A prognosis prediction method and system based on fusion of multi-modal features and clinical variables, a terminal and a storage medium

Abstract

The application relates to the technical field of data processing, and discloses a prognosis prediction method and system based on multi-modal feature and clinical variable fusion, a terminal and a storage medium, the system comprising a data acquisition module, an electroencephalogram signal preprocessing module, a feature extraction module, a feature screening module, a prediction model construction module and a prognosis prediction output module. The system acquires resting-state electroencephalogram signals of patients, pre-processes the electroencephalogram signals, extracts time-frequency features and brain region functional connection features, simultaneously combines various clinical variables, and constructs a multi-modal feature data set; subsequently, the features are reduced and screened by using a feature screening algorithm, and a machine learning classification model is used for training, so that the six-month functional outcomes of the patients are predicted, and an interpretable method is used to perform an interpretable operation on the prediction results of the model. The electroencephalogram local dynamic features and brain network connection features are fused, and the prognosis prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a prognostic prediction system, prognostic prediction method, terminal, and computer-readable storage medium based on the fusion of multimodal features and clinical variables. Background Technology

[0002] With the continued rise in stroke incidence, a large number of acute-phase patients experience varying degrees of impaired consciousness, and these patients are typically admitted to the intensive care unit. In clinical practice, physicians mainly rely on clinical behavior scales to assess patients' level of consciousness, and combine this with structural imaging examinations and basic physiological electrical signals to predict the long-term functional recovery prognosis.

[0003] However, the aforementioned existing technologies still have significant shortcomings. In the ICU (Intensive Care Unit) environment, patients are often affected by sedation, interference from life support equipment, and unstable conditions, making behavioral assessments difficult to accurately reflect their level of consciousness, and their reliability in the acute phase is insufficient. Structural imaging primarily reflects brain structural damage, while consciousness recovery is closely related to the functional integration of brain networks; relying solely on structural information is insufficient to accurately predict the potential for functional recovery. Traditional EEG analysis often focuses on single frequency bands or power characteristics, lacking characterization of the functional integration capacity of the whole brain network, resulting in limited predictive accuracy. Furthermore, existing research largely focuses on whether patients can regain consciousness, with less discussion on functional recovery, which is of greater concern to patients' families; and existing predictive models often mix samples from different etiologies, leading to poor model generalization ability due to differences in pathological mechanisms. More importantly, existing AI-based predictive methods are often black-box in nature, lacking interpretability, resulting in low trust in the predictive results among clinicians.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a prognostic prediction system and method based on the fusion of multimodal features and clinical variables, aiming to solve the problems of low accuracy in prognostic assessment and difficulty in early prediction for patients with acute stroke and impaired consciousness in existing technologies.

[0006] To achieve the above objectives, the present invention provides a prognostic prediction system based on the fusion of multimodal features and clinical variables. The prognostic prediction system based on the fusion of multimodal features and clinical variables includes: a data acquisition module, an EEG signal preprocessing module, a feature extraction module, a feature screening module, a prediction model construction module, and a prognostic prediction output module. The data acquisition module, the EEG signal preprocessing module, the feature extraction module, the feature filtering module, the prediction model construction module, and the prognosis prediction output module are connected in sequence, and the data acquisition module is also connected to the feature filtering module. The data acquisition module is used to collect the patient's resting-state electroencephalogram (EEG) signal data and clinical information; The EEG signal preprocessing module is used to preprocess the resting-state EEG signal to obtain high-quality EEG signal data; The feature extraction module is used to extract time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and to fuse the time-frequency domain features and the functional connectivity features to obtain a fused feature vector; The feature filtering module is used to filter the fused feature vector and the clinical variables in the clinical information to obtain a core feature set. The prediction model building module is used to train a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module is used to output the probability prediction value of the patient's functional recovery through the prognostic prediction model, and to sort and output the important features in the model decision-making process based on the interpretability method of game theory.

[0007] Optionally, in the prognostic prediction system based on the fusion of multimodal features and clinical variables, the data acquisition module includes an electroencephalogram (EEG) signal acquisition unit, a clinical information acquisition unit, and a data alignment unit. The EEG signal acquisition unit is used to acquire the patient's resting EEG signal data through an EEG recording device; The clinical information collection unit is used to collect the patient's clinical information, which includes age, gender, history of hypertension, history of diabetes, duration of impaired consciousness, and various sub-items of the CRS-R score. The duration of impaired consciousness is the number of days the patient was in a state of impaired consciousness from the onset of the disease to the time of EEG collection. The CRS-R score includes sub-indicators for hearing, vision, motor, speech, communication and arousal. The data alignment unit is used to align the resting-state EEG signal data with the clinical information at the patient level to form a multimodal initial dataset.

[0008] Optionally, in the prognostic prediction system based on the fusion of multimodal features and clinical variables, the EEG signal preprocessing module adopts a fully automated batch preprocessing process, which is based on the MNE library of the Python language and combines various artificial intelligence methods to directly standardize the raw EEG data. The standardization process includes, in sequence: bandpass filtering, removal of mains interference, resampling, removal of artifact signals, reference reconstruction, and independent component analysis to remove eye movement and electromyography interference.

[0009] Optionally, in the prognostic prediction system based on the fusion of multimodal features and clinical variables, the feature extraction module includes a time-frequency feature extraction unit, a functional connectivity feature extraction unit, and a feature fusion unit. The time-frequency feature extraction unit is used to extract time-frequency domain features from the high-quality EEG signal data. The time-frequency domain features include energy features, wavelet coefficients, and signal entropy values ​​to characterize the dynamic properties of local neuronal populations in the cortex. The functional connectivity feature extraction unit is used to filter the high-quality EEG signal data according to different frequency bands to obtain the EEG signal subsets corresponding to each frequency band. For each frequency band EEG signal subset, the coherence index between each pair of effective electrodes in the standard electrode layout is calculated. The coherence indices calculated for all frequency bands are flattened, redundancy is removed, and then spliced ​​to obtain the brain network functional connectivity features. The brain network functional connectivity features are used to quantify the information transmission and collaborative integration capabilities between different brain regions. The feature fusion unit is used to concatenate and fuse the time-frequency domain features with the functional connectivity features to obtain a fused feature vector; The time-frequency domain features characterize local neural dynamics, while the functional connectivity features characterize global brain network integration. The time-frequency domain features and the functional connectivity features are complementary in terms of neural information representation. The fused feature vector simultaneously contains information on local dynamic changes and global network integration to comprehensively characterize the brain functional phenotype of patients with acute-phase consciousness disorders.

[0010] Optionally, in the prognostic prediction system based on the fusion of multimodal features and clinical variables, the feature screening module includes a first-stage screening unit and a second-stage screening unit. The first-stage screening unit is used to receive the fused feature vector and the clinical variables in the clinical information, perform univariate statistical tests on each feature, calculate the significance level of the association between each feature and the prognostic label, and send the features whose significance level of association meets the preset requirements as significant features to the second-stage screening unit. The second-stage screening unit is used to receive the salient features and use a recursive feature elimination algorithm to perform dimensionality reduction screening on the salient features, and output a core feature set; The recursive feature elimination algorithm builds a model iteratively, removing the feature that contributes the least to the prediction in the current model in each iteration, until the number of remaining features meets a preset condition.

[0011] Optionally, in the prognostic prediction system based on the fusion of multimodal features and clinical variables, the prognostic prediction output module includes a probability prediction unit and an interpretable output unit. The probability prediction unit is used to output a probability prediction value of the patient's functional independence using the trained prognostic prediction model, and when the probability prediction value exceeds a preset threshold, it determines whether the patient's prognosis prediction result is a good prognosis or a bad prognosis, and presents the prediction result in a visual form. The interpretability output unit is used to calculate the marginal contribution of each input feature in the model decision-making process using the interpretability method of game theory, so as to obtain the contribution value and contribution direction of each feature to the prediction result. The interpretability output unit is also used to sort each feature according to the contribution value, output a feature importance ranking map, and output a force map for each predicted sample. In the force map, visualization elements are used to distinguish the contribution direction of each feature, and the contribution direction includes positive contribution direction and negative contribution direction. The feature importance ranking graph and the force graph are used to provide clinicians with evidence to support the model's predictions, thereby improving the acceptability of the model in clinical decision-making.

[0012] Furthermore, to achieve the above objectives, the present invention also provides a prognostic prediction method for a prognostic prediction system based on the fusion of multimodal features and clinical variables, wherein the prognostic prediction method includes: The data acquisition module collects the patient's resting-state electroencephalogram (EEG) signal data and clinical information; The EEG signal preprocessing module preprocesses the resting-state EEG signal to obtain high-quality EEG signal data; The feature extraction module extracts time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and fuses the time-frequency domain features and the functional connectivity features to obtain a fused feature vector; The feature filtering module performs feature filtering on the fused feature vector and the clinical variables in the clinical information to obtain a core feature set. The prediction model building module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module outputs the probability prediction value of the patient's functional recovery through the prognostic prediction model, and sorts and outputs the important features in the model decision-making process based on the interpretability method of game theory.

[0013] Optionally, the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables, wherein the prediction model construction module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model, specifically includes: The core feature set is used as the input variable, and the patient's six-month functional outcome is used as the label variable. The six-month functional outcome is based on the GOS-E score. When the GOS-E score is greater than the preset score, the six-month functional outcome is a good prognosis; otherwise, it is a poor prognosis. A random forest classifier is used for model training. The random forest classifier uses the default parameters of the scikit-learn library in Python and uses a majority voting mechanism to make decisions, so as to enhance its robustness to noise, outliers and missing features. The model parameters were optimized by five-fold cross-validation, and a stratified sampling strategy was adopted to ensure that the proportion of good prognoses and bad prognoses in each fold was consistent with the original dataset. After training, the prognostic prediction model was obtained.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a prognostic prediction program of a prognostic prediction system based on the fusion of multimodal features and clinical variables, stored in the memory and executable on the processor. When the prognostic prediction program of the prognostic prediction system based on the fusion of multimodal features and clinical variables is executed by the processor, it implements the steps of the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a prognostic prediction program of a prognostic prediction system based on the fusion of multimodal features and clinical variables, and when the prognostic prediction program of the prognostic prediction system based on the fusion of multimodal features and clinical variables is executed by a processor, it implements the steps of the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables as described above.

[0016] In this invention, the prognostic prediction system based on the fusion of multimodal features and clinical variables includes: a data acquisition module, an EEG signal preprocessing module, a feature extraction module, a feature filtering module, a prediction model construction module, and a prognostic prediction output module. The data acquisition module collects resting-state EEG signal data and clinical information from patients; the EEG signal preprocessing module preprocesses the resting-state EEG signals to obtain high-quality EEG signal data; the feature extraction module extracts time-frequency domain features and functional connectivity features from the high-quality EEG signal data and fuses these features to obtain a fused feature vector; the feature filtering module filters the fused feature vector and clinical variables from the clinical information to obtain a core feature set; the prediction model construction module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model; and the prognostic prediction output module outputs the probability prediction value of the patient's functional recovery through the prognostic prediction model and ranks and outputs important features in the model's decision-making process based on game theory interpretability methods. This invention improves the accuracy of prognostic prediction by fusing local dynamic features of EEG with brain network connectivity features. Attached Figure Description

[0017] Figure 1 This is the overall architecture diagram of the prognostic prediction system based on the fusion of multimodal features and clinical variables of this invention; Figure 2 This is a flowchart of a preferred embodiment of the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables of the present invention; Figure 3 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to user-analyzed data, user-stored data, user-displayed data, etc.) and signals involved in this invention are all information, data and signals authorized by the user or fully authorized by all parties; and the collection, use and processing of relevant information, data and signals comply with the laws, regulations and standards of relevant countries and regions.

[0020] To address the problems in existing technologies, this embodiment provides a prognostic prediction system based on the fusion of multimodal features and clinical variables, such as... Figure 1As shown, the prognostic prediction system based on the fusion of multimodal features and clinical variables includes: a data acquisition module, an EEG signal preprocessing module, a feature extraction module, a feature screening module, a prediction model construction module, and a prognostic prediction output module; The data acquisition module, the EEG signal preprocessing module, the feature extraction module, the feature filtering module, the prediction model construction module, and the prognosis prediction output module are connected in sequence, and the data acquisition module is also connected to the feature filtering module. The data acquisition module is used to collect the patient's resting-state electroencephalogram (EEG) signal data and clinical information; The EEG signal preprocessing module is used to preprocess the resting-state EEG signal to obtain high-quality EEG signal data; The feature extraction module is used to extract time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and to fuse the time-frequency domain features and the functional connectivity features to obtain a fused feature vector; The feature filtering module is used to filter the fused feature vector and the clinical variables in the clinical information to obtain a core feature set. The prediction model building module is used to train a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module is used to output the probability prediction value of the patient's functional recovery through the prognostic prediction model, and to sort and output the important features in the model decision-making process based on the interpretability method of game theory.

[0021] Understandably, this invention is used to predict functional recovery in patients with acute stroke and impaired consciousness. The system first collects resting-state EEG signals and key clinical information from patients during the acute phase to construct an initial multimodal dataset. Subsequently, a fully automated preprocessing procedure is performed on the raw EEG signals to eliminate various noises and artifacts present in the ICU environment, resulting in clean EEG data suitable for analysis. Based on this, the system extracts neurophysiological features from two dimensions: time-frequency domain features reflect the dynamic characteristics of local cortical neurons, while functional connectivity features characterize the information integration ability between different brain regions. These two types of features, along with clinical variables such as the patient's age, underlying diseases, and level of consciousness score, enter the feature selection stage. A two-stage dimensionality reduction strategy is used to eliminate redundant features, obtaining the core feature set with the highest predictive value. Then, the system uses this core feature set to train a random forest classifier, constructing a prognostic prediction model with the patient's functional independence outcome six months later as the prediction target. Finally, the system outputs individualized functional recovery probabilities and, using a game theory-based interpretability method, provides a ranking of the contributions of each feature to the prediction results, offering clinicians a visual basis for decision support. Through the above methods, the present invention achieves an objective, accurate, and interpretable assessment of the long-term functional prognosis of patients in the early stages of disease.

[0022] It should be noted that the multimodal prediction framework proposed in this invention is not only applicable to stroke patients with impaired consciousness, but can also be extended to patients with traumatic brain injury, hypoxic brain injury, and long-term vegetative state.

[0023] Furthermore, the data acquisition module includes an electroencephalogram (EEG) signal acquisition unit, a clinical information acquisition unit, and a data alignment unit; The EEG signal acquisition unit is used to acquire the patient's resting EEG signal data through an EEG recording device; The clinical information collection unit is used to collect the patient's clinical information, which includes age, gender, history of hypertension, history of diabetes, duration of impaired consciousness, and various sub-items of the CRS-R score. The duration of impaired consciousness is the number of days the patient was in a state of impaired consciousness from the onset of the disease to the time of EEG collection. The CRS-R score includes sub-indicators for hearing, vision, motor, speech, communication and arousal. The data alignment unit is used to align the resting-state EEG signal data with the clinical information at the patient level to form a multimodal initial dataset.

[0024] In this embodiment, the data acquisition module further includes an electroencephalogram (EEG) signal acquisition unit, a clinical information acquisition unit, and a data alignment unit.

[0025] The EEG signal acquisition unit is used to acquire the patient's resting-state EEG signal data through an EEG recording device. This resting-state EEG signal is usually acquired when the patient is quiet and has their eyes closed, which can avoid interference from task-induced signals, and is especially suitable for patients with acute-phase loss of consciousness in the ICU who cannot cooperate with instructions.

[0026] The clinical information acquisition unit is used to collect the patient's clinical information, including age, gender, history of hypertension, history of diabetes, duration of altered consciousness, and CRS. The R score comprises various sub-items. Among them, the duration of altered consciousness is defined as the number of days from the onset of illness to the time of EEG recording during which the patient remained in a state of altered consciousness; CRS... The R score comprises sub-indicators for six aspects: hearing, vision, motor function, speech, communication, and arousal. These clinical variables and EEG characteristics are complementary in predicting functional recovery.

[0027] The data alignment unit is used to align the resting-state EEG signal data with the clinical information at the patient level, forming a multimodal initial dataset. Through the alignment operation, it ensures that each EEG signal accurately corresponds to the clinical information of the same patient, providing standardized data for subsequent feature extraction and model training.

[0028] Furthermore, the EEG signal preprocessing module adopts a fully automated batch preprocessing process, which is based on the MNE library of the Python language and combines various artificial intelligence methods to directly standardize the raw EEG data. The standardization process includes, in sequence: bandpass filtering, removal of mains interference, resampling, removal of artifact signals, reference reconstruction, and independent component analysis to remove eye movement and electromyography interference.

[0029] In this embodiment, the EEG signal preprocessing module employs a fully automated batch preprocessing workflow. This workflow requires no manual intervention and can directly standardize the raw EEG data, making it particularly suitable for large-scale, interference-prone EEG signals collected in an ICU environment.

[0030] Specifically, this batch preprocessing workflow is built on the MNE (Magnetoencephalography, a tool for analyzing brain signals in Python) library in Python and integrates a variety of mature artificial intelligence methods. These include an adaptive threshold algorithm used in the artifact removal stage, which can dynamically identify and suppress common electromyography, electrocardiography, and motion artifacts in the ICU environment; and in the Independent Component Analysis (ICA) stage, the algorithm automatically identifies and removes independent components related to eye movements and electromyography, eliminating the need for manual identification of artifact components.

[0031] Furthermore, the standardization process is performed in the following order: first, bandpass filtering is performed to preserve the neurophysiological frequency bands related to conscious activity; then, mains interference is removed to eliminate power frequency noise; next, resampling is performed to standardize the signal sampling rate; then, an adaptive algorithm is used to remove various artifact signals; next, reference reconstruction is performed to standardize the signal reference standard; finally, independent component analysis is used to remove residual eye movement and electromyography interference. After the above processing, the raw EEG signal is transformed into high-quality, clean, and analyzable EEG data, providing a reliable data foundation for subsequent feature extraction.

[0032] Through the fully automated preprocessing process described above, this invention significantly reduces the reliance on the experience of professional operators in EEG data preprocessing, improves the efficiency and consistency of batch processing, and ensures the quality of EEG signals in the complex environment of the ICU, laying the foundation for the accuracy and stability of the prognostic prediction model.

[0033] Furthermore, the feature extraction module includes a time-frequency feature extraction unit, a functional connection feature extraction unit, and a feature fusion unit; The time-frequency feature extraction unit is used to extract time-frequency domain features from the high-quality EEG signal data. The time-frequency domain features include energy features, wavelet coefficients, and signal entropy values ​​to characterize the dynamic properties of local neuronal populations in the cortex. The functional connectivity feature extraction unit is used to filter the high-quality EEG signal data according to different frequency bands to obtain the EEG signal subsets corresponding to each frequency band. For each frequency band EEG signal subset, the coherence index between each pair of effective electrodes in the standard electrode layout is calculated. The coherence indices calculated for all frequency bands are flattened, redundancy is removed, and then spliced ​​to obtain the brain network functional connectivity features. The brain network functional connectivity features are used to quantify the information transmission and collaborative integration capabilities between different brain regions. The feature fusion unit is used to concatenate and fuse the time-frequency domain features with the functional connectivity features to obtain a fused feature vector; The time-frequency domain features characterize local neural dynamics, while the functional connectivity features characterize global brain network integration. The time-frequency domain features and the functional connectivity features are complementary in terms of neural information representation. The fused feature vector simultaneously contains information on local dynamic changes and global network integration to comprehensively characterize the brain functional phenotype of patients with acute-phase consciousness disorders.

[0034] It is understood that the feature extraction module of the present invention specifically includes a time-frequency feature extraction unit, a functional connectivity feature extraction unit, and a feature fusion unit, used to extract feature vectors from high-quality EEG signals that can comprehensively reflect the brain's functional state.

[0035] The time-frequency feature extraction unit is responsible for extracting time-frequency domain features from high-quality EEG signal data. These features include energy features (such as Teager-Kaiser energy), wavelet coefficients, and signal entropy values. These time-frequency domain features are primarily used to characterize the dynamic properties of local cortical neuronal populations, such as the excitability level of local neurons, the complexity of information processing, and the integrity of local neural circuits. Through time-frequency analysis, this invention can sensitively detect whether cortical function in patients with acute-phase altered consciousness is severely suppressed or shows early signs of recovery.

[0036] The functional connectivity feature extraction unit analyzes the collaborative relationships between brain regions from the perspective of different frequency bands. The specific processing flow is as follows: First, high-quality EEG signals are filtered according to different frequency bands (δ band: 0.5-4Hz, θ band: 4-8Hz, α band: 8-13Hz, β band: 13-30Hz, γ band: >30Hz) to obtain subsets of EEG signals corresponding to each band. Then, for each frequency band, the coherence index between each pair of effective electrodes in the standard electrode layout is calculated. This index quantifies the degree of signal synchronization between two brain regions. Finally, the coherence indices calculated for all frequency bands are flattened into a one-dimensional vector, and redundant information is removed before concatenation to form the brain network functional connectivity features. These brain network functional connectivity features can effectively quantify the information transmission capacity and collaborative integration state between different brain regions, reflecting the level of distributed information processing performed by the brain as a whole system.

[0037] The feature fusion unit concatenates and fuses the time-frequency domain features obtained from the time-frequency feature extraction unit with the functional connectivity features obtained from the functional connectivity feature extraction unit, generating a unified fused feature vector. Since time-frequency domain features primarily characterize local neural dynamic states, while functional connectivity features focus on the global brain network integration state, the two are naturally complementary in representing neural information. The fused feature vector simultaneously contains information on local dynamic changes and global network integration, thus enabling a more comprehensive and accurate representation of the brain functional phenotype of patients with acute-phase disorders of consciousness.

[0038] Furthermore, the feature filtering module includes a first-stage filtering unit and a second-stage filtering unit; The first-stage screening unit is used to receive the fused feature vector and the clinical variables in the clinical information, perform univariate statistical tests on each feature, calculate the significance level of the association between each feature and the prognostic label, and send the features whose significance level of association meets the preset requirements as significant features to the second-stage screening unit. The second-stage screening unit is used to receive the salient features and use a recursive feature elimination algorithm to perform dimensionality reduction screening on the salient features, and output a core feature set; The recursive feature elimination algorithm builds a model iteratively, removing the feature that contributes the least to the prediction in the current model in each iteration, until the number of remaining features meets a preset condition.

[0039] It is understood that the feature screening module of the present invention adopts a two-stage screening strategy, specifically including a first-stage screening unit and a second-stage screening unit, which are used to screen out the core feature set with the most predictive value from high-dimensional fused feature vectors and clinical variables.

[0040] The first-stage screening unit receives the fused feature vector from the feature fusion unit and clinical variables from the clinical information collected by the data acquisition module. This unit independently performs univariate statistical tests on each feature, calculating the significance of the association between each feature and the prognostic label. Specifically, based on the data distribution characteristics of the features, methods such as t-tests and chi-square tests can be used to measure whether the difference between the features and the two groups of samples with good and poor prognoses is statistically significant. Subsequently, features with a significance level (p) that meets a preset threshold (e.g., p-value less than 0.05) are retained, thus initially eliminating features that are irrelevant to the prognostic outcome or have a very weak association. This stage effectively reduces the number of features and lowers the complexity of subsequent calculations.

[0041] The second-stage filtering unit is connected to the first-stage filtering unit and receives the features retained after the first-stage filtering. It then uses a recursive feature elimination algorithm for further dimensionality reduction filtering. In this algorithm, the system first trains a base model (such as a random forest or support vector machine) based on the current feature set. Then, based on the feature importance output by the model (such as the feature importance score of a random forest or the absolute value of the coefficients of a linear model), it identifies the feature with the smallest contribution to prediction in the current model and removes it. Subsequently, the system repeats the above iterative process on the remaining features; that is, in each iteration, the model is retrained, feature importance is evaluated, and the feature with the smallest contribution is removed, until the number of remaining features meets a preset condition (e.g., reaching a preset target number of features, or the model performance begins to decline). Finally, the second-stage filtering unit outputs the optimized core feature set.

[0042] Through a two-stage screening process, this invention effectively removes redundant and noisy features while retaining key prediction information, thereby improving the compactness of the feature set and the stability of model training. Furthermore, the recursive feature elimination algorithm shows good consistency with the subsequently used random forest model in feature importance evaluation, further enhancing the interpretability and predictive performance of the screening results.

[0043] Furthermore, the prediction model construction module is the core module for prognostic prediction in this invention. Based on the core feature set output by the feature selection module, it trains a machine learning classification model capable of distinguishing between good and bad prognoses. Specifically, this module first uses the selected core feature set as input variables, and simultaneously uses the patient's functional outcome six months later as a label variable. Unlike traditional endpoints that only focus on whether consciousness is restored, this invention uses the Extended Glasgow Outcome Scale (GOS-E) score as a quantitative standard for functional outcome: a GOS-E score greater than 4 is defined as a good prognosis, indicating that the patient can achieve basic functional independence; a GOS-E score less than or equal to 4 is defined as a bad prognosis, indicating that the patient still has significant functional impairment or dependence. This label definition is more in line with the concerns of patients and their families regarding their ability to live independently.

[0044] In terms of model selection, this invention employs a random forest classifier for training. Specifically, the scikit-learn library in Python is used, with a fixed random seed of 42, and the default parameter configurations from the library are adopted. The random forest model makes the final decision through a majority voting mechanism that integrates multiple decision trees. This mechanism makes the model insensitive to misclassifications by a single tree, maintaining relatively stable predictive performance even in the presence of noise, outliers, or missing features. Especially in the ICU environment, where EEG signals and clinical indicators often struggle to achieve ideal signal-to-noise ratios and data integrity, random forests do not require strict assumptions about feature distributions or rely on feature standardization. They naturally adapt to the differences in dimensions and statistical distributions between clinical indicators and EEG features, making them highly suitable for the application scenarios described in this invention.

[0045] To prevent overfitting and enhance generalization ability, the prediction model building module employs five-fold cross-validation to optimize model parameters. Specifically, the original dataset is randomly divided into five similarly sized subsets. One subset is used as the validation set, and the remaining four subsets are used as the training set. This training and validation process is repeated five times. During each training iteration, the model adjusts its parameters based on the current training set and evaluates its performance on the validation set. The optimal parameter configuration is then selected by combining the results of all five validations. Using five-fold cross-validation reduces the bias caused by a single data partitioning method, making the model performance evaluation more robust. For rigorous clinical prediction tasks, this step significantly improves the reliability of the model's prediction results.

[0046] Furthermore, the prognostic prediction output module includes a probability prediction unit and an interpretability output unit; The probability prediction unit is used to output a probability prediction value of the patient's functional independence using the trained prognostic prediction model, and when the probability prediction value exceeds a preset threshold, it determines whether the patient's prognosis prediction result is a good prognosis or a bad prognosis, and presents the prediction result in a visual form. The interpretability output unit is used to calculate the marginal contribution of each input feature in the model decision-making process using the interpretability method of game theory, so as to obtain the contribution value and contribution direction of each feature to the prediction result. The interpretability output unit is also used to sort each feature according to the contribution value, output a feature importance ranking map, and output a force map for each predicted sample. In the force map, visualization elements are used to distinguish the contribution direction of each feature, and the contribution direction includes positive contribution direction and negative contribution direction. The feature importance ranking graph and the force graph are used to provide clinicians with evidence to support the model's predictions, thereby improving the acceptability of the model in clinical decision-making.

[0047] It is understood that the prognostic prediction output module of the present invention includes a probability prediction unit and an interpretability output unit, which are responsible for generating the prediction results and interpreting the prediction process, respectively.

[0048] The probability prediction unit utilizes a trained prognostic prediction model to calculate the probability of a patient regaining functional independence from the input core feature set. This probability value, ranging from 0 to 1, reflects the likelihood of a good prognosis six months later. To facilitate clinical decision-making, the system presets a classification threshold (e.g., 0.5, adjustable based on clinical needs). When the probability prediction value exceeds this threshold, the patient's prognosis is considered good (i.e., likely to achieve functional independence); conversely, it is considered poor (i.e., significant functional impairment remains). The prediction results are presented visually, such as displaying text labels for good or poor prognosis on the system interface, supplemented by probability value bar charts or dashboards, allowing clinicians to intuitively and quickly obtain their judgments.

[0049] The interpretability output unit addresses the "black box" problem of machine learning models (the "black box" nature of existing AI prediction methods results in insufficient interpretability, leading to low trust in model predictions by clinicians), thereby increasing clinicians' confidence in the prediction results. This unit employs game theory-based interpretability methods, such as the Shapley Additive Explanations (SHAP), to calculate the marginal contribution of each input feature in the model's decision-making process. Specifically, this method allocates the difference between the model's predicted value and the baseline value according to the contribution of each feature to the prediction result, thus obtaining the contribution value (quantifying the degree of influence) and contribution direction (indicating whether the feature positively promotes a good prognosis or negatively inhibits a good prognosis) for each feature.

[0050] After obtaining the contribution value and direction of each feature, the interpretability output unit further performs the following two output tasks: First, it sorts the features from largest to smallest according to their contribution values, outputting a feature importance ranking chart. This chart is usually presented in the form of a horizontal bar chart, with features having larger contribution values ​​ranking higher. Clinicians can easily identify the most critical factors affecting patient prognosis prediction (such as functional connectivity strength in a certain frequency band, age, CRS-R score, etc.). Second, for each predicted sample (i.e., each patient), it outputs a force map. In this force map, visual elements such as color, position, or arrows are used to clearly distinguish the positive and negative contribution directions of each feature—for example, red indicates a positive contribution (driving towards a better prognosis), and blue indicates a negative contribution (driving towards a worse prognosis). The length of the feature bar represents the magnitude of the contribution value. Through the force map, doctors can understand which specific features led to a patient being predicted to have a good or bad prognosis, thus gaining a deeper understanding of the individualized prediction results.

[0051] As can be seen, the feature importance ranking plot output by the prognostic prediction module reveals the core features affecting the prognostic recovery of the patient population from a global perspective, while the output force plot explains the predictive basis of a single sample from an individual perspective. Together, they provide clinicians with evidence supporting the model's predictions, significantly improving the transparency and acceptability of the model in clinical decision-making.

[0052] like Figure 2 As shown, based on the prognostic prediction system based on the fusion of multimodal features and clinical variables described in the above embodiments, the present invention also provides a prognostic prediction method based on the fusion of multimodal features and clinical variables. The prognostic prediction method based on the fusion of multimodal features and clinical variables includes the following steps: Step S10: The data acquisition module acquires the patient's resting-state electroencephalogram (EEG) signal data and clinical information.

[0053] In this embodiment, the patient's resting-state EEG signals are first acquired using an EEG recording device. Acquisition is typically performed when the patient is quiet, eyes closed, and without strong external stimuli, and lasts for at least 5 minutes to obtain stable neurophysiological activity. For patients in the ICU who cannot cooperate by closing their eyes, an eye mask can be used or the environment can be kept dim to minimize visual interference with the EEG. Electrode placement follows a standard 10-20 system layout, recording signals from 19 effective leads. The sampling rate is set to at least 250Hz to ensure the accuracy of subsequent frequency domain analysis.

[0054] Simultaneously, the clinical information acquisition unit obtains the patient's clinical variables from electronic medical records or bedside inquiries, including age, gender, history of underlying diseases such as hypertension and diabetes, duration of impaired consciousness from the onset of illness to the time of data collection, and scores for each sub-item of the Coma Recovery Scale-Revised (CRS-R) (auditory, visual, motor, verbal, communicative, and arousal). These EEG signals and clinical information are linked using the patient's ID number as a unique identifier, together forming the initial multimodal dataset.

[0055] Step S20: The EEG signal preprocessing module preprocesses the resting-state EEG signal to obtain high-quality EEG signal data.

[0056] Understandably, raw EEG signals contain various noises and artifacts, and must undergo standardized preprocessing before they can be used for subsequent analysis.

[0057] This invention employs a fully automated batch preprocessing workflow, which is developed based on the MNE library in Python and incorporates various artificial intelligence methods. It can adapt to the complex signal characteristics in the ICU environment without human intervention.

[0058] Specifically, the process begins with a 0.5-45Hz bandpass filter to remove DC drift and high-frequency electromyography (EMG) interference. A notch filter is then used to remove 50Hz (or 60Hz) mains frequency interference. Next, the raw data is resampled to a uniform 250Hz to eliminate sampling rate differences between different devices. An adaptive thresholding algorithm is then used to identify and remove motion artifacts and ECG interference with excessive amplitude. Reference reconstruction is performed, converting the original reference to an average reference or bilateral mastoid reference to ensure signal consistency. Finally, independent component analysis (ICA) is used to automatically identify independent components related to eye movements, blinking, and EMG, and these components are set to zero before signal reconstruction, thus completely removing biological artifacts. The entire preprocessing process runs in batch mode, handling dozens or even hundreds of data points at a time. After these steps, noise in the raw EEG signal is significantly suppressed, preserving neurophysiological components related to conscious activity and brain network function, resulting in high-quality EEG signal data suitable for feature extraction.

[0059] Step S30: The feature extraction module extracts time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and fuses the time-frequency domain features and the functional connectivity features to obtain a fused feature vector.

[0060] In this embodiment, the feature extraction module comprises three units: a time-frequency feature extraction unit, a functional connectivity feature extraction unit, and a feature fusion unit. The time-frequency feature extraction unit calculates three local dynamic features for each lead's high-quality EEG signal: Teager-Kaiser energy features (reflecting instantaneous energy changes in neuronal firing), wavelet coefficients (decomposing the signal through discrete wavelet transform to extract details and approximation coefficients at different scales), and signal entropy values ​​(such as sample entropy or approximate entropy, measuring the regularity and complexity of the signal). These features characterize the excitability, information processing capacity, and circuit integrity of local cortical neuronal populations from different perspectives. The functional connectivity feature extraction unit filters the EEG signal according to different frequency bands (usually divided into δ, θ, α, β, γ) to obtain the signal subsets corresponding to each frequency band. At each frequency band, the coherence index between all 19 effective electrodes in a standard 10-20 electrode layout is calculated. This index estimates the synchronization degree of the two electrode signals through the cross-correlation power spectral density in the frequency domain. After obtaining the 19×19 coherence matrix for each frequency band, the upper or lower triangular portion is extracted (diagonal redundancy is removed), flattened into a one-dimensional vector, and the vectors of all frequency bands are concatenated to form a high-dimensional functional connectivity feature. The feature fusion unit directly concatenates the above time-frequency domain features with the functional connectivity feature to generate a fused feature vector that simultaneously contains local dynamics and global network integration information.

[0061] Step S40: The feature filtering module performs feature filtering on the fused feature vector and the clinical variables in the clinical information to obtain a core feature set.

[0062] Understandably, to avoid the curse of dimensionality and improve the model's generalization ability and interpretability, this embodiment first performs a correlation test between each feature in the fused feature vector and clinical variables (such as age, CRS-R score, etc.) and the patient's prognostic label (good or bad). Depending on the data distribution type of the features, a t-test or chi-square test is selected to calculate the test statistic and its corresponding p-value for each feature. Features with p-values ​​less than a preset significance level (e.g., 0.05) are retained, i.e., those features that show a significant difference between the good and bad prognostic groups. This stage can quickly eliminate a large number of noisy features irrelevant to the outcome.

[0063] Furthermore, a recursive feature elimination algorithm is employed: the features retained in the first stage are input into a base model (usually consistent with the final prediction model, i.e., a random forest), and after training, the importance score of each feature is obtained (based on the reduction in Gini impurity or the decrease in average accuracy). Subsequently, one or a batch of features with the lowest current importance score (i.e., the smallest contribution to prediction) are removed, and the above process is repeated on the remaining features until the number of remaining features reaches a preset target value. The features that are finally retained are the core feature set. This set not only eliminates redundant and irrelevant features but also retains the most predictive EEG and clinical indicators.

[0064] Step S50: The prediction model building module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model.

[0065] Specifically, the core feature set is used as the input variable, and the patient's six-month functional outcome is used as the label variable. The six-month functional outcome is based on the GOS-E score. When the GOS-E score is greater than the preset score, the six-month functional outcome is considered a good prognosis; otherwise, it is considered a poor prognosis.

[0066] In this embodiment, supervised learning is performed using a core feature set as input variables and the patient's functional outcome six months later as label variables. The functional outcome is assessed using the Extended Glasgow Outcome Scale (GOS-E): a GOS-E score greater than 4 (i.e., the patient can live independently at home without daily assistance) is defined as a good prognosis; a score less than or equal to 4 (i.e., requiring varying degrees of care or being in a vegetative state) is defined as a poor prognosis. This label definition differs from the traditional approach that only focuses on whether consciousness is restored, and is more closely aligned with the actual concerns of patients and their families regarding their ability to live independently.

[0067] Furthermore, a random forest classifier is used for model training. The random forest classifier uses the default parameters of the scikit-learn library in Python and makes decisions using a majority voting mechanism to enhance robustness to noise, outliers, and missing features.

[0068] Understandably, the model training employs a random forest classifier, specifically using the default parameter of 42 random seeds from the Python scikit-learn library. Random forests generate multiple decision trees through bootstrapping aggregation, each tree growing based on a different training subset and a random subset of features, ultimately determining the predicted class through majority voting. This ensemble mechanism makes the model highly robust to noise, outliers, and missing data, and eliminates the need for feature normalization or standardization, making it ideal for clinical data scenarios that mix continuous values ​​(such as age), rank values ​​(such as CRS-R scores), and high-dimensional EEG features.

[0069] Furthermore, the model parameters were optimized through five-fold cross-validation, and a stratified sampling strategy was adopted to ensure that the proportion of good and bad prognoses in each fold was consistent with the original dataset. After training, a prognostic prediction model was obtained.

[0070] In this embodiment, to prevent overfitting, five-fold cross-validation is used for parameter tuning: the dataset is divided into five equal parts, one part is used as the validation set and the remaining four parts are used as the training set, and this process is repeated five times. The average performance of the five validations (such as accuracy and AUC) is used to evaluate the effect of different parameter configurations. After cross-validation optimization, the trained random forest model, i.e., the prognostic prediction model, is finally saved. This model can output a probability value of a good prognosis for new samples in subsequent steps.

[0071] Step S60: The prognostic prediction output module outputs the probability prediction value of the patient's functional recovery through the prognostic prediction model, and sorts and outputs the important features in the model decision-making process based on the interpretability method of game theory.

[0072] In this embodiment, the core features of a new patient are input into a pre-trained random forest model. The model outputs a value between 0 and 1, representing the predicted probability of the patient achieving a good prognosis six months later. The system presets a classification threshold; if the predicted probability is greater than the threshold, it is considered a good prognosis; otherwise, it is considered a poor prognosis. The prediction results are presented in a visual form, such as displaying "Prediction: Good Prognosis (Probability 78%)" on the system interface, supplemented by a colored progress bar or dashboard, to facilitate quick understanding by clinicians.

[0073] Regarding interpretability of the output, the prognostic prediction output module employs a game theory-based interpretability method (Shapley sum interpretation). The core idea of ​​this method is to decompose the model's predicted value for a sample into the sum of the marginal contributions of each feature. Specifically, for each input feature, the SHAP algorithm simulates the changes in the model's output when that feature is present and absent, and comprehensively considers all possible feature combinations to ultimately obtain a unique Shapley value (contribution value) and contribution direction for each feature (positive contribution improves the prognosis, while negative contribution worsens it).

[0074] The output formats include two types: first, a global feature importance ranking chart, which averages the absolute values ​​of all features across all samples and sorts them from largest to smallest, displaying a bar chart showing which features have the greatest impact on prognostic judgment at the population level; second, a single-sample power chart, which shows the specific contribution value and direction of each feature for each patient, using color (e.g., red for positive, blue for negative) and length to visually represent how each feature drives the final prediction result. Through these interpretable outputs, clinicians can not only learn about the prediction results but also understand "why the prediction was made this way," thereby increasing their confidence in the model's predictions.

[0075] The beneficial effects of this invention are as follows: This invention provides an objective prediction system and method that can be implemented at the bedside in the acute phase. By integrating local dynamic features of EEG with brain network connectivity features, it improves the accuracy of prognostic prediction, provides quantitative basis for clinical decision-making, assists doctors in formulating treatment strategies and ethical decisions, and can output interpretable indicators such as the importance of model prediction features, providing clinicians with evidence support for model prognostic prediction and ranking of the importance of features affecting patient functional recovery.

[0076] Furthermore, such as Figure 3 As shown, based on the above-mentioned prognostic prediction system and method based on the fusion of multimodal features and clinical variables, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 3 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0077] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 may also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a prognostic prediction program 40 based on the fusion of multimodal features and clinical variables. This prognostic prediction program 40 can be executed by the processor 10 to implement the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables in this application.

[0078] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables.

[0079] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visualized patient interface. The components of the terminal communicate with each other via a system bus.

[0080] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a prognostic prediction program of a prognostic prediction system based on the fusion of multimodal features and clinical variables, and when the prognostic prediction program of the prognostic prediction system based on the fusion of multimodal features and clinical variables is executed by a processor, it implements the steps of the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables as described above.

[0081] In summary, this invention discloses a prognostic prediction method, system, terminal, and storage medium based on the fusion of multimodal features and clinical variables. The system includes: a data acquisition module, an EEG signal preprocessing module, a feature extraction module, a feature filtering module, a prediction model construction module, and a prognostic prediction output module. The data acquisition module collects resting-state EEG signal data and clinical information from patients. The EEG signal preprocessing module preprocesses the resting-state EEG signals to obtain high-quality EEG signal data. The feature extraction module extracts time-frequency domain features and functional connectivity features from the high-quality EEG signal data and fuses them to obtain a fused feature vector. The feature filtering module filters the fused feature vector and clinical variables from the clinical information to obtain a core feature set. The prediction model construction module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module outputs the probability prediction value of the patient's functional recovery through the prognostic prediction model and ranks and outputs important features in the model's decision-making process based on a game theory-based interpretability method. This invention improves the accuracy of prognostic prediction by integrating local dynamic features of electroencephalography (EEG) with brain network connectivity features.

[0082] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0083] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0084] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A prognostic prediction system based on the fusion of multimodal features and clinical variables, characterized in that, The prognostic prediction system based on the fusion of multimodal features and clinical variables includes: a data acquisition module, an EEG signal preprocessing module, a feature extraction module, a feature screening module, a prediction model construction module, and a prognostic prediction output module; The data acquisition module, the EEG signal preprocessing module, the feature extraction module, the feature filtering module, the prediction model construction module, and the prognosis prediction output module are connected in sequence, and the data acquisition module is also connected to the feature filtering module. The data acquisition module is used to collect the patient's resting-state electroencephalogram (EEG) signal data and clinical information; The EEG signal preprocessing module is used to preprocess the resting-state EEG signal to obtain high-quality EEG signal data; The feature extraction module is used to extract time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and to fuse the time-frequency domain features and the functional connectivity features to obtain a fused feature vector; The feature filtering module is used to filter the fused feature vector and the clinical variables in the clinical information to obtain a core feature set. The prediction model building module is used to train a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module is used to output a probability prediction value of the patient's functional recovery through the prognostic prediction model.

2. The prognostic prediction system based on the fusion of multimodal features and clinical variables according to claim 1, characterized in that, The data acquisition module includes an electroencephalogram (EEG) signal acquisition unit, a clinical information acquisition unit, and a data alignment unit. The EEG signal acquisition unit is used to acquire the patient's resting-state EEG signal data through an EEG recording device; The clinical information collection unit is used to collect the patient's clinical information, which includes age, gender, history of hypertension, history of diabetes, duration of impaired consciousness, and various sub-items of the CRS-R score. The duration of impaired consciousness is the number of days the patient was in a state of impaired consciousness from the onset of the disease to the time of EEG collection. The CRS-R score includes sub-indicators for hearing, vision, motor, speech, communication and arousal. The data alignment unit is used to align the resting-state EEG signal data with the clinical information at the patient level to form a multimodal initial dataset.

3. The prognostic prediction system based on the fusion of multimodal features and clinical variables according to claim 1, characterized in that, The EEG signal preprocessing module adopts a fully automated batch preprocessing process, which is based on the MNE library of the Python language and combines various artificial intelligence methods to directly standardize the raw EEG data. The standardization process includes, in sequence: bandpass filtering, removal of mains interference, resampling, removal of artifact signals, reference reconstruction, and independent component analysis to remove eye movement and electromyography interference.

4. The prognostic prediction system based on the fusion of multimodal features and clinical variables according to claim 1, characterized in that, The feature extraction module includes a time-frequency feature extraction unit, a functional connection feature extraction unit, and a feature fusion unit; The time-frequency feature extraction unit is used to extract time-frequency domain features from the high-quality EEG signal data. The time-frequency domain features include energy features, wavelet coefficients, and signal entropy values ​​to characterize the dynamic properties of local neuronal populations in the cortex. The functional connectivity feature extraction unit is used to filter the high-quality EEG signal data according to different frequency bands to obtain the EEG signal subsets corresponding to each frequency band. For each frequency band EEG signal subset, the coherence index between each pair of effective electrodes in the standard electrode layout is calculated. The coherence indices calculated for all frequency bands are flattened, redundancy is removed, and then spliced ​​to obtain the brain network functional connectivity features. The brain network functional connectivity features are used to quantify the information transmission and collaborative integration capabilities between different brain regions. The feature fusion unit is used to concatenate and fuse the time-frequency domain features with the functional connectivity features to obtain a fused feature vector; The time-frequency domain features characterize local neural dynamics, while the functional connectivity features characterize global brain network integration. The time-frequency domain features and the functional connectivity features are complementary in terms of neural information representation. The fused feature vector simultaneously contains information on local dynamic changes and global network integration to comprehensively characterize the brain functional phenotype of patients with acute-phase consciousness disorders.

5. The prognostic prediction system based on the fusion of multimodal features and clinical variables according to claim 1, characterized in that, The feature filtering module includes a first-stage filtering unit and a second-stage filtering unit; The first-stage screening unit is used to receive the fused feature vector and the clinical variables in the clinical information, perform univariate statistical tests on each feature, calculate the significance level of the association between each feature and the prognostic label, and send the features whose significance level of association meets the preset requirements as significant features to the second-stage screening unit. The second-stage screening unit is used to receive the salient features and use a recursive feature elimination algorithm to perform dimensionality reduction screening on the salient features, and output a core feature set; The recursive feature elimination algorithm builds a model iteratively, removing the feature that contributes the least to the prediction in the current model in each iteration, until the number of remaining features meets a preset condition.

6. The prognostic prediction system based on the fusion of multimodal features and clinical variables according to claim 1, characterized in that, The prognostic prediction output module includes a probability prediction unit and an interpretable output unit; The probability prediction unit is used to output a probability prediction value of the patient's functional independence using the trained prognostic prediction model, and when the probability prediction value exceeds a preset threshold, it determines whether the patient's prognosis prediction result is a good prognosis or a bad prognosis, and presents the prediction result in a visual form. The interpretability output unit is used to calculate the marginal contribution of each input feature in the model decision-making process using the interpretability method of game theory, so as to obtain the contribution value and contribution direction of each feature to the prediction result. The interpretability output unit is also used to sort each feature according to the contribution value, output a feature importance ranking map, and output a force map for each predicted sample. In the force map, visualization elements are used to distinguish the contribution direction of each feature, and the contribution direction includes positive contribution direction and negative contribution direction. The feature importance ranking graph and the force graph are used to provide clinicians with evidence to support the model's predictions, thereby improving the acceptability of the model in clinical decision-making.

7. A prognostic prediction method based on the prognostic prediction system based on the fusion of multimodal features and clinical variables as described in any one of claims 1-6, characterized in that, The prognostic prediction method includes: The data acquisition module collects the patient's resting-state electroencephalogram (EEG) signal data and clinical information; The EEG signal preprocessing module preprocesses the resting-state EEG signal to obtain high-quality EEG signal data; The feature extraction module extracts time-frequency domain features and functional connectivity features from the high-quality EEG signal data, and fuses the time-frequency domain features and the functional connectivity features to obtain a fused feature vector; The feature filtering module performs feature filtering on the fused feature vector and the clinical variables in the clinical information to obtain a core feature set. The prediction model building module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model. The prognostic prediction output module outputs the probability prediction value of the patient's functional recovery through the prognostic prediction model, and sorts and outputs the important features in the model decision-making process based on the interpretability method of game theory.

8. A prognostic prediction method based on the prognostic prediction system based on the fusion of multimodal features and clinical variables as described in any one of claims 7, characterized in that, The prediction model building module trains a machine learning classification model based on the core feature set to generate a prognostic prediction model, specifically including: The core feature set is used as the input variable, and the patient's six-month functional outcome is used as the label variable. The six-month functional outcome is based on the GOS-E score. When the GOS-E score is greater than the preset score, the six-month functional outcome is a good prognosis; otherwise, it is a poor prognosis. The model is trained using a random forest classifier. The random forest classifier uses the default parameters of the scikit-learn library in Python and makes decisions using a majority voting mechanism to enhance robustness to noise, outliers and missing features. The model parameters were optimized by five-fold cross-validation, and a stratified sampling strategy was adopted to ensure that the proportion of good prognoses and bad prognoses in each fold was consistent with the original dataset. After training, the prognostic prediction model was obtained.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a prognostic prediction program of a prognostic prediction system based on the fusion of multimodal features and clinical variables, which is stored in the memory and can run on the processor. When the prognostic prediction program of the prognostic prediction system based on the fusion of multimodal features and clinical variables is executed by the processor, it implements the steps of the prognostic prediction method of the prognostic prediction system based on the fusion of multimodal features and clinical variables as described in any one of claims 7-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a prognostic prediction program for a prognostic prediction system based on the fusion of multimodal features and clinical variables. When the prognostic prediction program for a prognostic prediction system based on the fusion of multimodal features and clinical variables is executed by a processor, it implements the steps of the prognostic prediction method for a prognostic prediction system based on the fusion of multimodal features and clinical variables as described in any one of claims 7-8.

Images