An intracranial pressure prediction and early warning system based on deep learning

Abstract

This invention relates to the field of artificial intelligence and medical monitoring integration, and in particular to a deep learning-based intracranial pressure prediction and early warning system. The system includes: a multi-source data acquisition layer for collecting multi-source data from patients; a multi-source data preprocessing layer for preprocessing the multi-source data to obtain standardized multi-source data; a knowledge graph construction layer for constructing a medical knowledge graph centered on patients and time windows using the standardized multi-source data; a deep learning prediction layer for obtaining intracranial pressure prediction results based on subgraphs of the medical knowledge graph from multiple consecutive time windows; and a user interaction layer for outputting the intracranial pressure prediction results and issuing early warnings. This invention, by constructing a medical knowledge graph and then using deep learning to predict intracranial pressure, leverages the dual advantages of deep learning in graph data and time-series data to achieve multimodal data fusion prediction of intracranial pressure, and can provide clinically understandable interpretations of the prediction results, improving the credibility of the early warnings.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and medical monitoring integration, and in particular to a deep learning-based intracranial pressure prediction and early warning system. Background Technology

[0002] Intracranial pressure (ICP) is an important indicator of brain health. Abnormally high ICP can lead to serious complications such as brain herniation and cerebral edema, and even be life-threatening. Therefore, accurate prediction of ICP changes and timely warning are crucial for the treatment and care of brain diseases.

[0003] Existing intracranial pressure prediction methods have the following main shortcomings: First, their ability to fuse multi-source data is limited. Traditional methods often rely on single-modality data (such as using only time-series physiological signals) or simply splicing multi-source features, making it difficult to capture the deep semantic associations and causal relationships between different modalities. Second, the "black box" nature of deep learning models limits their clinical acceptability. Although existing models can output prediction results, they cannot explain the dominant factors in the prediction, making it difficult for doctors to judge the reliability of the prediction results. Third, there is a lack of dynamic knowledge accumulation mechanisms. A large number of historical cases accumulated in clinical practice have not been effectively transformed into reusable structured knowledge, resulting in each patient's prediction model not being able to benefit from similar past cases. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a deep learning-based intracranial pressure prediction and early warning system.

[0005] To achieve the above objectives, this invention provides a deep learning-based intracranial pressure prediction and early warning system. The system includes: a multi-source data acquisition layer, a multi-source data preprocessing layer, a knowledge graph construction layer, a deep learning prediction layer, and a user interaction layer. The multi-source data acquisition layer is used to collect multi-source data from patients. The multi-source data preprocessing layer is data-connected to the multi-source data acquisition layer and is used to preprocess the multi-source data to obtain standardized multi-source data. The knowledge graph construction layer is data-connected to the multi-source data preprocessing layer and is used to construct a medical knowledge graph centered on the patient and time window using the standardized multi-source data. The deep learning prediction layer is data-connected to the knowledge graph construction layer and is used to obtain intracranial pressure prediction results based on subgraphs of the medical knowledge graph from multiple consecutive time windows. The user interaction layer is data-connected to the deep learning prediction layer and is used to output the intracranial pressure prediction results and issue early warnings based on the intracranial pressure prediction results. This invention constructs a medical knowledge graph and then uses deep learning to predict intracranial pressure. It leverages the dual advantages of deep learning in graph data and time-series data to achieve multimodal data fusion for intracranial pressure prediction and provides clinically understandable interpretations of the prediction results, thereby improving the credibility of early warnings.

[0006] Optionally, it further includes: a data storage layer, which is used to store the data acquired by the multi-source data acquisition layer, the multi-source data preprocessing layer, the knowledge graph construction layer and the deep learning prediction layer, and the data storage layer is also used to store historical medical knowledge graphs.

[0007] Optionally, the multi-source data includes time-series physiological data, clinical event data, patient static characteristics, and text data.

[0008] Optionally, the preprocessing includes: The time-series physiological data are filtered and denoised, outlier identification and removal are performed, missing value filling and resampling are performed, and then the statistical features of the time-series physiological data are extracted as standardized time-series physiological data. The clinical event data is standardized by terminology and supplemented with event attributes to obtain standardized clinical event data. Standardized patient static characteristics are obtained by unifying the data format and standardizing the terminology of patient static characteristics. The text data is then structured to obtain standardized text data.

[0009] Optionally, the standardized multi-source data is used to construct the medical knowledge graph, the nodes of which include patient nodes, time window nodes, physiological state nodes, clinical event nodes, and text feature nodes.

[0010] Optionally, the patient node is connected to different time window nodes, and the physiological state node, the clinical event node, and the text feature node are connected to the time window nodes.

[0011] Optionally, the deep learning prediction layer performs the following steps: Obtain subgraphs of the medical knowledge graph across multiple consecutive time windows, and use a graph attention network to obtain a sequence of graph embedding vectors; The graph is embedded into a vector sequence and input into a long short-term memory neural network to obtain the intracranial pressure prediction value within a future time window.

[0012] Optionally, the deep learning prediction layer obtains the dominant factors of the intracranial pressure prediction value through temporal attention and intra-graph attention.

[0013] Optionally, the deep learning prediction layer determines model uncertainty by discarding predictions multiple times, and determines data support based on similarity retrieval of the medical knowledge graph.

[0014] Optionally, the deep learning prediction layer calculates the difference between 1 and the model uncertainty, and uses the weighted sum of the calculation result and the data support as the confidence level of the intracranial pressure prediction value.

[0015] In summary, the present invention has at least the following beneficial effects: 1. This invention collects multi-source patient data and constructs a medical knowledge graph centered on patients and time windows, and then uses deep learning to predict intracranial pressure, achieving multimodal data fusion for intracranial pressure prediction and improving prediction accuracy.

[0016] 2. Based on the construction of a medical knowledge graph, this invention constructs a prediction model for intracranial pressure using graph attention networks and long short-term memory neural networks. This model can leverage the dual advantages of deep learning in graph data and time-series data to achieve intracranial pressure prediction, thereby improving the reliability of the prediction results.

[0017] 3. While predicting intracranial pressure, this invention obtains the dominant factors of intracranial pressure prediction value through temporal attention and intra-graphical attention, providing a clinically understandable interpretation of the prediction results and assisting in decision-making and intervention strategies.

[0018] 4. This invention uses the weighted sum of model uncertainty and data support as the confidence level of intracranial pressure prediction. This avoids the prediction model from being overconfident in "rare but real" conditions and also prevents the value of reliable prediction from being underestimated due to insufficient historical data. This multi-dimensional verification makes the confidence level judgment more comprehensive and robust. The final output to doctors is no longer a single conclusion of "whether this prediction is accurate or not", but a credibility assessment that has been double-verified, which is more in line with the actual needs of clinical decision support for prediction reliability. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the framework of a deep learning-based intracranial pressure prediction and early warning system according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the intracranial pressure prediction process of a deep learning-based intracranial pressure prediction and early warning system according to an embodiment of the present invention. Detailed Implementation

[0021] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0022] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0023] It should be noted in advance that, in one alternative embodiment, apart from being described independently, the same symbols or letters appearing in all formulas have the same meaning.

[0024] In one optional embodiment, please refer to Figure 1 This invention provides a deep learning-based intracranial pressure prediction and early warning system, the system comprising: The system comprises a multi-source data acquisition layer, a multi-source data preprocessing layer, a knowledge graph construction layer, a deep learning prediction layer, and a user interaction layer. Specifically, the multi-source data acquisition layer collects multi-source data from patients; the multi-source data preprocessing layer, connected to the multi-source data acquisition layer, preprocesses the multi-source data to obtain standardized multi-source data; the knowledge graph construction layer, connected to the multi-source data preprocessing layer, constructs a medical knowledge graph centered on the patient and time window using the standardized multi-source data; the deep learning prediction layer, connected to the knowledge graph construction layer, obtains intracranial pressure prediction results based on subgraphs of the medical knowledge graph from multiple consecutive time windows; the user interaction layer, connected to the deep learning prediction layer, outputs the intracranial pressure prediction results and issues warnings based on them; and a data storage layer stores the data acquired by the multi-source data acquisition layer, the multi-source data preprocessing layer, the knowledge graph construction layer, and the deep learning prediction layer, and also stores the historical medical knowledge graph used to train the prediction model.

[0025] Specifically, in this embodiment, the multi-source data acquired by the multi-source data acquisition layer includes time-series physiological data, clinical event data, patient static characteristics, and text data. Time-series physiological data includes measured intracranial pressure (ICP), arterial blood pressure (ABP), cerebral blood flow velocity, intracranial temperature, and central venous pressure (CVP), etc. Clinical event data includes medication records, surgical procedures, and nursing procedures, etc. Patient static characteristics include patient name, patient gender, patient ID, age, disease diagnosis, and Glasgow Coma Scale (GCS), etc. Text data refers to intracranial medical imaging reports. It should be noted that if capturing dynamic changes in the Glasgow Coma Scale, it should be processed as time-series physiological data.

[0026] In this embodiment, time-series physiological data are typically collected continuously by bedside monitoring equipment at a high sampling frequency. Specifically: intracranial pressure and intracranial temperature are continuously measured using an intracranial pressure monitor, and the results are transmitted wirelessly to the multi-source data acquisition layer; arterial blood pressure is continuously measured using an arterial catheter connected to a pressure sensor, or intermittently using a non-invasive cuff, but the latter has a lower measurement frequency, so invasive measurement methods are generally preferred, and the collected arterial blood pressure is transmitted wirelessly to the multi-source data acquisition layer; cerebral blood flow velocity is measured using transcranial Doppler ultrasound, and the results are transmitted wirelessly to the multi-source data acquisition layer; central venous pressure is measured by inserting a central venous catheter into the superior / inferior vena cava via subclavian vein, internal jugular vein, or femoral vein puncture, connecting a pressure sensor for real-time monitoring, and the results are transmitted wirelessly to the multi-source data acquisition layer; if historical time-series physiological data of different patients is to be obtained, historical time-series physiological data from the hospital monitoring system can be imported into the multi-source data acquisition layer via a data interface based on the HL7 v2.x transmission protocol using a data cable. In addition, the multi-source data acquisition layer may include a manual input interface, allowing medical staff to manually input some time-series physiological data with lower measurement frequencies into the multi-source data acquisition layer.

[0027] In this embodiment, clinical event data comes from the Hospital Information System (HIS), medical order system, and nursing records; these are discrete, timestamped events. Patient static characteristic data is collected upon patient admission and remains largely unchanged (or changes very slowly) over time; it can be obtained from the Hospital Information System. Text data is unstructured and can be obtained from the hospital's image archiving and communication systems. Clinical event data, patient static characteristics, and text data can all be imported into the multi-source data acquisition layer via data cables through the data interfaces of the respective systems. Of course, medical staff can also manually enter this data into the multi-source data acquisition layer.

[0028] Specifically, in this embodiment, after the multi-source data acquisition layer collects time-series physiological data, clinical event data, patient static characteristics, and text data, it transmits them to the multi-source data preprocessing layer. The multi-source data preprocessing layer then performs data preprocessing to obtain standardized multi-source data. This standardized multi-source data includes standardized time-series physiological data, standardized clinical event data, standardized patient static characteristics, and standardized text data.

[0029] In this embodiment, based on signal characteristics and clinical needs, there is a high demand for noise reduction of measured intracranial pressure, arterial blood pressure, and cerebral blood flow velocity, while the demand for noise reduction of intracranial temperature and central venous pressure is relatively low. Specifically, noise sources for measured intracranial pressure typically include impulse noise or random interference that may be introduced by patient agitation, suctioning, or turning over. Median filtering can be used to remove impulse noise and protect key features in the ICP waveform. Noise sources for arterial blood pressure typically include motion artifacts, vascular pulsation interference, or sensor drift. Low-pass filtering can be used to retain low-frequency blood pressure fluctuations (such as respiratory cycle effects) and filter out high-frequency noise (such as electromyographic interference). Noise sources for cerebral blood flow velocity typically include vascular wall vibration, probe movement, or physiological interference (such as heartbeat and respiration). Wavelet transform filtering can be used to separate blood flow signals from noise components. Noise sources for intracranial temperature typically include ambient airflow and thermal equilibrium delay of the temperature probe. Noise sources for central venous pressure typically include sensor position changes, tubing bubbles, or changes in patient position. Both of these time-series physiological data can be denoised using median filtering.

[0030] After denoising the time-series physiological data, the next step is outlier identification and removal. Intracranial temperature and central venous pressure are usually relatively stable, so outlier detection and removal can be performed. The principle is to identify and remove outliers, and then use spline interpolation to fill in missing values. Compared with intracranial temperature and central venous pressure, the measured values ​​of intracranial pressure, arterial blood pressure, and cerebral blood flow velocity are not stable. Therefore, a sliding window dynamic thresholding method (window length of 10-30 seconds, threshold of 2-3 times the standard deviation) is used to identify and remove outliers, and then spline interpolation is used to fill in missing values.

[0031] After identifying outliers and imputing missing values ​​in the time-series physiological data, the data is resampled at preset sampling intervals to align various time-series physiological data points temporally. The resampling results are then stored in the data storage layer. The sampling interval is typically set to 100-500 ms. Furthermore, the time-series physiological data after outlier identification and missing value imputation can be displayed using a user interaction layer. Medical staff can then verify the data and manually modify it. This approach aims to combine objective data with the subjective experience of medical staff, improving data accuracy and the accuracy of intracranial pressure prediction.

[0032] After resampling the time-series physiological data, the statistical characteristics of the time-series physiological data can be obtained and used as standardized time-series physiological data. These statistical characteristics include, but are not limited to, the mean, maximum, minimum, standard deviation, number of times the threshold is exceeded, peak amplitude, and trend slope of the time-series physiological data within the time window. The time window length is set to 30 minutes and can be adjusted according to actual needs; the peak amplitude is the difference between the maximum and minimum values ​​of the time-series physiological data; the trend slope is the rate of change of the time-series physiological data per unit time, reflecting the "steepness" of the data's rise or fall, and can be calculated using linear regression or least squares methods, specifically employing existing technical means.

[0033] In this embodiment, the preprocessing operations for clinical event data include terminology standardization and event attribute supplementation. Terminology standardization refers to uniformly mapping names expressing the same meaning to standard entity names. Specifically, this involves extracting semantic features of terms using the RoBERTa model and matching them to standard terms using cosine similarity calculations to eliminate ambiguity in medical terminology and ensure that the same concept in the knowledge graph corresponds to only one node, thereby improving data consistency, query accuracy, and inference reliability. For example, "cerebral edema after cerebral hemorrhage," "intracranial hypertension syndrome," and "intracranial hypertension" are uniformly mapped to the standard entity name "intracranial hypertension." Event attribute supplementation refers to adding specific content to clinical events. Generally, hospital information systems, medical order systems, and nursing records contain detailed records of clinical events, so supplementation is unnecessary. However, if specific content of a clinical event is missing, it is supplemented as "unknown." For example, the specific content of the clinical event "administration" is "mannitol administration," but the dosage is missing. Since the hospital information system, medical order system, and nursing records do not record the dosage, it is supplemented as "dosage unknown." After completing terminology standardization and event attribute supplementation of the clinical event data, standardized clinical event data is obtained.

[0034] In this embodiment, the preprocessing operations for patient static features typically include data format standardization and terminology standardization. Terminology standardization primarily addresses the issue of using different terms to describe the same disease during diagnosis, while data format standardization focuses on age. The original data may contain different recording formats for age; for example, some may record it directly in "years," while others may include specific "months." In such cases, the age unit needs to be standardized to "years," for example, recording 6 months as 0.5 years. After completing the preprocessing of patient static features, standardized patient static features are obtained.

[0035] In this embodiment, intracranial medical imaging reports are typically stored as PDF files. Therefore, this embodiment uses PDF-Extract-Kit to extract the text data into structured information in JSON format, obtaining standardized text data. The information output by PDF-Extract-Kit includes the patient, examination time, and imaging features (such as midline structures, offset to the left by approximately 0.5 cm).

[0036] After obtaining standardized multi-source data, for any given patient, the standardized multi-source data is allocated to a corresponding time window based on its timestamp and stored in the data storage layer in the format of "patient-time window-standardized multi-source data" to facilitate the subsequent construction of a medical knowledge graph. In addition, the original multi-source data and intermediate data generated during the acquisition of standardized multi-source data (such as the resampling results of time-series physiological data) can also be stored under the corresponding time window.

[0037] In other alternative embodiments, other data besides time-series physiological data can be extracted directly from the patient's medical records. However, since the time granularity of the medical records is coarse, and events are usually described in "day" or "fuzzy time" (such as "this morning"), a larger time window length needs to be set.

[0038] Specifically, in this embodiment, the knowledge graph construction layer uses the standardized multi-source data to construct the medical knowledge graph. The nodes of the medical knowledge graph include patient nodes, time window nodes, physiological state nodes, clinical event nodes, and text feature nodes. Patient nodes are connected to different time window nodes through "belonging" relationships, meaning one patient corresponds to multiple time windows. Physiological state nodes are connected to their corresponding time window nodes through "owning" relationships, clinical event nodes through "occurrence" relationships, and image feature nodes through "presentation" relationships. Physiological state nodes include measured intracranial pressure nodes, arterial blood pressure nodes, cerebral blood flow velocity nodes, intracranial temperature nodes, and central venous pressure nodes. Clinical event nodes include medication record nodes, surgical operation nodes, and nursing operation nodes, and different medications, different surgical operations, and different nursing operations should each have their own nodes. Text feature nodes can also be called image feature nodes, and different image features need to have their own nodes.

[0039] More specifically, the node attributes of the patient node include patient name, patient gender, patient ID, age, disease diagnosis, and Glasgow Coma Scale score; the node attribute of the time window node is the time range; the node attribute of the physiological state node is the corresponding standardized time-series physiological data; the node attribute of the clinical event node is the specific content of the event, such as drug dosage, surgical site, and specific nursing procedures; and the node attribute of the text feature node is the corresponding image feature, such as midline structure offset and hematoma volume.

[0040] Furthermore, since the node attributes of the medical knowledge graph include not only numerical data but also textual data—for example, dosage is numerical data, while surgical site is textual data—it is necessary to encode the textual node attributes so that the deep learning model can use this data to predict intracranial pressure. Specifically, to enable the subsequent graph attention network to obtain graph embedding vectors from the medical knowledge graph, this embodiment uses an embedding layer to map textual attributes into dense vectors of fixed dimensions, which are then used as part of the initial node representation in the graph attention network's computation. In addition, to avoid large numerical features dominating gradient updates and weakening the influence of small numerical features due to excessive differences in numerical ranges during model training, this embodiment uses a minimax normalization method to normalize all numerical node attributes, thereby obtaining the final medical knowledge graph for predicting intracranial pressure, which is then stored in the data storage layer.

[0041] Specifically, in this embodiment, the deep learning prediction layer predicts intracranial pressure based on a prediction model, which is constructed using a Graph Attention Network (GAT) and a Long Short-Term Memory (LSTM) neural network. The deep learning prediction layer performs the following steps: S1. Obtain the medical knowledge graph subgraphs of multiple consecutive time windows, and use a graph attention network to obtain the graph embedding vector sequence.

[0042] In this embodiment, a medical knowledge graph subgraph for a time window is a local knowledge network centered on the node of that time window, containing all physiological states, clinical events, imaging features, and static features of the patient within that window. By inputting the medical knowledge graph subgraphs of the patient's most recent consecutive time windows into a graph attention network, the graph embedding vectors of the corresponding time windows can be obtained. The graph embedding vectors of multiple consecutive time windows can then form a graph embedding vector sequence.

[0043] S2. The graph is embedded into a vector sequence and input into a long short-term memory neural network to obtain the intracranial pressure prediction value within a future time window.

[0044] In this embodiment, a graph embedding vector sequence is input into a long short-term memory neural network, which then outputs a predicted intracranial pressure value for a future time window. The long short-term memory neural network incorporates a temporal attention mechanism, namely LSTM-Attention, enabling the deep learning prediction layer to capture the dominant factors in the predicted intracranial pressure value through temporal attention and in-graph attention.

[0045] Specifically, when multiple consecutive time windows of the medical knowledge graph subgraphs are input into the graph attention network, the intra-graph attention mechanism calculates the importance weights of each neighboring node for that time window node within each subgraph. Essentially, this process identifies which physiological parameters, clinical events, or imaging features contribute most to the overall state of each time window. Subsequently, the graph embedding vectors of these time windows are fed into a long short-term memory network with a temporal attention layer. The temporal attention mechanism calculates the contribution weights of each time window to the final prediction result along the time dimension. By combining these two levels of attention weights, the system can not only pinpoint "which time window is most critical" but also trace "which specific factors played a dominant role" within that window, ultimately forming a complete explanatory path from "critical time window" to "critical factors." These attention weights are automatically learned during model training, requiring no manual annotation, and are directly output during prediction, thus enabling interpretable analysis of intracranial pressure prediction results.

[0046] Furthermore, the deep learning prediction layer determines model uncertainty by discarding predictions multiple times, and determines data support based on similarity retrieval of the medical knowledge graph.

[0047] Specifically, model uncertainty is derived using the Monte Carlo dropout method. During the training phase of the prediction model, both the graph attention network and the long short-term memory network employ a random node dropout mechanism (i.e., dropout), enabling the model to generalize to different network structures during learning. During the prediction phase, the dropout mechanism remains enabled, and the same input is subjected to multiple independent forward propagations (e.g., 50 times), each yielding a predicted intracranial pressure value. Because the randomly dropped nodes differ with each propagation, the model's prediction results exhibit slight fluctuations. The predicted intracranial pressure values ​​for different node drops are recorded, and their mean and standard deviation are calculated. The calculated mean is used as the final intracranial pressure prediction result, while the standard deviation represents the model uncertainty for this prediction. Lower model uncertainty indicates higher certainty for the prediction, while higher uncertainty indicates internal disagreement within the model regarding the prediction result, resulting in lower certainty.

[0048] When obtaining data support, the medical knowledge graph subgraph of the current time window (hereinafter referred to as the current medical knowledge graph subgraph) is used as the query condition. The most similar historical medical knowledge graph subgraph is retrieved from the complete medical knowledge graph. During the retrieval, the physiological state nodes, clinical event nodes, image feature nodes, and their relational structures of the two medical knowledge graph subgraphs are compared, and a similarity score is calculated. The similarity score satisfies the following relationship: Where S is the similarity score. , and For similarity weights, , and The values ​​are 0.3, 0.5, and 0.2 respectively. To determine the similarity of nodes, For node attribute similarity, For structural similarity.

[0049] More specifically, the methods for obtaining node composition similarity, node attribute similarity, and structural similarity are as follows: 1. Node composition similarity Extract the node type sets of two medical knowledge graph subgraphs and calculate the Jaccard similarity coefficient between the two node type sets as the node composition similarity of the two medical knowledge graph subgraphs.

[0050] 2. Node attribute similarity Obtain the node type set of each of the two medical knowledge graph subgraphs, and take the intersection of the two node type sets to obtain the common node type set of the two medical knowledge graph subgraphs.

[0051] For each node in the common node type set, the node attributes of that node are obtained from both medical knowledge graph subgraphs. Following the order of the various nodes in the common node type set, the node attributes of each node in the same medical knowledge graph subgraph are arranged sequentially to form a one-dimensional long vector, which serves as the attribute similarity calculation vector for that medical knowledge graph subgraph. The attribute similarity calculation vector can be expressed as: in, Calculate vectors for attribute similarity of subgraphs in a medical knowledge graph. Let i be the node attribute of the i-th type of node in the set of common node types. , where m is the number of node types in the common node type set. Obtain the attribute similarity calculation vectors corresponding to the two medical knowledge graph subgraphs, and calculate the cosine similarity between the two attribute similarity calculation vectors, which is used as the node attribute similarity between the two medical knowledge graph subgraphs.

[0052] 3. Structural similarity Extract the edge sets of each of the two medical knowledge graph subgraphs and calculate the Jaccard similarity coefficient between the two edge sets as the structural similarity between the two medical knowledge graph subgraphs.

[0053] The similarity score threshold is set to 0.8, and historical medical knowledge graph subgraphs with similarity scores higher than the current medical knowledge graph subgraph are used as reference subgraphs, with the time windows corresponding to these reference subgraphs serving as reference time windows. Next, it is determined whether the overall intracranial pressure (ICP) situation is the same within at least one subsequent time window. In this embodiment, this determination is based on the cosine similarity of measured ICP values, as this method is simple and quick. Specifically, for any reference time window, the sequence of measured ICP values ​​(already resampled) within at least one subsequent time window is extracted and concatenated into a similarity judgment vector along the time dimension (if only one time window is involved, concatenation is unnecessary, and the sequence of measured ICP values ​​within that time window is directly used as the similarity judgment vector). Then, the cosine similarity of similarity judgment vectors between different reference time windows is calculated, and a similarity threshold of 0.8 is set. If the cosine similarity of two similarity judgment vectors is not less than the similarity threshold, the overall intracranial pressure situation of the corresponding two reference time windows is considered to be the same; otherwise, they are considered to be different. In this way, the number of reference time windows with the same overall intracranial pressure situation can be obtained. However, it should be noted that if reference time windows with the same overall intracranial pressure situation are counted as a group of similar time windows, then there may be multiple groups of similar time windows. Finally, the data support is calculated according to the following relationship: Where D represents the data support level. As a penalty item, This represents the number of reference time windows in a set of similar time windows. This represents the total number of reference time windows. The penalty term can be set as the ratio of the actual number to the ideal number within the reference time window, with a maximum value of 1. If the actual number within the reference time window exceeds the ideal number, the penalty term is set to 1. If the actual number within the reference time window is 0, the data support is set to a smaller value, such as 0.2.

[0054] The process of acquiring data support essentially utilizes knowledge graphs as a repository of clinical experience, providing references for current predictions through historical examples.

[0055] After calculating the model uncertainty and data support, the difference between 1 and the model uncertainty is calculated. The weighted sum of the calculation result and the data support is used as the confidence level of the intracranial pressure prediction, with initial weights of 0.7 and 0.3, respectively. Where C represents the confidence level and MU represents the model uncertainty.

[0056] Specifically, in this embodiment, after obtaining the predicted intracranial pressure value, the prediction result is transmitted to the user interaction layer and displayed through a digital display interface. Simultaneously, the user interaction layer compares the predicted intracranial pressure value with the normal intracranial pressure range. If the predicted intracranial pressure value is not within the normal range, a voice warning is issued. For example, when the predicted intracranial pressure value exceeds the upper limit of the normal intracranial pressure range, a voice prompt stating "Intracranial hypertension is imminent" is issued.

[0057] The user interaction layer is connected to the deep learning prediction layer to output intracranial pressure prediction results and issue warnings based on the intracranial pressure prediction results.

[0058] For easier understanding, please refer to Figure 2 This embodiment briefly describes the operation flow of the system, which includes: 1. Collecting multi-source data of patients; 2. Preprocessing the multi-source data to obtain standardized multi-source data; 3. Using the standardized multi-source data to construct a medical knowledge graph centered on patients and time windows, encoding the attributes of text-type nodes, and normalizing the attributes of numerical-type nodes; 4. Inputting the subgraphs of the medical knowledge graph of the patient's most recent consecutive time windows into the prediction model to obtain the predicted value of intracranial pressure and the dominant factors in the next time window; 5. Calculating the model uncertainty and data support, and then outputting the weighted sum of the model uncertainty and data support as the confidence level of the predicted value of intracranial pressure.

[0059] It should be noted that in some cases, the actions described in the specification can be performed in different orders and still achieve the desired results. In this embodiment, the order of steps is given only to make the embodiment clearer and easier to explain, and not to limit it.

[0060] In summary, this invention has at least the following beneficial effects: By collecting multi-source patient data and constructing a medical knowledge graph centered on the patient and time window, this invention achieves intracranial pressure prediction based on deep learning, realizing multimodal data fusion for intracranial pressure prediction and improving prediction accuracy; Based on the constructed medical knowledge graph, this invention builds an intracranial pressure prediction model using graph attention networks and long short-term memory neural networks, leveraging the dual advantages of deep learning on graph and time-series data to achieve intracranial pressure prediction, thus improving the reliability of the prediction results; While predicting intracranial pressure, this invention also utilizes time-series attention and... The dominant factor in intracranial pressure prediction is the attention within the model, providing a clinically understandable interpretation of the prediction results and supporting decision-making intervention strategies. This invention uses the weighted sum of model uncertainty and data support as the confidence level of the intracranial pressure prediction. This avoids the prediction model from being overconfident in "rare but real" states, and also prevents the underestimation of the value of reliable predictions due to insufficient historical data. This multi-dimensional verification makes the confidence level judgment more comprehensive and robust. The final output to doctors is no longer a single conclusion of "whether this prediction is accurate or not", but a credibility assessment that has been double-verified, which is more in line with the actual needs of clinical decision support for prediction reliability.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A deep learning-based intracranial pressure prediction and early warning system, characterized in that, include: Multi-source data acquisition layer, multi-source data preprocessing layer, knowledge graph construction layer, deep learning prediction layer, and user interaction layer; The multi-source data acquisition layer is used to acquire multi-source data from patients; The multi-source data preprocessing layer is connected to the multi-source data acquisition layer and is used to preprocess the multi-source data to obtain standardized multi-source data. The knowledge graph construction layer is connected to the multi-source data preprocessing layer and is used to construct a medical knowledge graph centered on patients and time windows using the standardized multi-source data. The deep learning prediction layer is data-connected to the knowledge graph construction layer, and is used to obtain intracranial pressure prediction results based on medical knowledge graph subgraphs in multiple consecutive time windows. The user interaction layer is connected to the deep learning prediction layer to output intracranial pressure prediction results and issue warnings based on the intracranial pressure prediction results.

2. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that, Also includes: The data storage layer is used to store the data acquired by the multi-source data acquisition layer, the multi-source data preprocessing layer, the knowledge graph construction layer, and the deep learning prediction layer. The data storage layer is also used to store historical medical knowledge graphs.

3. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that: The multi-source data includes time-series physiological data, clinical event data, patient static characteristics, and text data.

4. The intracranial pressure prediction and early warning system based on deep learning according to claim 3, characterized in that, The preprocessing includes: The time-series physiological data are filtered and denoised, outlier identification and removal are performed, missing value filling and resampling are performed, and then the statistical features of the time-series physiological data are extracted as standardized time-series physiological data. The clinical event data is standardized by terminology and supplemented with event attributes to obtain standardized clinical event data. Standardized patient static characteristics are obtained by unifying the data format and standardizing the terminology of patient static characteristics. The text data is then structured to obtain standardized text data.

5. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that: The medical knowledge graph is constructed using the standardized multi-source data. The nodes of the medical knowledge graph include patient nodes, time window nodes, physiological state nodes, clinical event nodes, and text feature nodes.

6. The intracranial pressure prediction and early warning system based on deep learning according to claim 5, characterized in that: The patient node is connected to different time window nodes, and the physiological state node, the clinical event node, and the text feature node are connected to the time window nodes.

7. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that, The deep learning prediction layer performs the following steps: Obtain subgraphs of the medical knowledge graph across multiple consecutive time windows, and use a graph attention network to obtain a sequence of graph embedding vectors; The graph is embedded into a vector sequence and input into a long short-term memory neural network to obtain the intracranial pressure prediction value within a future time window.

8. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that: The deep learning prediction layer obtains the dominant factors for the intracranial pressure prediction value through temporal attention and intra-graph attention.

9. The intracranial pressure prediction and early warning system based on deep learning according to claim 1, characterized in that: The deep learning prediction layer determines model uncertainty by discarding predictions multiple times, and determines data support based on similarity retrieval of the medical knowledge graph.

10. A deep learning-based intracranial pressure prediction and early warning system according to claim 9, characterized in that: The deep learning prediction layer calculates the difference between 1 and the model uncertainty, and uses the weighted sum of the calculation result and the data support as the confidence level of the intracranial pressure prediction value.

Images