A medical monitoring-oriented physiological data prediction system construction method and system

Abstract

The application provides a physiological data prediction system construction method and system for medical care, which is used for predicting future physiological data of a patient according to historical physiological data of the patient. The system introduces a manifold distortion field composed of step distortion and variable distortion, and uses the manifold distortion field to guide the embedding coding and weighted aggregation process, effectively avoiding the negative influence of the manifold distortion problem caused by the missing and asynchronous sampling of the historical physiological data of the patient in the medical care scene on the prediction result. The technical scheme of the application does not need to interpolate and forcibly align the historical physiological data of the patient, so that the real information expressed by the historical physiological data of the patient is retained, the accurate prediction of the future physiological index value of the patient is realized, and the probability of the medical staff misjudging the physiological state of the patient is reduced.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare, specifically to the field of human physiological data prediction, and more specifically, to a method and system for constructing a physiological data prediction system for medical monitoring. Background Technology

[0002] In smart healthcare scenarios such as intensive care units (ICUs), patient vital sign monitoring (e.g., continuous heart rate and blood pressure monitoring) and clinical laboratory data (e.g., intermittent blood biochemical indicators) constitute typical irregular multivariate time series (IMTS). In real-world medical environments, due to differences in sampling frequencies among different medical devices, occasional detachment of wearable sensors, and event-driven blood testing mechanisms, the recorded patient physiological data often exhibits extremely high missing rates. For example, in typical ICU datasets, the combined missing rate of vital signs and laboratory indicators typically exceeds 85%, sometimes reaching as high as 96%. This extreme irregularity manifests as "structural missingness" of the same physiological indicator along the time axis and "cross-variable asynchrony" between different medical devices, posing significant modeling challenges for predicting patient vital signs, leading to prediction bias and misjudgments of patient physiological states by medical staff. To address this irregularity in clinical data, existing research and mainstream technologies almost entirely rely on the Temporal Representation Construction (TRC) paradigm. The core idea of ​​this paradigm is to force irregular clinical observations to be mapped and aligned onto a unified time grid to adapt to standard deep learning architectures. Currently, existing TRC technologies mainly fall into the following two categories:

[0003] The first category is based on direct interpolation and imputation strategies. These schemes attempt to bridge sparse vital sign observations through zero-filling or continuous interpolation. However, in the extremely sparse conditions of the ICU setting, direct interpolation inevitably introduces a large number of artificial artifacts. For example, directly filling "blood pressure loss" caused by equipment detachment with 0 or attenuating it to the mean can cause the vital sign prediction model to fundamentally misjudge the patient's condition, incorrectly treating "high-risk signal loss" as "safe physiological state".

[0004] The second category is alignment and aggregation strategies. These approaches attempt to force the alignment of irregular laboratory and monitoring time steps through time window aggregation. However, when dealing with severe asynchrony across devices (such as heart rate sampling every second while blood is drawn once a day), rigid time aggregation cannot correctly match the true physiological correspondence between asynchronous signs, thus forcibly stuffing disparate clinical events into incompatible semantic units, effectively severing the temporal correlation within the patient's true physiological system.

[0005] From the theoretical perspective of the manifold hypothesis, the true physiological evolution trajectory of a patient is a potentially continuous manifold. However, sensor detachment and irregular blood draws disrupt this temporal uniformity, distorting the geometric structure of physiological data and producing "Series-Level Manifold Distortion." Traditional TRC techniques not only fail to address this distortion but also force irregular, asynchronous physiological indicator samples onto a uniform time grid through interpolation or alignment strategies. This involves simply filling in "missed" data and forcibly aligning time points, thereby disrupting the inherent semantic consistency of the physiological indicator data, exacerbating the distortion, and leading to severe biases in predicting future physiological indicators.

[0006] In summary, existing physiological data prediction schemes based on the TRC paradigm distort the true information expressed by patients' historical physiological data during the prediction process, resulting in inaccurate predictions of patients' future physiological indicators and causing medical staff to misjudge patients' physiological status.

[0007] It should be noted that the background information presented here is only for illustrating relevant information about the present invention to aid in understanding the technical solution of the present invention, and does not imply that the relevant information is necessarily prior art. The relevant information was submitted and disclosed together with the present invention, and should not be considered prior art unless there is evidence that the relevant information was disclosed before the filing date of the present invention. Summary of the Invention

[0008] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for constructing a physiological data prediction system for medical monitoring.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] According to a first aspect of the present invention, a method for constructing a physiological data prediction system for medical monitoring is provided. The system is used to predict a patient's future physiological data based on the patient's historical physiological data. The historical physiological data records multiple physiological indicators of the patient, each corresponding to parameter values ​​that are non-uniformly distributed over time. The construction method includes:

[0011] T1. Construct an initial system, the initial system comprising:

[0012] An initial step-level distortion-aware embedder is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in the historical physiological data are encoded separately, and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period.

[0013] The initial spatiotemporal encoder is used to: extract the corresponding temporal enhancement features for each distortion-sensing embedding, wherein each temporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological indicator and all other physiological indicators in the corresponding time period.

[0014] An initial variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Here, interval irregularity describes the degree of uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuation describes the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features.

[0015] An initial prediction decoder is used to predict a patient's future physiological data based on all variable-level features.

[0016] T2. Obtain the training set, which includes multiple training samples. Each training sample includes the patient's historical physiological data and future physiological data labels. The future physiological data labels include multiple physiological indicators and the actual parameter values ​​of each physiological indicator at multiple future prediction points.

[0017] T3. Using the patient's historical physiological data as input and the predicted parameter values ​​of each type of physiological indicator at multiple future prediction points as output, the initial system is trained to convergence using a training set to obtain a physiological data prediction system for medical monitoring; wherein, the training process updates the system parameters by minimizing the preset loss.

[0018] This scheme can achieve at least the following beneficial technical effects: it enables the constructed system to measure the integrity of global physiological data records through coverage collapse and time interval anomalies, and to capture the reliability of local data records through the irregularity of time intervals and fluctuations in local continuity. This allows the system to directly recover the true physiological evolution semantics of patients from their highly distorted and incomplete original historical physiological data, thus breaking away from the "forced filling alignment" mechanism of traditional schemes and avoiding its disruption of the temporal uniformity of manifold trajectories, distortion of potential geometric structures, and destruction of the inherent semantic consistency of time series, making the prediction results more accurate.

[0019] Optionally, in the initial step-level distortion-aware embedder, the predetermined step-level distortion calculation method is as follows:

[0020]

[0021]

[0022] in, Representing the Coverage collapse during a specific time period; The total number of categories representing physiological indicators; Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? The parameter values ​​of physiological indicators, when recorded. It is 1 if it is true, otherwise it is 0. Representing the The time intervals of these time periods are abnormal; Representing the A time period; Representing the The data centroid for a time period is the average value of all parameter values ​​of all categories of physiological indicators at any given time within that time period. Representing the The centroid of the data for a given time period; It represents a very small positive number.

[0023] Optionally, the initial step-level distortion-aware embedder encodes all classes of physiological indicators in the historical physiological data separately and fuses the collapse and time interval anomalies to obtain the distortion-aware embedding of each class of physiological indicator in each time period in the following manner:

[0024]

[0025] in, Representing the Physiological indicators in the first Distortion perception embedding within a time period; This represents a parameter value embedder used to map physiological indicator parameter values ​​to high-dimensional vectors; Representing the Physiological indicators in the first Parameter values ​​within a time period; Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? The parameter values ​​of physiological indicators, when recorded. It is 1 if it is true, otherwise it is 0. The time embedder is used to map the time corresponding to the physiological index parameter values ​​into a high-dimensional vector. Representing the Within the time period, the first The time corresponding to the parameter values ​​of physiological indicators; Represents the first linear transformation The step-level distortion scalar for each time period is obtained as follows:

[0026]

[0027] in, This represents the weight matrix in a linear transformation operation. This represents the bias vector in a linear transformation operation. Representing the The step-level distortion scalar for each time period is obtained as follows:

[0028]

[0029] in, , These represent different balance coefficients; Representing the Coverage collapse during a specific time period; Representing the The time intervals for each period are abnormal.

[0030] Optionally, the initial spatiotemporal encoder extracts the corresponding temporal enhancement features for each distortion-aware embedding in the following manner:

[0031]

[0032] in, Representing the Physiological indicators in the first Temporal enhancement features within a time period; represent Encoder; The total number of segments representing the time period; Representing the Physiological indicators in the first The refined distortion-aware embedding for each time period is obtained as follows:

[0033]

[0034] in, Represents the ReLU function activator. The weight matrix representing the first fully connected layer of the feature refinement tool. The weight matrix of the second fully connected layer representing the feature refinement tool. Representing the Physiological indicators in the first Distortion-aware embedding within a time period.

[0035] Optionally, the initial spatiotemporal encoder extracts the spatiotemporal enhancement features corresponding to each temporal enhancement feature in the following manner:

[0036]

[0037] in, Representing the Physiological indicators in the first Spatiotemporal enhancement features within a time period This represents a graph neural network with an adaptive adjacency matrix. Representing the Physiological indicators in the first Temporal enhancement features within a time period The total number of categories representing physiological indicators.

[0038] Optionally, in the initial variable-level distortion weighted aggregator, the predetermined variable-level distortion calculation method is as follows:

[0039]

[0040]

[0041] in, Representing the Irregularity of intervals for physiological indicators; Represents a record index based on physiological indicator parameter values. The standard deviation is calculated based on the dimensions. Representing the Physiological indicators The time corresponding to the parameter value of each record; Representing the Physiological indicators The time corresponding to the parameter value of each record; Represents the smallest positive number; Representing the Physiological indicators in the first Local continuous fluctuations within a time period; Represents a radius of The standard deviation is calculated within a local time window, where the radius is... A local time window is defined as a time period centered on a point in time, extending forward and backward along the time axis. Formed over a period of time; Representative of the first The records within a local time window centered on the [number]th time period The number of parameter values ​​for physiological indicators.

[0042] Optionally, in the initial variable-level distortion weighted aggregator, the predetermined weighted aggregation method is:

[0043]

[0044] in, Representing the Variable-level characteristics of physiological indicators; This represents a graph neural network with an adaptive adjacency matrix; The total number of categories representing physiological indicators; Representing the Intermediate variable-level features of physiological indicators are obtained in the following way:

[0045]

[0046] in, The total number of time periods; Representing the Physiological indicators in the first Spatiotemporal enhancement features within a time period; Representing the Physiological indicators in the first The normalized aggregate weights corresponding to each time period are obtained as follows:

[0047]

[0048] in, Representing the Physiological indicators in the first Aggregate weights corresponding to each time period; Representing the Physiological indicators in the first The aggregate weights corresponding to each time period are obtained as follows:

[0049]

[0050] in, Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? The parameter values ​​of physiological indicators, when recorded. It is 1 if it is true, otherwise it is 0. Representing the Physiological indicators in the first The variable-level distortion weights corresponding to each time period are obtained as follows:

[0051]

[0052] in, Representing the Irregularity of intervals for physiological indicators; Representing the Physiological indicators in the first Local continuous fluctuations within a time period; , These represent different balance coefficients.

[0053] Optionally, the initial prediction decoder predicts the patient's future physiological data based on all variable-level features in the following manner:

[0054]

[0055] in, Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point Represents a multilayer perceptron. Representing the Variable-level characteristics of physiological indicators Representing the The first of the physiological indicators The time corresponding to each prediction point.

[0056] Optionally, in T3, the preset loss is:

[0057]

[0058] in, Represents the pre-defined loss. The total number of categories representing physiological indicators. Represents the total number of prediction points. Represents the index of the predicted point. Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point Representing the The first of the physiological indicators The actual parameter values ​​corresponding to each prediction point.

[0059] This scheme can achieve at least the following beneficial technical effects: by constructing a loss function using the mean squared error method, it can better match the characteristics of irregular distribution of historical physiological data and missing or sparse records of some physiological indicators, thereby reducing the prediction error of the final system.

[0060] According to a second aspect of the present invention, a physiological data prediction system for medical monitoring is provided, constructed based on the method described in the first aspect of the present invention, for predicting the future physiological data of a patient based on the patient's historical physiological data, wherein the historical physiological data records multiple physiological indicators of the patient, each physiological indicator corresponding to parameter values ​​that are non-uniformly distributed on a time axis, and the system includes:

[0061] A step-level distortion-aware embedding is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in historical physiological data are encoded separately and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period.

[0062] The spatiotemporal encoder is used to: extract the corresponding temporal enhancement features of each distortion sensing embedding, wherein each temporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological indicator and all other physiological indicators in the corresponding time period.

[0063] A variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Interval irregularity describes the uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuation describes the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features.

[0064] A predictive decoder used to predict a patient's future physiological data based on all variable-level features.

[0065] According to a third aspect of the present invention, a method for predicting physiological data for medical monitoring is provided, the method comprising: S1, acquiring historical physiological data of a patient; S2, inputting the patient's historical physiological data into a physiological data prediction system for medical monitoring as described in the second aspect of the present invention for processing, to obtain predicted future physiological data of the patient.

[0066] Compared with existing technologies, the advantages of this invention are as follows: By introducing a manifold distortion field composed of step-level and variable-level distortions, and using the manifold distortion field to guide the embedding encoding and weighted aggregation process, the negative impact of manifold distortion caused by missing records and asynchronous sampling of patient historical physiological data in medical monitoring scenarios on prediction results is effectively avoided. This scheme eliminates the need for interpolation and forced alignment, thus preserving the true information expressed by the patient's historical physiological data, enabling accurate prediction of future physiological indicators and reducing the probability of medical staff misjudging the patient's physiological state. Attached Figure Description

[0067] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0068] Figure 1 This is a schematic diagram of the initial system architecture of a method for constructing a physiological data prediction system for medical monitoring according to an embodiment of the present invention;

[0069] Figure 2This is a schematic diagram of the ManiDF framework upon which the method for constructing a physiological data prediction system for medical monitoring is based according to an embodiment of the present invention.

[0070] Figure 3 This is a schematic diagram comparing the predicted results of future physiological data from the comparative experiment according to the present invention;

[0071] Figure 4 This is a schematic diagram comparing the prediction results of future meteorological data from the comparative experiment based on the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0073] As mentioned in the background section, existing physiological data prediction schemes based on the TRC paradigm distort the true information of patients' historical physiological data during the prediction process, resulting in inaccurate predictions of future physiological indicators and causing medical staff to misjudge patients' physiological status.

[0074] When researching the aforementioned problems, the inventors discovered that as the maximum sampling interval of patients' physiological indicators increases, rigid grid-based forced alignment inevitably leads to metric distortion, causing the collapse of the manifold's geometric structure. Taking the APN model (an advanced predictive model within the TRC paradigm) as an example, when processing extremely sparse and irregularly sampled data, the learned feature representations fail to maintain the original geometric structure of the data: in visualizations, the distribution tends to cluster towards the center or become completely scattered, losing its original clear spatial structural integrity, resulting in data collapse. The reason for this data collapse is that models like APN employ a rigid grid-based forced alignment strategy, forcibly aligning and splicing points with extremely irregular time intervals. Theoretical analysis shows that as the actual sampling interval of physiological indicators increases, this forced alignment method inevitably leads to metric distortion, directly destroying the potential geometric continuity of the physiological indicator data, thus causing data collapse. In summary, existing technologies fundamentally treat irregular missing patterns as a kind of "data defect (noise)" that needs to be smoothed and corrected, ignoring the structural information that these missing patterns themselves carry, determining the overall semantics of the physiological indicators.

[0075] Therefore, a physiological data prediction scheme based on the TRC paradigm to completely resolve the aforementioned manifold distortions from highly sparse and asynchronous raw physiological data will face a dilemma:

[0076] The first challenge is quantifying distortion features without reconstruction: if the existing TRC paradigm is used for implicit masking or reconstruction, an artificial pseudo-time step will inevitably be introduced on the basis of the original physiological index data; therefore, the model must have the ability to directly capture this inherent geometric deformation from the sparse original physiological index data.

[0077] The second challenge is the integration of distortion features: even with quantified distortion, the reliability of semantics for different categories of completely asynchronous physiological indicators varies greatly, making it extremely difficult to adaptively integrate this distortion information into deep representation learning. Improper handling can lead to the model failing to capture manifold distortions, thus ignoring sparse but crucial information.

[0078] To overcome this dilemma, the inventors proposed a novel framework called ManiDF (Manifold DistortionField), which transforms the paradigm of predicting irregular time series such as physiological indicators from "implicit masking" to "explicit distortion-aware restoration." Specifically, ManiDF constructs a computable explicit manifold distortion field that decomposes the geometrical disruption caused by irregular missing patterns into two complementary granularities: step-level distortion (measuring the integrity of global time slices of physiological indicator data) and variable-level distortion (capturing the reliability of local channels in physiological indicator data). Secondly, to address the challenge of fusing distortion features, the inventors designed a distortion-aware representation and aggregation mechanism. This mechanism injects step-level distortion during time encoding to guide the restoration of the manifold representation and utilizes variable-level distortion for weighted downsampling during variable aggregation, thereby adaptively adjusting the contribution of different recorded physiological indicator parameter values. It should be noted that ManiDF can effectively recover the intrinsic semantics directly from highly distorted physiological indicator data, thus significantly outperforming existing state-of-the-art baseline models.

[0079] The purpose of this invention is to address the inherent shortcomings of existing methods for predicting irregular multivariate time series physiological indicators, thereby improving the accuracy of predicting patients' future physiological indicators and enabling medical staff to avoid misjudging patients' physiological states. Therefore, based on the aforementioned ManiDF framework, this invention provides a method for constructing a physiological data prediction system for medical monitoring.

[0080] In summary, this system is used to predict a patient's future physiological data based on the patient's historical physiological data. This historical physiological data records multiple physiological indicators, each corresponding to parameter values ​​that are non-uniformly distributed over time. The construction method includes:

[0081] T1. Construct the initial system, such as Figure 1 As shown, the initial system includes:

[0082] An initial step-level distortion-aware embedder is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in the historical physiological data are encoded separately, and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period.

[0083] The initial spatiotemporal encoder is used to: extract the corresponding temporal enhancement features for each distortion-sensing embedding, wherein each temporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological indicator and all other physiological indicators in the corresponding time period.

[0084] An initial variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Here, interval irregularity describes the degree of uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuation describes the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features.

[0085] An initial prediction decoder is used to predict a patient's future physiological data based on all variable-level features.

[0086] T2. Obtain the training set, which includes multiple training samples. Each training sample includes the patient's historical physiological data and future physiological data labels. The future physiological data labels include multiple physiological indicators and the actual parameter values ​​of each physiological indicator at multiple future prediction points.

[0087] T3. Using the patient's historical physiological data as input and the predicted parameter values ​​of each type of physiological indicator at multiple future prediction points as output, the initial system is trained to convergence using a training set to obtain a physiological data prediction system for medical monitoring; wherein, the training process updates the system parameters by minimizing the preset loss.

[0088] This invention abandons the traditional forced completion and alignment TRC paradigm and proposes a prediction scheme based on ManiDF. This scheme shifts the processing paradigm from the traditional "implicit masking and forced alignment" to "explicit distortion-aware recovery," addressing the "distortion feature quantification dilemma" and "distortion feature fusion learning dilemma" caused by extremely high missing rates and asynchronous behavior across physiological indicators. This enables accurate prediction of highly distorted, irregular physiological indicators in healthcare scenarios. Specifically, the constructed system measures the integrity of global physiological data records through coverage collapse and time interval anomalies, and captures the reliability of local data records through the irregularity of time intervals and fluctuations in local continuity. This allows the system to directly recover the patient's true physiological evolution semantics from highly distorted and incomplete original historical physiological data, thus breaking away from the traditional "forced completion and alignment" mechanism. This avoids disrupting the temporal uniformity of manifold trajectories, distorting potential geometric structures, and destroying the inherent semantic consistency of time series, resulting in more accurate predictions.

[0089] To better understand the present invention, the present invention will be described in detail below with reference to specific embodiments.

[0090] According to one embodiment of the present invention, the present invention solution is based on the ManiDF framework, referring to... Figure 2 The main pipeline of the framework comprises four main processing stages: Stage 1 acquires the step-level distortion-aware embedding. Based on given historical physiological data of the patient, this stage calculates the step-level distortion caused by overlay collapse and time interval anomalies, and injects it into the distortion embedding layer through linear projection to explicitly represent the credibility of various recorded physiological indicators in different time periods. Stage 2 encodes the step-level distortion-aware embedding using a spatiotemporal encoder to extract spatiotemporal enhancement features of historical physiological indicator data. Stage 3 performs variable-level distortion weighted aggregation. This stage calculates variable-level distortion and uses it to complete distortion aggregation, enabling the system to adaptively reduce the weight of high-distortion regions in the physiological data during aggregation. Stage 4 maps the recovered manifold representation to the target space using a predictive decoder to predict the patient's future physiological data.

[0091] The following section provides a detailed explanation of the specific working principles and technical details of each stage.

[0092] According to one embodiment of the present invention, still referring to Figure 2In Phase 1, the first thing to do is to have Patient historical physiological data modeling for similar physiological indicators An irregular sparse time record set of a single physiological indicator:

[0093]

[0094] in, Represents the entire dataset. Representing the The first of the physiological indicators One record, Represents the index of physiological indicator categories. The total number of categories representing physiological indicators. Representing the The first of the physiological indicators Each record corresponds to a specific time. Representing the The first of the physiological indicators The parameter values ​​corresponding to each record Representing the The number of valid records of physiological indicators.

[0095] Next, step-level distortion is calculated to measure the reliability of various physiological indicator data across different time periods. Step-level distortion is calculated based on two components: coverage collapse and time interval anomalies. Coverage collapse describes the degree of missing records for all categories of physiological indicator parameters within a time period, while time interval anomalies describe the degree of anomalousness in the recording rhythm of all categories of physiological indicator parameters compared to the preceding adjacent time period. Specifically, coverage collapse is calculated as follows:

[0096]

[0097] in, Representing the Coverage collapse during a specific time period; Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first Observed Mask in historical physiological data. Did the first time period record the first one? The parameter values ​​of physiological indicators, when recorded. It is 1 if it is true, otherwise it is 0.

[0098] The calculation method for time interval anomalies is as follows:

[0099]

[0100] in, Representing the The time intervals of these time periods are abnormal; Representing the A time period; Representing the The data centroid for a time period is the average value of all parameter values ​​of all categories of physiological indicators at any given time within that time period. Representing the The centroid of the data for a given time period; It represents a very small positive number.

[0101] After obtaining the cover collapse and time interval anomalies, the step-level distortion scalar is further calculated based on these two results. :

[0102]

[0103] in, Representing the Step-level distortion scalar for each time period , These represent different balance coefficients, which are important in practical applications. , The value of can be adaptively determined through hyperparameter optimization methods, such as performing a grid search on the historical validation set.

[0104] Then, to enable the system to better perceive distortion, the scalar distortion is projected into the model space through a linear transformation:

[0105]

[0106] in, Represents the first linear transformation Step-level distortion scalar for each time period This represents the weight matrix in a linear transformation operation. This represents the bias vector in a linear transformation operation.

[0107] Finally, the distortion-perceived embeddings for each type of physiological indicator in each time period are obtained as follows:

[0108]

[0109] in, Representing the Physiological indicators in the first Distortion perception embedding within a time period; This represents a parameter value embedder used to map physiological indicator parameter values ​​to high-dimensional vectors; Representing the Physiological indicators in the first Parameter values ​​within a time period (Observed Value); The time embedder represents the time value that maps physiological indicator parameter values ​​to a high-dimensional vector. Representing the Within the time period, the first The time corresponding to the parameter values ​​of physiological indicators. It is broadcast to all physiological indicator categories within the same time period.

[0110] According to one embodiment of the present invention, still referring to Figure 2 Phase 2's task is to embed the distortion perception obtained in Phase 1 into the spatiotemporal encoder to jointly capture the temporal dynamics of each type of physiological indicator and the dependencies between various physiological indicators. Specifically:

[0111] First, a feature refinement tool is applied to refine the distortion-aware embedding:

[0112]

[0113] in, Representing the Physiological indicators in the first Refined distortion-perceived embedding within a specific time period Represents the ReLU function activator. The weight matrix representing the first fully connected layer of the feature refinement tool. The weight matrix represents the second fully connected layer of the feature refinement tool.

[0114] Subsequently, the time dependence of physiological indicators is modeled along the slice dimension (a slice is a record of a type of physiological indicator over a period of time) using Transformer layers:

[0115]

[0116] in, Representing the Physiological indicators in the first Temporal enhancement features within a time period; represent Encoder; The total number of segments representing the time period.

[0117] Next, a graph neural network (GNN) with an adaptive adjacency matrix is ​​used to further capture the interactions between various physiological indicators based on the previous step:

[0118]

[0119] in, Representing the Physiological indicators in the first Spatiotemporal enhancement features within a time period This represents a graph neural network with an adaptive adjacency matrix. Representing the Physiological indicators in the first Temporal enhancement features within a time period The total number of categories representing physiological indicators.

[0120] According to one embodiment of the present invention, still referring to Figure 2 After completing spatiotemporal encoding, Stage 3 aggregates the temporal representation into variable-level features using distortion awareness. This is done to calculate the variable-level distortion weights. This invention combines two concepts: interval irregularity and local continuity fluctuation. Interval irregularity describes the uneven distribution of parameter values ​​of a class of physiological indicators along the time axis, while local continuity fluctuation describes the fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a given time period. Specifically, interval irregularity is first calculated as follows:

[0121]

[0122] in, Representing the Irregularity of intervals for physiological indicators; Represents a record index based on physiological indicator parameter values. The standard deviation is calculated based on the dimensions. Representing the Physiological indicators The time corresponding to the parameter value of each record; Representing the Physiological indicators The time corresponding to the parameter value of each record; It represents a very small positive number.

[0123] Local continuity fluctuations are calculated as follows:

[0124]

[0125] in, Representing the Physiological indicators in the first Local continuous fluctuations within a time period; Represents a radius of The standard deviation is calculated within a local time window, where the radius is... A local time window is defined as a time period centered on a point in time, extending forward and backward along the time axis. Formed over a period of time; Representative of the first The records within a local time window centered on the [number]th time period The number of parameter values ​​for physiological indicators.

[0126] Next, define the weights for variable-level distortion:

[0127]

[0128] in, Representing the Physiological indicators in the first The variable-level distortion weights corresponding to each time period , These represent different balance coefficients.

[0129] Subsequently, aggregate weights are calculated based on the weights with variable-level distortion:

[0130]

[0131] in, Representing the Physiological indicators in the first The aggregate weight corresponds to each time period. This weight informs the system to reduce the weight of data in locations with high distortion or where physiological parameter values ​​are not recorded.

[0132] Next, the aggregation weights are normalized along the time dimension:

[0133]

[0134] in, Representing the Physiological indicators in the first The normalized aggregate weights corresponding to each time period Representing the Physiological indicators in the first The aggregate weights corresponding to each time period.

[0135] Then, the spatiotemporal enhancement features are applied using normalized aggregation weights. Perform weighted aggregation:

[0136]

[0137] in, Representing the Intermediate variable level characteristics of physiological indicators.

[0138] Finally, graph neural networks are applied to capture global relational context at the variable-level features:

[0139]

[0140] in, Representing the Variable-level characteristics of physiological indicators This represents a graph neural network with an adaptive adjacency matrix.

[0141] According to one embodiment of the present invention, still referring to Figure 2 After obtaining the variable-level features, stage 4 predicts the patient's future physiological data based on these features. The prediction method involves inputting the variable-level features into the prediction decoder (ForecastingLayer), which then queries the target layer. Make predictions; among them, Representing the The first of the physiological indicators The time corresponding to each prediction point. This process is represented as:

[0142]

[0143] in, Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point This represents a multilayer perceptron.

[0144] The present invention, through the above stages 1-4, enables the prediction of a patient's future physiological data based on the patient's historical physiological data, and this process does not introduce any artificial interpolation or alignment, and does not distort the original semantics expressed by the historical physiological data.

[0145] According to one embodiment of the present invention, during the system construction process, the loss can be designed in various ways (such as cross-entropy, mean squared error, etc.), and the preferred method is to train the system end-to-end by minimizing the mean squared error (MSE) at all prediction points, that is:

[0146]

[0147] in, Represents the pre-defined loss. Represents the total number of prediction points. Represents the index of the predicted point. Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point Representing the The first of the physiological indicators The true parameter values ​​corresponding to each prediction point. By constructing a loss function using the mean squared error method, the irregular distribution of historical physiological data and the missing or sparse records of some physiological indicators can be better matched, thereby reducing the prediction error of the final system.

[0148] According to one embodiment of the present invention, the present invention also provides a physiological data prediction system for medical monitoring based on the construction method of the present invention, the system comprising:

[0149] A step-level distortion-aware embedding is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in historical physiological data are encoded separately and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period.

[0150] The spatiotemporal encoder is used to: extract the corresponding temporal enhancement features of each distortion sensing embedding, wherein each temporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological indicator and all other physiological indicators in the corresponding time period.

[0151] A variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Interval irregularity describes the uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuation describes the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features.

[0152] A predictive decoder used to predict a patient's future physiological data based on all variable-level features.

[0153] According to one embodiment of the present invention, the present invention also provides a method for predicting physiological data for medical monitoring, the prediction method comprising: S1, acquiring the patient's historical physiological data; S2, inputting the patient's historical physiological data into a physiological data prediction system for medical monitoring as described in the present invention for processing, to obtain predicted future physiological data for the patient.

[0154] Finally, the inventors verified the effectiveness of the invention through comparative experiments. For example... Figure 3 As shown, this comparative experiment uses the APN model and the physiological data prediction system for medical monitoring of the present invention to predict future physiological data based on the same medical monitoring input data (derived from the publicly available intensive care medical dataset, MIMIC), and compares the prediction results. Figure 3 The results comparison shows that the prediction results of this invention are significantly closer to the true values ​​than those of the APN model.

[0155] Furthermore, it should be noted that the ManiDF framework of this invention can be used not only to construct predictive systems for medical monitoring, but also for scenarios such as weather forecasting (predicting climate change) and autonomous driving (predicting road conditions). (See also...) Figure 4 The weather forecasting system built based on the ManiDF concept of this invention also achieved better forecasting results when making forecasts based on meteorological input data (derived from the US Historical Climate Data Set, USHCN).

[0156] In summary, the key point of this invention lies in proposing a computable manifold distortion field, along with a corresponding distortion-aware representation and aggregation mechanism. Specifically, it includes:

[0157] Constructing a manifold distortion field for dual-granularity quantification: This invention explicitly quantifies geometrical damage caused by irregular missing values ​​from two complementary granularities: step-level distortion and variable-level distortion. Step-level distortion measures the integrity of global time slices in physiological indicator data through overlay collapse and time interval anomalies; variable-level distortion captures the reliability of local channels in physiological indicator data through the irregularity of time intervals and fluctuations in local continuity.

[0158] Step-level distortion-aware embedding: During the system's encoding phase, this invention explicitly injects the calculated scalar step-level distortion into the distortion-aware embedding to guide the system to perceive the degree of global degradation (unreliability) caused by missing historical physiological data records and abnormal recording rhythms, and to perform semantic recovery of the manifold representation.

[0159] Variable-level distortion weighted aggregation: This invention introduces an aggregation mechanism adjusted by variable-level distortion. After the data is processed by the spatiotemporal encoder, this mechanism uses variable-level distortion to perform weighted downsampling on the variable aggregation process, adaptively reducing the weight of highly distorted physiological indicator data and locations lacking physiological indicator numerical records, so as to guide the system to refer more to reliable data areas when making predictions.

[0160] In summary, the physiological data prediction system construction method for medical monitoring of the present invention effectively avoids the negative impact of manifold distortion caused by missing records and asynchronous sampling of patient historical physiological data on the prediction results in medical monitoring scenarios by introducing a manifold distortion field composed of step-level and variable-level distortions and using the manifold distortion field to guide the embedding encoding and weighted aggregation process. This scheme eliminates the need for interpolation and forced alignment, thus preserving the true information expressed by the patient's historical physiological data, enabling accurate prediction of the patient's future physiological indicators and reducing the probability of medical staff misjudging the patient's physiological state.

[0161] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0162] This invention can be a computer program product, which mainly refers to a software product that implements this solution through a computer program.

[0163] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a physiological data prediction system for medical monitoring, the system being used to predict a patient's future physiological data based on the patient's historical physiological data, wherein the historical physiological data records multiple physiological indicators of the patient, each physiological indicator corresponding to parameter values ​​that are non-uniformly distributed along a time axis, characterized in that... The construction method includes: T1. Construct an initial system, the initial system comprising: An initial step-level distortion-aware embedder is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in the historical physiological data are encoded separately, and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period. The initial spatiotemporal encoder is used to: extract the corresponding temporal enhancement features of each distortion-sensing embedding, wherein each temporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the variation pattern of the corresponding physiological indicator parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological indicator and all other physiological indicators in the corresponding time period. An initial variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Interval irregularity describes the uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuations describe the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features. An initial prediction decoder is used to predict a patient's future physiological data based on all variable-level features; T2. Obtain the training set, which includes multiple training samples. Each training sample includes the patient's historical physiological data and future physiological data labels. The future physiological data labels include multiple physiological indicators and the true parameter values ​​of each physiological indicator at multiple future prediction points. T3. Using the patient's historical physiological data as input and the predicted parameter values ​​of each type of physiological indicator at multiple future prediction points as output, the initial system is trained to convergence using a training set to obtain a physiological data prediction system for medical monitoring; wherein, the training process updates the system parameters by minimizing the preset loss.

2. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, In the initial step-level distortion-aware embedder, the predetermined step-level distortion calculation method is as follows: in, Representing the Coverage collapse during a specific time period; The total number of categories representing physiological indicators; Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? Physiological index parameter values, when recorded It is 1 if it is true, otherwise it is 0. Representing the The time intervals of these time periods are abnormal; Representing the A time period; Representing the The data centroid for a time period is the average value of all parameter values ​​of all categories of physiological indicators at any given time within that time period. Representing the The centroid of the data for a given time period; It represents a very small positive number.

3. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, The initial step-level distortion-aware embedder encodes all classes of physiological indicators in historical physiological data and fuses them to cover collapse and time interval anomalies to obtain the distortion-aware embedding of each class of physiological indicator in each time period in the following manner: in, Representing the Physiological indicators in the first Distortion perception embedding within a time period; This represents a parameter value embedder used to map physiological indicator parameter values ​​to high-dimensional vectors; Representing the Physiological indicators in the first Parameter values ​​within a time period; Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? Physiological index parameter values, when recorded It is 1 if it is true, otherwise it is 0. The time embedder is used to map the time corresponding to the physiological index parameter values ​​into a high-dimensional vector. Representing the Within the time period, the first The time corresponding to the parameter values ​​of physiological indicators; Represents the first linear transformation The step-level distortion scalar for each time period is obtained as follows: in, This represents the weight matrix in a linear transformation operation. This represents the bias vector in a linear transformation operation. Representing the The step-level distortion scalar for each time period is obtained as follows: in, , These represent different balance coefficients; Representing the Coverage collapse during a specific time period; Representing the The time intervals for each period are abnormal.

4. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, The initial spatiotemporal encoder extracts the corresponding temporal enhancement features for each distortion-aware embedding in the following manner: in, Representing the Physiological indicators in the first Temporal enhancement features within a time period; represent Encoder; The total number of segments representing the time period; Representing the Physiological indicators in the first The refined distortion-aware embedding for each time period is obtained as follows: in, Represents the ReLU function activator. The weight matrix representing the first fully connected layer of the feature refinement tool. The weight matrix of the second fully connected layer representing the feature refinement tool. Representing the Physiological indicators in the first Distortion perception embedding within a time period.

5. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, The initial spatiotemporal encoder extracts the spatiotemporal enhancement features corresponding to each temporal enhancement feature in the following manner: in, Representing the Physiological indicators in the first Spatiotemporal enhancement features within a time period This represents a graph neural network with an adaptive adjacency matrix. Representing the Physiological indicators in the first Temporal enhancement features within a time period The total number of categories representing physiological indicators.

6. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, In the initial variable-level distortion weighted aggregator, the predetermined variable-level distortion calculation method is as follows: in, Representing the Irregularity of intervals for physiological indicators; Represents a record index based on physiological indicator parameter values. The standard deviation is calculated based on the dimensions. Representing the Physiological indicators The time corresponding to the parameter value of each record; Representing the Physiological indicators The time corresponding to the parameter value of each record; Represents the smallest positive number; Representing the Physiological indicators in the first Local continuous fluctuations within a time period; Represents a radius of The standard deviation is calculated within a local time window, where the radius is... A local time window is defined as a time period centered on a point in time, extending forward and backward along the time axis. Formed over a period of time; Representative of the first The records within a local time window centered on the [number]th time period The number of parameter values ​​for physiological indicators.

7. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, In the initial variable-level distortion weighted aggregator, the predetermined weighted aggregation method is: in, Representing the Variable-level characteristics of physiological indicators; This represents a graph neural network with an adaptive adjacency matrix; The total number of categories representing physiological indicators; Representing the Intermediate variable-level features of physiological indicators are obtained in the following way: in, The total number of time periods; Representing the Physiological indicators in the first Spatiotemporal enhancement features within a time period; Representing the Physiological indicators in the first The normalized aggregate weights corresponding to each time period are obtained as follows: in, Representing the Physiological indicators in the first Aggregate weights corresponding to each time period; Representing the Physiological indicators in the first The aggregate weights corresponding to each time period are obtained as follows: in, Representing the Within the time period, the first A visibility mask for physiological indicators, which identifies the first physiological indicator in historical data. Did the first time period record the first one? The parameter values ​​of physiological indicators, when recorded. It is 1 if it is true, otherwise it is 0. Representing the Physiological indicators in the first The variable-level distortion weights corresponding to each time period are obtained as follows: in, Representing the Irregularity of intervals for physiological indicators; Representing the Physiological indicators in the first Local continuous fluctuations within a time period; , These represent different balance coefficients.

8. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, The initial prediction decoder predicts the patient's future physiological data based on all variable-level features in the following way: in, Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point Represents a multilayer perceptron. Representing the Variable-level characteristics of physiological indicators Representing the The first of the physiological indicators The time corresponding to each prediction point.

9. The method for constructing a physiological data prediction system for medical monitoring according to claim 1, characterized in that, In T3, the preset loss is: in, Represents the pre-defined loss. The total number of categories representing physiological indicators. Represents the total number of prediction points. Represents the index of the predicted point. Representing the The first of the physiological indicators The prediction parameter values ​​corresponding to each prediction point Representing the The first of the physiological indicators The actual parameter values ​​corresponding to each prediction point.

10. A physiological data prediction system for medical monitoring constructed based on the method of any one of claims 1-9, used to predict the future physiological data of a patient based on the patient's historical physiological data, wherein the historical physiological data records multiple physiological indicators of the patient, each physiological indicator corresponding to parameter values ​​that are non-uniformly distributed on a time axis, characterized in that, The system includes: A step-level distortion-aware embedding is used to: calculate coverage collapse and time interval anomalies based on historical physiological data according to a predetermined step-level distortion calculation method. Coverage collapse describes the degree of missing records of all categories of physiological indicator parameter values ​​in a time period, and time interval anomalies describe the degree of anomaly in the recording rhythm of all categories of physiological indicator parameter values ​​compared to the previous adjacent time period. All categories of physiological indicators in historical physiological data are encoded separately and coverage collapse and time interval anomalies are fused to obtain the distortion-aware embedding of each category of physiological indicator in each time period. The spatiotemporal encoder is used to: extract the corresponding temporal enhancement features of each distortion sensing embedding, wherein each temporal enhancement feature reflects the change pattern of the corresponding physiological index parameter value acquired in the corresponding time period on the complete time axis; and further extract the corresponding spatiotemporal enhancement features based on each temporal enhancement feature, wherein each spatiotemporal enhancement feature reflects the change pattern of the corresponding physiological index parameter value acquired in the corresponding time period on the complete time axis, as well as the interaction pattern between the corresponding physiological index and all other physiological indices in the corresponding time period. A variable-level distortion weighted aggregator is used to: calculate interval irregularity and local continuity fluctuations based on historical physiological data according to a predetermined variable-level distortion calculation method. Interval irregularity describes the uneven distribution of parameter values ​​of a class of physiological indicators on the time axis, and local continuity fluctuation describes the degree of fluctuation in the frequency of parameter value recordings of a class of physiological indicators over a time period. Through a predetermined weighted aggregation method, each interval irregularity and local continuity fluctuation is integrated into its corresponding spatiotemporal enhancement feature to obtain multiple variable-level features. A predictive decoder used to predict a patient's future physiological data based on all variable-level features.

11. A method for predicting physiological data for medical monitoring, characterized in that, The prediction method includes: S1. Obtain the patient's historical physiological data; S2. Input the patient's historical physiological data into the physiological data prediction system for medical monitoring as described in claim 10 for processing, so as to obtain the predicted future physiological data of the patient.

12. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-9, 11.

Images