress fluctuation characteristics characterize the intensity and duration of the abnormal stress-induced electrical discharge response in the central nervous system of the target child after being subjected to external nursing procedures (such as suctioning and turning over). The greater the stress fluctuation characteristics, the more severe the overreaction to external stimuli and the slower the subsidence.

[0038] Because transient EEG voltage values ​​at a single sampling moment may originate from random electromyographic interference or extremely short, sporadic lightning flashes, they cannot reflect the sustained state of the nervous system. However, the cumulative EEG voltage values ​​within the post-intervention time window represent the total stress energy released by the child's brain after external physical intervention, from the initial irritation response to the eventual return to calm, and the duration of this process. Only by determining the stress fluctuation characteristics based on the cumulative EEG voltage values ​​within the post-intervention time window can the absolute amplitude shift of EEG fluctuations be effectively integrated with the duration of the abnormal state. This allows for a reliable conversion of long-sequence AC fluctuation signals into a single stable scalar characteristic, thereby truly and objectively quantifying the overall overreaction load borne by the brain in response to external physical stimuli.

[0039] The normal equivalence value represents the percentage of remaining brain tissue within the physical space of a microbrain tissue unit (voxel) relative to normal brain tissue. The normal equivalence value ranges from 0 to 1 within a closed interval. A normal equivalence value closer to 1 indicates a denser physiological structure and a signal transduction state closer to the baseline level of normal brain tissue, without substantial ischemic or hypoxic damage. Conversely, a normal equivalence value closer to 0 indicates significant pathological changes, necrosis, or severe edema in the tissue structure of the corresponding microbrain tissue unit, suggesting a substantial loss of its normal physiological basis for carrying out neural activity.

[0040] The voxel grayscale values ​​in the three-dimensional matrix of the brain from MRI objectively reflect the water content and tissue density of brain tissue. The grayscale values ​​of brain injury lesions deviate significantly from the baseline level of normal brain tissue. By quantifying the degree to which the grayscale value of each voxel deviates from the baseline level of normal brain tissue, the extent of lesion damage of each voxel can be assessed and mapped to a normal equivalent value. This transforms visual image features into digital node attributes, providing the underlying data foundation for subsequent quantification of the spatial fragmentation distribution of abnormal regions.

[0041] Preferably, in some implementations of the present invention, the stress fluctuation characteristics at the current triggering moment are obtained based on the cumulative EEG voltage values ​​in the post-intervention time window sequence, including: The stress fluctuation characteristics at the current triggering time are obtained by multiplying the absolute value of the difference between the EEG voltage value at each sampling moment in the post-intervention time window sequence and the preset EEG baseline value, and then multiplying the result with the time interval between any two adjacent sampling moments.

[0042] The preset EEG baseline value represents the relative baseline EEG activity level of the target child before this nursing intervention. The absolute value of the difference between the EEG voltage value at each sampling time and the preset EEG baseline value represents the transient absolute deviation and transient abnormal discharge intensity of the electrical signal in the target child's brain caused by external stimulation from that relative baseline level at each sampling time. The sum of the absolute values ​​of the differences between the EEG voltage values ​​at all sampling times and the preset EEG baseline value represents the direct linear sum of all transient abnormal discharge intensities in the target child throughout the entire post-intervention time window, reflecting the total scale of the target child's brain's sustained abnormal activity during stimulation.

[0043] Since the summation value is directly affected by the sampling frequency of different EEG devices (the higher the sampling frequency, the larger the summation value), multiplying the sum by a fixed time interval can eliminate the dimensional interference caused by the difference in the hardware parameters of the sampling devices. This effectively integrates the absolute amplitude of the transient discharge with the duration of the abnormal irritation state, so that the obtained stress fluctuation characteristics can accurately quantify the total stress energy and overall overreaction load released by the target child's brain in response to external physical stimuli.

[0044] As can be seen from the data acquisition module 101, the fixed frequency of the bedside EEG monitor in this embodiment is set to 250 Hz, so the time interval between any two adjacent sampling times is 0.004 seconds.

[0045] In one specific implementation of this invention, the method for obtaining the preset EEG baseline value is as follows: taking the current triggering time as an example, the current triggering time is used as the data extraction time of the original continuous EEG signal. The original EEG signal of a preset duration (e.g., 600 seconds) is extracted backwards in the historical time direction as the original pre-intervention sequence. The original pre-intervention sequence is then filtered, smoothed, and cleaned according to the method of the data acquisition module 101 to obtain the pre-intervention time window sequence. All EEG voltage values ​​in the pre-intervention time window sequence are sorted in ascending order, and the median of the sorted EEG voltage values ​​is extracted as the preset EEG baseline value at the current triggering time. In this embodiment, the preset EEG baseline value is 25 microvolts.

[0046] Preferably, in some implementations of the present invention, analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the three-dimensional matrix of the MRI brain includes: Based on the degree of deviation between the gray value of each voxel in the three-dimensional matrix of the MRI brain and the preset reference gray value, the lesion equivalent value of each voxel is determined. The normal equivalent value of each voxel is determined based on the negative correlation coefficient of the pathological equivalent value of each voxel.

[0047] The preset reference grayscale baseline represents the baseline level and standard imaging characteristics of normal brain tissue in a relatively normal state during the early stages of the disease in the target child. The degree of deviation of the grayscale value of each voxel from the preset reference grayscale baseline represents the range of change in water content or cell density within the corresponding micro-brain tissue unit, reflecting the severity of pathological changes such as edema and necrosis in the corresponding brain tissue region due to ischemia and hypoxia (i.e., lesion equivalence value). The larger the lesion equivalence value, the deeper the degree of substantial lesion in the corresponding brain tissue region.

[0048] Specifically, the absolute difference between the grayscale value of each voxel in the current triggering time's 3D MRI brain matrix and a preset reference grayscale benchmark is divided by the preset reference grayscale benchmark to obtain the lesion equivalence value for each voxel. The result of dividing the absolute difference between the grayscale value of each voxel and the preset reference grayscale benchmark by the preset reference grayscale benchmark reflects the degree of deviation of the grayscale value of each voxel from the preset reference grayscale benchmark.

[0049] Since the lesion equivalence value is a negative indicator, the larger the value, the more severe the brain tissue lesion damage. Therefore, in this embodiment, the difference between 1 and the lesion equivalence value of each voxel is used as the normal equivalence value of each voxel. The negative correlation coefficient corresponding to the lesion equivalence value is obtained through this subtraction operation, so as to realize the transformation of the lesion equivalence value representing the degree of damage into the normal equivalence value representing the degree of tissue integrity. This provides consistent and positive basic data support for the comprehensive feature analysis module, so that the accumulation of normal equivalence values ​​can truly reflect the basic fault tolerance capacity reserve of the target child's brain at the current triggering time. Moreover, the joint distribution characteristics of the normal equivalence values ​​of adjacent voxels can accurately represent the connectivity integrity of normal brain tissue in three-dimensional space, so that the obtained comprehensive risk index can objectively and accurately quantify the actual brain injury risk of the target child.

[0050] In this embodiment, before analyzing the normal equivalent value of each voxel, voxels with a gray value of 0 (representing a pure black area of ​​the scan background) in the 3D matrix of the MRI brain at the current trigger time are removed.

[0051] In one specific implementation of this invention, the method for obtaining the preset reference grayscale benchmark is as follows: Extract the first MRI brain image created upon admission of the target child; remove voxels with a grayscale value of zero from the 3D matrix of the MRI brain image in this image to obtain an effective voxel set. Sort all grayscale values ​​in the effective voxel set in ascending order and extract the grayscale value located at the 90th percentile as the preset reference grayscale benchmark. In this embodiment, the preset reference grayscale benchmark obtained is 1850.

[0052] Furthermore, after analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of the voxels in the three-dimensional matrix of the MRI brain, the process also includes: If the normal equivalent value of any voxel in the 3D matrix of the brain from the MRI at the current trigger moment is less than or equal to 0, it means that the microbrain tissue unit corresponding to that voxel has undergone extremely severe substantial necrosis or extreme edema, and its physiological function of carrying neural activity has been almost completely lost. In this case, in order to prevent zero-value avalanche collapse or negative algebraic deadlock when calculating the feature product of each voxel with adjacent voxels in the subsequent calculation, the normal equivalent value of that voxel is updated to a preset minimum positive number; if the normal equivalent value of that voxel is greater than 0, it means that the microbrain tissue unit corresponding to that voxel still has some basic physiological structure or is in a relatively normal state, and its calculated value is within a reasonable physical definition range, then the normal equivalent value of that voxel is retained.

[0053] In one specific implementation of this invention, the preset minimum positive number is set to 0.01.

[0054] The comprehensive feature analysis module 103 is used to integrate the stress fluctuation characteristics of the current triggering time and the historical triggering time, the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the MRI brain, and the cumulative situation of the normal equivalent values ​​of voxels in the three-dimensional matrix of the MRI brain at the current triggering time, to obtain the comprehensive risk index at the current triggering time.

[0055] The comprehensive risk index characterizes the actual risk of neurological damage to the brain when subjected to external nursing interventions. The higher the comprehensive risk index, the weaker the target child's tolerance to interference based on the existing brain pathological structure, and the higher the risk of brain damage worsening in the short term.

[0056] The stress fluctuation characteristics at the current and historical trigger times reflect the scale and continuous evolution of abnormal discharges in the nervous system of the target child's brain when subjected to continuous external care stimuli, characterizing the dynamic stress load actually borne by the brain at the functional level; the joint distribution characteristics of the normal equivalent values ​​of each voxel and adjacent voxels reflect the connectivity integrity of normal brain tissue in three-dimensional space and the degree of fragmentation of lesions, characterizing the topological network structure of the brain's neural signal transmission pathways; the cumulative situation of the normal equivalent values ​​of voxels in the three-dimensional matrix of the brain in MRI reflects the total equivalent physical volume of undamaged normal brain cells in the child's brain, characterizing the basic fault tolerance capacity reserve of the child's brain at the physical structural level to resist external stimuli and maintain stable neural function.

[0057] The comprehensive risk index obtained by integrating the three can constrain the dynamic abnormal fluctuations in the time dimension with the physical structure base in the spatial dimension. It filters out false EEG fluctuation noise amplified by the fragmentation of the spatial distribution of normal brain tissue, and improves the robustness of the output risk assessment index when facing high-frequency routine clinical nursing interference.

[0058] Preferably, such as Figure 2 As shown, in some implementations of the present invention, the comprehensive feature analysis module 103 includes a first feature unit 201, a second feature unit 202, a third feature unit 203, and a comprehensive feature unit 204.

[0059] The first feature unit 201 is used to perform a normalization operation on the stress fluctuation characteristics of the current triggering time based on the stress fluctuation characteristics of the current triggering time and the historical triggering time within a preset backtracking window, so as to obtain the standard stress characteristics of the current triggering time.

[0060] The standard stress characteristic represents the dynamic stress load experienced by the target child at the current triggering moment, and its relative intensity level in the recent overall historical trend, with a value ranging from 0 to 1.

[0061] Because the stress fluctuation characteristics in the time dimension (unit: microvolt-second) are completely different from the physical dimensions of the normal equivalent value in the spatial dimension, and the absolute values ​​vary greatly, directly inputting the heterogeneous absolute values ​​into downstream cross-modal joint computation would cause the large-scale features to completely overwhelm the small-scale features. Therefore, by utilizing the stress fluctuation characteristics of the current trigger time and historical trigger times within a preset backtracking window, a normalization operation is performed on the stress fluctuation characteristics of the current trigger time to avoid numerical scale imbalance in cross-modal computation.

[0062] Specifically, the maximum value of the stress fluctuation characteristics between the current trigger time and historical trigger times within the preset backtracking window is taken as the maximum stress value of the window. The stress fluctuation characteristics at the current trigger time are divided by the sum of the maximum stress value of the window and the preset fluctuation benchmark to obtain the standard stress characteristics at the current trigger time. The preset fluctuation benchmark represents the background fluctuation scale of the target child's EEG in a resting state without external nursing intervention. Through the above normalization method including the benchmark constant constraint, local relative characteristics can be anchored to an objective physical base, and the absolute physical scale difference between continuous low fluctuations and continuous high fluctuations can be accurately preserved, ultimately outputting the standard stress characteristics in the range of 0 to 1.

[0063] In one specific implementation of this invention, the length of the preset retrospective window is set to 72 hours. The preset fluctuation benchmark is obtained as follows: during the resting period of the target child without external nursing intervention, a resting time window sequence of equal length to the post-intervention time window sequence is extracted according to the method of data acquisition module 101. The method for calculating stress fluctuation characteristics in this embodiment is used to process the EEG voltage values ​​in the resting time window sequence, and the obtained stress fluctuation characteristics are used as the preset fluctuation benchmark. In this embodiment, the preset fluctuation benchmark is 18000 microvolts per second.

[0064] In this embodiment of the invention, if the total number of trigger moments within the preset backtracking window is less than 3, it indicates that the number of available trigger moment samples is too small to support statistically significant normalization analysis. In this case, the current preset backtracking window is suspended, the evaluation and analysis of the current trigger moment is abandoned, and the system continues to listen for and wait for the next trigger moment until the total number of trigger moments within the preset backtracking window is greater than or equal to 3. Then, the normalization operation and subsequent analysis process are resumed.

[0065] The second feature unit 202 is used to determine the standard lesion spatial features at the current triggering time based on the joint distribution features of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the MRI brain at the current triggering time and the historical triggering time within the preset backtracking window.

[0066] The standard lesion spatial features represent the degree of spatial fragmentation of the abnormal ischemic lesion distribution in the target child at the current triggering moment, and its relative quantitative level in recent historical evolution, reflecting the relatively fragmented state of the neural signal transmission network cut by the lesion, with values ​​ranging from 0 to 1. The joint distribution characteristics of each voxel and the normal equivalent values ​​of adjacent voxels reflect the connectivity integrity of normal brain tissue in three-dimensional space and the degree of fragmentation of the ischemic lesion region.

[0067] Since the brain's ability to regulate itself in response to external stimuli depends not only on the total volume of the remaining normal brain cells, but also on whether these normal brain cells are spatially connected, highly fragmented lesions can sever normal neural regulatory pathways, causing small-scale stimuli to trigger violent discharge fluctuations. Therefore, it is necessary to analyze the standard lesion spatial characteristics at the current triggering time based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the 3D matrix of the MRI brain at the current and historical triggering times within the preset retrospective window. This will filter out amplitude clutter interference caused by differences in the spatial distribution of lesions and improve the logical stability and robustness of subsequent calculations of comprehensive risk indicators.

[0068] Furthermore, based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the 3D matrix of the MRI brain at the current and historical trigger times within the preset retrospective window, the standard lesion spatial characteristics at the current trigger time are determined, including: Multiply the normal equivalent value between each voxel in the 3D matrix of the brain at any trigger time within the preset backtracking window and each adjacent voxel to obtain the effective product between each voxel and each adjacent voxel. The ratio of the sum of all effective products at any given triggering time to the total number of effective products is used as the lesion spatial feature at any given triggering time. Based on the lesion spatial features at the current trigger time and the historical trigger time within the preset backtracking window, a normalization operation is performed on the lesion spatial features at the current trigger time to obtain the standard lesion spatial features at the current trigger time.

[0069] The effective product between each voxel and each of its neighboring voxels represents the probability of cooperation and joint carrying capacity of any two adjacent voxel units simultaneously being in a relatively normal state within a small spatial locality.

[0070] Specifically, in this embodiment, taking any voxel in the three-dimensional matrix of the brain at any trigger time within the preset backtracking window as an example, this voxel is taken as the central voxel. The well-known three-dimensional 26-neighborhood determination criterion is applied to extract the 26 surrounding voxels adjacent to the central voxel in the three-dimensional physical space. The product of the normal equivalent value between the central voxel and each of the 26 surrounding voxels is calculated to obtain the 26 effective products of the central voxel.

[0071] Specifically, if the central voxel is located at the edge of the three-dimensional matrix of the brain in MRI or at the boundary of brain tissue, and the number of adjacent voxels is less than 26, then only the actual adjacent voxels are extracted, and the product of the normal equivalent value between the central voxel and each actual adjacent voxel is calculated to obtain the actual effective product.

[0072] The sum of all effective products represents the total scale of all adjacent microscopic tissue connections with normal signal transmission capabilities within the three-dimensional space of the target child's whole brain. The ratio of this to the total number of effective products is used as the spatial feature of the lesion. This can eliminate the dimensional and order-of-magnitude differences caused by the absolute volume scale of MRI scans, objectively quantify the average connectivity integrity of the child's whole brain neural conduction network and the degree of fragmentation of abnormal lesions, and ensure that the spatial distribution features extracted at different triggering times have consistent comparability.

[0073] The spatial characteristics of lesions reflect the overall fragmentation degree of the brain's neural signal transmission topology network caused by the abnormal ischemic lesion area within the target child's brain. A larger spatial characteristic indicates a higher proportion of contiguous normal brain cells, a more concentrated distribution of abnormal ischemic lesions, and less severe disruption of the overall neural transmission network, resulting in lower spatial fragmentation. Conversely, a smaller spatial characteristic indicates a chaotic state where normal and diseased brain cells are numerous and intersecting in three-dimensional space, with the abnormal ischemic lesions scattered like loose sand, fragmenting the originally complete neural signal transmission network, resulting in higher spatial fragmentation.

[0074] In order to unify the numerical scale with the standard stress characteristics and avoid dimensional imbalance during subsequent cross-modal joint calculations, it is necessary to analyze the standard lesion spatial characteristics at the current trigger time by combining the lesion spatial characteristics at the current trigger time and the historical trigger times within the preset backtracking window, and convert the absolute volume into a dimensionless relative ratio.

[0075] Specifically, the maximum value of the lesion spatial features at the current trigger time and the historical trigger time within the preset backtracking window is taken as the maximum value of the lesion; the minimum value of the lesion spatial features at the current trigger time and the historical trigger time within the preset backtracking window is taken as the minimum value of the lesion; the maximum and minimum values ​​of the lesion are used to perform maximum and minimum value normalization on the lesion spatial features at the current trigger time to obtain the standard lesion spatial features at the current trigger time.

[0076] It should be understood that the method of normalizing the original data using maximum and minimum value normalization is a well-known technique and will not be elaborated further.

[0077] Furthermore, if the obtained standard lesion spatial feature is equal to 0, the standard lesion spatial feature is updated to a preset minimum positive constant (such as 0.001) to ensure that when it is used as a constraint denominator in the formula for calculating the comprehensive risk index, it does not cause the system to crash due to the phenomenon of division by zero.

[0078] Specifically, if the maximum value of the lesion within the preset backtracking window equals the minimum value, it indicates that the spatial fragmentation of the distribution of abnormal ischemic lesions in the brain of the target child remains absolutely stable at different trigger times within the entire preset backtracking window, without any evolution or diffusion. In this case, the denominator of the traditional maximum-minimum normalization formula will be zero. To prevent division-by-zero deadlock during normalization calculation, the spatial characteristics of the lesion at the current trigger time are directly used as the standard spatial characteristics of the lesion to ensure that subsequent cross-modal joint calculations can proceed smoothly and stably.

[0079] The third feature unit 203 is used to analyze the standard normal features at the current triggering time based on the accumulation of normal equivalent values ​​of all voxels in the three-dimensional matrix of the MRI brain at the current triggering time.

[0080] Standard normal features represent the remaining normal brain cell tissue volume of the target child at the current triggering time, and the relative preservation level in the initial admission state of the target child, with a value ranging from 0 to 1. The cumulative normal equivalent values ​​of all voxels in the 3D matrix of the brain from MRI represent the total equivalent physical volume of the currently undamaged normal brain cell tissue in the target child's brain.

[0081] Since the brain's physical resilience and tolerance to external stimuli depend on the total physical volume of undamaged normal brain cells, the function of a single micro-brain unit cannot reflect the overall stress-bearing base of the nervous system. Furthermore, the absolute value generated by the accumulation of normal equivalent values ​​of all brain voxels is extremely large, and it differs greatly from the stress fluctuation characteristics of the time dimension in terms of physical dimensions and numerical scale. Directly incorporating it into calculations would cause the massive spatial features to completely overwhelm the temporal features. Therefore, it is necessary to analyze the standard normal characteristics at the current triggering moment based on the accumulation of normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain from the MRI at the current triggering moment, to ensure the logical stability and robustness of subsequent calculations of comprehensive risk indicators.

[0082] Furthermore, based on the cumulative normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain from the MRI at the current triggering time, the standard normal characteristics at the current triggering time are analyzed, including: The sum of the normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain at the current triggering time is taken as the equivalent normal feature. The ratio of the equivalent normal feature to the preset capacity benchmark is used as the standard normal feature at the current triggering time.

[0083] The equivalent normal characteristics represent the total equivalent physical volume of undamaged normal brain cells in the target child's brain at the corresponding trigger time. This reflects the brain's basic tolerance capacity reserve at the physical structural level to resist external stimuli and maintain stable neural function. The preset capacity benchmark represents the theoretical maximum effective tissue volume of healthy brain tissue in the target child's initial state upon admission. Dividing the equivalent normal characteristics at the current trigger time by the preset capacity benchmark yields the standard normal characteristics at the current trigger time. This division operation converts the absolute capacity value into a relative proportion of normal brain cell tissue preserved within the 0-1 range, eliminating individual data differences caused by the congenital brain size of different children.

[0084] In one specific implementation of this invention, the method for obtaining the preset capacity benchmark is as follows: The first MRI brain image created upon admission of the target child is obtained. Voxels with a grayscale value of 0 are removed from the three-dimensional matrix of the MRI brain image in this image, resulting in an effective voxel set. The total number of effective voxels is multiplied by the theoretical upper limit of the normal equivalent value (1), and the result is used as the preset capacity benchmark. In this embodiment, the preset capacity benchmark is 286000.

[0085] The comprehensive feature unit 204 is used to fuse the standard stress features, standard lesion spatial features and standard normal features to obtain a comprehensive risk index.

[0086] The comprehensive risk index represents the actual risk of neurological damage and tolerance to interference faced by the target child's brain when subjected to external nursing interventions.

[0087] The comprehensive risk index, which integrates standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics, is cross-modal constrained from three dimensions: stress load at the dynamic functional level, conduction pathway integrity at the topological network level, and basic fault tolerance capacity at the static physical structure level. This effectively filters out false EEG discharge noise caused solely by differences in local lesion location or breaks in normal conduction pathways, and avoids irregular jumps in risk assessment results due to abnormalities in a single modality of data. This improves the computational robustness and output objectivity and accuracy of the multimodal heterogeneous data fusion assessment system.

[0088] Furthermore, by integrating the aforementioned standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics, a comprehensive risk index is obtained, including: The ratio of the square of the standard stress feature to the product of the standard lesion spatial feature and the standard normal feature is used as a comprehensive risk indicator.

[0089] Using the square of the standard stress characteristics as the numerator is to exponentially amplify the dynamic abnormal fluctuations in the time dimension, making the system more sensitive to severe stress-induced abnormal discharges, thus highlighting the substantial functional impact of external physical intervention on the child's brain. Using the product of the standard lesion spatial characteristics and the standard normal characteristics as the denominator is because the brain's resistance to external stimuli depends on both sufficient reserves of normal brain cell tissue and a complete spatial connectivity network; the product of these two constitutes the combined load-bearing base of the child's overall physical structure. A nonlinear algebraic penalty mechanism for spatial constraints in time is constructed using ratio calculation. Its underlying physical logic is that when the target child's normal brain tissue is sharply reduced or the lesion space is highly fragmented, leading to numerous disruptions in conduction pathways (i.e., the denominator is extremely small), even a tiny external stimulus causing slight EEG fluctuations will be amplified exponentially by this division mechanism. This calculation method can adaptively correct the surface EEG amplitude using the fragility of the underlying organizational structure, thereby objectively quantifying the risk of collapse in an extremely fragile brain under routine clinical care.

[0090] The prediction result output module 104 is used to determine the prognosis score based on the numerical characteristics of the comprehensive risk index; the prognosis score is used to predict the degree of brain injury in newborns.

[0091] The prognostic score represents the severity of the current brain injury in the target child and the probability of the condition worsening in the short term.

[0092] The comprehensive risk index is a numerical value representing the underlying absolute risk characteristics, integrating the dynamic stress load over time and the spatial physical bearing capacity. Determining prognostic scores based on the numerical characteristics of the comprehensive risk index transforms abstract absolute characteristic values ​​into relative risk scales, achieving unified dimensions for multimodal data and alignment with individual assessment benchmarks.

[0093] Preferably, in some implementations of the embodiments of the present invention, determining the prognostic score based on the numerical characteristics of the comprehensive risk index includes: Based on the relative relationship between the comprehensive risk index and the preset full-scale risk benchmark, the prognostic score is obtained.

[0094] The preset full-scale risk benchmark represents the theoretical upper limit and extreme threshold of physical risk that the target child's brain and nervous system can withstand under routine clinical intervention stimulation. The comprehensive risk index at the current trigger moment characterizes the actual degree of neurological damage risk faced by the brain under external nursing intervention at this time. The prognostic score, obtained based on the relative relationship between the comprehensive risk index at the current trigger moment and the preset full-scale risk benchmark, represents the percentage degree to which the target child's absolute neurological damage risk at the current trigger moment approaches the theoretical collapse threshold. This operation eliminates the abstract numerical values ​​and lack of upper boundary characteristics of the underlying comprehensive risk index generated by calculation, transforming it into an intuitive watermark scale between 0 and 100 points. The higher the prognostic score, the greater the probability of the target child suffering irreversible brain damage at this time, providing clinicians with a unified dimensional early warning indicator reflecting the deterioration of the condition.

[0095] Specifically, the ratio of the comprehensive risk index to the preset full-scale risk benchmark is multiplied by 100% to obtain the prognostic score; the ratio of the comprehensive risk index to the preset full-scale risk benchmark reflects the relative relationship between the comprehensive risk index and the preset full-scale risk benchmark.

[0096] Specifically, if the calculated prognostic score is greater than 100, the prognostic score is directly set to 100.

[0097] In one specific implementation of this invention, the method for obtaining the preset full-scale risk benchmark is as follows: extract historical comprehensive risk index records of all children with severe hypoxic-ischemic encephalopathy pre-stored in the local database of this system, and use the maximum value among them as the preset full-scale risk benchmark. In this embodiment, the preset full-scale risk benchmark is set to 150.0.

[0098] Specifically, if historical comprehensive risk indicator records are missing in the local database (e.g., when the system is first deployed and sufficient historical cases have not yet been accumulated), the pre-configured theoretical extreme value constant (e.g., 100.0) will be used directly as the preset full-scale risk benchmark. This theoretical extreme value constant represents the most extreme risk-bearing limit set in the system's underlying mathematical model.

[0099] Furthermore, the prediction result output module 104 is also used for: The data fitting algorithm is used to process the prognostic scores of the current trigger time and the historical trigger time within the preset backtracking window to obtain the score change rate and the basic score; Based on the score change rate and the base score, a linear calculation is performed for each future triggering time to obtain the prognostic score for each future triggering time.

[0100] Because a prognostic score at a single trigger point only reflects the brain injury risk status of the target child at a particular instant, in a real clinical setting, a single prognostic score is easily affected by local perturbations and will fluctuate occasionally. Furthermore, the pathological evolution of neonatal brain injury is a continuous and dynamic process, and its short-term deterioration or improvement has objective continuity and evolutionary inertia over time. Therefore, this embodiment predicts several prognostic scores for a future period based on several prognostic scores over a historical period. This not only effectively filters out errors caused by single abnormal perturbations and extracts the rate of change of scores representing the true slope of disease progression and the baseline score, but also upgrades the lagging post-hoc assessment to a prospective pre-hoc warning, helping medical staff to anticipate the deterioration of the child's brain injury and gain valuable time for emergency medical intervention.

[0101] Specifically, in this embodiment, the trigger time within the preset backtracking window is used as the independent variable, the prognostic score corresponding to each trigger time is used as the dependent variable, the least squares method is used for linear fitting, the slope of the fitted line is used as the rate of change of the score, and the intercept of the fitted line is used as the basic score.

[0102] It should be understood that the method of using the least squares method to fit a univariate linear regression and to calculate the slope and intercept of the equation is a well-known conventional data processing technique in this field, and will not be elaborated further.

[0103] Essentially, it constructs a univariate linear function with the trigger time as the independent variable and the prognostic score as the dependent variable. The specific linear calculation process is as follows: multiply each trigger time by the rate of change of the score, then add the product to the base score, and the sum is the prognostic score for each trigger time.

[0104] Furthermore, using the current triggering time as the baseline, a preset prediction time interval is extended forward along the physical time axis. Within this prediction time interval, multiple discrete moments are selected at equal intervals according to a preset time step, serving as each future triggering moment. The product of each future triggering moment and the rate of change of the score is added to the base score to obtain the prognostic score for each future triggering moment.

[0105] The preset time step eliminates the bias caused by uneven time scale, and can provide uniform, smooth and continuous extrapolation prediction sequences.

[0106] In one specific implementation of this invention, the preset prediction duration interval is set to 24 hours; the preset time step is set to 1 hour.

[0107] In this embodiment of the invention, the system constructs a two-dimensional visualization coordinate system. The horizontal axis of the two-dimensional visualization coordinate system corresponds to the time process in the physical world, and the vertical axis represents the prognostic score ranging from 0 to 100 points. The prognostic scores at the current triggering time and historical triggering times are plotted as scatter points in the two-dimensional visualization coordinate system and connected by a solid line; the prognostic scores at future triggering times are plotted as scatter points in the two-dimensional visualization coordinate system and connected by a dashed line, outputting a complete dynamic curve that includes historical evolution and future warnings. This complete dynamic curve is used to intuitively and comprehensively display the real-time fluctuation and evolutionary inertia of the brain injury risk of the target child to clinical medical staff, providing digital decision-making basis for medical staff to adjust nursing intervention strategies and deploy medical rescue plans in advance.

[0108] In summary, this invention obtains the stress fluctuation characteristics at the current triggering time based on the cumulative EEG voltage values ​​in the post-intervention time window sequence at the current triggering time; it analyzes the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the 3D MRI brain matrix at the current triggering time; and it integrates the stress fluctuation characteristics of the current triggering time and historical triggering times, the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the 3D MRI brain matrix, and the cumulative status of the normal equivalent values ​​of voxels in the 3D MRI brain matrix at the current triggering time to obtain a comprehensive risk index for the current triggering time and determine the prognostic score for the current triggering time.

Claims

1. A multimodal fusion AI-based assessment and prognostic prediction system for neonatal brain injury severity, characterized in that, The system includes: The data acquisition module is used to acquire the post-intervention time window sequence and the MRI brain three-dimensional matrix of the target child at the current triggering time; the post-intervention time window sequence includes the EEG voltage value at each sampling time; the MRI brain three-dimensional matrix includes the gray value of each voxel; The basic feature analysis module is used to obtain the stress fluctuation characteristics at the current triggering moment based on the cumulative EEG voltage values ​​in the post-intervention time window sequence; and to analyze the normal equivalent value of each voxel based on the data characteristics of the gray values ​​of voxels in the MRI brain three-dimensional matrix. The comprehensive feature analysis module is used to integrate the stress fluctuation characteristics of the current triggering time and the historical triggering time, the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the MRI brain, and the cumulative situation of the normal equivalent values ​​of voxels in the three-dimensional matrix of the MRI brain at the current triggering time, to obtain the comprehensive risk index at the current triggering time. The prediction result output module is used to determine the prognostic score based on the numerical characteristics of the comprehensive risk index; the prognostic score is used to predict the degree of neonatal brain injury.

2. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, The stress fluctuation characteristics at the current triggering moment are obtained based on the cumulative EEG voltage values ​​in the post-intervention time window sequence, including: The stress fluctuation characteristics at the current triggering time are obtained by multiplying the absolute value of the difference between the EEG voltage value at each sampling moment in the post-intervention time window sequence and the preset EEG baseline value, and then multiplying the result with the time interval between any two adjacent sampling moments.

3. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, The step of analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the three-dimensional matrix of the MRI brain includes: Based on the degree of deviation between the gray value of each voxel in the three-dimensional matrix of the MRI brain and the preset reference gray value, the lesion equivalent value of each voxel is determined. The normal equivalent value of each voxel is determined based on the negative correlation coefficient of the pathological equivalent value of each voxel.

4. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, The combined distribution characteristics of stress fluctuations at the current and historical trigger times, the normal equivalent values ​​of each voxel in the 3D MRI brain matrix, and the cumulative normal equivalent values ​​of voxels in the 3D MRI brain matrix at the current trigger time are used to obtain a comprehensive risk index for the current trigger time, including: Based on the stress fluctuation characteristics of the current trigger time and the historical trigger time within the preset backtracking window, a normalization operation is performed on the stress fluctuation characteristics of the current trigger time to obtain the standard stress characteristics of the current trigger time. Based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the brain at the current trigger time and the historical trigger time within the preset retrospective window, the standard lesion spatial characteristics at the current trigger time are determined. Based on the cumulative normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain at the current triggering time, analyze the standard normal characteristics at the current triggering time. By integrating the standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics, a comprehensive risk index is obtained.

5. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 4, characterized in that, The determination of standard lesion spatial features at the current triggering time, based on the joint distribution characteristics of the normal equivalent values ​​of each voxel and its adjacent voxels in the three-dimensional matrix of the MRI brain at the current triggering time and at historical triggering times within a preset retrospective window, includes: Multiply the normal equivalent value between each voxel in the 3D matrix of the brain at any trigger time within the preset backtracking window and each adjacent voxel to obtain the effective product between each voxel and each adjacent voxel. The ratio of the sum of all effective products at any given triggering time to the total number of effective products is used as the lesion spatial feature at any given triggering time. Based on the lesion spatial features at the current trigger time and the historical trigger time within the preset backtracking window, a normalization operation is performed on the lesion spatial features at the current trigger time to obtain the standard lesion spatial features at the current trigger time.

6. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 4, characterized in that, The analysis of standard normal characteristics at the current triggering time, based on the cumulative normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain from the MRI at the current triggering time, includes: The sum of the normal equivalent values ​​of all voxels in the three-dimensional matrix of the brain at the current triggering time is taken as the equivalent normal feature. The ratio of the equivalent normal feature to the preset capacity benchmark is used as the standard normal feature at the current triggering time.

7. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 4, characterized in that, The fusion of the standard stress characteristics, standard lesion spatial characteristics, and standard normal characteristics yields a comprehensive risk index, including: The ratio of the square of the standard stress feature to the product of the standard lesion spatial feature and the standard normal feature is used as a comprehensive risk indicator.

8. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, The determination of the prognostic score based on the numerical characteristics of the comprehensive risk index includes: Based on the relative relationship between the comprehensive risk index and the preset full-scale risk benchmark, the prognostic score is obtained.

9. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, The prediction result output module is also used for: The data fitting algorithm is used to process the prognostic scores of the current trigger time and the historical trigger time within the preset backtracking window to obtain the score change rate and the basic score; Based on the score change rate and the base score, a linear calculation is performed for each future triggering time to obtain the prognostic score for each future triggering time.

10. The multimodal fusion AI assessment and prognostic prediction system for neonatal brain injury severity according to claim 1, characterized in that, After analyzing the normal equivalent value of each voxel based on the data characteristics of the grayscale values ​​of voxels in the three-dimensional matrix of the MRI brain, the process further includes: If the normal equivalent value of any voxel in the 3D matrix of the MRI brain is less than or equal to 0, then the normal equivalent value of any voxel is updated to a preset minimum positive number; if the normal equivalent value of any voxel in the 3D matrix of the MRI brain is greater than 0, then the normal equivalent value of any voxel is retained.

Images