Post-stroke depression risk prediction system based on cerebral small vascular disease and inflammatory markers

Abstract

The invention discloses a post-stroke depression risk prediction system based on cerebral small vascular diseases and inflammatory markers, and relates to the technical field of computer-aided engineering, the post-stroke depression risk prediction system comprises: a data acquisition module acquires CSVD image data, serum inflammatory markers and clinical baseline information; the data preprocessing module processes image denoising registration, fills up marker missing values and encodes clinical information; the CSVD feature extraction module extracts image omics features and quantifies severity; the inflammation marker module calculates statistical characteristics and inflammation intensity; the multi-dimensional fusion module integrates features and eliminates redundancy; the risk prediction module is trained by using an improved attention CNN-LSTM model; the result output module visualizes risk and intervention suggestions; the model dynamic optimization module updates parameters by incremental learning. According to the method, multi-source key data are integrated, the prediction accuracy and generalization ability are improved, clinical interpretation and dynamic adaptability are achieved, early recognition of post-stroke depression is assisted, and patient prognosis is improved.

Description

Technical Field

[0001] This invention relates to the field of computer-aided engineering technology, and in particular to a post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers. Background Technology

[0002] Stroke has become a major disease threatening human health worldwide, with ischemic stroke accounting for a very high proportion. The incidence rate in my country far exceeds the world average. With the maturation of treatment techniques such as thrombolysis and thrombectomy, the survival rate of ischemic stroke patients has significantly improved. However, post-stroke depression, as a common complication, severely impacts patients' rehabilitation and quality of life. Post-stroke depression is mainly characterized by depressed mood, loss of interest, and slowed thinking. It not only hinders the recovery of neurological and cognitive functions but also significantly increases disability and mortality rates, placing a heavy burden on patients' families and society. In clinical practice, medical staff often focus more on patients' physical symptoms, while the emotional abnormalities of post-stroke depression are easily masked during short hospital stays. Traditional diagnostic methods rely on scale assessments, which suffer from strong subjectivity, assessment lag, and a high risk of missed diagnoses, making early and accurate identification difficult.

[0003] The development of machine learning technology has provided new pathways for disease risk prediction and has been applied in various medical scenarios. However, it still faces many limitations in the prediction of post-stroke depression. Existing prediction models mostly rely on single types of data or only integrate partial clinical baseline information, failing to fully combine the key roles of cerebral small vessel disease and inflammatory markers. Studies have confirmed that cerebral small vessel disease, by disrupting brain tissue structure and neurotransmitter transmission, and inflammatory markers, by affecting neuroendocrine function and monoamine neurotransmitter levels, are both closely related to the occurrence of post-stroke depression. However, there is currently no technology that deeply integrates the multidimensional imaging features of cerebral small vessel disease burden with the dynamic changes of inflammatory markers into the prediction model. Furthermore, existing models suffer from insufficient feature extraction, lack systematic mining of radiomics features and marker interaction features, and have insufficient generalization ability after training, failing to adapt to differences in equipment and data distribution across different medical centers. They also lack dynamic optimization mechanisms based on new case data, making it difficult to meet the needs of long-term clinical applications.

[0004] Furthermore, the clinical applicability of existing predictive models needs improvement. Most models only output risk probabilities, lacking clear risk level classifications and individualized intervention suggestions, making it difficult for healthcare professionals to quickly develop targeted plans. Some models are "black box" structures, unable to explain the mechanisms of action of various risk factors, reducing clinical trust. Simultaneously, the models do not consider risk differences at different stages of stroke or among different populations, and their fixed warning thresholds make them difficult to adapt to diverse clinical scenarios. These issues hinder the widespread clinical application of existing technologies and fail to effectively address the core pain points of early identification and delayed intervention in post-stroke depression. There is an urgent need to build a risk prediction system that integrates multi-source key data and possesses high accuracy and strong applicability. Summary of the Invention

[0005] The present invention proposes a post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers, comprising the following modules; The data acquisition module collects imaging data related to cerebral small vessel disease in stroke patients through medical imaging equipment, collects serum inflammatory marker data through clinical testing equipment, and collects patients' clinical baseline information at the same time. The data preprocessing module performs denoising, registration, and standardization on the acquired image data, fills in missing values ​​in the inflammatory marker detection data using multiple interquartile range method, identifies and corrects outliers using interquartile range method, encodes and transforms clinical baseline information, and converts categorical variables into numerical variables. The cerebral small vessel disease feature extraction module employs radiomics methods to extract morphological, textural, and grayscale histogram features from preprocessed image data. Figure 3 Class characteristics are used to quantify the severity score of cerebral small vessel disease, and a comprehensive severity index is calculated by combining the weights of various imaging indicators. The inflammatory biomarker feature processing module extracts statistical features from the standardized inflammatory biomarker data, calculates the detection value, dynamic change rate, and ratio of each biomarker to the normal reference value, and analyzes the correlation between biomarkers to construct an inflammatory response intensity index. The multi-dimensional feature fusion module aligns and normalizes the features of cerebral small vessel disease, inflammatory markers, and clinical baselines, and integrates multi-source data using a feature-level fusion strategy to eliminate dimensional differences and redundant information between features. The risk prediction model module is built on an improved attention mechanism CNN-LSTM hybrid model. The CNN network extracts the local spatial correlation information of features, the LSTM network captures the temporal dependency of features, and the attention mechanism adaptively allocates the feature weights of each dimension. The results output module outputs the probability of developing post-stroke depression, risk level, ranking of key risk factors, and individualized risk assessment report in a visual format, while also generating clinical intervention recommendations. The model dynamic optimization module regularly collects new clinical case data, updates model parameters using incremental learning algorithms, and adjusts feature weights and prediction thresholds based on clinical feedback to continuously improve the model's generalization ability and prediction accuracy.

[0007] Furthermore, it also includes a deep integration module for clinical baseline information. This module quantifies patients' cognitive function scores, emotional state baseline scores, sleep quality index, and social support level. The cognitive function scores are standardized and converted using the Montreal Cognitive Assessment Scale scores, the emotional state baseline is graded and coded using the Self-Rating Depression Scale scores, the sleep quality index is calculated using the Pittsburgh Sleep Quality Index, and the social support level is quantified using the Perceived Social Support Scale.

[0008] Furthermore, it also includes a feature selection and importance ranking module. This module uses recursive feature elimination combined with random forest algorithm to select the fused multi-dimensional features, calculates the Gini coefficient and information gain value of each feature, eliminates redundant and low-contribution features according to importance ranking, and retains the core features that have a significant impact on the prediction results. Feature selection reduces the computational load of the model.

[0009] Furthermore, it also includes a multi-factor weighted fusion module, which achieves deep feature integration by constructing a multi-dimensional feature weighted fusion formula, the formula being: Where F is the fused comprehensive feature vector, This represents the weighting coefficient for characteristics of cerebral small vessel disease. These are the feature weighting coefficients for inflammatory markers. The weighting coefficients represent the interaction characteristics between cerebral small vessel disease and inflammatory markers. These are the weighting coefficients for clinical baseline characteristics, and + + + =1, The number of cases is characteristic of cerebral small vessel disease. For the first Normalized coefficients of individual characteristics of cerebral small vessel disease For the first Quantitative values ​​of individual characteristics of cerebral small vessel disease. The number of inflammatory marker characteristics, For the first Normalization coefficients of inflammatory marker characteristics For the first Quantitative values ​​of the characteristics of an inflammatory marker For the first The characteristics of small vessel disease of the brain and the first The interaction coefficients of the characteristics of each inflammatory marker The number of clinical baseline characteristics, For the first Normalization coefficients of clinical baseline characteristics, For the first Quantification of clinical baseline characteristics.

[0010] Furthermore, it also includes a model interpretation module, which uses a combination of SHAP value analysis and partial dependency graphs to explain the degree and direction of influence of each core feature on the prediction results, generate a feature contribution heatmap, and output an interpretation report in clinically understandable language.

[0011] Furthermore, it also includes a risk warning threshold dynamic adjustment module. This module dynamically adjusts the risk warning threshold according to the intervention needs of different medical scenarios and the differences in population characteristics. It adopts the receiver operating characteristic curve combined with the Youden index maximization principle to set personalized warning thresholds for different stages of stroke and different age groups and underlying disease conditions of patients. When the patient's predicted risk probability exceeds the corresponding threshold, the warning prompt is automatically triggered, and the threshold adjustment process and basis are recorded at the same time.

[0012] Furthermore, it also includes a risk level classification module, which converts the predicted risk probability into a clinically applicable risk level through a risk level calculation model. The formula is as follows: ,in The risk level of post-stroke depression. For risk probability weighting coefficients, The probability of depression is output by the model. This is a weighted coefficient for the severity of cerebral small vessel disease. The comprehensive severity index of cerebral small vessel disease. This is the weighting coefficient for the intensity of the inflammatory response. As an indicator of the intensity of the inflammatory response, This is the rounding function. and The function is used to limit the risk level to the range of 1-5.

[0013] Furthermore, it also includes a cross-center data adaptation module. This module uses a domain-adaptive algorithm to eliminate data distribution shifts caused by differences in equipment and testing methods between different medical centers. It aligns the feature distributions of different centers by minimizing the maximum mean difference, constructs a cross-center data calibration model, and adaptively adjusts the input data of new centers. At the same time, it uses a transfer learning strategy to initialize the model with training data from existing centers and fine-tunes the model parameters by combining labeled data from new centers.

[0014] Furthermore, it also includes a real-time monitoring and updating module, which supports dynamic monitoring of patients' cerebral small vessel disease status and inflammatory marker levels, regularly collects patients' follow-up image data and laboratory test data, automatically inputs updated feature information into the system, recalculates the probability and level of depression risk, generates a risk change trend curve, and records the key driving factors of risk changes.

[0015] Furthermore, it also includes a clinical decision support module. This module generates individualized intervention recommendations based on risk prediction results and risk levels, combined with the latest clinical guidelines. For patients with very low risk, it recommends routine health education and regular follow-up; for patients with low risk, it recommends psychological counseling and lifestyle intervention; for patients with intermediate risk, it recommends combined psychological intervention and inflammation control treatment; for patients with high risk, it recommends standardized antidepressant drug treatment and cognitive behavioral therapy; and for patients with very high risk, it recommends multidisciplinary collaborative diagnosis and treatment and close monitoring.

[0016] Compared with existing technologies, the beneficial effects of this invention are: At the data integration and feature mining level, the system innovatively integrates imaging data of cerebral small vessel disease, serum inflammatory marker data, and clinical baseline information to achieve comprehensive coverage of key multi-source data. Through professional image processing techniques, it extracts multi-dimensional features of cerebral small vessel disease, including morphology and texture. The system analyzes the detection values, dynamic change rates, and interaction relationships of inflammatory markers, and combines this with quantitative processing of clinical baseline information, overcoming the limitations of existing models that rely on a single data source. The multi-dimensional feature fusion module, through scientific weight allocation and interactive feature mining, eliminates data redundancy and dimensional differences, fully releasing the synergistic value of multi-source data. This provides solid data support for accurate prediction, making the prediction results more reflective of the patient's true risk status.

[0017] In terms of model performance and adaptability, the CNN-LSTM hybrid model based on an improved attention mechanism can capture both the local spatial correlations of features and uncover temporal dependencies, adaptively allocating feature weights and significantly improving the accuracy and stability of predictions. The feature selection module removes low-contribution features, reducing computational load and further improving model efficiency. The cross-center data adaptation module, through domain adaptation algorithms and transfer learning strategies, eliminates data distribution biases across different medical centers, significantly enhancing the model's versatility and enabling stable application in diverse clinical scenarios. The model dynamic optimization module employs incremental learning algorithms, periodically integrating new case data to update parameters, continuously improving the model's generalization ability and avoiding performance degradation due to long-term use.

[0018] In terms of clinical applicability and intervention guidance, the system offers rich and intuitive outputs, providing not only risk probabilities and grading but also ranking key risk factors and individualized assessment reports, enabling healthcare professionals to quickly identify high-risk triggers in patients. The model interpretation module uses visualization to clearly define the mechanisms of action of each feature, breaking down the "black box" barrier and enhancing clinical trust. The dynamic adjustment module for risk warning thresholds and the risk grading mechanism adapt to the clinical needs of different stages and populations, providing healthcare professionals with clear intervention directions. The clinical decision support module generates differentiated intervention recommendations based on the latest guidelines, covering the entire risk spectrum from health education to multidisciplinary collaborative diagnosis and treatment, facilitating early and precise intervention and effectively improving patient prognosis. The real-time monitoring and update module supports dynamic tracking of patient conditions, allowing for timely adjustments to intervention plans, further improving treatment outcomes, and providing strong technical support for the early prevention and control of post-stroke depression, possessing significant clinical value and social significance. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers proposed in this invention. Figure 2 Line graph showing the dynamic changes in the patient's depression risk level during the follow-up period; Figure 3 A bar chart showing the importance of core predictive features. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 3 A post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers, comprising the following modules; The data acquisition module collects imaging data related to cerebral small vessel disease in stroke patients through medical imaging equipment, including the volume of white matter high signal, the number of cerebral microbleeds, the degree of enlargement of perivascular spaces, and the number and location of lacunar infarcts. It also collects serum inflammatory marker data through clinical testing equipment, including interleukin-6, C-reactive protein, and homocysteine. At the same time, it collects the patient's clinical baseline information, including age, gender, stroke type, onset time, past medical history, and neurological deficit score. The data preprocessing module performs denoising, registration, and standardization on the acquired image data. It uses a nonlocal mean filtering algorithm to remove image noise, a rigid registration technique to unify the image spatial coordinate system, and Z-score standardization to convert the image gray values ​​to a uniform distribution range. It uses multiple interpolation to fill missing values ​​in the inflammatory marker detection data, identifies and corrects outliers using the interquartile range method, encodes and converts clinical baseline information, and converts categorical variables into numerical variables. The cerebral small vessel disease feature extraction module uses radiomics methods to extract features from preprocessed image data, including morphological features, texture features, and gray-level histogram features. Morphological features include the sphericity and surface area-to-volume ratio of high signal in white matter. Texture features include the energy, contrast, and correlation of the gray-level co-occurrence matrix. Gray-level histogram features include the mean, standard deviation, and skewness. At the same time, the module quantifies the severity score of cerebral small vessel disease and calculates a comprehensive severity index by combining the weights of various image indicators. The inflammatory biomarker feature processing module extracts statistical features from the standardized inflammatory biomarker data, calculates the detection value, dynamic change rate, and ratio of each biomarker to the normal reference value, and analyzes the correlation between biomarkers to construct an inflammatory response intensity index. The multi-dimensional feature fusion module aligns and normalizes the features of cerebral small vessel disease, inflammatory markers, and clinical baselines, and integrates multi-source data using a feature-level fusion strategy to eliminate dimensional differences and redundant information between features. The risk prediction model module is built on an improved attention mechanism CNN-LSTM hybrid model. The CNN network extracts the local spatial correlation information of features, the LSTM network captures the temporal dependency of features, the attention mechanism adaptively assigns the weights of features in each dimension, the model training process uses cross-validation to optimize hyperparameters, and an early stopping strategy is used to prevent overfitting. The results output module outputs the probability of developing post-stroke depression, risk level, ranking of key risk factors, and individualized risk assessment report in a visual form, while also generating clinical intervention recommendations. The model dynamic optimization module regularly collects new clinical case data, updates model parameters using incremental learning algorithms, and adjusts feature weights and prediction thresholds based on clinical feedback to continuously improve the model's generalization ability and prediction accuracy.

[0024] This invention also includes a clinical baseline information deep integration module, which quantifies the patient's cognitive function score, emotional state baseline score, sleep quality index, and social support level. The cognitive function score is standardized and converted using the Montreal Cognitive Assessment Scale score, the emotional state baseline is graded and coded using the Self-Rating Depression Scale score, the sleep quality index is calculated using the Pittsburgh Sleep Quality Index, and the social support level is quantified using the Perceived Social Support Scale.

[0025] This invention also includes a feature selection and importance ranking module. This module uses recursive feature elimination combined with a random forest algorithm to select the fused multi-dimensional features, calculates the Gini coefficient and information gain value of each feature, and eliminates redundant and low-contribution features according to importance ranking, retaining the core features that have a significant impact on the prediction results. The core features include white matter high signal volume and number of cerebral microbleeds related to cerebral small vessel disease, interleukin-6 level and C-reactive protein dynamic change rate related to inflammatory markers, and age and neurological deficit score related to clinical baseline. Feature selection reduces the computational load of the model and improves the prediction speed and accuracy.

[0026] This invention also includes a multi-factor weighted fusion module, which achieves deep feature integration by constructing a multi-dimensional feature weighted fusion formula, the formula being: Where F is the fused comprehensive feature vector, This is a weighting coefficient for the characteristics of cerebral small vessel disease, with a value ranging from 0.3 to 0.4. This represents the characteristic weighting coefficient of inflammatory markers, with a value ranging from 0.25 to 0.35. The weighting coefficient for the interaction characteristics between cerebral small vessel disease and inflammatory markers ranges from 0.15 to 0.25. This is the weighting coefficient for clinical baseline characteristics, with a value ranging from 0.05 to 0.15. + + + =1, The number of cases is characteristic of cerebral small vessel disease. For the first Normalized coefficients of individual characteristics of cerebral small vessel disease For the first Quantitative values ​​of individual characteristics of cerebral small vessel disease. The number of inflammatory marker characteristics, For the first Normalization coefficients of inflammatory marker characteristics For the first Quantitative values ​​of the characteristics of an inflammatory marker For the first The characteristics of small vessel disease of the brain and the first The interaction coefficients of the characteristics of each inflammatory marker The number of clinical baseline characteristics, For the first Normalization coefficients of clinical baseline characteristics, For the first Quantification of clinical baseline characteristics.

[0027] This invention also includes a model interpretation module, which uses a combination of SHAP value analysis and partial dependency graphs to explain the degree and direction of influence of each core feature on the prediction results, generate a feature contribution heatmap, and identify key pathogenic factors in high-risk patients, such as the positive promoting effect of excessive white matter high signal volume on the risk of depression in cerebral small vessel disease, and the risk-enhancing effect of elevated interleukin-6 levels among inflammatory markers. At the same time, it outputs an interpretation report in clinically understandable language.

[0028] This invention also includes a risk warning threshold dynamic adjustment module. This module dynamically adjusts the risk warning threshold according to the intervention needs of different medical scenarios and the differences in population characteristics. It adopts the receiver operating characteristic curve combined with the Youden index maximization principle to set personalized warning thresholds for different stages of stroke, such as the acute phase and recovery phase, as well as for different age groups and underlying disease conditions of patients. When the patient's predicted risk probability exceeds the corresponding threshold, a warning prompt is automatically triggered, and the threshold adjustment process and basis are recorded at the same time.

[0029] This invention also includes a risk level classification module, which converts the predicted risk probability into a clinically applicable risk level through a risk level calculation model, using the following formula: ,in The risk level for post-stroke depression is graded on a scale of 1 to 5. This is the risk probability weighting coefficient, with a value ranging from 0.5 to 0.6. The probability of depression output by the model, taking a value between 0 and 1. This is a weighting coefficient for the severity of cerebral small vessel disease, with a value ranging from 0.2 to 0.3. The comprehensive severity index for cerebral small vessel disease ranges from 0 to 10. This is a weighting coefficient for the intensity of the inflammatory response, with a value ranging from 0.1 to 0.2. This is an indicator of the intensity of the inflammatory response, with a value ranging from 0 to 5. This is the rounding function. and The function is used to limit the risk level to the range of 1-5, where level 1 is very low risk, level 2 is low risk, level 3 is medium risk, level 4 is high risk, and level 5 is very high risk. Different risk levels correspond to different clinical intervention plans, providing clear action guidelines for medical staff.

[0030] This invention also includes a cross-center data adaptation module. This module uses a domain-adaptive algorithm to eliminate data distribution shifts caused by differences in equipment and detection methods between different medical centers. It aligns the feature distributions of different centers using the maximum mean difference minimization method, constructs a cross-center data calibration model, and adaptively adjusts the input data from new centers to ensure that its distribution is consistent with the model training data. At the same time, it uses a transfer learning strategy to initialize the model using training data from existing centers and fine-tunes the model parameters using a small amount of labeled data from new centers, thereby improving the system's versatility and stability in multi-center clinical applications.

[0031] This invention also includes a real-time monitoring and updating module. This module supports dynamic monitoring of the patient's cerebral small vessel disease status and inflammatory marker levels, regularly collects the patient's follow-up imaging data and laboratory test data, automatically inputs updated feature information into the system, recalculates the probability and level of depression risk, generates a risk change trend curve, and records key driving factors of risk changes, such as a decrease in inflammatory marker levels leading to a decrease in risk level and a progression of cerebral small vessel disease leading to an increase in risk level. This provides medical staff with a dynamic tracking tool for changes in the patient's condition and supports timely adjustments to early intervention plans.

[0032] This invention also includes a clinical decision support module. Based on risk prediction results and risk levels, this module generates individualized intervention recommendations in conjunction with the latest clinical guidelines. For patients with very low risk, routine health education and regular follow-up are recommended; for patients with low risk, psychological counseling and lifestyle intervention are recommended; for patients with moderate risk, combined psychological intervention and inflammation control treatment are recommended; for patients with high risk, standardized antidepressant drug treatment and cognitive behavioral therapy are recommended; and for patients with very high risk, multidisciplinary collaborative diagnosis and treatment and close monitoring are recommended.

[0033] The following two examples further illustrate specific embodiments of the present invention: Example 1: Application of post-stroke depression risk prediction in acute ischemic stroke patients in tertiary hospitals This embodiment targets acute ischemic stroke patients admitted to the neurology department of a tertiary hospital. The patients are within 2 weeks of onset and are ≥60 years old. A post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers is applied to verify the integration of multi-source data, deep feature fusion and model prediction accuracy. All technical modules are covered. The sample size is set at 300 cases according to the clinical study design requirements. The depression risk prediction and assessment of patients are completed from 7 to 30 days after admission.

[0034] 1. Data acquisition module in operation The data acquisition module used a 3.0T MRI device to acquire imaging data of the patient's cerebral small vessel disease. The scanning sequences included T1WI, T2WI, DWI, FLAIR, and SWI, with specific parameters as follows: 3D-T1 sequence TR=8ms, TE=3.2ms, matrix 256×256mm, voxel 1mm³; FLAIR sequence TR=8000ms, TE=120ms, slice thickness 5mm. The acquired imaging parameters included the volume of high signal in white matter, the number of cerebral microbleeds (counted in SWI sequence, low signal areas <10mm in diameter), the degree of enlargement of perivascular spaces (number of EPVS in the basal ganglia), and the number and location of lacunar infarcts (low signal lesions in DWI sequence, diameter <20mm).

[0035] Serum inflammatory markers were collected using clinical testing equipment: fasting venous blood was drawn within 24 hours of admission, and interleukin-6 (IL-6) was detected by ELISA (sensitivity 0.1 pg / mL); C-reactive protein (CRP) was detected by latex immunoturbidimetry (range 0.1-200 mg / L); and homocysteine ​​(Hcy) was detected by cyclic enzymatic assay (limit of detection 5 μmol / L). Simultaneously, baseline clinical information was collected: age, sex, stroke type (determined according to TOAST classification), onset time, past medical history (hypertension, diabetes, coronary artery disease, etc., defined according to clinical diagnostic criteria), and neurological deficit score (NIHSS score, assessed within 24 hours of admission, total score 0-42).

[0036] The clinical baseline information deep integration module supplemented the collection of baseline data on cognitive function, emotional state, sleep quality, and social support level. Cognitive function was assessed using the MoCA scale, with a maximum score of 30, and standardized to a 0-1 interval by dividing the actual score by 30. Emotional state baseline was assessed using the Self-Rating Depression Scale (SDS), with a maximum score of 100, graded as <53, 53-62, 63-72, and >72, coded as 1-4. Sleep quality was assessed using the PSQI index, with a maximum score of 21, and the actual score was used. Social support level was assessed using the Perceived Social Support Scale (MSPSS), with a maximum score of 84, and standardized to a 0-1 interval by dividing the actual score by 84.

[0037] 2. Data Preprocessing Module Operation Image data preprocessing: A nonlocal mean filtering algorithm was used to remove noise, with a filter window size of 7×7 and a similarity threshold of 0.1. Rigid registration was used to unify all images to the MNI152 standard spatial coordinate system, with registration errors controlled within 1mm. Z-score normalization was used to process the image grayscale values; the calculation formula is as follows: ,in The original grayscale value. The grayscale mean is... The standard deviation is used to convert gray values ​​to a uniform distribution range of -3 to 3 using this formula.

[0038] Inflammatory marker data processing: Missing values ​​were filled using multiple imputation with a missing rate of <5% and 5 imputations. Outliers were identified using the interquartile range method, and quartiles Q1 and Q3 were calculated. Outliers were defined as less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR, where IQR = Q3 - Q1. Outliers were corrected using the nearest neighbor mean replacement method. IL-6, CRP, and Hcy values ​​were standardized to the 0-1 range by dividing the actual value by the upper limit of the reference value, which was set as IL-6 < 7 pg / mL, CRP < 10 mg / L, and Hcy < 15 μmol / L.

[0039] Clinical baseline information processing: Categorical variables were converted into numerical variables. Gender was assigned as male=0 and female=1. Hypertension and diabetes were assigned as none=0 and present=1. Continuous variables, including age and NIHSS score, were standardized to the 0-1 interval using Min-Max to ensure uniformity of feature dimensions.

[0040] 3. Feature extraction and processing module execution The cerebral small vessel disease feature extraction module uses radiomics methods to extract features: morphological features are calculated by determining the sphericity to surface area-to-volume ratio of white matter hypersignal. The formula for sphericity calculation is as follows: ,in For volume, The equivalent diameter; texture features are extracted using the gray-level co-occurrence matrix to determine energy, contrast, and correlation. The energy calculation formula is as follows: The formula for calculating contrast is: The correlation calculation formula is: The matrix distance was set to 1, and the angle was set to 0° / 45° / 90° / 135°. The mean, standard deviation, and skewness of the gray-level histogram features were calculated. At the same time, the comprehensive severity index of cerebral small vessel disease was quantified: according to the clinical imaging assessment standards, ≥2 points on the FLAIR sequence Fazekas scale were counted as 1 point, ≥1 lacunar infarct was counted as 1 point, ≥1 deep cerebral microbleed was counted as 1 point, and >10 EPVS in the basal ganglia were counted as 1 point, with a total score of 0-4 points. The index was calculated by dividing the total score by 4 and then standardizing it to the 0-10 range.

[0041] The inflammatory marker characteristic processing module calculates the detected value, dynamic change rate, and ratio to normal reference value for each marker. The detected value is a standardized result. The dynamic change rate is the difference between the detected value at 7 days and 24 hours after admission divided by the 24-hour detected value. The ratio to normal reference value is the actual value divided by the upper limit of the reference value. Pearson correlation analysis is used to calculate the correlation between markers. The correlation coefficient between IL-6 and CRP is r=0.62. An inflammatory response intensity index is constructed, calculated using the following formula: The value is scaled to the 0-5 range by multiplying by 5.

[0042] The feature selection and importance ranking module uses a combination of recursive feature elimination and random forest algorithm to select features: the number of decision trees is set to 100, the cross-validation fold is 5, and the Gini coefficient of each feature is calculated. Among the features of cerebral small vessel disease, the Gini coefficient of white matter high signal volume is 0.18 and the number of cerebral microbleeds is 0.15. Among the features of inflammatory markers, IL-6 level is 0.16 and CRP dynamic change rate is 0.14. Among the clinical baseline features, age is 0.12 and NIHSS score is 0.11. Features with a Gini coefficient <0.1 are removed according to importance ranking, and 12 core features are retained, reducing the model's computational load by 30%.

[0043] 4. Feature Fusion and Model Training The multi-dimensional feature fusion module adopts a feature-level fusion strategy. First, it normalizes all core features to the 0-1 range through Min-Max standardization, and then substitutes them into the multi-dimensional feature weighted fusion formula: Setting parameters: =0.35 (weight of cerebral small vessel disease characteristics) =0.3 (weight of inflammatory marker characteristics) =0.2 (interaction feature weight) =0.15 (weight of clinical baseline characteristics). =4 (Core features of cerebral small vessel disease: volume of high signal in white matter, number of cerebral microbleeds, degree of EPVS, and comprehensive severity index). All are 0.25 (equal weight normalization). The values ​​are 0.6, 0.4, 0.3, and 0.5 (standardized quantized values), respectively. =4 (Core features of inflammatory markers: IL-6 level, CRP dynamic change rate, Hcy ratio, inflammatory response intensity S), Both are 0.25. The values ​​are 0.5, 0.3, 0.4, and 0.6 respectively. Set to 0.1 (unified interaction coefficient); =4 (Core clinical baseline features: age, NIHSS score, MoCA score, SDS score). Both are 0.25. The values ​​are 0.7, 0.4, 0.6, and 0.3 respectively. The calculation process is as follows: ; ; ; . The fused integrated feature vector F = 0.4659 is used as the model input.

[0044] The risk prediction model module constructs an improved attention mechanism CNN-LSTM hybrid model: The CNN network uses 3 convolutional layers with 3×3 kernels of 32, 64, and 128 neurons respectively, combined with 2 pooling layers using max pooling with a stride of 2 to extract local spatial correlation features; the LSTM network has 2 hidden layers with 128 and 64 neurons respectively, capturing temporal dependencies; the attention mechanism calculates the weights of each feature through fully connected layers, with 0.38 for cerebral small vessel disease features, 0.32 for inflammatory markers, and 0.3 for clinical baseline. Model training uses 5-fold cross-validation to optimize hyperparameters, with a learning rate of 0.001 and a batch size of 32. An early stopping strategy is set so that training stops if the validation set loss does not decrease for three consecutive rounds to prevent overfitting.

[0045] 5. Results Output and Model Optimization The results output module displays the following in a visual format: Risk Probability (model output F corresponds to a probability of 0.48) and Risk Level (substituted into the risk level calculation model): set up =0.55 (risk probability weight) =0.25 (weight of severity of cerebral small vessel disease) =0.15 (weight of inflammatory response intensity) =0.48 (risk probability) =5 (Comprehensive Severity Index, standardized 0.5 × 10) =3 (Intensity of inflammatory response), the calculation process is as follows: The risk level is 2 (low risk); key risk factors are ranked (high signal volume in white matter > IL-6 level > NIHSS score); individualized assessment report (including image feature screenshots and biomarker detection curves); clinical intervention recommendations (psychological counseling + lifestyle intervention).

[0046] The model interpretation module uses SHAP value analysis to generate a feature contribution heatmap: white matter high signal volume SHAP value 0.18 (positive contribution), IL-6 level 0.15 (positive), NIHSS score 0.12 (positive), MoCA score -0.09 (negative), identifying high-risk triggers; the clinical language output interpretation report states: "Increased white matter high signal volume (Fazekas scale 2 points) and elevated IL-6 level (3.5 pg / mL) may increase the risk of depression, while normal cognitive function (MoCA score 24 points) can reduce the risk."

[0047] The model dynamic optimization module collects new case data every 3 months, 50 cases each time, and uses an incremental learning algorithm to update the model parameters: freeze the underlying parameters of CNN-LSTM, fine-tune the weights of the attention layer and fully connected layer, and reduce the learning rate to 0.0005 to ensure that the model's generalization ability continues to improve.

[0048] 6. Application effect data Table 1 compares the predictive performance of depression risk in patients with acute ischemic stroke in tertiary hospitals:

[0049] Table 1 shows that existing single clinical models rely solely on baseline information, resulting in low feature coverage and an inability to integrate imaging and inflammatory data, leading to insufficient sensitivity and specificity. Existing imaging models only integrate cerebral small vessel disease data, lacking synergy between inflammatory and clinical features, and exhibiting moderate predictive performance. The system of this invention, through deep fusion of multi-source features, simultaneously incorporates three core data categories: imaging, inflammation, and clinical data, significantly improving AUC, sensitivity, and specificity. It offers comprehensive feature coverage and strong clinical interpretability. This advantage stems from the mining of interactive features between cerebral small vessel disease and inflammatory biomarkers, the weighting of key features through attention mechanisms, and the model interpretation module's solution to the "black box" problem, making the prediction results more accurate and easier for clinicians to understand and accept, effectively meeting the clinical needs of early depression risk screening in patients with acute ischemic stroke.

[0050] Example 2: Application of post-stroke depression risk prediction in patients recovering from ischemic stroke in community hospitals This embodiment targets patients in the recovery period of ischemic stroke who are followed up in the neurology department of a community hospital. The patients are 1-3 months after the onset of the disease and are ≥60 years old. The risk prediction system of this invention is applied to verify the cross-center data adaptation, real-time monitoring and clinical decision support functions, covering all technical modules, with a sample size of 200 cases, and completes dynamic risk assessment and intervention guidance for patients during the 3-month follow-up period.

[0051] 1. Cross-center data adaptation and collection The cross-center data adaptation module addresses the equipment differences between community hospitals and tertiary hospitals: Community hospitals use 1.5T MRI equipment with image data voxel sizes of 1.5mm³, resulting in resolution differences. A domain-adaptive algorithm is used to calculate the maximum mean difference (MMD), aligning the image feature distributions of different equipment, reducing the MMD value from 0.28 to 0.05. A cross-center data calibration model is constructed, performing grayscale stretching on the 1.5T MRI image data with a stretching coefficient of 1.2, and employing the Sobel operator for edge enhancement to align it with the distribution of the 3.0T data from tertiary hospitals. Furthermore, the methods for detecting inflammatory markers in community hospitals—immunoturbidimetric assay (CRP) and chemiluminescence immunoassay (IL-6)—differ from those in tertiary hospitals. A transfer learning strategy is employed, initializing the model using 1000 labeled data points from tertiary hospitals and fine-tuning the model parameters using 20 labeled data points from community hospitals, with a learning rate of 0.0001 and 10 iterations, to eliminate the impact of differences in detection methods.

[0052] The data acquisition module collects data periodically: imaging data (T2WI and FLAIR sequences of 1.5T MRI), inflammatory markers (fasting venous blood, detection frequency consistent with Example 1), and clinical baseline update information (NIHSS score and MoCA score changes) are collected once a month; the real-time monitoring and update module automatically stores each collection of data and generates time series features, such as the monthly average change rate of IL-6 of -0.12.

[0053] 2. Real-time monitoring and feature updates The real-time monitoring and update module dynamically tracks the patient's status: At the first month follow-up, the patient's white matter high signal volume remained unchanged, with a quantified value of 0.6; the IL-6 level decreased from 3.5 pg / mL to 2.8 pg / mL, with a dynamic change rate of -0.2; and the NIHSS score decreased from 4 to 2. At the second month follow-up, the white matter high signal volume remained stable at 0.6; the IL-6 level decreased to 2.5 pg / mL, with a dynamic change rate of -0.11; and the NIHSS score was 1. The system automatically updated the feature information and recalculated the comprehensive feature vector F=0.32, which is lower than the initial calculation of 0.4659; the risk probability was recalculated to 0.35.

[0054] The risk warning threshold dynamic adjustment module sets personalized thresholds for patients in the recovery period: based on the receiver operating characteristic curve and using the Youden exponent maximization principle, the warning threshold for patients in the recovery period is adjusted from 0.5 in the acute phase to 0.4. When the patient's predicted risk probability exceeds the corresponding threshold, a warning prompt is automatically triggered; in this case, the patient's risk probability is 0.35 < 0.4, so no warning is triggered.

[0055] 3. Clinical decision support and model optimization The clinical decision support module generates intervention recommendations based on risk level: In the first month, the patient's risk level is level 2 (low risk), and psychological counseling and lifestyle intervention are recommended. The frequency of psychological counseling is once a week for 30 minutes each time, and the lifestyle intervention includes 30 minutes of exercise per day and a low-salt diet. In the second month, the patient's risk level drops to level 1 (very low risk), and routine health education and regular follow-up are recommended. The frequency of health education is once a month, and the follow-up interval is adjusted to once every 2 months.

[0056] The model dynamic optimization module integrates new case data from community hospitals every 6 months, with 40 cases each time. Incremental learning is used to update the model: 30 cases from the new data are used for parameter fine-tuning, freezing the LSTM layer parameters and only updating the weights of the CNN layer and attention layer; 10 cases are used for model validation, which improves the model's AUC value in the community hospital scenario from 0.82 to 0.86, and continuously enhances its adaptability.

[0057] 4. Application effect data Table 2 shows the prediction and intervention effects of depression risk in patients recovering from ischemic stroke in community hospitals:

[0058] Table 2 shows that the performance of the non-adapted cross-center model significantly decreased when applied in community hospitals due to the lack of handling of differences in equipment and testing methods, failing to accurately capture patient risk characteristics. The model without real-time monitoring could only assess risk once, unable to track changes in the patient's condition during the recovery period, resulting in fixed intervention recommendations that did not match the actual progression of the patient's condition. The system of this invention eliminates differences in equipment and testing methods through a cross-center data adaptation module, ensuring application stability in primary healthcare institutions. The real-time monitoring module dynamically updates risk data, reflecting the trend of changes in the patient's condition. The clinical decision support module generates intervention recommendations matching the recovery period based on changes in risk level, better meeting patient needs. Therefore, patient compliance is significantly improved, ultimately reducing the incidence of depression. This result verifies the practicality of the system in community hospital scenarios, solving the problems of difficult cross-center application and insufficient dynamic intervention of existing models. It provides an effective technical tool for primary healthcare institutions to carry out post-stroke depression prevention and control, and facilitates the implementation of hierarchical medical treatment.

[0059] Reference Figure 2This line graph highlights the clinical value of the real-time monitoring and intervention guidance module of this invention. Patients without intervention experienced a slow decline in risk level due to the lack of targeted measures, remaining at low risk after 12 weeks and consistently above intermediate risk. In contrast, patients intervened using the system of this invention saw their risk level gradually decrease through dynamic monitoring of risk changes, reaching extremely low risk after 12 weeks, and the rate of decline was significantly faster than in the non-intervention group. This data validates the effectiveness of the system's "dynamic monitoring-precise intervention" closed loop, helping medical staff flexibly optimize intervention strategies based on changes in patient risk, shortening the duration of high risk, and reducing the probability of PSD (post-traumatic stress disorder).

[0060] Reference Figure 3 The bar chart visually presents the contribution of each feature to the prediction of depression risk, confirming the effectiveness of the feature screening strategy of this invention. White matter high signal volume and serum IL-6 level are the top two features, with importance values ​​of 0.22 and 0.19 respectively, indicating that structural damage and inflammatory response in cerebral small vessel disease are the core mechanisms driving PSD, consistent with the core design logic of "CSVD + inflammatory markers" in this invention. Clinical and imaging features such as NIHSS score and the number of cerebral microbleeds follow closely behind, while basic features such as age have lower importance, suggesting that clinical intervention should focus on improving cerebral small vessel disease and controlling inflammation. This clear ranking of feature importance helps medical staff quickly identify key triggers in high-risk patients, providing precise targets for individualized intervention.

[0061] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers, characterized in that, Includes the following modules; The data acquisition module collects imaging data related to cerebral small vessel disease in stroke patients through medical imaging equipment, collects serum inflammatory marker data through clinical testing equipment, and collects patients' clinical baseline information at the same time. The data preprocessing module performs denoising, registration, and standardization on the acquired image data, fills in missing values ​​in the inflammatory marker detection data using multiple interquartile range method, identifies and corrects outliers using interquartile range method, encodes and transforms clinical baseline information, and converts categorical variables into numerical variables. The feature extraction module for cerebral small vessel disease uses radiomics to extract three types of features—morphology, texture, and grayscale histogram—from the preprocessed image data, quantifies the severity score of cerebral small vessel disease, and calculates a comprehensive severity index by combining the weights of various image indicators. The inflammatory biomarker feature processing module extracts statistical features from the standardized inflammatory biomarker data, calculates the detection value, dynamic change rate, and ratio of each biomarker to the normal reference value, and analyzes the correlation between biomarkers to construct an inflammatory response intensity index. The multi-dimensional feature fusion module aligns and normalizes the features of cerebral small vessel disease, inflammatory markers, and clinical baselines, and integrates multi-source data using a feature-level fusion strategy to eliminate dimensional differences and redundant information between features. The risk prediction model module is built on an improved attention mechanism CNN-LSTM hybrid model. The CNN network extracts the local spatial correlation information of features, the LSTM network captures the temporal dependency of features, and the attention mechanism adaptively allocates the feature weights of each dimension. The results output module outputs the probability of developing post-stroke depression, risk level, ranking of key risk factors, and individualized risk assessment report in a visual format, while also generating clinical intervention recommendations. The model dynamic optimization module regularly collects new clinical case data, updates model parameters using incremental learning algorithms, and adjusts feature weights and prediction thresholds based on clinical feedback to continuously improve the model's generalization ability and prediction accuracy.

2. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a deep integration module for clinical baseline information. This module quantifies patients’ cognitive function scores, emotional state baseline scores, sleep quality index, and social support level. The cognitive function scores are standardized and converted using the Montreal Cognitive Assessment Scale scores, the emotional state baseline is graded and coded using the Self-Rating Depression Scale scores, the sleep quality index is calculated using the Pittsburgh Sleep Quality Index, and the social support level is quantified using the Perceived Social Support Scale.

3. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a feature selection and importance ranking module. This module uses recursive feature elimination combined with random forest algorithm to select multi-dimensional features after fusion, calculates the Gini coefficient and information gain value of each feature, eliminates redundant features and low contribution features according to importance ranking, and retains core features that have a significant impact on the prediction results. Feature selection reduces the computational load of the model.

4. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a multi-factor weighted fusion module, which achieves deep feature integration by constructing a multi-dimensional feature weighted fusion formula, the formula being: Where F is the fused comprehensive feature vector, This represents the weighting coefficient for characteristics of cerebral small vessel disease. These are the feature weighting coefficients for inflammatory markers. The weighting coefficients represent the interaction characteristics between cerebral small vessel disease and inflammatory markers. These are the weighting coefficients for clinical baseline characteristics, and + + + =1, The number of cases is characteristic of cerebral small vessel disease. For the first Normalized coefficients of individual characteristics of cerebral small vessel disease For the first Quantitative values ​​of individual characteristics of cerebral small vessel disease. The number of inflammatory marker characteristics, For the first Normalization coefficients of inflammatory marker characteristics For the first Quantitative values ​​of the characteristics of an inflammatory marker For the first The characteristics of small vessel disease of the brain and the first The interaction coefficients of the characteristics of each inflammatory marker The number of clinical baseline characteristics, For the first Normalization coefficients of clinical baseline characteristics, For the first Quantification of clinical baseline characteristics.

5. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a model interpretation module, which uses a combination of SHAP value analysis and partial dependency graphs to explain the degree and direction of influence of each core feature on the prediction results, generate a feature contribution heatmap, and output an interpretation report in clinically understandable language.

6. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a risk warning threshold dynamic adjustment module. This module dynamically adjusts the risk warning threshold according to the intervention needs of different medical scenarios and the differences in population characteristics. It adopts the receiver operating characteristic curve combined with the Youden index maximization principle to set personalized warning thresholds for different stages of stroke and different age groups and underlying disease conditions of patients. When the patient's predicted risk probability exceeds the corresponding threshold, the warning prompt is automatically triggered, and the threshold adjustment process and basis are recorded at the same time.

7. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a risk level classification module, which converts the predicted risk probability into a clinically applicable risk level through a risk level calculation model. The formula is as follows: ,in The risk level of post-stroke depression. For risk probability weighting coefficients, The probability of depression is output by the model. This is a weighted coefficient for the severity of cerebral small vessel disease. The comprehensive severity index of cerebral small vessel disease. This is the weighting coefficient for the intensity of the inflammatory response. As an indicator of the intensity of the inflammatory response, This is the rounding function. and The function is used to limit the risk level to the range of 1-5.

8. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a cross-center data adaptation module, which uses a domain-adaptive algorithm to eliminate data distribution shifts caused by differences in equipment and testing methods between different medical centers. It aligns the feature distributions of different centers by minimizing the maximum mean difference, builds a cross-center data calibration model, and adaptively adjusts the input data of new centers. At the same time, it uses a transfer learning strategy to initialize the model with training data from existing centers and fine-tunes the model parameters by combining labeled data from new centers.

9. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a real-time monitoring and updating module, which supports dynamic monitoring of patients' cerebral small vessel disease status and inflammatory marker levels, regularly collects patients' follow-up image data and laboratory test data, automatically inputs updated feature information into the system, recalculates the probability and level of depression risk, generates a risk change trend curve, and records the key driving factors of risk changes.

10. The post-stroke depression risk prediction system based on cerebral small vessel disease and inflammatory markers according to claim 1, characterized in that, It also includes a clinical decision support module, which generates individualized intervention recommendations based on risk prediction results and risk levels, combined with the latest clinical guidelines. For patients with very low risk, it recommends routine health education and regular follow-up; for patients with low risk, it recommends psychological counseling and lifestyle intervention; for patients with intermediate risk, it recommends combined psychological intervention and inflammation control treatment; for patients with high risk, it recommends standardized antidepressant drug treatment and cognitive behavioral therapy; and for patients with very high risk, it recommends multidisciplinary collaborative diagnosis and treatment and close monitoring.

Images