Mild behavioral disorder cognitive decline prediction method based on multi-modal image fusion

Abstract

The invention discloses a mild behavioral disorder cognitive decline prediction method based on multi-modal image fusion, and relates to the technical field of mild cognitive impairment prediction. According to the method, feature vectors of different modes are spliced to form a joint feature matrix, redundant information is eliminated through principal component analysis, and main variation components are reserved; inputting the feature set after dimension reduction into a prediction model based on machine learning, optimizing hyper-parameters through cross validation, and training a classifier so as to identify a cognitive decline risk mode of the patient with the mild behavioral disorder; multi-modal image data are received in real time through a standardized API interface, a prediction algorithm is automatically executed, and a report including individualized risk scores, confidence intervals, risk level classification and interpretability analysis is generated; the report is output in a structured format, so that a clinician can quickly refer to the report in the diagnosis process, meanwhile, the report is supported to be exported in a PDF or electronic health record format, and seamless integration with an existing medical information system is ensured.

Description

Technical Field

[0001] This invention relates to the field of mild cognitive impairment prediction technology, specifically to a method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion. Background Technology

[0002] Mild cognitive impairment is a transitional state between normal adulthood and dementia. It is characterized by cognitive impairment, but not as severe as dementia. Generally, individuals with mild cognitive impairment can live independently and are self-sufficient. The prevalence of mild cognitive impairment is approximately 15.5%, which may progress to dementia. This high conversion rate is based on a certain time frame; some people may never develop dementia and remain in the early stages of dementia, i.e., mild cognitive impairment.

[0003] A search revealed a high-precision intelligent auxiliary diagnostic method for early diagnosis of type 2 diabetes mellitus (T2DM) cognitive impairment due to brain aging (CN112837807B). This application discloses a method addressing the urgent need for classification and statistical analysis of T2DM cognitive impairment. It collects clinical data and brain MRI images of patients with T2DM, performs image preprocessing and data augmentation, and utilizes neural network methods to automatically detect T2DM cognitive impairment and output the results. However, this application is limited by relying solely on voxel-based morphological analysis, which is insufficient and leads to a low recognition rate. Therefore, we propose a multimodal image fusion-based method for predicting cognitive decline in mild behavioral disorders. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0006] A method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion includes the following steps:

[0007] Step 1: Data Acquisition. Collect multimodal brain imaging data from patients with mild behavioral disorders.

[0008] Step 2: Data preprocessing. The acquired multimodal brain imaging data is preprocessed, including motion correction, spatial standardization, intensity normalization, denoising, registration, and standardization, to ensure consistent data quality.

[0009] Step 3: Feature extraction. Key features are extracted from the preprocessed images, covering multimodal information such as structure, function, and metabolism. Brain region volume, functional connectivity strength, and metabolic activity indicators are extracted.

[0010] Step 4: Multimodal fusion. The extracted features are fused using a feature-level fusion strategy. After dimensionality reduction through principal component analysis, the features are input into the prediction model.

[0011] Step 5: Build a model, construct a machine learning-based cognitive decline prediction model, and train and validate it using a fusion feature set;

[0012] Step 6: Predict the model output. Apply the trained model to predict the risk of cognitive decline in patients with mild behavioral disorders and output the prediction results.

[0013] Step 7: Performance evaluation. The trained model is evaluated, including key metrics such as accuracy, recall, and F1 score, to verify the model's reliability and generalization ability.

[0014] Step 8: Model optimization. Optimize model parameters based on evaluation results, such as adjusting the learning rate or regularization coefficient, and use grid search technology to improve prediction accuracy.

[0015] Step 9: Risk prediction. Integrate the optimized model into the clinical decision support system to achieve automated risk prediction and output a detailed cognitive decline probability report.

[0016] Step 10: Model update. Regularly update the model using newly acquired multimodal brain imaging data to ensure the timeliness and adaptability of the prediction method.

[0017] Furthermore, the multimodal imaging in the data acquisition includes structural magnetic resonance imaging, functional magnetic resonance imaging, and positron emission tomography (PET) data. Structural magnetic resonance imaging is used to capture volume changes in gray and white matter of the brain, functional magnetic resonance imaging is used to analyze functional connectivity patterns in resting or task states, and PET is used to monitor brain metabolic activities such as glucose uptake or amyloid deposition.

[0018] Furthermore, the data acquisition also includes quality control of the image data, specifically including automatic detection of head motion artifacts, quantitative evaluation of signal-to-noise ratio, verification of image registration accuracy, and elimination of abnormal data.

[0019] Furthermore, spatial standardization in the data preprocessing adopts a template-based registration method, using the MNI 152 standard brain template.

[0020] Furthermore, the functional connectivity strength in the feature extraction is obtained by calculating the time-series correlation of resting-state fMRI.

[0021] Furthermore, in the multimodal fusion, the prediction model is a support vector machine classifier, and the kernel function is a radial basis function.

[0022] Furthermore, the performance evaluation also includes using receiver operating characteristic curves and area under the curve to comprehensively quantify the model's discrimination ability, and combining confusion matrix analysis to determine false positive and false negative rates.

[0023] Furthermore, the model optimization process employs a Bayesian optimization algorithm to automatically search for optimal hyperparameters, such as regularization coefficients and kernel function parameters, in order to improve prediction accuracy and reduce the risk of overfitting.

[0024] Furthermore, the risk prediction output report integrates the patient's individualized risk score, confidence interval, and visualization charts, facilitating clinical interpretation.

[0025] Furthermore, the model update uses an online learning mechanism to periodically and incrementally update the model parameters, and verifies the model's timeliness based on new data, ensuring that the prediction method continues to adapt to clinical needs.

[0026] The beneficial effects of this invention are as follows:

[0027] 1. This invention concatenates feature vectors from different modalities to form a joint feature matrix, uses principal component analysis to eliminate redundant information and retain the main variable components; the dimensionality-reduced feature set is input into a machine learning-based prediction model, and the hyperparameters are optimized and the classifier is trained through cross-validation to identify the cognitive decline risk patterns of patients with mild behavioral disorders; at the same time, the feature importance assessment method is applied to analyze the contribution of each modality to enhance the interpretability of the model.

[0028] 2. This invention receives multimodal image data in real time through a standardized API interface, automatically executes prediction algorithms, and generates a report that includes individualized risk scores, confidence intervals, risk level classifications, and interpretability analysis. The report is output in a structured format, making it easy for clinicians to quickly refer to during the diagnostic process. It also supports exporting to PDF or electronic health record formats, ensuring seamless integration with existing medical information systems. In addition, the system has a built-in verification module that continuously monitors prediction performance and triggers alarms when anomalies occur, in order to maintain the clinical reliability and usability of the prediction results.

[0029] 3. The model update mechanism of this invention incrementally learns new data quarterly, uses Bayesian optimization to adjust parameters, monitors AUC and F1 scores to maintain accuracy >95%, and applies differential privacy to protect sensitive information. The entire process emphasizes interpretability and robustness, reveals the role of biomarkers through feature importance analysis, and uses feedback loops to optimize the model to adapt to changes in patient subgroups. Ultimately, it achieves efficient and reliable cognitive decline risk assessment and intervention management in clinical decision support systems. Attached Figure Description

[0030] Figure 1 This is a flowchart of the process of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0032] Please see Figure 1 This invention provides a method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion, comprising the following steps:

[0033] Step 1: Data Acquisition. Collect multimodal brain imaging data from patients with mild behavioral disorders, including structural magnetic resonance imaging (sMRI), functional magnetic resonance imaging (fMRI), positron emission tomography (PET), and diffusion tensor imaging (DTI) data. Perform the data using standard clinical scanning equipment (such as a 3T MRI scanner) and follow established imaging protocols to ensure data quality and consistency.

[0034] Step 2: Data Preprocessing. Preprocessing of the acquired multimodal brain imaging data includes motion correction, spatial normalization, intensity normalization, denoising, registration, and standardization to ensure consistent data quality. Motion correction employs head motion parameter estimation and correction algorithms (such as the MCFLIRT tool in FSL software) to reduce artifacts caused by slight patient movements during scanning. Spatial normalization aligns brain structures from different subjects in the same coordinate space by registering images to a standard brain template (e.g., MNI 152 space). Intensity normalization uses histogram matching or z-score normalization to eliminate signal intensity differences between different scanners or sessions. Denoising uses Gaussian filtering or wavelet transform techniques to remove high-frequency noise and preserve key brain region features. Registration involves precise alignment between multimodal images (e.g., linear or nonlinear transformations), such as registering functional magnetic resonance imaging (fMRI) data to structural fMRI images. Standardization includes adjusting image resolution and voxel size to a uniform standard (e.g., 2mm). 3 The processing effect was verified by using voxel resolution and applying quality control indicators (such as root mean square error assessment).

[0035] Step 3: Feature Extraction. Key features are extracted from the preprocessed images, covering multimodal information such as structure, function, and metabolism. Brain region volume, functional connectivity strength, and metabolic activity indicators are extracted. Feature extraction includes quantifying brain region volume and cortical thickness using structural magnetic resonance imaging (sMRI), for example, automatically segmenting key regions such as the hippocampus and frontal lobe using FreeSurfer software; calculating functional connectivity strength from functional magnetic resonance imaging (fMRI), constructing functional connectivity matrices (such as default mode networks and salience networks) based on resting-state data, and evaluating local consistency or low-frequency amplitude indicators; extracting metabolic activity indicators, such as standardized uptake value (SUV) or glucose metabolic rate, through positron emission tomography (PET) to reflect brain region energy consumption; simultaneously, integrating diffusion tensor imaging (DTI) data to extract white matter fiber tract integrity and anisotropy fraction; furthermore, feature fusion techniques (such as graph-based fusion or multi-kernel learning methods) are applied to integrate multimodal features and generate a comprehensive feature vector to enhance the predictive ability of subsequent models.

[0036] Step 4: Multimodal fusion. The extracted features are fused using a feature-level fusion strategy. After dimensionality reduction through principal component analysis, the features are input into the prediction model. Feature vectors from different modalities are concatenated to form a joint feature matrix. Principal component analysis is used to eliminate redundant information and retain the main variable components. The dimensionality-reduced feature set is input into a machine learning-based prediction model (such as support vector machine or random forest). Hyperparameters are optimized through cross-validation, and a classifier is trained to identify cognitive decline risk patterns in patients with mild behavioral disorders. Simultaneously, feature importance assessment methods (such as permutation importance or SHAP value) are applied to analyze the contribution of each modality, enhancing the model's interpretability.

[0037] Step 5: Build a model and construct a machine learning-based cognitive decline prediction model. Train and validate the model using a fusion feature set. Specifically, employ algorithms such as Support Vector Machines (SVM) or Random Forests, optimizing hyperparameters (e.g., kernel parameters of SVM or tree depth of Random Forest) through k-fold cross-validation. Fit the model to the training set to learn cognitive decline risk patterns. Subsequently, evaluate model performance on an independent test set, calculating metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic (AUC) curve to ensure model generalization ability. Furthermore, based on feature importance analysis results, further adjust the model structure to prevent overfitting, and verify prediction reliability using a confusion matrix and calibration curve.

[0038] Step 6: Predictive Model Output. Apply the trained model to predict the cognitive decline risk of patients with mild behavioral disorders and output the prediction results. Specifically, input the new multimodal image feature set into the optimized model, calculate the individual cognitive decline risk probability value (e.g., probability score in the range of 0-1), and generate a prediction report, including risk classification (e.g., high risk, medium risk, low risk), confidence interval, and corresponding threshold (e.g., cut-off value set based on ROC curve). Furthermore, use the feature importance assessment results (e.g., SHAP value) to explain the contribution of key image modalities, enhance the interpretability of the prediction, and display the output results through visualization tools (e.g., risk heatmap or decision curve) to facilitate intervention decisions by clinicians.

[0039] Step 7: Performance Evaluation. The trained model is evaluated, including calculating key metrics such as accuracy, recall, and F1 score to verify the model's reliability and generalization ability. Sensitivity, specificity, and area under the receiver operating characteristic (AUC) curve are further calculated to ensure stable performance on different datasets. Simultaneously, the distribution of prediction results is analyzed using a confusion matrix, and the degree of calibration of probability predictions is assessed using calibration curves, thereby comprehensively verifying the model's predictive reliability.

[0040] Step 8: Model optimization. Based on the evaluation results, optimize model parameters, such as adjusting the learning rate or regularization coefficient, and use grid search technology to improve prediction accuracy. Introduce random search or Bayesian optimization methods to efficiently explore the hyperparameter space, while combining cross-validation strategies to ensure the generalization performance of the optimization process on independent datasets. In addition, to address the risk of overfitting, apply early stopping mechanisms or ensemble learning techniques (such as random forests or gradient boosting) to enhance model stability and prediction consistency. Finally, re-evaluate the performance of the optimized model (as described in Step 7), confirm the improvement of key indicators, and generate an optimization report to facilitate subsequent clinical deployment.

[0041] Step 9: Risk Prediction. The optimized model is integrated into the clinical decision support system to achieve automated risk prediction and output a detailed cognitive decline probability report. The system receives multimodal image data in real time through a standardized API interface, automatically executes prediction algorithms, and generates a report that includes individualized risk scores, confidence intervals, risk level classifications, and interpretability analysis. The report is output in a structured format for easy reference by clinicians in the diagnostic process, and can also be exported to PDF or electronic health record formats to ensure seamless integration with existing medical information systems. In addition, the system has a built-in validation module that continuously monitors prediction performance and triggers alarms when anomalies occur to maintain the clinical reliability and usability of the prediction results.

[0042] Step 10: Model Update. The model is updated regularly using newly acquired multimodal brain imaging data to ensure the timeliness and adaptability of the prediction method. Specifically, the update process employs incremental learning techniques or periodic full model retraining, combined with cross-validation to evaluate the performance after the update, in order to maintain prediction accuracy. At the same time, an automated data cleaning module is integrated to handle noise and outliers in the new data, and the update history is recorded through a log system for easy traceability and auditing. In addition, a feedback mechanism is established to dynamically adjust the update frequency based on actual clinical prediction results, ensuring that the model continuously adapts to changes in the patient population.

[0043] In this embodiment, preferably, the multimodal images acquired during data acquisition include structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET) data. SMRI is used to capture volume changes in the gray and white matter of the brain, fMRI is used to analyze functional connectivity patterns in resting or task states, and PET is used to monitor brain metabolic activities such as glucose uptake or amyloid deposition. These image data are integrated through feature-level fusion methods, including using principal component analysis for dimensionality reduction or deep learning networks to extract shared features, to enhance the comprehensiveness and robustness of the input information of the prediction model.

[0044] In this embodiment, preferably, data acquisition also includes quality control of image data, specifically including automatic detection of head motion artifacts, quantitative evaluation of signal-to-noise ratio, verification of image registration accuracy, and exclusion of abnormal data, to ensure that all acquired image data meets preset quality standards. For example, structural magnetic resonance imaging needs to ensure the accuracy of gray and white matter segmentation, functional magnetic resonance imaging needs to confirm time series stability, and positron emission tomography needs to verify the reliability of metabolic activity signals, thereby improving the robustness and repeatability of subsequent feature fusion and model prediction.

[0045] In this embodiment, preferably, spatial standardization in data preprocessing employs a template-based registration method using the MNI 152 standard brain template. This method achieves precise alignment between individual brain images and the template through affine transformation or nonlinear deformation field algorithms to eliminate spatial positional differences between different subjects. Specifically, the registration process includes initial linear registration to adjust global position and size, subsequent nonlinear registration to optimize local structural matching, and evaluation of registration accuracy through mutual information or correlation indices to ensure that gray matter and white matter segmentation regions are consistent in the standard space. This provides a spatially unified input data foundation for multimodal feature fusion. The preprocessing stage also integrates head motion correction and intensity normalization steps to reduce the impact of scanner differences and motion artifacts, further improving the stability and reliability of subsequent feature extraction and model prediction.

[0046] In this embodiment, preferably, the functional connectivity strength in feature extraction is obtained by calculating the time-series correlation of resting-state fMRI. Specifically, the Pearson correlation coefficient is used to assess the correlation between time series of different brain regions to generate a whole-brain functional connectivity matrix; dynamic functional connectivity changes are captured by sliding window analysis, and thresholding is applied to eliminate noise effects to enhance feature stability. The functional connectivity matrix is ​​further fused with structural imaging features (such as gray matter volume) and metabolic features (such as PET-normalized uptake values) in a multimodal manner, and dimensionality reduction is performed using principal component analysis or independent component analysis to extract a high-dimensional feature subset, thereby improving the prediction accuracy and robustness of the subsequent model for cognitive decline in mild behavioral disorders.

[0047] In this embodiment, preferably, the prediction model in the multimodal fusion is a support vector machine classifier, and the kernel function is a radial basis function. This model optimizes hyperparameters (such as the penalty coefficient C and kernel parameter γ) using a grid search method and employs k-fold cross-validation to evaluate the model's generalization ability, thereby effectively handling nonlinear classification problems in high-dimensional feature spaces. During training, the feature subset after multimodal fusion is input into the model, and the classification performance is analyzed using the confusion matrix and ROC curve, further improving the prediction accuracy and robustness for cognitive decline in mild behavioral disorders.

[0048] In this embodiment, preferably, the performance evaluation also includes using receiver operating characteristic curves and area under the curve to comprehensively quantify the model's discriminative ability, and combining confusion matrix analysis to determine false positive and false negative rates. Key biomarkers are identified by ranking feature importance (e.g., based on support vector machine weights or permutation importance), providing interpretable evidence for the pathological mechanisms of mild behavioral disorders.

[0049] In this embodiment, preferably, the Bayesian optimization algorithm is used to automatically search for optimal hyperparameters, such as regularization coefficients and kernel function parameters, during model optimization to improve prediction accuracy and reduce the risk of overfitting. Compared to grid search, Bayesian optimization efficiently explores the hyperparameter space by constructing a Gaussian process surrogate model and using iterative optimization strategies (such as aiming to improve the acquisition function), significantly reducing computational overhead and avoiding local optima. Simultaneously, the algorithm adaptively adjusts the search range based on prior knowledge, ensuring convergence to the global optimum within a finite number of iterations. This further enhances the model's generalization ability and robustness in predicting cognitive decline in mild behavioral disorders. During optimization, the validation set loss function and overfitting indicators (such as the learning curve) are monitored in real time, dynamically adjusting model complexity and providing a more stable and interpretable hyperparameter configuration for multimodal fusion features.

[0050] In this embodiment, preferably, the risk prediction output report integrates the patient's individualized risk score, confidence interval, and visualization charts for easy clinical interpretation. The confidence interval is estimated using Bootstrap resampling technology, reflecting the range of uncertainty in the prediction results and enhancing the reliability of clinical decision-making. The visualization charts include a risk distribution heatmap, a time-series trend chart, and a feature contribution bar chart, intuitively displaying the impact of key biomarkers. Furthermore, the report provides dynamic interactive functions, allowing clinicians to adjust parameters for scenario simulation and automatically generate personalized intervention suggestions based on electronic health records, improving diagnostic and treatment efficiency. The entire output process strictly adheres to privacy protection protocols to ensure data security, and is exported in PDF or Web format using standardized templates for easy multi-platform sharing and longitudinal tracking.

[0051] In this embodiment, preferably, the model update employs an online learning mechanism (such as stochastic gradient descent or an adaptive optimizer) to incrementally update model parameters quarterly, effectively capturing data distribution drift and changes in the clinical environment. Simultaneously, based on newly acquired multimodal imaging and electronic health record data, the model's timeliness is periodically verified using metrics such as cross-validation, AUC-ROC curves, and F1 scores, ensuring that the prediction accuracy remains above 95% and that generalization ability is stable. Furthermore, this mechanism supports dynamic adjustment of the learning rate and feature weights, combined with clinical feedback loops (such as doctor-annotated misdiagnosis cases) to optimize the model, enabling it to continuously adapt to the needs of different patient subgroups and emerging disease patterns. The update process strictly adheres to privacy protection protocols, employing differential privacy technology to anonymize new data, preventing the leakage of sensitive information, and an automated monitoring system reports model performance metrics in real time, facilitating rapid iteration and deployment, ultimately improving the long-term reliability and clinical application value of the prediction method.

[0052] The working principle and usage of this invention are as follows: Based on the collaborative processing of multimodal image data and machine learning modeling, it captures early cognitive decline patterns in mild behavioral disorders by integrating structural, functional, and metabolic information. The usage process begins with standardized data acquisition, obtaining structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET) data, ensuring image quality through head motion detection, signal-to-noise ratio assessment, and outlier exclusion. In the preprocessing stage, spatial normalization is performed, aligning individual brain images to the MNI 152 template, and intensity normalization is applied to reduce scanner variability. The feature extraction module automatically quantifies key brain region volumes, functional connectivity matrices, and metabolic activity indicators, such as segmenting the hippocampus using FreeSurfer, calculating Pearson correlation coefficients to construct a default pattern network, and extracting standardized uptake values.

[0053] Multimodal fusion employs feature-level concatenation and principal component analysis for dimensionality reduction, eliminating redundancy and generating a comprehensive feature vector, which is then input into a support vector machine or random forest classifier. Model training utilizes k-fold cross-validation to optimize hyperparameters, such as kernel parameters and regularization coefficients, and performance is evaluated using confusion matrix, ROC curve, and AUC metrics. In the prediction phase, it outputs individualized risk scores (0-1 probability values), confidence intervals, and risk level classifications, incorporating SHAP values ​​to explain key feature contributions and generating visual reports such as risk heatmaps. During clinical integration, the model receives new data in real-time via an API interface, automatically executes prediction algorithms, and outputs structured reports (PDF or electronic health record format) to support physician decision-making and intervention.

[0054] The model update mechanism incrementally learns new data quarterly, employs Bayesian optimization to adjust parameters, and monitors AUC and F1 scores to maintain accuracy >95%. Differential privacy protection is also applied to safeguard sensitive information. The entire process emphasizes interpretability and robustness, revealing the role of biomarkers through feature importance analysis and utilizing feedback loops to optimize the model to adapt to changes in patient subgroups. Ultimately, this enables efficient and reliable cognitive decline risk assessment and intervention management within a clinical decision support system.

[0055] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion, characterized in that, Includes the following steps: Step 1: Data Acquisition. Collect multimodal brain imaging data from patients with mild behavioral disorders. Step 2: Data preprocessing. The acquired multimodal brain imaging data is preprocessed, including motion correction, spatial standardization, intensity normalization, denoising, registration, and standardization, to ensure consistent data quality. Step 3: Feature extraction. Key features are extracted from the preprocessed images, covering multimodal information such as structure, function, and metabolism. Brain region volume, functional connectivity strength, and metabolic activity indicators are extracted. Step 4: Multimodal fusion. The extracted features are fused using a feature-level fusion strategy. After dimensionality reduction through principal component analysis, the features are input into the prediction model. Step 5: Build a model, construct a machine learning-based cognitive decline prediction model, and train and validate it using a fusion feature set; Step 6: Predict the model output. Apply the trained model to predict the risk of cognitive decline in patients with mild behavioral disorders and output the prediction results. Step 7: Performance evaluation. The trained model is evaluated, including key metrics such as accuracy, recall, and F1 score, to verify the model's reliability and generalization ability. Step 8: Model optimization. Optimize model parameters based on evaluation results, such as adjusting the learning rate or regularization coefficient, and use grid search technology to improve prediction accuracy. Step 9: Risk prediction. Integrate the optimized model into the clinical decision support system to achieve automated risk prediction and output a detailed cognitive decline probability report. Step 10: Model update. Regularly update the model using newly acquired multimodal brain imaging data to ensure the timeliness and adaptability of the prediction method.

2. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The multimodal imaging data acquired includes structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), and positron emission tomography (PET) data. SMRI is used to capture volume changes in gray and white matter of the brain, fMRI is used to analyze functional connectivity patterns in resting or task states, and PET is used to monitor brain metabolic activities such as glucose uptake or amyloid deposition.

3. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The data acquisition also includes quality control of the image data, specifically including automatic detection of head motion artifacts, quantitative evaluation of signal-to-noise ratio, verification of image registration accuracy, and elimination of abnormal data.

4. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, Spatial standardization in the data preprocessing adopts a template-based registration method, using the MNI 152 standard brain template.

5. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The functional connectivity strength in the feature extraction is obtained by calculating the time-series correlation of resting-state fMRI.

6. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, In the multimodal fusion, the prediction model is a support vector machine classifier, and the kernel function is a radial basis function.

7. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The performance evaluation also includes using receiver operating characteristic curves and area under the curve to comprehensively quantify the model's discrimination ability, and combining confusion matrix analysis to analyze false positive and false negative rates.

8. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The model optimization process employs a Bayesian optimization algorithm to automatically search for optimal hyperparameters, such as regularization coefficients and kernel function parameters, in order to improve prediction accuracy and reduce the risk of overfitting.

9. The method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The risk prediction output report integrates the patient's individualized risk score, confidence interval, and visualization charts, facilitating clinical interpretation.

10. A method for predicting cognitive decline in mild behavioral disorders based on multimodal image fusion according to claim 1, characterized in that, The model update uses an online learning mechanism to periodically and incrementally update the model parameters, and verifies the model's timeliness based on new data, ensuring that the prediction method continues to adapt to clinical needs.

Images