Machine learning based early warning method for neurofunctional deterioration in the human brain

Abstract

The present application relates to the technical field of neurological deterioration early warning, and particularly relates to a brain neurological deterioration early warning method based on machine learning, comprising: acquiring multi-modal clinical features of a plurality of target objects; reconstructing head region images of each target object to determine key lesion features; determining corresponding preliminary deterioration degrees based on the key lesion features of each target object to classify each target object; inputting the multi-modal clinical features of each target object into a preset deterioration prediction model to obtain deterioration prediction results of each target object, inputting the prediction results into a preset SHAP interpretation model to obtain contribution prediction results of each target object; determining whether each target object meets early warning standards based on the probability qualified rate and the contribution qualified rate of each classification group to generate early warning prompt information. The present application can realize quality evaluation of model prediction results, thereby improving early warning prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of early warning technology for deterioration of neurological function, and in particular to an early warning method for deterioration of human brain neurological function based on machine learning. Background Technology

[0002] With the accelerating aging of the population, the incidence and mortality rates of traumatic brain injury (TBI) in the elderly have increased significantly. Among elderly TBI cases, mild TBI accounts for a large proportion, and its initial clinical manifestations are often mild. Patients frequently present with clear consciousness and preserved verbal communication abilities, and the clinical symptoms are relatively subtle, easily leading to insufficient clinical assessment and delayed treatment. However, some elderly patients with mild TBI experience neurological deterioration. Neurological deterioration refers to the rapid progression from a mild state where the patient can initially communicate verbally to a pathological process of deepening consciousness impairment within a short period (usually within 24 hours of admission), which can lead to severe disability or even death.

[0003] In recent years, with the rapid development of artificial intelligence technology, machine learning has demonstrated significant advantages in medical prediction models. However, existing machine learning-based medical early warning methods suffer from the following technical problems: Existing methods lack mechanisms for evaluating the quality of the prediction results themselves. Traditional machine learning models directly use the output prediction probability as the basis for clinical decision-making without verifying the reliability of the prediction results. Due to the characteristics of medical data, such as high noise, high missing data, and limited sample size, the predictive stability of the model varies significantly across different time points and patient groups. Therefore, how to evaluate the quality of the prediction probability of machine learning models and dynamically identify substandard cases is an urgent technical problem to be solved. Summary of the Invention

[0004] Therefore, this invention provides an early warning method for the deterioration of human brain neural function based on machine learning, in order to overcome the problem that the existing technology lacks an evaluation mechanism for the quality of model prediction results, making it difficult to guarantee the accuracy of early warning predictions.

[0005] To achieve the above objectives, the present invention provides an early warning method for the deterioration of human brain neural function based on machine learning, comprising: Acquire multimodal clinical features of several target subjects, including basic clinical features, clinical coagulation features, CT imaging features, and treatment-related features; Based on the CT imaging features, the head region images of each target object are reconstructed, and the key lesion areas corresponding to each target object are determined to identify key lesion features. Based on the key pathological characteristics of each target object, the corresponding preliminary judgment of the degree of deterioration is determined so as to classify each target object and obtain several classification groups, wherein each classification group includes at least one target object. The multimodal clinical features of each target object are input into a preset deterioration prediction model to obtain the deterioration prediction results of each target object. The deterioration prediction results are then input into a preset SHAP interpretation model to obtain the contribution prediction results of each target object. The deterioration prediction results include the predicted risk probability of several time nodes within a preset period, and the contribution prediction results include the contribution distribution of the multimodal clinical features of the target object to the deterioration prediction results. The probability qualification rate of a single classification group is determined based on the deterioration prediction results of each target object in a single classification group, and the contribution qualification rate of a single classification group is determined based on the contribution prediction results of each target object in a single classification group. Based on the probability pass rate and contribution pass rate corresponding to each classification group, determine whether each target object meets the early warning standard, and generate early warning information.

[0006] Further, identifying the key lesion areas of any of the target objects includes: The head region image of the target object is divided into regions to obtain normal regions and abnormally extended regions; The boundary of the abnormal expansion region is divided based on the edge pixel features of the abnormal expansion region in order to determine the key lesion region corresponding to the target object.

[0007] Further, determining the probability pass rate corresponding to a single classification group includes: The probability change representation value of each target object is determined based on the predicted risk probability of each target object in a single classification group at each time node within a preset period; Based on the probability change representation value of each target object in a single classification group, determine whether the deterioration prediction result of each target object in the classification group conforms to the preset change trend, so as to identify a number of labeled target objects and a number of unlabeled target objects; The probability pass rate corresponding to the classification group is determined based on the number of labeled target objects in a single classification group and the deterioration prediction results.

[0008] Further, determining the contribution qualification rate corresponding to a single classification group includes: Based on the contribution prediction results of each target object in a single classification group, several key features corresponding to each target object in the classification group are determined; The contribution qualification rate of a single classification group is determined based on the contribution distribution of each key feature corresponding to each target object in a single classification group.

[0009] Further, determining whether each of the target objects meets the early warning criteria includes: Based on the probability qualification rate and contribution qualification rate corresponding to each of the classification groups, several key classification groups and several predicted anomaly classification groups are determined, wherein any key classification group includes several key objects, and any predicted anomaly classification group includes several predicted anomaly objects. Based on the deterioration prediction results of each key object, the deterioration probability of each key object is determined in order to determine whether each key object meets the early warning criteria. Based on the contribution prediction results of each predicted abnormal object, several abnormal characteristics of each predicted abnormal object are determined to determine whether each predicted abnormal object meets the early warning criteria.

[0010] Further, determining whether any of the aforementioned key objects meets the early warning criteria includes: If the probability of deterioration of any of the key objects is greater than a first preset probability, the key object is determined to meet the early warning criteria.

[0011] Further, determining whether each of the predicted abnormal objects meets the early warning criteria includes: If all the predicted abnormal objects have the same abnormal characteristics, then each of the predicted abnormal objects is determined to meet the early warning criteria.

[0012] Furthermore, several labeled target objects and several unlabeled target objects are identified, including: If the probability change characterization value of any target object in a single classification group is less than a preset change characterization value, the deterioration prediction result of the target object is determined to conform to the preset change trend, and the target object is identified as a marked target object. If the probability change representation value of any target object in a single classification group is greater than or equal to a preset change representation value, it is determined that the deterioration prediction result of the target object does not conform to the preset change trend, and the target object is identified as an unlabeled target object.

[0013] Furthermore, several key features corresponding to each target object in a single classification group are determined, including: The contribution differences of each feature are determined based on the contribution prediction results of each target object in a single classification group; Based on the differences in the contribution of each feature, several key features corresponding to each target object in the classification group are determined.

[0014] Further, determining the probability change representation value of any of the target objects includes: The probability difference between adjacent time nodes is determined based on the predicted risk probability of the target object at each time node within a preset period. The probability change representation value of the target object is determined based on the change in the probability difference between adjacent time nodes.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: by reconstructing the head region image of the target object based on the CT imaging characteristics of the target object and identifying the key lesion features of key lesion areas, the preliminary judgment of the deterioration degree of the target object is determined, and the target object is initially classified. Target objects with similar lesion degrees are grouped together, and the deterioration prediction results of the target objects in the same group are analyzed to determine the overall probability qualification rate of the corresponding classification group. In addition, SHAP feature interpretability analysis is performed on the deterioration prediction results to determine the contribution distribution of each multimodal clinical feature, thereby analyzing the overall contribution qualification rate of the corresponding classification group. By comprehensively considering the single-group probability qualification rate and contribution qualification rate, it is determined whether each target object meets the warning criteria. This enables quality assessment of the model prediction results and improves the accuracy of warning prediction.

[0016] Furthermore, by segmenting the head region image of the target object to obtain corresponding normal and abnormal expansion regions, the accuracy of subsequent lesion region localization is improved. Boundary segmentation is performed using the edge pixel features of the abnormal expansion region, and the lesion region boundaries are calibrated hierarchically to ultimately pinpoint the lesion region. This improves the reliability of subsequent deterioration analysis and further enhances the accuracy of early warning prediction. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the early warning method for the deterioration of human brain neural function based on machine learning, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of determining the key lesion area of ​​any of the target objects in an embodiment of the present invention. Figure 3 A flowchart illustrating the process of determining the probability pass rate corresponding to a single classification group in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the process of determining the contribution qualification rate for a single classification group in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0019] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] Please see Figure 1The diagram shown is a flowchart illustrating an early warning method for the deterioration of human brain neural function based on machine learning, according to an embodiment of the present invention. The early warning method for the deterioration of human brain neural function based on machine learning provided in this embodiment includes: Step S1: Obtain multimodal clinical features of several target objects, including basic clinical features, clinical coagulation features, CT imaging features, and treatment-related features; In this embodiment, the basic clinical characteristics include demographic characteristics, basic physical characteristics, neurological function scores, and basic comorbidity characteristics. Demographic characteristics include age, sex, and body mass index; basic physical characteristics include admission temperature, heart rate, systolic blood pressure, diastolic blood pressure, and blood oxygen saturation; neurological function scores include the initial NIHSS neurological deficit score and Glasgow Coma Scale (GCS); basic comorbidities include hypertension, diabetes, hyperlipidemia, history of stroke, and coronary heart disease. Comorbidities are quantified using a binary method, with a corresponding medical history marked as 1 and no corresponding medical history marked as 0. The basic clinical characteristics of the target subject can be extracted from the hospital's electronic medical record system and admission assessment forms. Clinical coagulation characteristics include prothrombin time, activated partial thromboplastin time, thrombin time, fibrinogen, D-dimer, and platelet count. Fasting venous blood biochemical test data within 24 hours of admission can be retrieved to select specific coagulation function test indicators as clinical coagulation characteristics. CT imaging features include basic image parameters and global texture features. Basic image parameters include intracranial CT values, mean skull grayscale value, ventricular volume, and total brain tissue volume. Global texture features include image grayscale entropy, pixel contrast, and grayscale homogeneity. This allows for the acquisition of raw DICOM image data from the first thin-section plain CT scan of the head upon admission of the target subject. Treatment-related features include whether pre-hospital antihypertensive intervention was implemented, whether anticoagulants were administered, whether dehydrating agents were used, initial fluid resuscitation volume upon admission, invasive procedure records, and sedation drug dosage. The duration of basic symptomatic treatment and preoperative pretreatment are also included. Understandably, practitioners can set specific feature dimensions based on actual conditions and perform data preprocessing. Structured data can be directly extracted, while unstructured text data can be extracted using a combination of natural language processing and rule-based information extraction methods.

[0021] Step S2: Reconstruct the head region image of each target object based on the CT imaging features, and determine the key lesion area corresponding to each target object to determine the key lesion features; Please see Figure 2 The diagram illustrates a flowchart of determining the key lesion area of ​​any of the target objects according to an embodiment of the present invention. Specifically, in step S2, determining the key lesion area of ​​any of the target objects includes: Step S21: Divide the head region image of the target object into regions to obtain normal regions and abnormal extended regions; Step S22: Based on the edge pixel features of the abnormal expansion region, the boundary of the abnormal expansion region is divided to determine the key lesion region corresponding to the target object.

[0022] In this embodiment, for any target object, the head region image of the target object can be reconstructed based on CT imaging characteristics. In practical applications, the gray-level reference range of normal brain tissue, the physiological gray-level thresholds of the skull and ventricles can be pre-entered. All pixels in the head image are traversed, and pixels whose gray-level falls within the physiological reference range and satisfy the brain tissue connectivity characteristics are clustered as normal regions. Regions other than normal regions are identified as abnormal expansion regions. Abnormal expansion regions include the lesion itself, secondary edema around the lesion, and compensatory tissue deformation regions. Redundant regions need to be refined based on edge pixel features to locate the true lesion region. In practical applications, the Canny edge detection operator can be called to extract the region contour pixels. Edge pixel features include gray-level gradient features, texture entropy features, and edge rings with smooth curvature. Edge rings are secondary edema and compensatory tissue deformation regions around the lesion, which can be directly removed. Spline interpolation fitting is performed on the selected and retained inner lesion contour pixels to complete the discontinuous contours of the tomography and generate closed boundaries to obtain the key lesion region.

[0023] Understandably, key lesion features include lesion location, mean HU value of the largest cross-section of the lesion, three-dimensional volume of the lesion, boundary irregularity, lesion infiltration depth, and lesion offset from the midline.

[0024] This invention improves the accuracy of subsequent lesion location by dividing the head region image of the target object into corresponding normal and abnormal expansion regions. Boundary delineation is performed using the edge pixel features of the abnormal expansion region, and the lesion region boundaries are calibrated hierarchically to ultimately pinpoint the lesion region. This enhances the reliability of subsequent deterioration analysis and further improves the accuracy of early warning prediction.

[0025] Step S3: Based on the key lesion characteristics of each target object, determine the corresponding preliminary judgment of the degree of deterioration, so as to classify each target object and obtain several classification groups, wherein each classification group includes at least one target object. In this embodiment, a deterioration degree scoring system can be constructed based on an expert system or a deterioration degree assessment model can be constructed to analyze the key change characteristics of each target object in order to obtain the corresponding preliminary judgment of the deterioration degree. For example, the preliminary judgment of the deterioration degree level can be divided in advance, including mild deterioration, moderate deterioration, and severe deterioration, and corresponding feature thresholds can be set for each deterioration degree level to classify the target objects.

[0026] Step S4: Input the multimodal clinical features of each target object into the preset deterioration prediction model to obtain the deterioration prediction results of each target object, and input the deterioration prediction results into the preset SHAP interpretation model to obtain the contribution prediction results of each target object. The deterioration prediction results include the predicted risk probability of several time nodes within the preset period, and the contribution prediction results include the contribution distribution of the multimodal clinical features of the target object to the deterioration prediction results. In this embodiment, the input of the preset deterioration prediction model is the multimodal clinical features of any target object, and the output is the predicted risk probability of the object at several time nodes within a preset period. The preset deterioration prediction model is a temporal fusion multi-branch neural network model. For example, the XGBoost model can be used, which internally splits feature branches and performs branch encoding on clinical basic features, coagulation temporal features, imaging lesion features, and treatment-related features respectively. It is trained based on data in the historical feature library and outputs the predicted risk probability of intracranial disease deterioration at each time node. The risk probability value ranges from 0 to 1. The risk probabilities corresponding to all time nodes are integrated to generate a temporal prediction sequence, which serves as the deterioration prediction result for the target object. The deterioration judgment criteria are the occurrence of any event such as disease progression and aggravation, expansion of intracranial lesions, or worsening of neurological deficits, avoiding prediction bias caused by a single outcome judgment. The specific time node selection should be determined in combination with the deterioration pattern of human brain neurological function and clinical follow-up nodes.

[0027] Understandably, the input to the pre-defined SHAP interpretation model is the multimodal clinical features of any target object and the corresponding deterioration prediction results. The output is the contribution distribution of each feature. The pre-defined SHAP interpretation model is constructed based on the feature attribution interpretation method of game theory Shapley value. For each deterioration prediction result, the contribution of each feature to the deterioration prediction result (SHAP value) is calculated. A positive contribution indicates that the corresponding feature increases the risk of disease deterioration, a negative contribution indicates that the corresponding feature inhibits disease deterioration, and zero contribution indicates that the feature has no significant impact on the evolution of the disease.

[0028] Step S5: Determine the probability qualification rate of a single classification group based on the deterioration prediction results of each target object in a single classification group, and determine the contribution qualification rate of a single classification group based on the contribution prediction results of each target object in a single classification group. Please see Figure 3 The diagram illustrates the process of determining the probability pass rate for a single classification group according to an embodiment of the present invention. Specifically, in step S5, determining the probability pass rate for a single classification group includes: Step S51: Determine the probability change representation value of each target object based on the predicted risk probability of each target object in a single classification group at each time node within a preset period. Specifically, in step S51, determining the probability change representation value of any of the target objects includes: Step S511: Determine the probability difference between adjacent time nodes based on the predicted risk probability of the target object at each time node within a preset period. Step S512: Determine the probability change representation value of the target object based on the change in probability difference between adjacent time nodes.

[0029] In this embodiment, for any target object, the probability change characterization value can reflect the fluctuation of the probability difference between adjacent time nodes of the target object within a preset period. The larger the probability change characterization value, the more drastic the fluctuation of the probability difference between adjacent time nodes of the target object within the preset period, and the more unstable the disease evolution. For example, the predicted risk probability difference between each time node of the target object and the previous time node within the preset period is calculated, and the variance of each probability difference is calculated as the probability change characterization value of the target object.

[0030] Step S52: Based on the probability change characterization value of each target object in a single classification group, determine whether the deterioration prediction result of each target object in the classification group conforms to the preset change trend, so as to identify a number of labeled target objects and a number of unlabeled target objects; Specifically, in step S52, several marked target objects and several unmarked target objects are determined, including: If the probability change characterization value of any target object in a single classification group is less than a preset change characterization value, the deterioration prediction result of the target object is determined to conform to the preset change trend, and the target object is identified as a marked target object. If the probability change representation value of any target object in a single classification group is greater than or equal to a preset change representation value, it is determined that the deterioration prediction result of the target object does not conform to the preset change trend, and the target object is identified as an unlabeled target object.

[0031] Step S53: Determine the probability pass rate corresponding to the classification group based on the number of marked target objects in a single classification group and the deterioration prediction result.

[0032] In this embodiment, the probability pass rate can reflect the consistency of disease evolution of each target object within the corresponding classification group. The higher the probability pass rate, the better the consistency of the predicted risk evolution pattern within the group. For example, for any classification group, the ratio of the number of marked target objects in the classification group to the total number of target objects in the classification group is determined as the first ratio, and the ratio of the minimum predicted risk probability to the maximum predicted risk probability of any target object is determined as the corresponding second ratio. The second ratios of each target object are sorted, and the ratio of the minimum second ratio to the maximum second ratio is calculated and determined as the correction coefficient. The product of the first ratio and the correction coefficient is determined as the probability pass rate corresponding to the classification group.

[0033] Please see Figure 4 The diagram illustrates the process of determining the contribution qualification rate for a single classification group according to an embodiment of the present invention. Specifically, in step S5, determining the contribution qualification rate for a single classification group includes: Step S54: Based on the contribution prediction results of each target object in a single classification group, determine several key features corresponding to each target object in the classification group; Specifically, in step S54, several key features corresponding to each target object in a single classification group are determined, including: Step S541: Determine the contribution difference of each feature based on the contribution prediction results of each target object in a single classification group; Step S542: Determine several key features corresponding to each target object in the classification group based on the differences in contribution of each feature.

[0034] In this embodiment, for any classification group, the SHAP contribution of multimodal features of all target objects within the classification group is traversed. The mean and standard deviation of the contribution of each type of single feature among all target objects in the group are calculated. The feature standard deviation is defined as the feature contribution difference. The smaller the contribution difference, the more similar the pathogenic risk factors of patients in the same group; the larger the contribution difference, the higher the dispersion of the risk-driving mechanism of patients with the same condition. A difference threshold is set, and features with contribution differences below the difference threshold are identified as key features. In actual implementation, the difference threshold can be set based on the actual situation or the mean contribution difference.

[0035] Step S55: Determine the contribution qualification rate of a single classification group based on the contribution distribution of each key feature corresponding to each target object in a single classification group.

[0036] In this embodiment, for any classification group, the contribution qualification rate is determined by statistically analyzing the overlap of the contribution distribution of each key feature. For example, a contribution distribution curve can be constructed, with features as the horizontal axis and contribution as the vertical axis, to construct the contribution distribution curve of each target object. The closed area enclosed by the key features at the left and right ends, the horizontal axis, and the curve is taken as the contribution area of ​​the corresponding target object. The ratio of the overlapping area of ​​the contribution areas of each target object to the average area of ​​the contribution areas of each target object is taken as the contribution qualification rate.

[0037] Step S6: Determine whether each target object meets the early warning criteria based on the probability pass rate and contribution pass rate corresponding to each classification group, so as to generate early warning information.

[0038] Specifically, in step S6, determining whether each target object meets the warning criteria includes: Step S61: Based on the probability qualification rate and contribution qualification rate corresponding to each classification group, determine a number of key classification groups and a number of predicted anomaly classification groups, wherein any key classification group includes a number of key objects and any predicted anomaly classification group includes a number of predicted anomaly objects. In this embodiment, a probability pass rate threshold and a contribution pass rate threshold are preset. For any classification group, if the probability pass rate of the classification group is greater than the probability pass rate threshold and the contribution pass rate is greater than the contribution pass rate threshold, then the classification group is determined as a critical classification group. Classification groups whose probability pass rate or contribution pass rate does not meet the threshold conditions are determined as predicted abnormal classification groups. In practical applications, the probability pass rate threshold can be set based on the actual situation or the average probability pass rate that has passed the pass verification in historical data, and the contribution pass rate threshold can be set based on the actual situation or the average contribution pass rate that has passed the pass verification in historical data.

[0039] Step S62: Determine the deterioration probability of each key object based on the deterioration prediction results of each key object, so as to determine whether each key object meets the early warning criteria; In this embodiment, for any key object, the average predicted risk probability of several time nodes within a preset period is determined as the deterioration probability of the target object.

[0040] Specifically, in step S62, determining whether any of the key objects meets the warning criteria includes: If the probability of deterioration of any of the key objects is greater than a first preset probability, the key object is determined to meet the early warning criteria.

[0041] Understandably, implementers can pre-set early warning standards to obtain a first preset probability. The higher the first preset probability, the higher the requirement that the probability of the key object deteriorating reaches the early warning standard.

[0042] Step S63: Based on the contribution prediction results of each predicted abnormal object, determine several abnormal characteristics of each predicted abnormal object to determine whether each predicted abnormal object meets the early warning criteria.

[0043] In this embodiment, for any predicted abnormal object, the average feature contribution of each predicted abnormal object is calculated. If the feature contribution of any predicted abnormal object is greater than the corresponding average feature contribution, then the feature is determined as the abnormal feature corresponding to the predicted abnormal object.

[0044] Specifically, in step S63, determining whether each of the predicted abnormal objects meets the early warning criteria includes: If all the predicted abnormal objects have the same abnormal characteristics, then each of the predicted abnormal objects is determined to meet the early warning criteria.

[0045] This invention reconstructs the head region image of the target object based on the CT imaging features of the target object and identifies key lesion features in key lesion areas to determine the preliminary degree of deterioration of the target object, thus achieving preliminary classification of the target object. Target objects with similar lesion degrees are grouped together, and the deterioration prediction results of the target objects in the same group are analyzed to determine the overall probability qualification rate of the corresponding classification group. SHAP feature interpretability analysis is performed on the deterioration prediction results to determine the contribution distribution of each multimodal clinical feature, thereby analyzing the overall contribution qualification rate of the corresponding classification group. By comprehensively considering the probability qualification rate of a single group and the contribution qualification rate, it is determined whether each target object meets the warning criteria. This enables quality assessment of the model prediction results and improves the accuracy of warning prediction.

[0046] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A machine learning-based early warning method for the deterioration of human brain neural function, characterized in that, include: Acquire multimodal clinical features of several target subjects, including basic clinical features, clinical coagulation features, CT imaging features, and treatment-related features; Based on the CT imaging features, the head region images of each target object are reconstructed, and the key lesion areas corresponding to each target object are determined to identify key lesion features. Based on the key pathological characteristics of each target object, the corresponding preliminary judgment of the degree of deterioration is determined so as to classify each target object and obtain several classification groups, wherein each classification group includes at least one target object. The multimodal clinical features of each target object are input into a preset deterioration prediction model to obtain the deterioration prediction results of each target object. The deterioration prediction results are then input into a preset SHAP interpretation model to obtain the contribution prediction results of each target object. The deterioration prediction results include the predicted risk probability of several time nodes within a preset period, and the contribution prediction results include the contribution distribution of the multimodal clinical features of the target object to the deterioration prediction results. The probability qualification rate of a single classification group is determined based on the deterioration prediction results of each target object in a single classification group, and the contribution qualification rate of a single classification group is determined based on the contribution prediction results of each target object in a single classification group. Based on the probability pass rate and contribution pass rate corresponding to each classification group, determine whether each target object meets the early warning standard, and generate early warning information.

2. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 1, characterized in that, Identifying the key lesion areas of any of the target objects includes: The head region image of the target object is divided into regions to obtain normal regions and abnormally extended regions; The boundary of the abnormal expansion region is divided based on the edge pixel features of the abnormal expansion region in order to determine the key lesion region corresponding to the target object.

3. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 2, characterized in that, Determining the probability pass rate for a single classification group includes: The probability change representation value of each target object is determined based on the predicted risk probability of each target object in a single classification group at each time node within a preset period; Based on the probability change representation value of each target object in a single classification group, determine whether the deterioration prediction result of each target object in the classification group conforms to the preset change trend, so as to identify a number of labeled target objects and a number of unlabeled target objects; The probability pass rate corresponding to the classification group is determined based on the number of labeled target objects in a single classification group and the deterioration prediction results.

4. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 3, characterized in that, Determining the contribution qualification rate for a single classification group includes: Based on the contribution prediction results of each target object in a single classification group, several key features corresponding to each target object in the classification group are determined; The contribution qualification rate of a single classification group is determined based on the contribution distribution of each key feature corresponding to each target object in a single classification group.

5. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 4, characterized in that, Determining whether each of the aforementioned target objects meets the early warning criteria includes: Based on the probability qualification rate and contribution qualification rate corresponding to each of the classification groups, several key classification groups and several predicted anomaly classification groups are determined, wherein any key classification group includes several key objects, and any predicted anomaly classification group includes several predicted anomaly objects. Based on the deterioration prediction results of each key object, the deterioration probability of each key object is determined in order to determine whether each key object meets the early warning criteria. Based on the contribution prediction results of each predicted abnormal object, several abnormal characteristics of each predicted abnormal object are determined to determine whether each predicted abnormal object meets the early warning criteria.

6. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 5, characterized in that, Determining whether any of the aforementioned key objects meets the early warning criteria includes: If the probability of deterioration of any of the key objects is greater than a first preset probability, the key object is determined to meet the early warning criteria.

7. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 6, characterized in that, Determining whether each of the predicted abnormal objects meets the early warning criteria includes: If all the predicted abnormal objects have the same abnormal characteristics, then each of the predicted abnormal objects is determined to meet the early warning criteria.

8. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 7, characterized in that, Identify several labeled target objects and several unlabeled target objects, including: If the probability change characterization value of any target object in a single classification group is less than a preset change characterization value, the deterioration prediction result of the target object is determined to conform to the preset change trend, and the target object is identified as a marked target object. If the probability change representation value of any target object in a single classification group is greater than or equal to a preset change representation value, it is determined that the deterioration prediction result of the target object does not conform to the preset change trend, and the target object is identified as an unlabeled target object.

9. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 8, characterized in that, Determine several key features corresponding to each target object in a single classification group, including: The contribution differences of each feature are determined based on the contribution prediction results of each target object in a single classification group; Based on the differences in the contribution of each feature, several key features corresponding to each target object in the classification group are determined.

10. The early warning method for the deterioration of human brain neural function based on machine learning according to claim 9, characterized in that, Determining the probability change representation value of any of the target objects includes: The probability difference between adjacent time nodes is determined based on the predicted risk probability of the target object at each time node within a preset period. The probability change representation value of the target object is determined based on the change in the probability difference between adjacent time nodes.

Images