A cerebral hemorrhage prognosis analysis system fusing images and clinical data

Abstract

The application discloses a cerebral hemorrhage prognosis analysis system fusing images and clinical data, constructs a multi-modal time sequence correlation data group by aligning time points when collecting cerebral hemorrhage image data and clinical diagnosis and treatment data; carries out hematoma region segmentation and surrounding structure compression analysis, extracts hematoma morphological features and space-occupying effect features to form an image risk feature set; carries out index fluctuation analysis and collaborative abnormal pattern recognition based on clinical data, screens risk sensitive indexes having a linkage relationship to form a clinical risk factor set; carries out consistency detection on the two types of risk features and traces causes of deviation, determines modal weight configuration and executes differentiated fusion to generate a fusion prognosis feature vector; carries out two-dimensional risk stratification of functional prognosis and survival prognosis, evaluates prediction confidence in combination with modal consistency and feature stability, and outputs a structured prognosis analysis report with a confidence level, so as to provide an interpretable decision support for cerebral hemorrhage patient prognosis evaluation.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data. Background Technology

[0002] Intracerebral hemorrhage is a common acute and critical condition in neurosurgery, characterized by its sudden onset, rapid progression, and highly variable prognosis. Accurate assessment of the prognosis of patients with intracerebral hemorrhage is crucial for developing appropriate treatment plans and optimizing the allocation of medical resources. Currently, in clinical practice, doctors primarily rely on cranial CT images to observe the location and extent of the hemorrhage, combined with the patient's level of consciousness, vital signs, and other clinical manifestations, to make empirical judgments. This assessment method has limitations such as strong subjectivity and significant individual variability, making it difficult to arrive at standardized and quantifiable prognostic conclusions.

[0003] Existing prognostic studies for intracerebral hemorrhage often focus on single data sources, analyzing only imaging features such as hemorrhage volume and lesion size, or only clinical indicators such as neurological function scores at admission. However, the prognosis of intracerebral hemorrhage is influenced by both imaging findings and clinical status, and single-modal analysis struggles to capture the complementary and corroborative relationships between these two types of information. When imaging features and clinical manifestations point to inconsistent prognostic conclusions, how to rationally integrate conflicting information and assess the reliability of the prediction results is a core challenge facing current prognostic assessment technologies. Summary of the Invention

[0004] This invention discloses a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data. It aims to achieve the joint extraction of imaging risk features and clinical risk factors through temporal integration and collaborative analysis of multimodal data, and to perform adaptive weight fusion based on intermodal consistency detection. Finally, it outputs a stratified prognostic conclusion with confidence assessment, providing a reliable quantitative reference for clinical decision-making.

[0005] The first aspect of this invention proposes a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data. This system performs the following steps to achieve prognostic analysis of cerebral hemorrhage:

[0006] Collect imaging data and clinical diagnosis and treatment data of cerebral hemorrhage, and perform time-point alignment of the acquisition data and the clinical diagnosis and treatment data to construct a time-series related data group;

[0007] Based on the time-series correlated data set, the hematoma region is segmented to obtain the spatial distribution of the hematoma. The spatial distribution of the hematoma is used to perform compression analysis of the surrounding structures to obtain structural offset features. Based on the spatial distribution of the hematoma and the structural offset features, hematoma expansion risk features are extracted to form an image risk feature set.

[0008] Based on the time-series correlated data set, clinical diagnosis and treatment indicator fluctuation analysis is performed to obtain clinical dynamic characteristics. Collaborative abnormal pattern recognition is performed on the clinical dynamic characteristics to obtain linkage risk markers. Based on the linkage risk markers, risk-sensitive indicators are screened to form a set of clinical risk factors.

[0009] The image risk feature set and the clinical risk factor set are subjected to risk orientation consistency detection to obtain modality consistency score. Based on the modality consistency score, the modality weight configuration is determined. Based on the modality weight configuration, differential fusion is performed to generate fusion prognostic feature vector.

[0010] Based on the fused prognostic feature vector and the modality consistency score, a prognostic assessment is performed to obtain stratified prognostic indicators and prediction confidence levels. Based on the stratified prognostic indicators and the prediction confidence levels, a prognostic analysis report is output.

[0011] A second aspect of this invention provides a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data, comprising:

[0012] The data acquisition module is used to acquire cerebral hemorrhage imaging data and clinical diagnosis and treatment data, and to perform acquisition time point alignment on the cerebral hemorrhage imaging data and the clinical diagnosis and treatment data to construct a time-series associated data group;

[0013] The image analysis module is used to segment the hematoma region based on the time-series correlated data group to obtain the spatial distribution of the hematoma, perform compression analysis of the surrounding structures through the spatial distribution of the hematoma to obtain structural displacement features, and extract hematoma expansion risk features based on the spatial distribution of the hematoma and the structural displacement features to form an image risk feature set.

[0014] The clinical analysis module is used to perform clinical diagnosis and treatment indicator fluctuation analysis based on the time-series associated data group to obtain clinical dynamic characteristics, perform collaborative abnormal pattern recognition on the clinical dynamic characteristics to obtain linkage risk markers, and screen risk-sensitive indicators based on the linkage risk markers to form a set of clinical risk factors.

[0015] The fusion processing module is used to perform risk orientation consistency detection on the image risk feature set and the clinical risk factor set to obtain a modality consistency score, determine the modality weight configuration based on the modality consistency score, and perform differential fusion to generate a fusion prognostic feature vector according to the modality weight configuration;

[0016] The prognostic assessment module is used to perform prognostic assessment based on the fused prognostic feature vector and the modality consistency score to obtain stratified prognostic indicators and prediction confidence, and output a prognostic analysis report based on the stratified prognostic indicators and the prediction confidence.

[0017] The beneficial effects of this invention are reflected in the following points: 1. After the acquisition time points of cerebral hemorrhage imaging data are aligned, a temporal correlation is formed with clinical data. Hematoma region segmentation not only extracts the volume and location information of the hemorrhage foci, but also identifies active hemorrhage markers through internal density stratification analysis. The compression analysis of surrounding structures quantifies the degree of midline shift and ventricular deformation. The imaging risk feature set constructed in this way covers three levels: hematoma morphology, density heterogeneity, and mass effect. Compared with traditional methods that rely solely on hemorrhage volume, it has a more comprehensive risk characterization capability. 2. Clinical indicator fluctuation analysis breaks through the limitations of isolated indicator monitoring. By constructing a linkage matrix through temporal correlation analysis between indicators, synchronously deteriorating indicator combinations are screened out and their pathological mechanism association is verified. The identification of this synergistic abnormality pattern reveals a clinically significant linkage pattern such as blood pressure fluctuation combined with decreased consciousness. The resulting clinical risk factor set focuses on sensitive indicators that truly affect prognosis rather than all monitoring parameters. 3. Modal consistency detection establishes a conflict identification and cause tracing mechanism for situations where images and clinical information contradict each other. The weight configuration is dynamically adjusted according to the credibility of each modality data rather than using a fixed ratio. The prediction confidence integrates two factors: modal fusion quality and feature temporal stability. This makes the final output stratified prognostic index accompanied by a credibility label, which clinicians can use to judge the reference value of the prediction results and decide whether supplementary evaluation is needed. Attached Figure Description

[0018] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0019] Unless otherwise specified, the same reference numerals in different figures represent the same or similar technical features, and different reference numerals may be used to represent the same or similar technical features.

[0020] Figure 1 This is a flowchart illustrating a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data, as described in this invention.

[0021] Figure 2 This is a schematic diagram of hematoma region segmentation and structural compression analysis according to the present invention.

[0022] Figure 3 This is a structural block diagram of a brain hemorrhage prognostic analysis system that integrates imaging and clinical data according to the present invention.

[0023] Wherein: 1-Skull; 2-Brain parenchyma; 3-Hematoma core area; 4-Mixed density area; 5-Active hemorrhage marker; 6-Hematoma boundary; 7-Midline structure; 8-Compressed lateral ventricle; 9-Contralateral lateral ventricle; 10-Midline offset value; 11-Ventricular deformation area; 12-Hematoma adjacent brain tissue area; 13-Normal midline position. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, devices, and circuits are omitted so as not to obscure the description of this application with unnecessary detail.

[0025] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] The technical solutions of the embodiments of this application will be described below.

[0028] like Figure 1 As shown, this embodiment of the invention provides a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data, and performs prognostic analysis by executing steps S110 to S150:

[0029] Step S110: Collect cerebral hemorrhage imaging data and clinical diagnosis and treatment data, and perform time-point alignment of the cerebral hemorrhage imaging data and clinical diagnosis and treatment data to construct a time-series related data group.

[0030] Specifically, imaging and clinical data of cerebral hemorrhage were collected. Cranial CT scans served as the primary source of imaging data, with a slice thickness set to a combination of 5mm conventional slices and 1mm thin-slice reconstruction, covering the entire brain region from the skull base to the skull top. The imaging data also included CT angiography sequences using contrast agent bolus tracking technology, with an arterial phase delay of 20-25 seconds and a venous phase delay of 50-60 seconds, used to assess active bleeding signs. The imaging data was stored in DICOM format and labeled with patient identification and examination timestamps. Clinical data was retrieved from the hospital's information management platform, encompassing three categories: vital sign monitoring records, laboratory test results, and neurological function assessment scales. Vital sign monitoring records included continuous monitoring values ​​of blood pressure, heart rate, and transcutaneous oxygen saturation. Blood pressure was recorded at least once per hour in the clinical data, with a high-frequency monitoring mode of once every 15 minutes for critically ill patients. Laboratory test results included coagulation function indicators in the clinical data, with a focus on the international normalized ratio (INR) and activated partial thromboplastin time (APT). Neurological function assessments used the Glasgow Coma Scale and the National Institutes of Health Stroke Scale (NIHSS). Clinical data required assessments at admission, 6 hours, 24 hours, and 72 hours after symptom onset. Imaging data of cerebral hemorrhage were linked to clinical data using a unique patient identifier.

[0031] A time-series correlated data set was constructed by aligning the acquisition time points of cerebral hemorrhage imaging data and clinical diagnostic data. Since there is a difference in acquisition interval between the CT examination timestamp and the clinical indicator recording timestamp, a unified time reference benchmark needs to be established for time point alignment of cerebral hemorrhage imaging data and clinical diagnostic data. The time-series correlated data set uses the onset time as the zero point, converting the original timestamps of cerebral hemorrhage imaging data and clinical diagnostic data into time offsets relative to the onset time. The time offset is expressed in hours and retains one decimal place. The time point matching of cerebral hemorrhage imaging data and clinical diagnostic data adopts the nearest neighbor principle. Data from the same time point is considered when the time difference between the CT examination time and the clinical indicator recording time is less than 30 minutes; otherwise, it is marked as asynchronous data. For example, blood pressure values ​​acquired within 25 minutes after a patient's CT examination can be paired with that CT image, while laboratory results acquired 45 minutes after the examination need to be marked as asynchronous to avoid incorrectly associating data from different time points during rapidly changing stages of the patient's condition. The temporal correlation dataset organizes successfully matched intracerebral hemorrhage imaging data and clinical diagnosis and treatment data into a temporal slice structure. Each temporal slice contains CT imaging features and corresponding clinical indicator values ​​at that time point. The hyperacute phase (0-6 hours after onset) typically contains 2-3 temporal slices, and this period has the highest data density in the temporal correlation dataset, reflecting multimodal information of the rapid disease progression. The hyperacute phase of intracerebral hemorrhage is a high-incidence period for hematoma expansion; approximately 30% of patients experience hematoma volume increase within 6 hours of onset. High-density data acquisition during this period is crucial for timely detection of hematoma expansion. The temporal correlation dataset ultimately forms a multimodal data sequence based on the patient and time axis, with the sequence length covering the 72-hour observation window after onset.

[0032] Step S120: Hematoma region segmentation is performed based on time-series correlated data to obtain hematoma spatial distribution. Compression analysis of surrounding structures is performed through hematoma spatial distribution to obtain structural displacement features. Based on hematoma spatial distribution and structural displacement features, hematoma expansion risk features are extracted to form an image risk feature set.

[0033] In some embodiments, the step of segmenting the hematoma region based on the temporal correlation data set to obtain the spatial distribution of the hematoma includes: performing intracranial high-density region detection on the brain hemorrhage image data in the temporal correlation data set to obtain candidate hematoma regions; performing internal density layering analysis on the candidate hematoma regions to obtain the hematoma core region and mixed density region; performing high-density spot localization identification on the mixed density region to identify active hemorrhage markers; and constructing the spatial distribution of the hematoma based on the hematoma core region, the mixed density region, and the active hemorrhage markers.

[0034] Intracranial high-density region detection was performed on intracranial high-density images from the temporal correlation dataset to identify candidate hematoma regions. Voxels with CT values ​​exceeding 60 HU within the brain parenchyma were marked as high-density candidate voxels, and high-density voxel screening was performed on CT images from each temporal slice in the temporal correlation dataset. Thin-slice reconstruction sequences from the temporal correlation dataset were used for fine detection of high-density regions; a 1 mm slice thickness could detect smaller high-density lesions compared to a 5 mm conventional slice. After connected component analysis, high-density candidate voxels formed several independent high-density regions. Isolated regions with an area less than 50 mm² were removed as noise, and the remaining high-density regions constituted candidate hematoma regions. Candidate hematoma regions may contain non-hematoma high-density structures such as calcifications and postoperative metal artifacts, requiring differentiation based on both location and morphological features. Basal ganglia calcifications maintained stable location and morphology across multiple temporal slices in the temporal correlation dataset, while hematoma regions may change in volume or density over time; therefore, calcifications could be excluded from the candidate hematoma regions. The candidate hematoma region records the spatial coordinate range, average density value, and volume of each high-density region. The detection results of multiple time-series slices in the time-series correlated data group are stored in chronological order.

[0035] Internal density stratification analysis was performed on candidate hematoma areas to identify the hematoma core and mixed density regions. In the acute phase, hematomas are not homogeneous; coagulated blood clots and uncoagulated blood exhibit different CT density values. The density distribution within the candidate hematoma area reflects the pathological state of the hemorrhage. Voxels with CT values ​​greater than 70 HU within the candidate hematoma area were classified as high-density layers, while those with CT values ​​between 50 and 70 HU were classified as medium-density layers. The hematoma core, composed of the largest connected regions within the high-density layers, represents the main coagulated portion of the hematoma and is typically located in the center. Regions within the candidate hematoma area where high-density and medium-density layers intersect are defined as mixed density regions. The presence of mixed density regions indicates active bleeding within the hematoma or that the hematoma has not yet fully coagulated. The volume ratio of the hematoma core to the mixed density region is recorded as an indicator of hematoma maturity; a ratio less than 2 indicates poor hematoma stability. This hematoma maturity indicator is incorporated into the imaging risk feature set to predict the risk of hematoma expansion. When the mixed density area is distributed in a crescent shape at the edge of the hematoma, it indicates that the hematoma is still expanding. When it is distributed in a patchy pattern inside the hematoma, it indicates rebleeding within the hematoma. After density stratification, the candidate hematoma area is completely divided into two sub-regions: the hematoma core area and the mixed density area. The spatial union of the two sub-regions is the extent of the hematoma body.

[0036] High-density spots were used to locate and identify active hemorrhage markers in mixed-density areas. Point signs on CT images are an independent risk factor for predicting hematoma expansion, manifesting as focal high-density spots with a diameter of 1-2 mm within mixed-density areas. Isolated voxel clusters within mixed-density areas with CT values ​​exceeding 20 HU of the hematoma's average density were considered candidates for high-density spots, with the connected volume of these clusters limited to 1-10 mm³ to exclude large areas of high density. Morphological characteristics of the high-density spots were used to differentiate between true point signs and artifacts; only spots with regular morphology, clear boundaries, and no surrounding radiosclerosis artifacts were included as active hemorrhage markers. High-density spots detected in mixed-density areas were sorted from highest to lowest density value; the spot with the highest density value was most likely to represent an active hemorrhage, and its coordinates were used as the primary marker for active hemorrhage. Active hemorrhage markers were stored as point sets, with each marker containing two attributes: three-dimensional coordinates and spot diameter. When no matching high-density spots were detected within a mixed-density area, the active hemorrhage marker set was empty and marked as negative. When the number of spots in the active hemorrhage marker exceeds 3, it indicates multifocal active hemorrhage, and the risk of hematoma expansion is significantly increased.

[0037] The spatial distribution of the hematoma is constructed based on the hematoma core region, the mixed density region, and the active hemorrhage markers. For example... Figure 2 As shown, in the head CT image, the skull 1 surrounds the brain parenchyma 2, forming a closed cavity. The hematoma lesion is located inside the brain parenchyma 2, presenting as a high-density image. Spatial integration of the three types of data adopts a hierarchical label coding strategy. The hematoma core area 3 is located in the center of the hematoma and presents as a uniform high density, labeled with a voxel value of 1; the mixed density area 4 is located at the edge of the hematoma and presents as a non-uniform density, labeled with a voxel value of 2; the active hemorrhage marker 5 appears as a focal high-density spot within the mixed density area 4, corresponding to a voxel label value of 3. The three-dimensional label matrix of the hematoma spatial distribution maintains the same spatial coordinate system as the original CT image, with a voxel size of 0.5mm × 0.5mm × 1mm. The boundary transition area between the hematoma core area 3 and the mixed density area 4 is smoothed using morphological dilation operations to eliminate jagged artifacts at the segmentation boundary. The hematoma boundary 6 marks the outer contour of the hematoma spatial distribution. When the coordinates of the points marked with active hemorrhage 5 are mapped to the voxel grid of the hematoma spatial distribution, voxels within a spherical region centered on the spot's center and with the spot's diameter as its diameter are all assigned a label value of 3. The hematoma spatial distribution is accompanied by a metadata table recording the overall statistical characteristics of the hematoma, including the total volume, the volume of the hematoma core region 3, the volume of the mixed density region 4, and the number of active hemorrhage markers 5. When the mixed density region 4 accounts for more than 30% of the total volume in the hematoma spatial distribution, it is marked as an unstable hematoma, indicating the need for close monitoring of hematoma volume changes. The temporal slices of the hematoma spatial distribution are stored sequentially according to the time axis of the temporally correlated data sets, forming a complete spatiotemporal record of the hematoma morphological evolution.

[0038] In some embodiments, the step of performing surrounding structural compression analysis based on the spatial distribution of the hematoma to obtain structural offset features includes: locating the brain tissue region adjacent to the hematoma based on the spatial distribution of the hematoma; performing midline structural displacement quantification on the brain tissue region adjacent to the hematoma to obtain a midline offset value; performing ventricular compression morphology analysis based on the spatial distribution of the hematoma to obtain a ventricular deformation index; and performing compression severity grading on the midline offset value and the ventricular deformation index to generate structural offset features.

[0039] Based on the spatial distribution of the hematoma, the brain tissue region adjacent to the hematoma can be located. For example... Figure 2 As shown, the brain parenchyma 2 extending 15 mm outward from the hematoma boundary 6 is defined as the hematoma-adjacent brain tissue region 12. The boundary of the hematoma voxels in the spatial distribution of the hematoma is extracted and expanded outward after morphological gradient operation. The boundary expansion operation of the spatial distribution of the hematoma uses a spherical structuring element with a radius set to 15 mm to cover the typical range of the edema zone around the hematoma. The intersection of the expanded region with the mask of brain parenchyma 2 is taken, and the part of the expansion range that exceeds the boundary of brain parenchyma 2 is excluded. The remaining region constitutes the hematoma-adjacent brain tissue region 12. The hematoma-adjacent brain tissue region 12 is further subdivided into two sub-regions according to anatomical location: the para-midline region and the para-ventricular region. The compression manifestations of different sub-regions have different clinical significance. When the spatial distribution of the hematoma is located in the basal ganglia, the hematoma-adjacent brain tissue region 12 mainly involves the internal capsule and thalamus. Compression damage in these areas is closely related to motor dysfunction. The ratio of the volume of the brain tissue region 12 adjacent to the hematoma to the volume of the hematoma in the spatial distribution of the hematoma is defined as the space-occupying effect coefficient. The larger the space-occupying effect coefficient, the wider the range of surrounding tissue involvement caused by a unit volume of hematoma. The space-occupying effect coefficient is included in the imaging risk feature set to assess the severity of the space-occupying effect. The area within the brain tissue region 12 adjacent to the hematoma with a CT value lower than that of normal brain parenchyma 2 represents the edema zone around the hematoma. The ratio of the edema zone volume to the hematoma volume reflects the degree of secondary brain injury.

[0040] Midline shift values ​​were obtained by quantifying the displacement of midline structures in the brain tissue region adjacent to the hematoma. The septum pellucidum, as a landmark anatomical structure of the midline, directly reflects the asymmetrical distribution of intracranial pressure when its position shifts. Whether the brain tissue region adjacent to the hematoma involves midline structures determines the probability of midline shift. Midline shift measurement was performed at the level of the lateral ventricle body, where the septum pellucidum is most clearly displayed. The measurement benchmark for midline shift was the midpoint of the line connecting the anteroposterior diameter of the inner table of the skull. When the brain tissue region adjacent to the hematoma involves the septum pellucidum, the septum pellucidum shifts to the opposite side of the hematoma, and the shift distance is calculated using image coordinates. Midline shift values ​​are recorded in millimeters, with positive values ​​indicating a shift to the right and negative values ​​indicating a shift to the left, the direction of shift being opposite to the side of the hematoma in its spatial distribution. A midline shift value exceeding 10 mm indicates a severe mass effect and is one of the important indications for neurosurgical intervention. Different subregions within the brain tissue adjacent to the hematoma contribute differently to midline shift. Compression in the periventricular region directly causes midline shift, while compression in the periventricular region indirectly affects midline position through cerebrospinal fluid circulation obstruction. The changing trends of midline shift values ​​in different time-series slices reflect the dynamic evolution of the mass effect, with a progressive increase in shift values ​​indicating disease progression.

[0041] For example, the step of performing ventricular compression morphology analysis based on the spatial distribution of the hematoma to obtain the ventricular deformation index includes: locating the contours of the bilateral ventricular regions based on the spatial distribution of the hematoma; comparing the morphology of the compressed side and the contralateral side of the bilateral ventricular region contours to obtain bilateral asymmetry; analyzing the degree of collapse of the compressed side contour in the bilateral ventricular region contours to obtain a deformation degree value; and generating the ventricular deformation index based on the bilateral asymmetry and the deformation degree value.

[0042] The contours of the bilateral ventricles were located based on the spatial distribution of the hematoma. The lateral ventricles appear as symmetrically distributed, butterfly-shaped low-density areas on CT images, and the spatial distribution of the hematoma provides a reference for the spatial relationship between the hematoma and the ventricles. A cerebrospinal fluid density threshold of 0-15 HU was set; regions in the CT images where the hematoma spatial distribution met this density range and were located within the anatomical position of the ventricles were extracted as candidate ventricles. The left and right lateral ventricles were separated by the midline, with the sagittal plane containing the midpoint of the anteroposterior diameter of the inner table of the skull as the boundary; the left side of the boundary was the left ventricle, and the right side was the right ventricle. The contours of the bilateral ventricles were extracted from the segmentation mask using an edge detection algorithm. The high threshold for Canny edge detection was set to 100, and the low threshold to 50. The contour of the ventricle on the same side as the hematoma spatial distribution was marked as the compressed side contour, and the contour of the contralateral ventricle was marked as the reference side contour. The two contours were stored in the corresponding fields of the bilateral ventricle region contours. The contours of the bilateral ventricles were extracted at three typical levels: the body, the trigone, and the temporal horn. The body level was used to assess lateral compression, the trigone level to assess posterior morphology, and the temporal horn level to assess inferior morphology. When the hematoma spatially affects different ventricle regions, the compression morphology of the bilateral ventricle contours varies; compression of the body often presents as an arc-shaped depression, while compression of the temporal horn often presents as a tubular narrowing.

[0043] Bilateral asymmetry was obtained by comparing the morphology of the compressed side and the contralateral side of the bilateral ventricular region contours. Normally, the bilateral lateral ventricles are mirror-symmetrically distributed. Unilateral compression caused by the space-occupying effect disrupts this symmetry, and the degree of asymmetry in the bilateral ventricular region contours reflects the severity of compression. Morphological comparison first calculated the area of ​​the bilateral ventricular contours. The area of ​​the compressed side was denoted as A_c, and the area of ​​the reference side as A_r. The area difference rate was calculated as (A_r - A_c) / A_r. Bilateral asymmetry was defined as the weighted sum of the area difference rate and the shape difference rate of the two ventricles, with a weight of 0.7 for the area difference rate and a weight of 0.3 for the shape difference rate. The shape difference rate was obtained by calculating the Hausdorff distance between the two contours. After mirror flipping, the compressed side contour was registered with the reference side contour using point set matching. The maximum mismatch distance was divided by the transverse diameter of the reference side ventricle to obtain the standardized shape difference rate. Bilateral asymmetry is measured in ranges from 0 to 1, where 0 indicates complete symmetry and 1 indicates extreme asymmetry or complete disappearance of the compressed side. The asymmetry of the bilateral ventricular region contours at three typical levels is calculated, and the maximum value is taken as the final bilateral asymmetry, reflecting the morphological changes at the most severely compressed level. A bilateral asymmetry exceeding 0.4 indicates a significant unilateral mass effect, requiring a comprehensive assessment of intracranial pressure distribution in conjunction with midline shift.

[0044] The degree of collapse and deformation was analyzed on the compressed side of the bilateral ventricle contour to obtain deformation values. Collapse of the compressed ventricle contour manifests as a shortened transverse diameter and morphological distortion. Geometric features of the compressed side contour were used to quantify the degree of collapse. The transverse diameter of the compressed side ventricle was compared with previous normal CT images of the same patient or an age-matched standard reference value. The shortening rate was calculated as (D_ref - D_c) / D_ref, where D_ref is the reference transverse diameter and D_c is the current transverse diameter. The deformation value considered both the shortening rate and the change in contour roundness. Roundness was defined as 4π × area / perimeter², with reduced roundness due to compression indicating a tendency towards elongation. The rate of change in roundness of the compressed side contour was calculated as (C_ref - C_c) / C_ref, where C_ref is the roundness of the reference side and C_c is the roundness of the compressed side. The deformation degree value is calculated using the formula F = 0.6 × R_d + 0.4 × R_c, where R_d is the rate of shortening of the transverse diameter and R_c is the rate of change in roundness. Both indicators are standardized within the range of 0-1. A deformation degree value exceeding 0.5 is considered moderate to severe collapse, at which point the compressed ventricle is significantly smaller than its normal shape, and further deterioration may lead to complete ventricular occlusion. A deformation degree value of 1 indicates that the compressed ventricle can no longer be identified as a usable cavity, corresponding to a loss of the compressed side's contour in the bilateral ventricular region outline.

[0045] The ventricular deformation index (I) is generated based on bilateral asymmetry and deformation degree values. These two indicators reflect the compression state of the ventricles from different perspectives: bilateral asymmetry emphasizes the comparison between the two sides, while the deformation degree value emphasizes the morphological changes on the compressed side itself. The ventricular deformation index I integrates the two indicators using a weighted fusion strategy. The weight of both bilateral asymmetry and deformation degree values ​​is set to 0.5, and the fusion formula is I = 0.5 × A_s + 0.5 × F_v, where A_s is the bilateral asymmetry and F_v is the deformation degree value. The ventricular deformation index ranges from 0 to 1, with a higher value indicating more severe ventricular deformation. 0 indicates normal ventricular morphology, and 1 indicates complete disappearance of the compressed ventricle. Significant separation between bilateral asymmetry and deformation degree values ​​requires special attention; for example, a high bilateral asymmetry but a low deformation degree value suggests simultaneous compression of both ventricles or the presence of congenital ventricular asymmetry. The ventricular deformation index is divided into four levels based on the numerical range: less than 0.2 indicates normal or slight deformation, 0.2-0.4 indicates mild deformation, 0.4-0.7 indicates moderate deformation, and greater than 0.7 indicates severe deformation. The ventricular deformation index and its grading results, along with the original bilateral asymmetry and deformation degree data, are output together for subsequent comprehensive assessment of the severity of compression.

[0046] Structural offset features are generated by grading the severity of compression based on midline offset values ​​and ventricular deformation index. For example... Figure 2As shown, the normal midline position 13 is represented by a dashed line. Midline structure 7 shifts to the contralateral side due to the hematoma's mass effect. The midline shift value 10 is the horizontal distance between midline structure 7 and the normal midline position 13. The compressed lateral ventricle 8 is located on the same side as the hematoma and is directly compressed. The ventricular deformation area 11 marks the most severely compressed area. The contralateral lateral ventricle 9 has a relatively normal morphology and serves as an assessment reference. The two indicators are integrated into a unified assessment of compression severity through a weighted comprehensive score. The weight of the midline shift value is set at 0.6, and the weight of the ventricular deformation index is set at 0.4. The midline shift value is first normalized, converted to a standardized score within the range of 0-1 with a reference upper limit of 15mm. Shifts exceeding 15mm are truncated to 1. The ventricular deformation index, already a standardized score within the range of 0-1, is directly used in the weighted fusion calculation. Structural displacement features were categorized into four levels based on a comprehensive score: 0-0.25 for mild compression, 0.25-0.5 for moderate compression, 0.5-0.75 for severe compression, and 0.75-1 for extremely severe compression. The midline displacement value and the raw values ​​of the ventricular deformation index, along with the compression level, were collectively incorporated into the structural displacement features. Changes in the structural displacement features across different time-series slices were used to assess the evolution of the mass effect; an increase in the compression level across two consecutive time-series slices indicated rapid disease progression.

[0047] Based on the spatial distribution and structural displacement characteristics of hematoma, a risk feature set for hematoma expansion was formed by extracting hematoma expansion risk features. Hematoma expansion is a major cause of early neurological deterioration in cerebral hemorrhage, and the rate of change in hematoma volume between adjacent time-series slices in the spatial distribution of hematoma is the core indicator for predicting hematoma expansion. Significant hematoma expansion is defined as an increase in hematoma volume exceeding 33% or an absolute increase exceeding 6 ml within 6 hours after onset in the spatial distribution of hematoma. This criterion is included in the risk threshold field of the image risk feature set. Point signs, island signs, and mixed signs from CT plain scans are extracted from the original images corresponding to the spatial distribution of hematoma. Point signs appear as high-density spots with a diameter of 1-2 mm within the hematoma; island signs appear as multiple separated small hematomas at the hematoma edge; and mixed signs appear as uneven density within the hematoma. Among the structural displacement features, the rate of midline shift progression is included in the image risk feature set as a quantitative indicator of worsening mass effect. An increase in midline shift exceeding 3 mm between adjacent time-series slices indicates rapid aggravation of the mass effect. In the structural displacement feature, a ventricular deformation index exceeding 0.5 is marked as a moderate to severe lesion effect, and this marker is included in the imaging risk feature set as a grading criterion for structural compression features. Hematoma maturity indicators, unstable hematoma markers, and the volumes of the hematoma core and mixed density regions in the hematoma spatial distribution are simultaneously included in the imaging risk feature set to assess the stability of the hematoma. The ratio of the lesion effect coefficient to the edema zone volume in the structural displacement feature set is included in the imaging risk feature set to assess the degree of secondary brain injury. The imaging risk feature set organizes data in feature vector form, containing 18 feature items across four dimensions: hematoma morphology, hematoma density, hematoma evolution, and structural compression. The raw values ​​in the hematoma spatial distribution and structural displacement features are standardized before being written into the imaging risk feature set. The standardization method uses Z-score transformation to ensure that all features have the same numerical scale.

[0048] Step S130: Based on the time-series associated data group, perform clinical diagnosis and treatment indicator fluctuation analysis to obtain clinical dynamic characteristics, perform collaborative abnormal pattern identification on the clinical dynamic characteristics to obtain linkage risk markers, and screen risk-sensitive indicators based on linkage risk markers to form a set of clinical risk factors.

[0049] Specifically, clinical dynamic characteristics were obtained by performing fluctuation analysis of clinical indicators based on the time-series correlation dataset. Vital signs such as blood pressure, heart rate, and level of consciousness showed significant fluctuations during the acute phase of cerebral hemorrhage. The clinical indicator values ​​of each time-series slice in the time-series correlation dataset formed the data basis for the analysis. The temporal changes in systolic blood pressure in the time-series correlation dataset were extracted using the sliding window method, with a window width set to 2 hours. The difference between the maximum and minimum systolic blood pressure values ​​within the window was defined as the blood pressure fluctuation amplitude. For example, if a patient's systolic blood pressure fluctuated from 150 mmHg to 190 mmHg within a 2-hour window, the fluctuation amplitude reached 40 mmHg, indicating unstable blood pressure control and an increased risk of recurrent cerebral hemorrhage or hematoma expansion. A decrease of more than 2 points in the Glasgow Coma Scale score in adjacent time-series slices of the time-series correlation dataset was considered a deterioration in level of consciousness. The temporal gradient of the score change reflected the rate of progression of neurological function impairment. A patient with a Glasgow Coma Scale score of 12 upon admission who dropped to 9 points after 6 hours (a decrease of 3 points) suggested potential hematoma expansion or worsening cerebral edema, requiring immediate repeat cranial CT scan. The Clinical Dynamics feature integrates the fluctuation parameters of various indicators into a structured data table. Row indices correspond to the monitored indicator names, and column indices correspond to the fluctuation feature types, including fluctuation amplitude and rate of change. Dynamic changes in coagulation function indicators are recorded separately in the Clinical Dynamics feature. An international normalized ratio exceeding 1.4 or an activated partial thromboplastin time exceeding 1.5 times the normal upper limit is marked as an abnormal coagulation state, which is included in the assessment criteria for linked risk markers. In the time-series correlated data set, the sampling interval for laboratory test results is typically longer than the vital signs monitoring interval. The Clinical Dynamics feature calculates fluctuation parameters separately for indicators with different sampling frequencies. The Clinical Dynamics feature ultimately includes 12 vital signs fluctuation indicators, 8 laboratory test change indicators, and 4 neurological function score evolution indicators, totaling a 24-dimensional feature vector.

[0050] In some embodiments, the step of performing collaborative abnormal pattern recognition to obtain linkage risk markers for the clinical dynamic features includes: extracting multidimensional indicator change vectors from the clinical dynamic features; performing inter-indicator temporal correlation analysis on the multidimensional indicator change vectors to obtain an indicator linkage matrix; screening synchronous deterioration indicator combinations based on the indicator linkage matrix to obtain collaborative abnormal patterns; and generating linkage risk markers based on the collaborative abnormal patterns.

[0051] Multidimensional indicator change vectors are extracted from clinical dynamic features. The numerical change sequences of each monitoring indicator within the observation period constitute the basic elements of the vector. The 24-dimensional features in the clinical dynamic features are expanded along the time axis to form a multidimensional time series. The time series data of three indicators in the clinical dynamic features—blood pressure fluctuation amplitude, heart rate change rate, and consciousness score decline value—are extracted first to form the core vector components for assessing cardiovascular and cerebrovascular status. Blood pressure fluctuation amplitude reflects the stability of cerebral perfusion pressure, heart rate change rate reflects autonomic nervous system regulation function, and consciousness score decline value directly reflects the degree of brain function impairment. The joint analysis of these three indicators can capture key pathological changes in the acute phase of cerebral hemorrhage. The standardization of indicator values ​​adopts the standard score method. Each component in the multidimensional indicator change vector is converted into a standardized value with a mean of 0 and a standard deviation of 1, eliminating the influence of differences in the dimensions of different indicators, so that the millimeter-mercury unit of blood pressure and the score unit can be compared and calculated on the same numerical scale. In clinical dynamic characteristics, indices with inconsistent sampling time points were unified to the same time grid through linear interpolation. Vital signs were collected hourly, while laboratory tests were collected every 6-8 hours. Interpolation aligned all indicators to a uniform time resolution, which was set to 30 minutes. The multidimensional indicator change vector contained 24 component values ​​at each standardized time point, forming a 24×T data matrix for the entire observation period, where T is the number of time points. The 72-hour observation window corresponds to T=144 time points. Missing values ​​in the multidimensional indicator change vector were filled using a forward imputation strategy, i.e., the missing position was filled with the most recent valid measurement value of the indicator. If an indicator was missing at the start of the observation, the population reference mean of that indicator was used for imputation.

[0052] Inter-indicator temporal correlation analysis is performed on the multidimensional indicator change vector to obtain the indicator linkage matrix. Different clinical indicators exhibit physiological or pathological correlations, and the synchronous change patterns of each component in the multidimensional indicator change vector reflect this inherent linkage. The Pearson correlation coefficient between any two indicator components in the multidimensional indicator change vector is calculated using the formula r = Σ(x_i - x_avg)(y_i - y_avg) / sqrt[Σ(x_i - x_avg)²·Σ(y_i - y_avg)²], where x_avg and y_avg are the mean values ​​of the two indicators over time, respectively. The correlation coefficient ranges from -1 to 1, with positive values ​​indicating a positive correlation (both indicators change in the same direction) and negative values ​​indicating a negative correlation (both indicators change in opposite directions). The indicator linkage matrix is ​​a 24×24 symmetric matrix, where matrix element (i,j) stores the correlation coefficient between the i-th and j-th indicators, and the diagonal elements are always 1. Cross-correlation analysis was performed on indicator pairs with time-lag correlations in the multidimensional indicator change vector. Correlation coefficients under different time lags were calculated, and the maximum value was entered into the indicator linkage matrix. The statistical significance of the correlation coefficients was evaluated using a t-test. Correlation coefficients with a p-value greater than 0.05 in the indicator linkage matrix were set to 0 to eliminate statistically insignificant weak correlations. The indicator linkage matrix was divided into three levels based on the absolute value of the correlation coefficient: strong correlation, moderate correlation, and weak correlation. An absolute value greater than 0.7 indicated a strong correlation, 0.4-0.7 indicated a moderate correlation, and less than 0.4 indicated a weak correlation. A correlation coefficient between systolic and diastolic blood pressure typically exceeding 0.9 indicates a strong physiological correlation, while a negative correlation coefficient of -0.75 between elevated systolic blood pressure and decreased consciousness score suggests a pathological correlation.

[0053] For example, the step of selecting synchronous deterioration indicator combinations based on the indicator linkage matrix to obtain a collaborative abnormal pattern includes: extracting strongly correlated indicator pairs from the indicator linkage matrix to obtain a highly correlated indicator group; performing a consistency judgment on the deterioration direction for the highly correlated indicator group to select co-deterioration indicator pairs; conducting clinical pathological mechanism correlation verification on the co-deterioration indicator pairs to obtain mechanism-related indicator clusters; and generating a collaborative abnormal pattern based on the mechanism-related indicator clusters.

[0054] Strongly correlated indicator pairs are extracted from the indicator linkage matrix to obtain highly correlated indicator groups. Indicator pairs with an absolute correlation coefficient exceeding a preset threshold constitute strong correlation candidates. The threshold for the indicator linkage matrix is ​​set to 0.7 to screen indicator combinations with significant statistical correlation. The coefficient of determination corresponding to this threshold is 0.49, indicating that approximately half of the variation between the two indicators can be explained by each other. The positions of elements in the indicator linkage matrix that meet the threshold condition are extracted and converted into indicator pair numbers. The element in the i-th row and j-th column corresponds to the pairing of indicator i and indicator j. Due to the symmetry of the matrix, only the upper triangular region needs to be scanned. After excluding diagonal elements and duplicate pairs, a list of highly correlated indicator pairs is formed, and the length of the list depends on the number of strongly correlated elements in the indicator linkage matrix. Systolic blood pressure and diastolic blood pressure usually show a high positive correlation in the indicator linkage matrix, with a correlation coefficient exceeding 0.9. This indicator pair is included in the highly correlated indicator group but labeled as a physiological correlation. Physiological correlation refers to an inherent correlation caused by normal physiological mechanisms rather than an abnormal correlation caused by pathological processes. Within the highly correlated indicator group, each indicator pair is sorted from highest to lowest correlation coefficient. This sorting is used to determine priority during screening, with higher correlation coefficients prioritized for pathological mechanism verification. Indicator pairs with an absolute negative correlation coefficient exceeding 0.7 in the indicator linkage matrix are also included in the highly correlated indicator group. Negative correlation indicates an inverse relationship between the two indicators; for example, a decrease in cerebral perfusion pressure during increased intracranial pressure is a typical example of negative correlation. The highly correlated indicator group typically consists of 10-30 pairs. If the group is too large, the correlation coefficient threshold can be increased to 0.75 for streamlined screening.

[0055] For highly correlated indicator groups, a consistency-based judgment of deterioration direction was used to screen for indicators exhibiting the same deterioration direction. The clinical significance of the indicator value changes determined the direction of deterioration; elevated blood pressure, decreased consciousness score, and prolonged clotting time were all defined as deterioration directions. Each indicator in the highly correlated indicator group was pre-labeled with its deterioration direction definition, based on clinical medical knowledge: an increase in value representing deterioration was labeled as positive deterioration, and a decrease in value representing deterioration was labeled as negative deterioration. The changing trends of the two indicators in each pair within the highly correlated indicator group were determined separately during the observation period. The trend was determined by the sign of the linear regression slope; a positive slope indicated an increase in indicator value over time, and a negative slope indicated a decrease. Indicator pairs exhibiting the same deterioration direction were considered to have the same deterioration direction when both trends conformed to their respective definitions. In the highly correlated indicator group, positively correlated indicators require both indicators to change in a deteriorating direction simultaneously, while negatively correlated indicators require one indicator to deteriorate while the other indicator changes in the opposite direction. For example, in the negatively correlated pair of elevated blood pressure and decreased heart rate, a change in blood pressure towards an increase and a change in heart rate towards a decrease satisfies the condition of co-deterioration. The screening results for co-deterioration indicator pairs are marked with information on the magnitude of change, defined as the relative percentage change in the indicator value from the observation start point to the end point. A larger magnitude of change indicates a more severe degree of deterioration. In the highly correlated indicator group, indicator pairs with physiological associations are marked as low priority even if they meet the co-deterioration condition. Co-deterioration indicator pairs prioritize combinations with pathological associations. When the co-deterioration indicator pair is empty, it indicates that no multi-indicator synergistic deterioration has occurred during the current observation period, and the overall condition is relatively stable.

[0056] To verify the correlation between the same deterioration index pairs and clinicopathological mechanisms, a cluster of mechanism-related indicators was obtained. Statistical correlation is not equivalent to causal association. The effectiveness of the same deterioration index pairs needs to be verified in conjunction with clinicopathological mechanisms to eliminate spurious correlations. Spurious correlations refer to two indicators that show a numerical correlation but lack an actual pathophysiological connection. The association between elevated blood pressure and increased intracranial pressure in the same deterioration index pairs conforms to the pathological mechanism of Cushing's reaction. Cushing's reaction is characterized by a triad of elevated blood pressure, bradycardia, and altered respiratory rhythm, which is a typical compensatory response to a rapid increase in intracranial pressure and often predicts impending brain herniation. This index pair was included in the cluster of mechanism-related indicators after mechanism verification. A pre-built pathological mechanism knowledge base stores the association rules of indicators related to cerebral hemorrhage. Knowledge base entries include four fields: trigger indicator, response indicator, association direction, and mechanism type. If a same deterioration index pair successfully matches a knowledge base entry, it is confirmed as mechanism-related. The cluster of mechanism-related indicators will be clustered according to the pathological mechanism type based on the verified index pairs. Multiple index pairs related to the same mechanism will be grouped into the same cluster. The clustering results integrate the scattered index pair relationships into a systematic description of the pathological process. The intracranial pressure elevation mechanism cluster comprises pairwise pairs of three indicators: elevated blood pressure, bradycardia, and decreased consciousness. Any combination of these three indicators in a pair of concordant deterioration indicators is also included in this cluster. Activation of this cluster suggests that the patient may be experiencing an intracranial pressure crisis. Each cluster within the mechanism-related indicator cluster is accompanied by a mechanism description text, explaining the pathophysiological interpretation of the linkage between indicators within that cluster. This mechanism description text is used to generate clinical prompts to assist physicians in understanding the clinical significance of abnormal linkages.

[0057] Synergistic abnormality patterns are generated based on mechanism-related indicator clusters. Each mechanism cluster represents a specific pathophysiological process. Each cluster within the mechanism-related indicator cluster is converted into a synergistic abnormality pattern. The conversion process encapsulates the structured data of the cluster into a standardized abnormality pattern description. The naming rule for synergistic abnormality patterns is a combination of mechanism type and severity. For example, the intracranial pressure increase mechanism cluster corresponds to an intracranial pressure increase pattern and is labeled with a severity level. The naming uses clinically understandable terminology to enable physicians to quickly identify the risk type. The number of indicators within a mechanism-related indicator cluster and the magnitude of indicator deterioration jointly determine the severity classification of the synergistic abnormality pattern. A number of indicators greater than 3 and an average deterioration exceeding 30% are classified as severe; a number of indicators of 2-3 and an average deterioration of 15%-30% are classified as moderate; and all other cases are classified as mild. The activation time of the synergistic abnormality pattern is recorded. The activation time is defined as the moment when the first indicator within the cluster begins to deteriorate. The timing of the activation time reflects the starting point of the pathological process and helps to retrospectively analyze the timeline of disease evolution. The output format of the co-anomaly pattern includes five fields: pattern code, pattern name, list of involved indicators, severity level, and activation time. The standardized output format ensures a unified data structure across different co-anomaly patterns, facilitating comparison and management. When the mechanism-related indicator cluster is empty, the co-anomaly pattern output is an empty set, indicating that no co-anomaly with a clear pathological mechanism has been detected.

[0058] Linked risk markers are generated based on synergistic anomaly patterns. The presence of synergistic anomalies indicates a risk of systemic disease deterioration. Each abnormal combination detected in the synergistic anomaly pattern generates a corresponding risk marker. Linked risk markers are stored in key-value pairs, where the key is the encoded identifier of the synergistic anomaly pattern, and the value is the risk weight of that pattern and the detection timestamp. The combination of blood pressure fluctuations and decreased consciousness in the synergistic anomaly pattern corresponds to a risk marker for increased intracranial pressure, with a weight set at 0.85. This marker indicates the need for urgent assessment for the presence of impending brain herniation. The combination of coagulation abnormalities and increased hematoma volume corresponds to a risk marker for hematoma expansion in the synergistic anomaly pattern, with a weight set at 0.90. This is a key indicator for early intervention. When multiple markers in the linked risk pattern are present simultaneously, a comprehensive risk score is calculated using the formula R = 1 - ∏(1 - w_i), where w_i is the risk weight of each marker. This formula is based on the combined probability calculation of independent risk events. When multiple risks are present simultaneously, the comprehensive risk is higher than any single risk. The comprehensive risk score is incorporated into the clinical risk factor set as a quantitative indicator of the overall risk level. When the collaborative anomaly mode is an empty set, the linkage risk marker is set to the baseline state, indicating that no obvious linkage anomalies have been detected. The linkage risk marker records the first detection time and duration; if the duration exceeds 2 hours, the marker is upgraded to a persistent linkage risk.

[0059] A clinical risk factor set was formed by screening risk-sensitive indicators based on linkage risk markers. Risk-sensitive indicators are clinical indicators involved in linkage anomalies and have independent predictive value for prognosis; all indicators involved in linkage risk markers are risk-sensitive candidates. Indicators involved in the top three weighted markers in linkage risk markers are prioritized for inclusion in the risk-sensitive indicator pool; these indicators exhibit significant linkage anomalies in the current case. The screening of risk-sensitive indicators also considers the modifiability of the indicators; among the linkage risk marker-related indicators, indicators that can be controlled through treatment, such as blood pressure and coagulation function, have higher priority than immutable indicators such as age and bleeding site. The clinical risk factor set selects the final included indicators from the risk-sensitive indicator pool, based on the product of the indicator's frequency of occurrence in the linkage risk markers and the corresponding marker's risk weight. When the linkage risk markers are at baseline, the clinical risk factor set includes general risk indicators related to the prognosis of cerebral hemorrhage, including the Glasgow Coma Scale score on admission, hematoma volume, and patient age. The clinical risk factor set is stored in feature vector form, containing three attribute fields: indicator name, current value, and risk weight. Indicators corresponding to persistent markers in the linked risk labeling are assigned high weights in the clinical risk factor set, indicating the need for continuous monitoring. The dimensionality of the clinical risk factor set is controlled at 8-12 indicators; excessive dimensionality increases the computational complexity of the fusion analysis and the risk of overfitting.

[0060] Step S140: Perform risk orientation consistency detection on the image risk feature set and the clinical risk factor set to obtain a modality consistency score, determine the modality weight configuration based on the modality consistency score, and perform differentiated fusion to generate a fusion prognostic feature vector according to the modality weight configuration.

[0061] In some embodiments, the step of performing risk orientation consistency detection on the image risk feature set and the clinical risk factor set to obtain a modal consistency score includes: performing risk severity quantification on the image risk feature set and the clinical risk factor set to obtain image risk scores and clinical risk scores, respectively; calculating the risk orientation deviation of the image risk scores and the clinical risk scores to obtain modal deviation values; conducting deviation cause analysis based on the modal deviation values ​​to identify differences in data source credibility; and generating a modal consistency score based on the differences in credibility between the modal deviation values ​​and the data source.

[0062] Risk severity quantification was performed on both the imaging risk feature set and the clinical risk factor set to obtain imaging risk scores and clinical risk scores, respectively. Independent scoring models were used for risk quantification in both feature sets. The 18 features in the imaging risk feature set were converted into a single risk score through weighted summation. The weights for the three core features—hematoma volume, hematoma expansion rate, and midline shift—were set to 0.15, 0.20, and 0.15, respectively, while the remaining 15 features shared the remaining weight of 0.50. The imaging risk score was calculated using the formula S_img=Σ(w_i×f_i), where w_i is the weight of the i-th feature, f_i is the standardized value of that feature, and the score range was normalized to 0-100. The 8-12 indicators in the clinical risk factor set were also calculated using a weighted summation method, with higher weights assigned to the Glasgow Coma Scale score and coagulation function indicators. The clinical risk score is calculated considering the degree of deterioration of the indicators; the further the indicator value deviates from the normal range, the higher the risk score it contributes. The degree of deviation is measured by the multiple of the standard deviation. The scoring model parameters for the imaging risk feature set and the clinical risk factor set are derived from the statistical results of previous prognostic studies of intracerebral hemorrhage. The model has been validated with multi-center data and has good predictive efficacy. The imaging risk score and the clinical risk score represent the risk assessment results of two dimensions: imaging manifestations and clinical status, respectively. Similar values ​​for both indicate that the risk judgments of the two modalities are consistent.

[0063] Modal deviation values ​​are obtained by calculating the risk orientation deviation between imaging risk scores and clinical risk scores. Ideally, the risk assessments of the two modalities should point to the same prognostic conclusion. The difference between imaging risk scores and clinical risk scores reflects the degree of inconsistency between the information of the two modalities. The modal deviation value is defined as the ratio of the absolute difference between the two scores to their means, and the calculation formula is D=|S_img-S_cli| / [(S_img+S_cli) / 2]×100%, where S_img is the imaging risk score and S_cli is the clinical risk score. When both scores are less than 10, the absolute difference |S_img-S_cli| is used instead of the relative difference to avoid the denominator being too small and causing calculation abnormalities. When the imaging risk score is 75 and the clinical risk score is 45, the modal deviation is calculated to be 50%, indicating a significant discrepancy in risk assessment between the two modalities. This suggests that CT imaging shows a large hematoma or signs of active bleeding, but the patient's level of consciousness and vital signs are relatively stable, possibly indicating an early stage of hematoma expansion where imaging changes have not yet fully reflected in clinical manifestations. Modal deviation is classified into four levels based on its numerical range: less than 15% is highly consistent, 15%-30% is basically consistent, 30%-50% is moderately divergent, and greater than 50% is severely divergent. When both the imaging risk score and the clinical risk score are close to 0 or close to 100, even a small absolute difference may result in a large modal deviation; in such cases, a comprehensive judgment based on the absolute difference is necessary.

[0064] Modal deviation analysis is conducted to identify discrepancies in data source reliability. Modal deviations can stem from various causes, such as differences in data quality, asynchronous data collection times, or rapid disease progression. When modal deviations reach moderate or higher levels, the underlying cause needs to be traced. Retrospective examination of the original data corresponding to the modal deviation value first assesses the quality of the cerebral hemorrhage imaging data. Motion artifacts, excessive slice thickness, or incomplete scanning range in CT images reduce the image reliability score. In agitated patients, motion artifacts on CT images due to lack of cooperation can obscure hematoma boundaries, making accurate volume measurement difficult; in such cases, the image reliability score should be reduced by 0.2-0.3. The completeness of clinical data is assessed to determine if key indicators are missing. Missing Glasgow Coma Scale scores, coagulation function indicators, or vital sign records reduce the clinical reliability score. The timeliness of cerebral hemorrhage imaging data is assessed, focusing on the interval between the CT scan and the current assessment. If the interval exceeds 6 hours and the patient's condition fluctuates, the image reliability score will be reduced accordingly. The standardization of clinical data is assessed, focusing on the collection conditions of laboratory tests and the operational status of vital sign monitoring equipment. The presence of hemolyzed specimens or alarms in monitoring equipment will result in a reduction in the clinical reliability score. Data source reliability difference is defined as the difference between the image reliability score and the clinical reliability score. A positive value indicates that the imaging data of cerebral hemorrhage is more reliable, while a negative value indicates that the clinical data is more reliable. The larger the absolute value, the more significant the difference in data quality between the two modalities. When both the modality deviation value and the absolute value of the data source reliability difference are large, it suggests that the data in the modality with lower reliability has assessment bias. When the absolute value of the data source reliability difference is small, but the modality deviation value is still large, it suggests that the deviation may stem from the complexity of the disease itself rather than data quality issues.

[0065] A modal consistency score is generated based on the modal deviation value and the difference in data source credibility. These two indicators jointly determine the reliability of modal fusion; the smaller the modal deviation value and the smaller the absolute value of the difference in data source credibility, the higher the modal consistency. The modal consistency score is calculated using a two-factor weighted model, with the formula A = w_d × (1 - D_norm) + w_c × (1 - C_norm), where A is the modal consistency score, D_norm is the normalized modal deviation value, C_norm is the normalized absolute value of the difference in data source credibility, and w_d and w_c are the weights of the two factors, respectively. The modal deviation value is normalized to the 0-1 range with a reference upper limit of 50%, and deviation values ​​exceeding 50% are truncated to 1. The weight w_d is set to 0.6 to reflect the dominant role of the deviation degree. The absolute value of the difference in data source credibility is normalized with a reference upper limit of 0.3, and the weight w_c is set to 0.4 to reflect the impact of data quality differences on fusion reliability. The modality consistency score ranges from 0 to 1. A higher value indicates greater consistency in the risk assessment results of the two modalities, and a higher reliability of the fusion result. When the modality deviation is at a high consistency level and the absolute value of the difference in data source reliability is small, the modality consistency score usually exceeds 0.85, indicating a high consistency state. In this case, the two modalities can be fused using similar weights. A modality consistency score between 0.5 and 0.7 indicates a moderate consistency state. During fusion, the roles of the dominant and auxiliary modalities need to be determined based on data reliability. A modality consistency score below 0.5 indicates significant conflict between modalities, with the two modalities providing contradictory prognostic information. The fusion strategy needs to be specially designed to avoid misleading conclusions from contradictory information. The modality consistency score is simultaneously output to the modality weight configuration module and the prediction confidence calculation module. The weight configuration determines the magnitude of the weight difference between modalities, while the confidence calculation reflects the overall reliability level of the fusion result.

[0066] In some embodiments, determining the modality weight configuration based on the modality consistency score includes: performing a modality conflict degree determination based on the modality consistency score to obtain a conflict level; performing a backtracking assessment of the data quality of each modality for the conflict level to obtain a modality credibility ranking; dividing the dominant modality and auxiliary modality according to the modality credibility ranking; and generating a modality weight configuration by performing weight differential allocation based on the dominant modality and the auxiliary modality.

[0067] Modality consistency scores are used to determine the degree of modality conflict and obtain the conflict level. The consistency score and the degree of conflict are inversely related; a lower consistency score indicates more severe intermodal conflict, requiring a differentiated fusion strategy. The conflict level is divided into four levels based on the numerical range of the consistency score: a score greater than 0.85 indicates no conflict, 0.70-0.85 indicates mild conflict, 0.50-0.70 indicates moderate conflict, and less than 0.50 indicates severe conflict. The thresholds for these four levels are determined based on the relationship curve between modality consistency and prediction accuracy in clinical prognostic studies. A consistency score of 0.78 corresponds to a mild conflict level. At this level, the two modalities can be fused using similar weights, with the weight difference controlled within 0.2. Mild conflict usually stems from minor differences at the data collection time or the normal fluctuation range of the disease condition. When the conflict level is moderate or severe, the deviation analysis results corresponding to the modality consistency score need to be included in the decision-making reference to determine which modality's information is more reliable. Moderate conflict indicates a substantial discrepancy in the risk assessment between the two modalities, while severe conflict indicates that the two modalities provide contradictory prognostic directions. When the modality consistency score is near the boundary value, the determination of the conflict level should be based on historical trends. Only when the scores of two consecutive time-series slices are both below the threshold is the conflict level confirmed, avoiding frequent jumps in the conflict level due to single score fluctuations. When the conflict level is severe, it indicates that the two modalities provide contradictory prognostic information. The interpretation of the fusion results requires special caution and is recommended to be combined with clinical experience for comprehensive judgment. If necessary, the original data should be reviewed or additional examinations should be conducted to clarify the true disease status.

[0068] Data quality backtesting was conducted for each modality based on its conflict level to obtain a modality reliability ranking. Data quality determines the reliability of modality information; data quality backtesting is necessary when the conflict level is moderate or severe to determine the more reliable modalities. Data quality assessment for the imaging modality includes three dimensions: image sharpness, scan completeness, and timeliness. Image sharpness is quantified using the signal-to-noise ratio (SNR). Scan completeness checks whether the entire lesion area is covered. Timeliness assesses the time interval between the CT scan and the current time point; if the interval exceeds 6 hours and the patient's condition fluctuates during that time, the reference value of the cerebral hemorrhage imaging data decreases. For clinical modalities with high conflict levels, data quality assessment focuses on the standardization of indicator collection, whether laboratory tests were performed under standard conditions, and whether there are abnormal values ​​due to equipment malfunctions in vital sign monitoring. The reliability of the clinical modality is significantly reduced when key indicators such as the Glasgow Coma Scale score or coagulation function indicators are missing. Data quality scores for both modalities were calculated separately and compared, with higher scores ranked first, forming a modality reliability ranking. The modality reliability ranking is stored in a list format, with the first position representing the most reliable modality and the second position representing the least reliable modality. When the conflict level is no conflict or mild conflict, the modality credibility ranking defaults to the order of imaging modality and then clinical modality, reflecting the fundamental role of imaging evidence in the prognostic assessment of intracerebral hemorrhage. In the modality credibility ranking, a difference of less than 0.1 between the scores of two modalities is marked as equivalent in credibility; in this case, the ranking result has little impact on the weighting. The conflict level and the modality credibility ranking together determine the bias of the fusion strategy; at higher conflict levels, modality information with higher credibility rankings is given priority.

[0069] Modal confidence ranking categorizes modalities into dominant and auxiliary modalities. This role determines the relative importance of each modality's information during the fusion process; the modality ranked first in confidence is designated as the dominant modality. The dominant modality provides the baseline risk assessment during fusion, with its feature vectors carrying the majority of the weight, typically accounting for 60%-80% of the fusion weight. The feature values ​​of the dominant modality constitute the main components of the fused prognostic feature vector. The modality ranked second in confidence is designated as the auxiliary modality, providing supplementary risk information to correct or validate the dominant modality's assessment. The distinction between dominant and auxiliary modalities is not fixed. As the disease progresses and data is updated, the modal confidence ranking may change, leading to a role reversal. For example, improved timeliness of brain hemorrhage imaging data after a follow-up CT scan may cause the imaging modality to shift from auxiliary to dominant. When the dominant modality is the imaging modality, the fusion results reflect more the impact of hematoma morphology and structural compression on prognosis. When imaging shows a midline shift exceeding 10 mm, even with relatively mild clinical symptoms, the fusion results tend to indicate a high-risk assessment, suggesting the need for surgical intervention. When the dominant modality is the clinical modality, the fusion results reflect more the impact of the patient's overall physiological state and organ function on prognosis. When modality reliability ranking shows that the reliability of two modalities is comparable, no explicit primary / secondary distinction is made; both modalities are labeled as equivalent modalities and a balanced weighting strategy is adopted. Although the secondary modality has a lower weight, it still plays an important role in supplementing information, especially when the dominant modality's information is incomplete or has assessment blind spots.

[0070] Modality weight configurations are generated based on differentiated weight allocation between the dominant and auxiliary modalities. The weight values ​​directly determine the contribution ratio of each modality's information to the fusion result; the difference in roles between the dominant and auxiliary modalities is reflected through weight differentiation. The modality weight configuration includes two values: the imaging modality weight *w_img* and the clinical modality weight *w_cli*. Their sum is always equal to 1 to ensure the normalization of the fusion result. When the dominant modality is the imaging modality, the baseline value of *w_img* is set to 0.65, and *w_cli* is correspondingly set to 0.35; when the dominant modality is the clinical modality, the weight configuration is reversed to *w_img*=0.35 and *w_cli*=0.65. The conflict level adjusts the baseline weights. For mild conflict, the baseline weights remain unchanged; for moderate conflict, the dominant modality weight increases to 0.70; for severe conflict, the auxiliary modality weight further decreases to 0.20, and the dominant modality weight increases accordingly to 0.80. Higher conflict levels rely more heavily on modality information with higher reliability. When both the dominant and auxiliary modalities are equivalent, the modality weights are configured as a balanced set of w_img=0.50 and w_cli=0.50. The modality weight configuration also includes a weight reset confidence field, which reflects the rationality of the current weight configuration. The confidence level is positively correlated with the modality consistency score. When the weight reset confidence level is below 0.6, an uncertainty warning is added to the metadata of the fused prognostic feature vector. The weight difference between the dominant and auxiliary modalities is controlled within a reasonable range. Even under severe conflict, the auxiliary modality weight is not lower than 0.15 to preserve its information contribution and prevent serious bias in the overall evaluation due to errors in a single modality's data.

[0071] Based on the modality weight configuration, differentiated fusion is performed to generate a fused prognostic feature vector. Weighted feature fusion linearly combines the feature vectors of the two modalities according to the configured weights. In the modality weight configuration, w_img and w_cli act on the imaging features and clinical features, respectively. The 18-dimensional feature vector from the imaging risk feature set and the 8-12-dimensional feature vector from the clinical risk factor set are first dimensionally aligned, and then expanded to the same dimension using zero-padding. The formula for calculating the fused prognostic feature vector is F = w_img × V_img + w_cli × V_cli, where V_img is the expanded imaging feature vector and V_cli is the expanded clinical feature vector. When the modality weight configuration is balanced, the fused prognostic feature vector is equivalent to the arithmetic mean of the two feature vectors; when the weights are differentiated, the fused prognostic feature vector is skewed towards the feature space of the dominant modality. During feature fusion, highly correlated feature terms are deredundant; if the hematoma volume in the imaging features and the hematoma volume in the clinical features are duplicated, they are merged into a single feature. After redundancy removal, the fusion prognostic feature vector typically has 20-25 dimensions, encompassing comprehensive information across four dimensions: hematoma morphology, structural compression, vital sign fluctuations, and neurological function scores. The weighted reliability information in the modal weight configuration is appended to the metadata of the fusion prognostic feature vector as a reference factor for calculating prediction confidence.

[0072] Step S150: Based on the fusion of prognostic feature vectors and modal consistency scores, perform prognostic assessment to obtain stratified prognostic indicators and prediction confidence levels, and output a prognostic analysis report based on the stratified prognostic indicators and prediction confidence levels.

[0073] In some embodiments, the step of obtaining stratified prognostic indicators and prediction confidence based on the fused prognostic feature vector and the modality consistency score includes: performing prognostic risk deconstruction on the fused prognostic feature vector to obtain functional prognostic scores and survival prognostic scores; performing prognostic risk stratification based on the functional prognostic scores and survival prognostic scores to obtain stratified prognostic indicators; performing feature stability analysis on the fused prognostic feature vector to obtain feature dispersion; and integrating the modality consistency score and the feature dispersion to generate prediction confidence.

[0074] Prognostic risk deconstruction was performed on the fused prognostic feature vector to obtain functional and survival prognostic scores. The prognostic assessment of intracerebral hemorrhage includes two dimensions: functional recovery and survival outcome. Different features in the fused prognostic feature vector contribute differently to these two dimensions. Structural features such as hematoma volume, midline shift, and ventricular deformation index have higher predictive weights for survival outcome, while Glasgow Coma Scale score and degree of neurological deficit have higher predictive weights for functional recovery. The calculation of the functional prognostic score selected a subset of features related to functional recovery from the fused prognostic feature vector, including four categories: baseline neurological function score, changes in consciousness, hemorrhage location, and patient age. Functional features in the fused prognostic feature vector were converted into a functional prognostic score of 0-100 using a logistic regression model. A lower score indicates a better prospect for functional recovery, while a higher score indicates a greater risk of disability. A functional prognostic score of 28 corresponds to a patient's basic independence in daily living activities at discharge, capable of independently performing basic activities such as dressing and eating. The survival prognostic score is calculated by selecting a subset of features related to survival outcome from the fused prognostic feature vector, including four categories of features: hematoma volume and expansion rate, severity of mass effect, coagulation function status, and stability of vital signs. The functional prognostic score and the survival prognostic score use the same numerical range for comprehensive assessment. There is a certain correlation between the two scores, but they are not completely consistent. Some patients may have a good survival prognosis but a poor functional prognosis.

[0075] Stratified prognostic indicators are obtained by stratifying prognostic risks based on functional and survival prognostic scores. A single score is insufficient to comprehensively depict prognostic status; the combination of functional and survival prognostic scores can distinguish different types of prognostic risk patterns. The stratified prognostic indicators employ a two-dimensional stratification strategy, classifying functional and survival prognostic scores into low-risk, intermediate-risk, and high-risk levels, resulting in nine prognostic combination types. Functional prognostic scores below 30 are classified as low-risk, 30-60 as intermediate-risk, and above 60 as high-risk; the same cutoff values ​​are used for survival prognostic score classification. The core outputs of the stratified prognostic indicators include prognostic risk level labels and a comprehensive prognostic index. Risk level labels are used for rapid classification, while the comprehensive prognostic index is used for fine-grained ranking. Low-risk combinations are labeled as having a good prognosis, high-risk combinations as having a poor prognosis, and mixed-risk combinations are labeled as having an intermediate or segregated prognosis, depending on the specific pattern. A separation pattern of high functional prognostic score and low survival prognostic score suggests that the patient may survive but suffer from severe disability. A survival prognostic score of 25 indicates a high probability of survival, but a functional prognostic score of 72 suggests a possible serious sequela such as hemiplegia or aphasia. The expected quality of life should be fully explained during doctor-patient communication. The comprehensive prognostic index is obtained by weighted averaging of the functional and survival prognostic scores. The two dimensions are set with equal weight by default, but the weighting can be adjusted according to the assessment purpose in specific clinical scenarios. Risk level labels in stratified prognostic indicators are color-coded for easy identification, and the comprehensive prognostic index is retained to two decimal places for precise comparison and prognosis tracking among patients.

[0076] Feature stability analysis was performed on the fused prognostic feature vector to obtain feature dispersion. The reliability of the prediction result depends not only on the feature values ​​themselves, but also on the stability of the features during the time-series observation period. Features with drastic fluctuations in the fused prognostic feature vector will reduce the certainty of the prediction. After extracting the values ​​of the fused prognostic feature vector from different time slices, the coefficient of variation for each dimension was calculated. The coefficient of variation is defined as the ratio of the standard deviation to the mean, reflecting the relative fluctuation of the feature. When the absolute value of the feature mean is less than a set threshold, the standard deviation is used instead of the coefficient of variation to avoid calculation anomalies. Feature dispersion is defined as the weighted average of the coefficients of variation of each dimension of the fused prognostic feature vector. The weight is proportional to the importance of each feature in the prognostic model, and the fluctuation of core features contributes more to the feature dispersion. In the fused prognostic feature vector, the hematoma volume increased from 25ml to 35ml during the observation period, with a coefficient of variation of approximately 0.23. The fluctuation of this feature indicates that the condition has not yet stabilized, and the uncertainty of the prediction result increases accordingly. The feature dispersion ranges from 0 to 1. The larger the value, the more drastic the feature fluctuation and the lower the stability of the prediction. When the fused prognostic feature vector contains only data from a single time point, the feature dispersion cannot be calculated based on temporal variation. In this case, the deviation of the feature value from the distribution of the reference population is used as a substitute indicator. A feature dispersion exceeding 0.4 indicates that the disease is in a rapid evolution phase, and the prediction result at the current time point may change significantly in the short term. A feature dispersion of 0.45 indicates that the hematoma volume or level of consciousness fluctuates drastically during the observation period. It is recommended to shorten the reassessment interval from 24 hours to 6 hours to capture changes in the disease in a timely manner.

[0077] The prediction confidence score is generated by integrating modality consistency score and feature dispersion. The reliability of the prediction result is jointly determined by the data fusion quality and feature stability; the modality consistency score reflects the former, and the feature dispersion reflects the latter. The prediction confidence score is calculated using a two-factor integration model, with the formula C = w_a × A + w_d × (1 - D), where C is the prediction confidence score, A is the modality consistency score, D is the feature dispersion, and w_a and w_d are the weights of the two factors, respectively. The weight w_a for the modality consistency score is set to 0.6, reflecting the fundamental impact of data fusion quality on prediction reliability; the weight w_d for feature dispersion is set to 0.4, reflecting the moderating effect of disease stability on prediction timeliness. The prediction confidence score ranges from 0 to 1; a higher value indicates a more reliable prediction result and greater clinical reference value. When the modality consistency score is 0.82 and the feature dispersion is 0.15, the calculated prediction confidence is 0.83, which is considered a high confidence level. This confidence level indicates a high degree of consistency between the imaging and clinical data and a relatively stable condition, making the prediction result a reliable basis for developing a treatment plan. When the prediction confidence is below 0.5, the reference value of the prediction result is limited, and it is recommended to supplement data collection or wait for the condition to stabilize before reassessing. Prediction confidence is also divided into three levels based on the numerical range: high confidence (greater than 0.75), medium confidence (0.5-0.75), and low confidence (less than 0.5). The modality consistency score and feature dispersion have different impact patterns on prediction confidence; the former reflects the reliability at the data level, while the latter reflects the certainty at the condition level. Both together determine the final clinical application value of the prediction.

[0078] A prognostic analysis report is generated based on stratified prognostic indicators and predictive confidence scores. The report integrates all results of the prognostic assessment into a structured document, with stratified prognostic indicators and predictive confidence scores presented as the core conclusions at the forefront. The title area of ​​the prognostic analysis report displays basic patient information, assessment time, and assessment type, labeled as a multimodal prognostic analysis fusing imaging and clinical data. Stratified prognostic indicators are presented in the report in dual forms: risk level labels and a comprehensive prognostic index. Risk level labels use intuitive color coding: green for good prognosis, yellow for moderate prognosis, and red for poor prognosis. Predictive confidence scores are displayed as percentages next to the risk level labels, along with the confidence level for quick interpretation. The details area of ​​the prognostic analysis report displays the calculation basis of the stratified prognostic indicators, including functional prognostic scores, survival prognostic scores, and their corresponding risk levels. The main contributing features of each score are visualized in bar chart form. The factors influencing predictive confidence scores are broken down and explained in the report, with specific values ​​for modal consistency scores and feature dispersion, and their contribution ratios to the confidence scores listed item by item. The prognostic analysis report generates personalized clinical recommendations based on a combination of stratified prognostic indicators and predicted confidence levels. For high-risk, high-confidence cases, aggressive intervention is recommended; for high-risk, low-confidence cases, close monitoring and reassessment are advised; and for cases with a characteristic dispersion exceeding 0.4, shorter reassessment intervals are recommended to promptly capture changes in the patient's condition. When stratified prognostic indicators show a prognostic segregation pattern, the report specifically highlights the difference between functional and survival prognoses and their clinical significance, assisting in doctor-patient communication and treatment goal setting. The prognostic analysis report supports version tracking, generating an independent report for each assessment and retaining historical records.

[0079] The following is a detailed description of a prognostic analysis system for cerebral hemorrhage that integrates imaging and clinical data, provided in an embodiment of this application. See also... Figure 3 , Figure 3 This diagram illustrates a structural block diagram of a brain hemorrhage prognostic analysis system 300 that integrates imaging and clinical data, as provided in an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The brain hemorrhage prognostic analysis system 300 that integrates imaging and clinical data, as provided in this embodiment of the application, includes:

[0080] Data acquisition module 301 is used to acquire cerebral hemorrhage imaging data and clinical diagnosis and treatment data, and to perform acquisition time point alignment on the cerebral hemorrhage imaging data and the clinical diagnosis and treatment data to construct a time-series associated data group;

[0081] Image analysis module 302 is used to segment the hematoma region based on the time-series correlated data group to obtain the spatial distribution of the hematoma, perform surrounding structure compression analysis through the spatial distribution of the hematoma to obtain structural displacement features, and extract hematoma expansion risk features based on the spatial distribution of the hematoma and the structural displacement features to form an image risk feature set.

[0082] Clinical analysis module 303 is used to perform clinical diagnosis and treatment indicator fluctuation analysis based on the time-series associated data group to obtain clinical dynamic characteristics, perform collaborative abnormal pattern recognition on the clinical dynamic characteristics to obtain linkage risk markers, and screen risk-sensitive indicators based on the linkage risk markers to form a set of clinical risk factors.

[0083] The fusion processing module 304 is used to perform risk orientation consistency detection on the image risk feature set and the clinical risk factor set to obtain a modality consistency score, determine the modality weight configuration based on the modality consistency score, and perform differential fusion to generate a fusion prognostic feature vector according to the modality weight configuration.

[0084] The prognostic assessment module 305 is used to perform prognostic assessment based on the fused prognostic feature vector and the modality consistency score to obtain stratified prognostic indicators and prediction confidence, and output a prognostic analysis report based on the stratified prognostic indicators and the prediction confidence.

[0085] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

[0086] The above embodiments are not an exhaustive list based on the present invention, and there may be many other embodiments not listed. Any substitutions and improvements made without departing from the concept of the present invention are within the protection scope of the present invention.

Claims

1. A system for prognostic analysis of cerebral hemorrhage that integrates imaging and clinical data, characterized in that, include: The data acquisition module is used to acquire cerebral hemorrhage imaging data and clinical diagnosis and treatment data, and to perform time-point alignment of the acquisition data and the clinical diagnosis and treatment data to construct a time-series related data group. The image analysis module is used to segment the hematoma region based on the time-series correlated data group to obtain the spatial distribution of the hematoma, perform compression analysis of the surrounding structures through the spatial distribution of the hematoma to obtain structural displacement features, and extract hematoma expansion risk features based on the spatial distribution of the hematoma and the structural displacement features to form an image risk feature set. The clinical analysis module is used to perform fluctuation analysis of clinical diagnosis and treatment indicators based on the time-series correlated data set to obtain clinical dynamic characteristics, and to perform collaborative abnormal pattern identification on the clinical dynamic characteristics to obtain linkage risk markers, including: extracting multidimensional indicator change vectors from the clinical dynamic characteristics; performing time-series correlation analysis between indicators on the multidimensional indicator change vectors to obtain an indicator linkage matrix; screening synchronous deterioration indicator combinations based on the indicator linkage matrix to obtain collaborative abnormal patterns; generating linkage risk markers based on the collaborative abnormal patterns; and screening risk-sensitive indicators based on the linkage risk markers to form a clinical risk factor set. The fusion processing module is used to perform risk orientation consistency detection on the image risk feature set and the clinical risk factor set to obtain a modality consistency score, including: performing risk severity quantification on the image risk feature set and the clinical risk factor set to obtain image risk scores and clinical risk scores respectively; calculating the risk orientation deviation of the image risk scores and the clinical risk scores to obtain modality deviation values; performing deviation cause analysis based on the modality deviation values ​​to identify data source credibility differences; generating a modality consistency score based on the modality deviation values ​​and the data source credibility differences; and determining modality weight configuration based on the modality consistency score, including: performing modality conflict degree determination based on the modality consistency score to obtain conflict levels; performing data quality backtracking assessment of each modality for the conflict levels to obtain modality credibility ranking; dividing the dominant modality and auxiliary modality according to the modality credibility ranking; performing weight differential allocation based on the dominant modality and the auxiliary modality to generate modality weight configuration; and performing differential fusion based on the modality weight configuration to generate a fusion prognostic feature vector. The prognostic assessment module is used to perform prognostic assessment based on the fused prognostic feature vector and the modality consistency score to obtain stratified prognostic indicators and prediction confidence scores. This includes: performing prognostic risk deconstruction on the fused prognostic feature vector to obtain functional prognostic scores and survival prognostic scores; performing prognostic risk stratification based on the functional prognostic scores and survival prognostic scores to obtain stratified prognostic indicators; performing feature stability analysis on the fused prognostic feature vector to obtain feature dispersion; integrating the modality consistency score and the feature dispersion to generate prediction confidence scores; and outputting a prognostic analysis report based on the stratified prognostic indicators and the prediction confidence scores.

2. The system according to claim 1, characterized in that, The step of segmenting the hematoma region based on the time-series correlated data group to obtain the spatial distribution of the hematoma includes: Perform intracranial high-density region detection on the brain hemorrhage imaging data in the time-series correlated data group to obtain candidate hematoma regions; Internal density stratification analysis was performed on the candidate hematoma region to obtain the hematoma core area and mixed density area; High-density spot localization and identification of active bleeding markers are performed on the mixed density areas; The spatial distribution of the hematoma is constructed based on the hematoma core region, the mixed density region, and the active bleeding markers.

3. The system according to claim 1, characterized in that, The method of obtaining structural displacement characteristics by performing compression analysis on surrounding structures through the spatial distribution of the hematoma includes: Based on the spatial distribution of the hematoma, the brain tissue region adjacent to the hematoma is located; Midline structural displacement quantization was performed on the brain tissue region adjacent to the hematoma to obtain the midline offset value; Based on the spatial distribution of the hematoma, morphological analysis of ventricular compression was performed to obtain the ventricular deformation index; The midline offset value and the ventricular deformation index are used to grade the severity of compression and generate structural offset features.

4. The system according to claim 3, characterized in that, The method of obtaining the ventricular deformation index by performing ventricular compression morphology analysis based on the spatial distribution of the hematoma includes: The contours of the bilateral ventricles were located based on the spatial distribution of the hematoma. The bilateral ventricular region contours were compared with the contralateral side to obtain bilateral asymmetry. The degree of collapse of the compressed side contour in the bilateral ventricle region is analyzed to obtain the degree of deformation value. The ventricular deformation index is generated based on the bilateral asymmetry and the deformation degree value.

5. The system according to claim 1, characterized in that, The step of obtaining coordinated anomaly patterns by filtering synchronously deteriorating indicator combinations based on the indicator linkage matrix includes: Extract strongly correlated indicator pairs from the indicator linkage matrix to obtain highly correlated indicator groups; For the highly correlated indicator group, a consistency judgment of the deterioration direction is performed to screen for indicator pairs that deteriorate in the same direction; The aforementioned co-deterioration indexes were used to verify the correlation between clinical and pathological mechanisms and obtain a cluster of related indicators. Based on the aforementioned mechanism, a collaborative anomaly pattern is generated using related indicator clusters.

Images