Noninvasive intracranial pressure prediction method and system and storage medium thereof

Abstract

The invention provides a non-invasive intracranial pressure prediction method and system and a storage medium, and the method comprises the following steps: obtaining a multi-modal vital sign original signal from a subject in a non-invasive manner, embedding a timestamp of the multi-modal vital sign original signal, and generating a high-dimensional time sequence feature of each vital sign; based on a preset reference sequence, associating the time sequence features with the original signals by adopting an attention mechanism to obtain alignment features of each vital sign; inputting the alignment features into a gating circulation unit network, and extracting hidden features of each vital sign; and splicing and fusing the hidden vital sign features, sending the fused hidden vital sign features into a classifier, and outputting a judgment result about whether the intracranial pressure of the subject rises or not. The clinical application value of a non-invasive monitoring technology is improved, and then the treatment effect and prognosis quality of a patient are improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical monitoring, and particularly relates to a non-invasive intracranial pressure prediction method and system and a storage medium thereof. BACKGROUND

[0002] In the field of medical vital sign monitoring, due to the increasing incidence of nervous system diseases (such as intracranial tumors, hydrocephalus, and cerebral hemorrhage) and craniocerebral injury caused by traffic accidents, the condition of abnormal increase of intracranial pressure is becoming increasingly common. The increase of intracranial pressure will seriously affect the blood circulation of brain tissue, leading to the decrease of cerebral perfusion pressure, the reduction of cerebral blood flow, and the limitation of venous return, and even causing the compression and displacement of brain tissue, and the brain hernia of severe patients, which threatens the safety of patients. Therefore, early and accurate monitoring and prediction of intracranial pressure changes have important clinical significance for the treatment decision and prognosis improvement of patients.

[0003] At present, the clinical intracranial pressure monitoring technology is mainly divided into two categories: invasive and non-invasive. Although the invasive monitoring method has high measurement accuracy, it has obvious clinical defects such as infection, bleeding and other complications, which limits its wide use. In contrast, the non-invasive monitoring method gradually becomes an important development direction of clinical intracranial pressure monitoring due to its high safety and easy implementation. However, the existing non-invasive monitoring technology generally relies on a single vital sign signal (such as measurement of cerebral blood flow velocity or optic nerve sheath diameter), which is easily affected by patient individual differences and measurement device limitations, resulting in difficult to achieve the error standard recommended by the American Association for Medical Instrumentation (AAMI), i.e. the error between the measured value and the actual intracranial pressure should be less than 10%, and the maximum error should be less than 2mmHg in the range of 0 to 20mmHg. Therefore, the accuracy and reliability of the current non-invasive intracranial pressure monitoring technology in clinical practice are still insufficient, and further improvement is needed.

[0004] In recent years, with the rapid development of deep learning technology and multi-modal signal fusion analysis method in the field of biomedicine, researchers have begun to try to use the comprehensive analysis of multiple vital signs to improve the accuracy and stability of intracranial pressure prediction. However, since medical vital sign data usually has different sampling frequencies and signal characteristics, for example, the electrocardiogram is sampled hundreds of times per second, while the blood pressure may be sampled only once per minute, how to effectively process and align these multi-modal data becomes a key technical problem. Although the conventional interpolation up-sampling method can maintain more original information, it is easy to introduce artificial noise and false features, which affects the prediction accuracy of the subsequent model; while the down-sampling method has low computational complexity, but it may ignore important short-term variation features and lose valuable information, which is difficult to meet the demand for accuracy and real-time in clinical practice.

[0005] In addition, there is a problem of feature redundancy between different modalities in multi-modal vital sign signal fusion. The traditional simple splicing fusion is easy to cause the generalization performance of the model to be reduced, cannot effectively utilize the independent and complementary information of each modality signal, and further limits the improvement of the prediction accuracy.

[0006] In summary, the existing non-invasive intracranial pressure prediction technology still has great challenges in the effective alignment and fusion of multi-modal vital signs, and cannot fully meet the comprehensive requirements of high precision, real-time performance, stability and generalization ability in clinical practice. SUMMARY

[0007] The purpose of the present application is to solve the problems in the above background art, and to provide a non-invasive intracranial pressure prediction method, system and storage medium, to improve the clinical application value of non-invasive monitoring technology, and to improve the treatment effect and prognosis quality of patients.

[0008] The technical solution adopted by the present application is: a non-invasive intracranial pressure prediction method, comprising the following steps: Obtain multi-modal vital sign original signals from the subject by a non-invasive method, and embed the time stamps to generate high-dimensional time sequence features of each vital sign; Based on a preset reference sequence, use an attention mechanism to associate the time sequence features with the original signals to obtain aligned features of each vital sign; Input the aligned features into a gated recurrent unit network to extract hidden features of each vital sign; Splice and fuse the hidden features of each vital sign, and send them to a classifier to output a determination result of whether the subject has elevated intracranial pressure.

[0009] In the above technical solution, the multi-modal vital signs include bioelectrical impedance, electrocardiogram, heart rate, respiratory rate, pulse rate, body temperature and / or blood oxygen saturation.

[0010] In the above technical solution, the attention mechanism is: calculate the similarity weight between the preset reference sequence and the modality time sequence features through the attention mechanism, and aggregate the original vital sign signals according to the obtained similarity weight to generate the aligned features.

[0011] In the above technical solution, the attention mechanism is: Embed the preset reference sequence as a query; embed the time sequence features of any vital sign as a key; and embed the original signal of the vital sign under the corresponding time stamp as a value; Through the attention mechanism, the value is weighted and aggregated according to the association score between the query and the key to obtain the aligned features of the vital sign.

[0012] In the technical solution, the reference sequence is obtained by mapping time points that are uniformly sampled in a normalization time interval [0, 1] according to a preset reference sequence length through a time embedding function.

[0013] In the technical solution, the attention mechanism, the gated recurrent unit network and the classifier all include trainable parameters. The method further includes: based on the multi-modal vital sign signals of multiple subjects as input and whether the invasive intracranial pressure measurement value at the corresponding moment exceeds a threshold as a label, training the trainable parameters by jointly minimizing the following two losses: The classification loss is used to measure the difference between the predicted result output by the classifier and the invasive intracranial pressure measurement value label. The orthogonal loss is used to constrain the orthogonality between the hidden features of each modality extracted by the gated recurrent unit network.

[0014] In the technical solution, in the training stage, for the hidden feature representations of any two modalities extracted by the gated recurrent unit network, all element-wise inner products between them are calculated to form an inner product matrix; after squaring the difference between the inner product matrix and a unit matrix of the same dimension at the corresponding positions and accumulating, the orthogonal loss value of the pair of modalities is obtained; the sum of the orthogonal loss values of all modality pairs is summed as the overall orthogonal loss; the overall orthogonal loss and the classification loss jointly constitute a joint loss function, which is used to optimize the trainable parameters in the attention mechanism, the gated recurrent unit network and the classifier.

[0015] In the technical solution, the gated recurrent unit network is composed of multiple sub-networks, and each sub-network receives aligned features of one modality and outputs hidden features of the modality.

[0016] The application also provides a non-invasive intracranial pressure prediction system for implementing the method in the technical solution, which includes: A signal acquisition module is configured to acquire multi-modal vital sign raw signals from a subject in a non-invasive manner, and to embed the time stamps of the raw signals to generate high-dimensional time sequence features of each vital sign. A signal alignment module is configured to associate the time sequence features with the raw signals by using an attention mechanism based on a preset reference sequence to obtain aligned features of each vital sign. A feature extraction module is configured to input the aligned features into a gated recurrent unit network to extract hidden features of each vital sign. An intracranial pressure determination module is configured to splice and fuse the hidden features of the vital signs and input them into a classifier to output a determination result of whether the subject has elevated intracranial pressure.

[0017] The application further provides a computer readable storage medium, which stores a computer program, and the computer program enables a computer to implement the non-invasive intracranial pressure prediction method in the above technical solution when executed by a processor.

[0018] The application has the following beneficial effects: the application collects multi-modal vital sign signals non-invasively, generates high-dimensional time sequence features by time embedding, effectively realizes signal alignment of different sampling frequencies based on a reference sequence attention mechanism, further extracts hidden features by using a gated recurrent unit network (GRU), and finally predicts the intracranial pressure state by splicing, fusing and classifying. The application effectively overcomes the problem of inconsistent sampling frequencies of different vital signs, avoids noise or information loss caused by traditional interpolation or down-sampling methods, and improves the accuracy and real-time performance of intracranial pressure prediction; the long-term dependence of time series is captured by the GRU network, which can better learn the law of the evolution of different vital signs over time and improve the prediction stability; the generalization performance and robustness of the prediction model are significantly improved by integrating multi-modal information, which is conducive to accurately determining the intracranial pressure rising trend and thus more timely clinical intervention.

[0019] Further, the application specifically limits the types of multi-modal vital signs to bioelectrical impedance, electrocardiogram, heart rate, respiratory rate, body temperature and / or blood oxygen saturation. The application is aimed at intracranial pressure changes, and these signals cover various physiological mechanisms and body function states, and can comprehensively reflect the conditions of multiple systems such as the nervous, circulatory, respiratory and metabolic systems of the patient, further improving the accuracy and comprehensiveness of prediction; the application provides a set of signals that are easy to obtain in clinical application, increasing the practicability of the method and the feasibility of clinical popularization.

[0020] Further, the application is based on an attention mechanism, and the similarity weight between the reference sequence and each modal time sequence feature is calculated to weight and aggregate the original vital sign signals, realizing fine alignment of data; dynamic alignment is realized by the similarity weight, more accurately capturing the correlation of vital signs under different sampling frequencies, and maximizing the retention of effective information in the original data; the dynamic and adaptive weighting method makes the alignment process more intelligent, improving the fault tolerance and robustness of the model to signal time differences.

[0021] Further, the application takes reference sequence embedding as a query (Query), vital sign time sequence features as a key (Key), and original signal values as a value (Value), calculates an association score, and then weights and aggregates to generate aligned features; the Q-K-V implementation details inside the attention mechanism are specified, which is conducive to stable implementation and further optimization of the method; and the original observation information of each vital sign can be effectively aggregated according to the association degree with the reference sequence, so that the effectiveness and representativeness of the aligned features are improved.

[0022] Further, the present application provides a concise and effective reference sequence generation strategy, ensuring the consistency and stability of the time alignment reference point of the attention mechanism; avoiding the randomness error caused by manual selection of time points, making the model training more stable and the results more reproducible.

[0023] Further, the present application improves the model prediction accuracy by guiding the model prediction task through the classification loss; reduces the redundant information between features by explicitly constraining the independence of each modal feature representation through the orthogonal loss, effectively improving the model generalization ability and feature utilization efficiency; jointly optimizing the two losses ensures that the model has both accuracy and robustness.

[0024] Further, the present application explicitly determines how to realize the orthogonality constraint of modal features through specific matrix operations, making the method easy to understand and implement; the explicit orthogonality constraint between modalities greatly reduces the correlation between modalities, improves the complementarity of features between different modalities, and further enhances the prediction accuracy and generalization ability.

[0025] Further, the present application explicitly determines the structure of the GRU network, i.e., it includes multiple sub-networks, each of which receives aligned features of one modality and independently outputs hidden features; through the modal-specific GRU sub-network, the time-dependent patterns of different vital signs can be fully mined and expressed; the independent network structure between modalities makes the training process more flexible, effectively avoiding the mutual interference of cross-modal features, and improves the feature extraction efficiency and quality of each modality; the subsequent multi-modal fusion can also fully utilize the independent and rich expression of each vital sign, further enhancing the overall prediction performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a method flowchart of the present application; Figure 2 is a reference sequence attention diagram of a specific embodiment; Figure 3 is a prediction model architecture diagram of a specific embodiment; Figure 4 is a data sample collection scheme diagram of a specific embodiment. DETAILED DESCRIPTION

[0027] The present application will be further described in detail below in conjunction with the drawings and specific embodiments, so as to facilitate a clear understanding of the present application, but they do not constitute a limitation on the present application.

[0028] Example 1 As shown in Figure 1 , the present application provides a non-invasive intracranial pressure prediction method, comprising the following steps: S1, acquiring multi-modal vital sign raw signals from the subject in a non-invasive manner and embedding their timestamps to generate high-dimensional time-series features for each vital sign; S2, based on a pre-set reference sequence, using an attention mechanism to associate the time-series features with the raw signals to obtain aligned features for each vital sign; S3, inputting the aligned features into a gated recurrent unit network to extract hidden features for each vital sign; S4, concatenating and fusing the hidden features of each vital sign and sending them to a classifier to output a determination result of whether the subject has elevated intracranial pressure.

[0029] Specifically, in step S1, the multi-modal vital signs include bioelectrical impedance, electrocardiogram, heart rate, respiratory rate, pulse rate, body temperature, and / or blood oxygen saturation, which are acquired in the following ways: Bioelectrical impedance amplitude and phase: measuring the bioelectrical impedance values of the head tissue in a non-invasive manner through an electrode array attached to the patient's head, reflecting the volume changes of the brain tissue or cerebrospinal fluid; Electrocardiogram (ECG) signal: acquired using standard ECG lead method, reflecting the patient's cardiac electrophysiological activity; Heart rate (HR) data: extracted from ECG waveform or photoplethysmography (PPG) waveform, reflecting the patient's heart rate; Respiratory rate (RR) data: calculated by a chest and abdominal respiration sensor or by extracting the respiratory waveform from the ECG signal, reflecting the patient's respiratory rate; Pulse rate (PR) data: measured by photoplethysmography (PPG), fixing a photoelectric sensor (commonly found in finger clip or ear clip pulse oximeter) on the fingertip or earlobe, measuring the slight fluctuation of blood volume caused by heartbeat through infrared / visible light irradiation of subcutaneous blood vessels; Body temperature (Temp) data: continuously collected by attaching a body surface temperature sensor, reflecting the patient's metabolic state; Blood oxygen saturation (SpO2) data: real-time collected by a finger clip pulse oximeter, reflecting the patient's blood oxygen concentration and circulation status.

[0030] Due to the large differences in sampling frequency and measurement range of each modal vital sign signal, the following preprocessing steps are performed on the collected raw data: (1) Data quality inspection and denoising Performing preliminary quality inspection on the collected raw vital sign data, specific methods include but are not limited to: Outlier identification: checking the reasonableness of the value range of the original collected signal, removing obvious outliers caused by equipment failure or patient motion artifacts.

[0031] Filtering and noise reduction: For signals with high-frequency noise, such as bioimpedance and ECG signals, a Butterworth low-pass filter or other conventional filter is used to reduce noise. (2) Signal missing value processing For local data missing due to temporary sensor shedding or device abnormalities, the specific processing method includes: Short-time missing (e.g. less than 5 seconds) is filled by linear interpolation or cubic spline interpolation; Long-time missing (e.g. greater than 5 seconds) data segment is deleted or interpolated by the mean of the stable segments before and after it; (3) Multi-modal timestamp synchronization Due to the significant difference in sampling frequency of different vital signs, for example, ECG signal may be as high as 512 times per second, while blood pressure or body temperature data is only once per minute, so synchronization is needed.

[0032] (4) Data segmentation and windowing The synchronized data is segmented in a sliding window manner (e.g. each window length is 2 minutes, and the sliding step is 30 seconds) to facilitate continuous real-time prediction: In each window, the data sequence of each modality signal is used for subsequent high-dimensional time embedding and feature extraction; The sliding window method ensures continuous output of real-time intracranial pressure state prediction results and captures the dynamic change trend of the disease.

[0033] Specifically, in step S1, a time embedding method based on position encoding is used to convert the timestamp of each time series into a fixed-dimensional embedding vector. Given timestamp t and time embedding dimension d e , the time embedding function is represented as:

[0034]

[0035] where i is the index of the embedding vector, PE(t) is the position encoding matrix with dimension (N, T, d e ), N is the number of samples, T is the number of time steps, and freq is the frequency parameter. Through the above formula, each timestamp t is mapped to a fixed-dimensional embedding vector.

[0036] This embodiment can integrate time position information into the model in a continuous and learnable way, while avoiding the numerical range difference problem that may be caused by directly using timestamps. In addition, through the combination of sine and cosine functions, the model can capture periodic changes at different time scales, thereby better handling long-term dependencies in time series data.

[0037] Since different vital signs (such as heart rate, blood pressure, blood oxygen saturation, etc.) in medical time series data usually have different sampling frequencies and time dependencies. For example, electrocardiogram may sample 512 times per second, while blood pressure may sample once per minute. This inconsistent sampling frequency will cause the length and dimension of the time series to be mismatched.

[0038] Therefore, in step S2, the reference sequence attention mechanism is used to align the feature dimensions of the multi-modal signal, and the vital sign data with different sampling frequencies are mapped to the same feature dimension, and then in step S3, the time sequence features after the reference sequence attention alignment of each modality signal are extracted by the gated recurrent unit (GRU); the extracted features of the multi-modal vital sign signal are spliced and fused, and input into the prediction output layer to obtain the non-invasive intracranial pressure increase prediction result.

[0039] Specifically, as shown in Figure 2 In step S2, in order to realize the unified alignment of the multi-modal vital sign signal, the embodiment constructs a group of "reference sequences" in each sliding time window, and maps them to a feature vector recognizable by the network through a time embedding function. The reference sequence (r) is input to the ATT module as the query r=[r1, r2, …, r k ] input to the ATT module, and the multi-modal time sequence features (X1, X2, …, X N ) are used as the source of key and value: the embedded time sequence features of each modality are mapped to the key of the ATT module, and the original signal points are mapped to the value. The black arrows in the figure show the full connection mapping between the query and the sample points of each modality. The ATT output aligned features (A1, A2, …, A N ). The ATT module is followed by a Linear layer, and then a GRU encoder extracts the hidden representation.

[0040] The specific implementation steps are as follows: 1. Determination of normalized time interval: set the sliding time window length used for each prediction as w (for example, 2 minutes), and then determine the actual time interval corresponding to each window. Map the actual time interval linearly to the normalized interval [0, 1] in order to construct a unified reference independent of the sampling frequency.

[0041] 2. Uniformly sampled reference time points: pre-select the reference sequence length parameter L (for example, 16 or 24) in the normalized interval [0, 1], which represents the number of time steps of the reference sequence. According to the L equally spaced division principle, generate L relative time points in the [0, 1] interval: the first point corresponds to 0, the last Lth point corresponds to 1, and the remaining points are distributed according to equal intervals.

[0042] Example: taking L=24 as an example, the distance between adjacent two reference time points is 1 / 23.

[0043] 3. Time embedding mapping: the normalized scalar values of the L reference time points are sequentially input into a pre-designed time embedding function as "time stamps". The time embedding function can adopt a known sine-cosine position encoding, or a learnable fully connected mapping. Common practices include: Sine-cosine encoding: a plurality of sets of sine and cosine functions are calculated for the numerical value of each reference time point, respectively, and the results are spliced into a fixed-dimensional vector; Learnable mapping: the reference time scalar is mapped to a high-dimensional embedding vector through one or more fully connected networks.

[0044] 4. Reference sequence embedding matrix construction: the L high-dimensional vectors output by the time embedding function are further stacked to obtain the reference sequence embedding matrix. The reference sequence embedding will be used as "query" in the attention alignment step, enabling the network to dynamically allocate weights on the time embedding features of different modal signals.

[0045] The above process can avoid artificial interpolation noise, does not directly interpolate low or high frequency signals, and reduces information distortion or noise introduction; the reference sequence length L can be flexibly configured according to performance requirements to control the scale of attention calculation; after normalization, it is independent of the specific sampling frequency and is suitable for multi-modal signals of any sampling rate; through learnable or fixed encoding high-dimensional embedding, the attention mechanism can more accurately locate the corresponding relationship of different modal signals on the time axis.

[0046] The generation process of the aligned features in step S2 specifically includes the following steps: 1. Preparation of query, key, and value Query (Query): use the reference sequence embedding matrix (size LxD_e) constructed above as the query set, and each row corresponds to a query vector of a reference time step; Key (Key): the time series feature sequence (length L, dimension D_e) obtained by time embedding of each modal vital sign signal in step S1 is used as the key set, which has the same row and column number as the query matrix; Value (Value): the original vital sign measurement value sequence (such as bioelectrical impedance amplitude, heart rate value, etc.) corresponding to each time step is collected into a value matrix (size LxD_v, where D_v is the channel number of the original signal value) along the time step.

[0047] 2. Similarity weight calculation By using the fixed reference sequence as the query (Query), the similarity between the time series and the reference sequence is directly calculated, thereby avoiding the information loss or noise introduction problem caused by interpolation or downsampling, and realizing the matching of feature dimensions by mapping the time series of different sampling frequencies to the dimension of the reference sequence.

[0048] For a given reference sequence As query and timestamp As key, respectively, time embedding is obtained And The calculation formula of attention output is:

[0049] Where is the weight vector of query, is the weight vector of key, is the time series observation value, and the attention score is normalized to a probability distribution by the softmax function. Finally, the attention weight is generated by weighted sum of attention score and time series value Aligning the vital sign time series of different sampling frequencies to the same dimension.

[0050] Although the traditional recurrent neural network (RNN) and its variants (such as LSTM and GRU) perform well in capturing time dependencies, their local window mechanism is difficult to capture global dependencies. In addition, there may be redundancy or correlation between different dimensions of multi-dimensional time series, and direct use of RNN may lead to model overfitting or performance degradation. In order to solve these problems, the embodiment introduces attention mechanism to dynamically select important time points, and combines GRU to capture local time patterns, thereby realizing efficient feature extraction of time series data.

[0051] In this embodiment, the scores of each row in the similarity score matrix are normalized by the softmax operation to obtain the attention weight matrix A, ensuring that the weight sum of each query on all keys corresponding to the row is 1.

[0052] 3. Weighted aggregation to generate aligned features Weighted sum of value matrix using attention weight matrix A: that is, multiplying A with value matrix to generate aligned feature matrix X; each row in the aligned feature matrix is the aligned feature fused from the original signal value according to the query-key matching relationship at the reference time step.

[0053] Specifically, in step S3, the following steps are specifically included: 1. In order to adapt the aligned features generated by attention alignment to the input dimension of GRU, the aligned feature matrix X of attention output is mapped to the embedding dimension So as to be used as GRU input subsequently. This mapping unifies the feature dimension of each time step to Dh.

[0054] 2. For each vital sign modality, an independent GRU subnetwork is constructed. Each subnetwork receives the mapped time-series input sequence and updates its hidden state in time steps: The GRU-based feature extraction formula is as follows:

[0055] wherein, is the hidden state calculated by GRU.

[0056] 3. After the end of the last time step of the sequence, the final hidden state of the modality GRU is extracted as the high-dimensional time-series feature representation of the modality in the entire time window. Each modality corresponds to output a Dh-dimensional hidden vector.

[0057] Specifically, in step S4, for N vital sign signals, each signal is obtained by -GRU feature extraction to obtain the corresponding feature representation In order to effectively fuse these features in multi-modal learning, the embodiment adopts the feature concatenation method to fuse the features of different modalities into a more expressive feature set. As shown in the following formula:

[0058] In this embodiment, the input of the classifier is the feature vector of the fused multi-modal vital signs After the vector is processed by several fully connected layers and activation layers, a binary classification prediction value is output, which is used to indicate whether intracranial hypertension (ICH) occurs at time step t.

[0059] Specifically, the prediction module is composed of three layers of fully connected layers (Linear) and two segments of ReLU activation (ReLU) alternately, and the mathematical expression is as follows:

[0060] wherein, Linear() represents a fully connected layer; ReLU() represents a Rectified Linear Unit activation function; the output of each layer is first subjected to ReLU activation to introduce nonlinearity, so that the model can fit more complex feature relationships; and the final output In the binary classification scenario, 1 indicates that ICH occurs, and the intracranial pressure is elevated; otherwise, 0 indicates normal.

[0061] Embodiment 2 As shown in Figure 3 The present application provides a non-invasive intracranial pressure prediction system for implementing the method described in the above technical solutions, comprising: The signal acquisition module is configured to acquire multi-modal vital sign original signals from the subject in a non-invasive manner, and embed time stamps of the original signals to generate high-dimensional time sequence features of each vital sign. The signal alignment module is configured to associate the time sequence features with the original signals based on a preset reference sequence by using an attention mechanism to obtain aligned features of each vital sign. The feature extraction module is configured to input the aligned features into a gated recurrent unit network to extract hidden features of each vital sign. The intracranial pressure determination module is configured to splice and fuse the hidden features of the vital signs, and input the hidden features into a classifier to output a determination result of whether the subject has an elevated intracranial pressure.

[0062] In this embodiment, the signal acquisition module processes the original time sequence by using a time window size w=2 (minutes) to capture local information of the historical sequence , that is, the historical 2-minute data of the multi-modal vital signs are used to predict the NICP of the next minute. The time stamp of any time sequence is processed by a sliding window and time embedding to obtain a vector as , and the time embedding dimension .

[0063] In the signal alignment module, a reference sequence is generated by uniform sampling in the interval as , and time embedding is performed on the reference sequence to obtain , where is the length of the reference sequence. The time embedding vector, the reference sequence, and the time sequence value are input into the attention mechanism part of the model to output attention weights . The value matrix is weighted and summed using the attention weight matrix A: that is, A is multiplied with the value matrix to generate an aligned feature matrix.

[0064] The feature extraction module inputs the aligned sequence features into a gated recurrent unit (GRU) network to extract hidden feature representations of each modality. The GRU network is designed to have an independent subnetwork for each modality to fully exploit the time sequence dependency of each modality signal.

[0065] The intracranial pressure determination module inputs the hidden features of each modality after splicing and fusion into a classification layer to obtain a real-time determination result of whether there is an elevated intracranial pressure: the determination result output by the classifier can be divided into “ICP elevated (such as ≥15 mmHg)” and “ICP normal”; The prediction result can be continuously and real-time output to facilitate quick response and decision intervention of clinical medical staff. The model parameter configuration of this embodiment is shown in Table 1.

[0066] Table 1 Model and training parameter configuration

[0067] To more reasonably evaluate the generalization ability of the model in the time series prediction task, the training process introduces a time series cross-validation strategy. Specifically, the samples are divided in a forward chaining manner, that is, in each fold, the data of the earlier time period is used as the training set, and the data of the subsequent time period is used as the validation set, ensuring that the training and validation data strictly follow the time sequence, so as to be more close to the prediction process in the real scene and avoid future information leakage. In addition, cross-validation also effectively improves the adaptability of the model to different time period data and improves the reliability and stability of the overall evaluation results.

[0068] Specifically, the training process of the system in the embodiment needs to include the following steps: 1. Data set construction and division As shown in Figure 4 , the synchronous monitoring data of multiple patients with craniocerebral injury in the ICU of a certain third-grade hospital is collected, including: heart rate (HR), respiratory rate (RR), oxygen saturation (SpO2), pulse rate (PR), and bioelectrical impedance amplitude and phase parameters, as well as simultaneous invasive intracranial pressure (ICP) measurement values.

[0069] The original signal is denoised, outliers are removed, and missing values are interpolated (linear interpolation for short missing, or mean filling for long missing).

[0070] The invasive ICP measurement value is classified according to the threshold of 15mmHg: ≥15mmHg is labeled as ICH class (label=1), and <15mmHg is labeled as normal class (label=0).

[0071] The preprocessed time series segments with labels are divided into training set and test set according to time sequence: Training set: 2023-09-20 16:56:00 to 2023-09-21 23:59:00; Test set: 2023-09-22 00:00:00 to 2023-09-22 07:56:00; The multi-modal time series data is segmented using a sliding window (window length w=2 minutes, step s=30 seconds), and each window generates a training sample.

[0072] 2. Model structure initialization Signal alignment module: preset reference sequence length L=24, uniformly sample 24 time points in the interval [0,1] and map them into embedding vectors (position encoding). Initialize attention layer parameters: query, key, value linear mapping matrix and bias.

[0073] Feature extraction module: for each modality, construct a separate GRU subnetwork with hidden unit number h=64. Initialize GRU weight parameters (update gate, reset gate and candidate activation matrix).

[0074] Intracranial pressure determination module: after feature concatenation, input into a three-layer fully connected layer, with hidden layer neuron numbers of 128 and 64 in turn; each layer is followed by ReLU activation. The output layer uses Sigmoid activation to generate a binary classification probability.

[0075] 3. Training process Adam optimizer is used with learning rate lr=1e-5 and L2 regularization weight-decay=1e-4.

[0076] Iterate the training set in batches of size M=32 for multiple rounds, and traverse all samples in each round.

[0077] At each training step: Forward propagation: perform attention alignment→GRU extraction→fusion classification, and calculate the predicted output.

[0078] Calculate the classification loss and orthogonal loss to form the joint loss.

[0079] In multi-modal vital sign data, the orthogonal loss can effectively control the correlation between different physiological signals, ensuring that the model extracts unique features from each signal, thereby improving prediction accuracy. To reduce feature redundancy and improve the generalization ability of the model, the model introduces an orthogonal loss function. The orthogonal loss function constrains the orthogonality between feature representations, enabling the model to learn more independent and meaningful feature representations.

[0080] For each pair of feature representations , calculate their Gram matrix: ; Then, calculate the difference between the Gram matrix and the identity matrix , to get the orthogonal loss:

[0081] The model optimizes the prediction task loss and the orthogonal loss simultaneously to form a joint loss function. The prediction task loss usually uses cross-entropy to measure the difference between the predicted results and the true labels, while the orthogonal loss is introduced as a regularization term, aiming to minimize both the prediction task loss and the orthogonal loss. The joint loss function can be represented as:

[0082] where, is the prediction task loss (e.g., cross-entropy), (ofspenalty) is a hyper-parameter that controls the weight of the orthogonality loss in the total loss. In this way, the model not only optimizes the prediction accuracy, but also constrains the correlation between features through the orthogonality loss, reduces redundancy, and improves prediction performance.

[0083] Backpropagation: Calculate the gradient based on the joint loss and update all trainable parameters.

[0084] Continue training until the training set loss no longer decreases significantly or the maximum number of iterations (e.g., 100 epochs) is reached.

[0085] Table 2 shows the results of intracranial pressure (ICP) prediction in this embodiment: Table 2 NICP model prediction experiment results

[0086] Based on the experimental data in Table 2, the prediction model of the present application based on the fusion of multi-modal vital signs shows high precision (Precision) and recall (Recall) on both the training set and the test set, effectively identifying normal ICP. On the training set, the precision is 98.57%, the recall is 98.55%, and the F1 value reaches 0.98; while the precision of the test set is 83.55%, the recall is 85.68%, and the F1 value is 0.84. This result shows that the model can better capture data features in the training phase and has certain generalization ability on the test set.

[0087] However, the precision and F1 value of the test set are lower than those of the training set, and this difference may be due to the way the data set is divided and the development stage of the patient's condition. First, since the test set and the training set are divided in chronological order, the test set may correspond to the stage after the patient has been treated. At this stage, the patient's condition may have changed significantly. In this case, the patient's ICP may tend to be normal or already normal. Therefore, in the training phase, the model is exposed to a higher proportion of ICH labels, resulting in the model learning more features related to ICH. However, at the test stage, due to the decrease in the proportion of ICH labels in the test set, the model's performance in identifying ICH decreases. In addition, this performance difference may also be related to changes in other vital signs.

[0088] As the disease progresses, the patient's heart rate (HR), respiratory rate (RR), oxygen saturation (SpO2), and pulse rate (PR) may change significantly, which may be associated with changes in ICP, thereby affecting the model's ability to identify ICH.

[0089] To verify the effectiveness of the non-invasive intracranial pressure prediction proposed in the present application, two classic machine learning models and one deep learning model are selected to construct a baseline for predicting increased intracranial pressure, and the performance of the NICP prediction method proposed in the present embodiment is compared. In the data preprocessing stage, all comparison models use a unified downsampling strategy to process multi-modal vital sign data: feature engineering is performed by extracting time domain statistics (including maximum, minimum, mean, median, variance, etc.) to align the input feature dimension. The NICP model does not need to go through this step and can directly input the original data sequence in the time window and use the reference sequence attention mechanism to complete the feature dimension alignment.

[0090] The baseline model and feature engineering used in the present embodiment are as follows: XGBoost (XGB): an optimized implementation of the gradient boosting decision tree framework, which improves performance through second-order gradient approximation and regularization strategies. It performs well in handling high-dimensional sparse features and is suitable for complex feature interaction modeling of medical data.

[0091] RandomForest (RF): an ensemble learning method that builds a group of decorrelated decision trees through Bootstrap sampling. It is robust to noisy data but has the risk of bias in feature importance evaluation.

[0092] LSTM: a variant of gated recurrent neural network that preserves long-term temporal dependencies through cell state. It is good at capturing continuous change patterns of vital signs.

[0093] For all baseline models, a downsampling strategy is used in the preprocessing stage: a. downsample the multi-modal raw signals in each sliding window; b. extract time domain statistical features (maximum, minimum, mean, median, variance, etc.) to construct a fixed-dimensional feature vector; c. input the feature vector to XGB, RF, and LSTM.

[0094] The NICP model does not require this feature engineering and directly inputs the original time series data and reference sequence attention alignment mechanism to complete feature alignment.

[0095] The comparison results are shown in Table 3 (the best value is bold and the suboptimal value is underlined). The overall performance evaluation includes accuracy (Acc), precision (Prec), recall (Rec), and F1 score.

[0096] Table 3 Comparison of model performance

[0097] In the traditional machine learning model, the RF model performs best on the training set, with Prec (98.32%), Rec (97.96%), and F1 value (0.98) being the highest (tied with the NICP model), showing a high fitting to the training data. However, on the test set, its Prec and Rec dropped to 81.64% and 80.42%, respectively, with an F1 value of 0.81, indicating limited generalization ability, but still better than XGB and LSTM models. XGB has a Prec and Rec of 95.52% and 95.65% on the training set, with an F1 value of 0.95, showing strong fitting ability. However, on the test set, the performance significantly decreased, with Prec dropping to 81.66% and Rec falling to 76.84%, and F1 value only 0.79, reflecting its obvious overfitting problem.

[0098] In terms of deep learning models, LSTM achieved a good performance on the training set with Prec 92.97%, Rec 93.61%, and F1 0.93, but its performance on the test set dropped, with Prec 81.41%, Rec 73.47%, and F1 0.77, showing a certain degree of overfitting. In contrast, the NICP model continued to maintain top-level performance on the training set (Prec 98.57%, Rec 98.55%, F1 0.98), and its Prec (83.55%) and Rec (85.68%) on the test set were also among the top, with an F1 value of 0.84. The overall performance was stable, with strong generalization ability, and its performance was close to or better than the Transformer model in multiple indicators.

[0099] Conclusion: The experimental results show that traditional machine learning models have strong fitting ability on the training set, but generally have weak generalization ability. Deep learning models, especially Transformer, show stronger generalization ability and application potential in handling complex feature relationships and long sequence data. The NICP model maintains high training performance while having excellent test performance, demonstrating good balance between precision and recall, and is one of the models with the best overall performance.

[0100] To verify the effectiveness of each module of the model, this embodiment conducts ablation experiments on the system to explore the importance of each component and feature in the multi-modal vital sign prediction model. The experiment covers three dimensions of model structure, regularization strategy, and feature selection, and the experimental results are shown in Table 4.

[0101] Table 4 Comparison of Ablation Experiment Results

[0102] The original model (i.e., the present embodiment) performs best on the training set, with a Precision (Prec) of 98.57%, a Recall (Rec) of 98.55%, and an F1 score of 0.98, and on the test set, a Prec of 83.55%, a Rec of 85.68%, and an F1 score of 0.84, with balanced overall performance.

[0103] After removing the GRU module, the performance on the training set decreases slightly, with an F1 score of 0.96, while the Prec on the test set increases slightly (83.69%) but the F1 score decreases to 0.83, indicating that the GRU module has a certain effect on improving the overall generalization ability. After removing the orthogonal loss, the F1 score on the training set decreases to 0.97, and the performance on the test set decreases slightly, indicating that the orthogonal loss helps to decouple the representation and improve the generalization ability during the training process. After replacing the GRU with an LSTM, the F1 score on the training set remains at 0.98, but the F1 score on the test set decreases to 0.80, indicating that the LSTM has the ability to model long sequences, but its performance is not as good as the GRU in this task, possibly because it has more parameters, leading to insufficient generalization performance.

[0104] Removing the bioimpedance feature causes the F1 score on the training set to decrease significantly to 0.93, and the F1 score on the test set also decreases to 0.81, indicating that this feature has a significant positive effect on the overall performance of the model, especially on the test set, where its contribution is particularly prominent.

[0105] After removing the heart rate (HR) feature, the F1 score on the training set remains 0.98, but the F1 score on the test set decreases to 0.82, indicating that this feature may introduce redundant or interfering information in the training set, but still has certain discriminative ability for the test set.

[0106] Removing the respiratory rate (RR) does not significantly affect the performance on the training set (F1 remains 0.98), but the F1 score on the test set decreases to 0.80, indicating that this feature has a certain effect on the test set.

[0107] After removing PR, the performance on the training set remains unchanged (F1 is 0.98), but the F1 score on the test set decreases to 0.83, further verifying its help for prediction accuracy in the test phase. After removing the SPO2 feature, the F1 score on the training set decreases to 0.97, and the F1 score on the test set also decreases to 0.81, indicating that it provides effective information for the training set, but its support for the test set is limited.

[0108] In summary, the model is significantly sensitive to different structures and features, especially the removal of bioimpedance and PR features, which causes a significant decrease in the F1 score on the test set to 0.81 and 0.83, respectively, indicating that they have a greater contribution to the generalization performance of the model.

[0109] Embodiment 3 The application further provides a computer readable storage medium, which stores a computer program, and the computer program enables a computer to implement the noninvasive intracranial pressure prediction method in the above technical solution when executed by a processor.

[0110] Embodiment 4 The application provides an electronic device, comprising a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the noninvasive intracranial pressure prediction method in the above technical solution.

[0111] Those skilled in the art should understand that the embodiments of the application can be provided as a method, a system, or a computer program product. Therefore, the application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the application can adopt a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0112] The application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices generate a device for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0113] These computer program instructions can also be stored in a computer readable memory that can guide the computer or other programmable data processing devices to work in a specific way, so that the instructions stored in the computer readable memory produce a product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0114] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1steps of the functions specified in the block or blocks. Figure 1 steps of the functions specified in the block or blocks.

[0115] That which is not described in detail in this specification is well known to those skilled in the art.

Claims

1. A non-invasive intracranial pressure prediction method, characterized by: The method comprises the following steps: obtaining multi-modal vital sign raw signals from a subject in a non-invasive manner, and embedding time stamps thereof to generate high-dimensional time sequence features of each vital sign; based on a preset reference sequence, using an attention mechanism to associate the time sequence features with the raw signals to obtain aligned features of each vital sign; inputting the aligned features into a gated recurrent unit network to extract hidden features of each vital sign; concatenating and fusing the hidden features of each vital sign, and sending them into a classifier to output a determination result of whether the subject has intracranial pressure rise.

2. The method of claim 1, wherein: The multi-modal vital signs include bioelectrical impedance, electrocardiogram, heart rate, respiratory rate, pulse rate, body temperature and / or blood oxygen saturation.

3. The method of claim 1, wherein: The attention mechanism is: calculating the similarity weight between the preset reference sequence and the time sequence features of each modality through the attention mechanism, and weighting and aggregating the original vital sign signals according to the obtained similarity weight to generate the aligned features.

4. The method of claim 3, wherein The attention mechanism is: embedding the preset reference sequence as a query; embedding the time sequence features of any vital sign as a key; embedding the original signal of the vital sign at the corresponding time stamp as a value; using the attention mechanism to weight and aggregate the values according to the association score between the query and the key to obtain the aligned features of the vital sign.

5. The method of claim 1, wherein: The reference sequence is obtained by mapping time points uniformly sampled in a normalized time interval [0, 1] according to a preset reference sequence length through a time embedding function.

6. The method of claim 1, wherein: The attention mechanism, the gated recurrent unit network and the classifier all contain trainable parameters; The method further comprises: based on the multi-modal vital sign signals of a plurality of subjects as input, and whether the invasive intracranial pressure measurement value at the corresponding moment exceeds the threshold as a label, training the trainable parameters by jointly minimizing the following two losses: classification loss: used to measure the difference between the prediction result output by the classifier and the invasive intracranial pressure measurement value label; orthogonal loss: used to constrain the orthogonality between the hidden features of each modality extracted by the gated recurrent unit network.

7. The method of claim 6, wherein: In the training stage, for the hidden feature representations of any two modalities extracted by the gated recurrent unit network, first calculate the pairwise inner product of all elements between them and form an inner product matrix; square the difference between the inner product matrix and the unit matrix of the same dimension at the corresponding positions and accumulate to obtain the orthogonal loss value of the pair of modalities; sum the orthogonal loss values of all pairs of modalities to obtain the overall orthogonal loss; the overall orthogonal loss and the classification loss jointly constitute a joint loss function for optimizing the trainable parameters in the attention mechanism, the gated recurrent unit network and the classifier.

8. The method of claim 6, wherein: The gated recurrent unit network is composed of a plurality of sub-networks, each of which receives the aligned features of one modality and outputs the hidden features of the modality.

9. A non-invasive intracranial pressure prediction system, characterized by: For implementing the method of any one of claims 1-8, comprising: a signal acquisition module for obtaining multi-modal vital sign raw signals from a subject in a non-invasive manner, and embedding time stamps thereof to generate high-dimensional time sequence features of each vital sign; a signal alignment module for using an attention mechanism to associate the time sequence features with the raw signals based on a preset reference sequence to obtain aligned features of each vital sign; The feature extraction module is configured to input the aligned features into a gated recurrent unit network to extract hidden features of each vital sign; The intracranial pressure determination module is configured to splice and fuse the hidden features of the vital signs and send them to a classifier to output a determination result of whether the subject has elevated intracranial pressure.

10. A computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, causes the processor to perform: The computer is caused to implement the noninvasive intracranial pressure prediction method of any one of claims 1-8.

Images