Hemodialysis process hypotension risk time physiological signal warning method and system

Abstract

The application discloses a time sequence physiological signal early warning method and system for low blood pressure risk in hemodialysis process, and the method comprises the following steps: constructing a real-time state evolution sequence of a target object in a dialysis process, calculating K sequence similarities of the real-time state evolution sequence, selecting the maximum similarity, extracting a personalized evolution window length based on the maximum similarity, backtracking a state evolution subsequence with a length equal to the personalized evolution window length, inputting the state evolution subsequence into a time sequence classification model, outputting a risk probability of a low blood pressure event, and if the risk probability is greater than a set early warning risk threshold, a low blood pressure risk early warning signal is sent out; the application identifies a cluster level critical moment when a group state evolution trajectory starts to significantly diverge from high consistency based on a condensation degree sequence, and finally determines a time interval between the cluster level critical moment and the latest sampling moment as a personalized evolution window length corresponding to the cluster, so that each cluster has a window scale matched with its own common evolution period.

Description

Technical Field

[0001] This invention relates to the field of time-series risk warning, specifically to a method and system for early warning of time-series physiological signals of hypotension risk during hemodialysis. Background Technology

[0002] During hemodialysis, patients often experience dynamic changes in various physiological parameters. Time-series analysis of continuously collected physiological signals can identify potential hypotension risk patterns. Current early warning methods primarily include threshold-based rule-based judgments and machine learning-based classification predictions.

[0003] However, the above methods still have significant limitations in practical applications. First, a fixed time window is difficult to adapt to individual differences in the rate of risk evolution among different dialysis subjects; that is, some subjects may rapidly enter a high-risk state in a short period of time, while others show a slow evolution trend, and a uniform time scale can easily lead to delayed early warning. Second, general classification models often treat all historical samples as homogeneous data for training, ignoring the fact that there may be multiple typical but different physiological state evolution paths before a hypotensive event, reducing the individual adaptability of state pattern matching.

[0004] Therefore, this invention provides a method and system for early warning of the risk of hypotension during hemodialysis based on time-series physiological signals. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for early warning of hypotension risk during hemodialysis based on temporal physiological signals. This method solves the technical problems mentioned in the background by introducing a personalized evolutionary window length and a risk warning within that window length.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for early warning of hypotension risk during hemodialysis based on time-series physiological signals, including:

[0008] S1. Construct a real-time state evolution sequence of the target object during the dialysis process;

[0009] S2. Calculate the K sequence similarities between the real-time state evolution sequence and the pre-constructed K full-process evolution template sequences;

[0010] The K full-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters, and the physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events.

[0011] S3. Select the maximum similarity among the K sequence similarities;

[0012] S4. If the maximum similarity is greater than the set evolutionary similarity threshold, then based on the full evolutionary template sequence corresponding to the maximum similarity, extract the associated personalized evolutionary window length.

[0013] S5. Using the latest sampling time as the endpoint, backtrack and extract a subsequence of state evolution with a length equal to the length of the personalized evolution window from the real-time state evolution sequence.

[0014] S6. Input the state evolution subsequence into a pre-trained time-series classification model and output the risk probability of a hypotension event occurring in the target object within the prediction interval; wherein, the time length of the prediction interval is equal to the length of the personalized evolution window;

[0015] S7. If the risk probability is greater than the set warning risk threshold, a low blood pressure risk warning signal is issued.

[0016] In some specific embodiments, constructing a real-time state evolution sequence of the target object during dialysis includes:

[0017] S1-1. Obtain N raw physiological parameters of the target object at the sampling time;

[0018] S1-2. Characterize the N original physiological parameters and concatenate them into a physiological state vector that represents the instantaneous physiological state during dialysis.

[0019] S1-3. Using the start time of dialysis as the zero point, continuously collect and update the physiological state vector of the target object at each sampling time.

[0020] S1-4. Arrange each physiological state vector in the time sequence of the sampling time to form a real-time state evolution sequence.

[0021] In some specific embodiments, the method for constructing the personalized evolution window length includes:

[0022] A1. Obtain the historical state evolution sequence of D hypotension events; where the latest sampling time of the historical state evolution sequence corresponds to the hypotension event.

[0023] A2. Anchor the pre-mutation vectors of each of the D historical state evolution sequences;

[0024] A3. Using the pre-mutation vector as the endpoint, extract D pre-mutation local sequences of fixed length H from its corresponding historical state evolution sequence in reverse order.

[0025] A4. Cluster the D pre-mutation local sequences to obtain K physiological state evolution clusters; where each physiological state evolution cluster contains G pre-mutation local sequences, and G is a variable that is equal or unequal;

[0026] A5. Construct the cohesion sequences of K physiological state evolution clusters;

[0027] A6. Identify the critical cohesion degree in the cohesion degree sequence;

[0028] A7. The sampling time corresponding to the anchor critical cohesion is taken as the cluster-level critical time; where the cluster-level critical time represents the time point at which the evolution trajectory of the population state within the cluster begins to diverge significantly.

[0029] A8. Calculate the time interval between the cluster-level critical moment and the latest sampling moment, and use it as the personalized evolution window length of the target evolving cluster;

[0030] A9. Traverse the K physiological state evolution clusters and repeat the calculation of the personalized evolution window length until the K personalized evolution window lengths are obtained.

[0031] In some specific embodiments, the historical state evolution sequence of D hypotension events is obtained, including:

[0032] A1-1. Select D hypotension events from the hypotension event database;

[0033] A1-2. Anchor the occurrence times of D hypotension events on the time axis and mark them as the latest sampling time;

[0034] A1-3. Starting from the dialysis initiation time and ending at each latest sampling time, obtain the physiological state vector of the hypotension event at all sampling times.

[0035] A1-4. Arrange the physiological state vectors of all sampling times in ascending chronological order to form D historical state evolution sequences.

[0036] In some specific embodiments, anchoring the pre-mutation vectors of each of the D historical state evolution sequences includes:

[0037] A2-1. In each historical state evolution sequence, calculate the similarity between adjacent physiological state vectors; where each similarity is associated with the sampling time of the previous physiological state vector;

[0038] A2-2. Arrange the similarities in ascending order according to the associated sampling times to form a similarity sequence, resulting in D similarity sequences;

[0039] A2-3. Identify the mutation similarity of each of the D similar sequences;

[0040] A2-4. Extract the adjacent physiological state vectors corresponding to the mutation similarity, and label the physiological state vector of the previous moment as the vector before mutation.

[0041] In some specific embodiments, the D pre-mutation local sequences are clustered to obtain K physiological state evolution clusters, including:

[0042] A4-1. Vectorize each pre-mutation local sequence to generate D pre-mutation evolution vectors;

[0043] A4-2. From the D pre-mutation evolution vectors, randomly select K pre-mutation evolution vectors as the first round cluster centers;

[0044] A4-3. For any pre-mutation evolution vector, calculate its vector distance to the K first-round cluster centers, and assign it to the first-round cluster center with the smallest vector distance.

[0045] A4-4. Traverse all pre-mutation evolution vectors until each pre-mutation evolution vector is assigned to the first-round cluster center with the minimum vector distance, forming K initial physiological state evolution clusters;

[0046] A4-5. Based on all the pre-mutation evolution vectors within each initial physiological state evolution cluster, calculate its cluster center vector and anchor the cluster center vector as the cluster center for the next round.

[0047] A4-6. Calculate the vector distance between the cluster centers of the next round and the cluster centers of the first round;

[0048] A4-7. If the vector distance is greater than the preset threshold, then all pre-mutation evolution vectors are iteratively allocated based on the cluster center in the next round.

[0049] A4-8. Perform the iterative allocation on the D pre-mutation evolution vectors respectively until the distance between all vectors is less than the preset threshold, and obtain K physiological state evolution clusters.

[0050] In some specific embodiments, the cohesion sequences of K physiological state evolution clusters are constructed, including:

[0051] A5-1. Within the K physiological state evolution clusters, select any one target evolution cluster;

[0052] A5-2. Starting from the latest sampling time of the G pre-mutation local sequences contained in the target evolutionary cluster, backtrack towards the historical direction at each sampling time, and extract G local evolutionary sub-sequences with each backtracking time as the endpoint.

[0053] A5-3. Calculate the F subsequence distances between each pair of the G local evolutionary subsequences; where F = G(G-1) / 2;

[0054] A5-4. Obtain the mean and standard deviation of the distances between F subsequences;

[0055] A5-5. Based on the mean and standard deviation of the distances between the F subsequences, calculate the cohesion that characterizes the population consistency of the G locally evolved subsequences.

[0056] A5-6. Arrange the cohesion corresponding to each retrospective moment in ascending time order to construct the cohesion sequence of the target evolution cluster;

[0057] A5-7. Traverse the K physiological state evolution clusters and repeat the cohesion sequence construction to form K cohesion sequences.

[0058] In some specific embodiments, the steps for constructing the full-evolution template sequence include:

[0059] B1. Local sequence of the cluster center of the anchored physiological state evolution cluster;

[0060] B2. Calculate the distances between the cluster center local sequence and the G pre-mutation local sequences within the cluster;

[0061] B3. Locate the optimal local sequence based on the minimum distance among the G local sequence distances;

[0062] B4. Anchor the historical state evolution sequence to which the optimal local sequence belongs, and use it as the whole evolution template sequence associated with the physiological state evolution cluster;

[0063] B5. Traverse the K physiological state evolution clusters and repeat the anchoring of the whole evolution template sequence to obtain K whole evolution template sequences.

[0064] This invention provides a method and system for early warning of time-series physiological signals of hypotension risk during hemodialysis, which has the following beneficial effects:

[0065] This invention locates the vector before mutation by anchoring the mutation point of the similarity between adjacent physiological state vectors in each historical state evolution sequence, and extracts the local sequence before mutation with the sampling time corresponding to the vector as the endpoint. This ensures that the extracted local sequences are all in the stage before the occurrence of hypotension event and the group state evolution trajectory still maintains high consistency. In contrast, if the latest sampling time (i.e. the time when the hypotension event occurs) is used as the endpoint for backtracking, it will inevitably include the state divergence of each object due to individual differences in the stage near the event.

[0066] Furthermore, all pre-mutation local sequences are clustered to form K physiological state evolution clusters. Each cluster contains several historical dialysis objects with similar evolutionary trends in the pre-mutation local stage, so that different types of high-risk evolutionary processes can be classified into different clusters according to their actual trajectory characteristics.

[0067] Furthermore, based on the pre-mutation local sequences contained in each physiological state evolution cluster, the cohesion of the local evolution subsequences of the population is calculated by tracing back from the latest sampling time step by step in the historical direction. Based on the cohesion sequence, the cluster-level critical moment when the population state evolution trajectory starts to diverge significantly from high consistency is identified. Finally, the time interval between the cluster-level critical moment and the latest sampling time is determined as the personalized evolution window length corresponding to the cluster, so that each cluster has a window scale that matches its own common evolution period.

[0068] Secondly, the present invention provides a time-series physiological signal early warning system for the risk of hypotension during hemodialysis, comprising:

[0069] The real-time sequence construction module is used to construct the real-time state evolution sequence of the target object during the dialysis process;

[0070] The similarity calculation module is used to calculate the K sequence similarities between the real-time state evolution sequence and the K pre-constructed full-process evolution template sequences;

[0071] The K whole-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters. The physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events.

[0072] The similarity selection module is used to select the maximum similarity among the K sequence similarities;

[0073] The evolution length extraction module is used to determine: if the maximum similarity is greater than the set evolution similarity threshold, then based on the full evolution template sequence corresponding to the maximum similarity, the associated personalized evolution window length is extracted;

[0074] The subsequence backtracking module is used to backtrack and extract a state evolution subsequence with a length equal to the personalized evolution window length from the real-time state evolution sequence, with the latest sampling time as the endpoint.

[0075] The risk probability output module is used to input the state evolution subsequence into a pre-trained time series classification model and output the risk probability of the target object experiencing a hypotension event within the prediction interval; wherein, the time length of the prediction interval is equal to the length of the personalized evolution window;

[0076] The risk warning module is used to determine whether a low blood pressure risk warning signal is issued if the risk probability is greater than a set warning risk threshold.

[0077] Compared with the prior art, the beneficial effects of the time-series physiological signal early warning system for low blood pressure risk during hemodialysis of the present invention are the same as those of the time-series physiological signal early warning method for low blood pressure risk during hemodialysis described above, so they will not be repeated here. Attached Figure Description

[0078] Figure 1 This is a flowchart illustrating the time-series physiological signal early warning method for the risk of hypotension during hemodialysis according to the present invention.

[0079] Figure 2 This is a schematic diagram illustrating the calculation process of the personalized evolution window length described in this invention;

[0080] Figure 3 This is a schematic diagram of the labeling process for the pre-mutation vector described in this invention;

[0081] Figure 4 This is a schematic diagram illustrating the construction process of the cohesion sequence described in this invention;

[0082] Figure 5 This is a schematic diagram of the anchoring process of the entire evolution template sequence described in this invention;

[0083] Figure 6 This is a structural block diagram of the time-series physiological signal early warning system for the risk of hypotension during hemodialysis according to the present invention. Detailed Implementation

[0084] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0085] Example 1: Please refer to Figure 1 This invention provides a method for early warning of hypotension risk during hemodialysis based on time-series physiological signals, comprising the following steps:

[0086] S1. Construct a real-time state evolution sequence of the target object during the dialysis process;

[0087] S2. Calculate the K sequence similarities between the real-time state evolution sequence and the pre-constructed K full-process evolution template sequences;

[0088] The K full-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters, and the physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events.

[0089] Specifically, when calculating sequence similarity, the sequence length of the real-time state evolution sequence is first obtained, which is the number of sampling times; then, a prefix subsequence with the same length as the sequence is extracted from the full evolution template sequence, and the sequence similarity between the prefix subsequence and the real-time state evolution sequence is calculated.

[0090] The formula for calculating sequence similarity is:

[0091] ;

[0092] in, Represents cosine similarity. The prefix subsequence representing the real-time state evolution sequence. This represents the corresponding length of the prefix subsequence of the entire evolution template sequence, where T is the sequence length, i.e., the number of sampling times; and Let represent the physiological state vectors at the t-th sampling time in the two sequences, respectively;

[0093] Sequence similarity is calculated by taking the cosine similarity between the physiological state vectors of two sequences at the same time point, and then averaging the results. This represents the overall similarity between the two time series sequences in the entire evolutionary trend. The value ranges from [−1, 1]. A larger value indicates that the two sequences are closer in direction and evolutionary pattern, while a smaller value indicates greater differences.

[0094] S3. Select the maximum similarity among the K sequence similarities;

[0095] S4. If the maximum similarity is greater than the set evolutionary similarity threshold, then based on the full evolutionary template sequence corresponding to the maximum similarity, extract the associated personalized evolutionary window length.

[0096] In this embodiment, the evolutionary similarity threshold is determined by cross-validation on a historical hypotension event dataset and is used to distinguish high-risk evolutionary patterns from normal physiological fluctuations.

[0097] S5. Using the latest sampling time as the endpoint, backtrack and extract a subsequence of state evolution with a length equal to the length of the personalized evolution window from the real-time state evolution sequence.

[0098] S6. Input the state evolution subsequence into a pre-trained time-series classification model and output the risk probability of a hypotension event occurring in the target object within the prediction interval; wherein, the time length of the prediction interval is equal to the length of the personalized evolution window;

[0099] Specifically, the time-series classification model preferably employs an LSTM or Transformer architecture; its pre-training steps follow a commonly used supervised training process in the machine learning field, including:

[0100] Positive sample construction: For each historical hypotension event, according to the physiological state evolution cluster to which it belongs, obtain the personalized evolution window length corresponding to the cluster; with the time of occurrence of the hypotension event as the endpoint, backtrack the personalized evolution window length, extract the state evolution subsequence of the corresponding time period as positive samples, and label them as 1;

[0101] Negative sample construction: During dialysis processes in which hypotension never occurred, a continuous time period of the same length as the positive sample was selected, and its state evolution subsequence was extracted as a negative sample and labeled as 0;

[0102] It should be noted that the time length of the prediction interval is equal to the length of the personalized evolution window because the supervision signal during model training is constructed based on the correspondence of "historical window length = future risk assessment window length".

[0103] In other words, the model learns the following mapping: "If the current state evolutionary subsequence matches a certain high-risk evolutionary pattern, then a hypotension event is likely to occur in the future within a time frame equal to the length of its input history."

[0104] If a prediction duration different from the personalized evolution window length is used during the inference phase (such as a fixed 10 minutes), the input-output temporal structure will be inconsistent with the training distribution, which will impair the reliability of the model's early warning.

[0105] S7. If the risk probability is greater than the set warning risk threshold, a low blood pressure risk warning signal is issued.

[0106] In this embodiment, the early warning risk threshold is determined by analyzing the risk probability output distribution of historical dialysis subjects, and is used to distinguish between high-risk evolution patterns and normal physiological fluctuations.

[0107] In this embodiment, a real-time state evolution sequence of the target object is constructed, and its maximum similarity with K full-process evolution template sequences is calculated. Then, the corresponding personalized evolution window length is extracted, and a historical state sub-sequence of equal length is input into the pre-trained time series classification model to output the probability of low blood pressure risk within the same time period in the future. When the probability exceeds the warning risk threshold set based on the historical data distribution, a low blood pressure risk warning signal is issued to achieve dynamic risk identification of the current dialysis process.

[0108] In this embodiment, step S1 specifically includes:

[0109] S1-1. Obtain N raw physiological parameters of the target object at the sampling time;

[0110] In this embodiment, the target object refers to the dialysis patient currently undergoing hemodialysis.

[0111] S1-2. Characterize the N original physiological parameters and concatenate them into a physiological state vector that represents the instantaneous physiological state during dialysis.

[0112] S1-3. Using the start time of dialysis as the zero point, continuously collect and update the physiological state vector of the target object at each sampling time.

[0113] S1-4. Arrange each physiological state vector in the time sequence of the sampling time to form a real-time state evolution sequence.

[0114] In this embodiment, by taking the start time of dialysis as the zero point of time, the physiological state vectors of each sampling time are arranged in chronological order to form a real-time state evolution sequence that characterizes the continuous physiological state changes of the target object during the current dialysis process.

[0115] Example 2: See Figures 2 to 5 The technical solution of this embodiment 2 differs from that of embodiment 1 in that it discloses the method for constructing the personalized evolution window length described in embodiment 1, the steps of which include:

[0116] A1. Obtain the historical state evolution sequence of D hypotension events; where the latest sampling time of the historical state evolution sequence corresponds to the hypotension event.

[0117] A2. Anchor the pre-mutation vectors of each of the D historical state evolution sequences;

[0118] A3. Using the pre-mutation vector as the endpoint, extract D pre-mutation local sequences of fixed length H from its corresponding historical state evolution sequence in reverse order.

[0119] Specifically, the fixed length is used as a backtracking time window to capture common features of high-risk patterns, which is determined within the range of 5 to 30 minutes through historical data verification.

[0120] A4. Cluster the D pre-mutation local sequences to obtain K physiological state evolution clusters; where each physiological state evolution cluster contains G pre-mutation local sequences, and G is a variable that is equal or unequal;

[0121] A5. Construct the cohesion sequences of K physiological state evolution clusters;

[0122] A6. Identify the critical cohesion degree in the cohesion degree sequence;

[0123] Specifically, the critical cohesion degree represents the critical point at which the evolutionary trajectory of the population state within a physiological state evolution cluster transitions from high consistency to significant divergence. This can be achieved by calculating the first-order difference between adjacent cohesion degrees and simultaneously comparing these differences with a set threshold; when the difference exceeds the set threshold, the corresponding cohesion degree is marked as the critical cohesion degree.

[0124] A7. The sampling time corresponding to the anchor critical cohesion is taken as the cluster-level critical time; where the cluster-level critical time represents the time point at which the evolution trajectory of the population state within the cluster begins to diverge significantly.

[0125] It should be noted that the cluster-level critical moment is not the moment when the hypotension event occurs (i.e., the latest sampling moment). Its actual meaning is: when performing a retrospective analysis on the historical dialysis objects included in this physiological state evolution cluster, when tracing back to this moment, the state evolution trajectories of each object begin to diverge significantly; continuing to trace back further, the state evolution patterns between different objects no longer have commonalities.

[0126] In other words, the time interval between the critical moment of this cluster and the moment of the hypotension event reflects the typical common evolutionary period characterized by this physiological state evolution cluster. Therefore, this time interval is defined as the personalized evolutionary window length of the corresponding cluster, used to determine the length of the historical state evolution subsequence to be extracted in real-time early warning, and simultaneously set the prediction interval for future hypotension risk.

[0127] A8. Calculate the time interval between the cluster-level critical moment and the latest sampling moment, and use it as the personalized evolution window length of the target evolving cluster;

[0128] A9. Traverse the K physiological state evolution clusters and repeat the calculation of the personalized evolution window length until the K personalized evolution window lengths are obtained.

[0129] In this embodiment, local sequences before mutation are extracted from historical hypotension events, and physiological state evolution clusters are formed by clustering. The critical moment of each cluster is determined based on the critical point of the cohesion sequence of each cluster. Finally, the time interval from the time of the hypotension event to the time of the occurrence of the event is used as the personalized evolution window length of the corresponding cluster.

[0130] In this embodiment, step A1 specifically includes:

[0131] A1-1. Select D hypotension events from the hypotension event database;

[0132] Specifically, the hypotension event database is an electronic medical record database containing continuously recorded physiological parameters of dialysis subjects during dialysis and labeled with clinically confirmed hypotension events.

[0133] A1-2. Anchor the occurrence times of D hypotension events on the time axis and mark them as the latest sampling time;

[0134] A1-3. Starting from the dialysis initiation time and ending at each latest sampling time, obtain the physiological state vector of the hypotension event at all sampling times.

[0135] A1-4. Arrange the physiological state vectors of all sampling times in ascending chronological order to form D historical state evolution sequences.

[0136] In this embodiment, clinically confirmed hypotension events are selected from the hypotension event database, and all physiological state vectors from the start of dialysis to the time of the event are arranged in chronological order to form D historical state evolution sequences.

[0137] In this embodiment, step A2 specifically includes:

[0138] A2-1. In each historical state evolution sequence, calculate the similarity between adjacent physiological state vectors; where each similarity is associated with the sampling time of the previous physiological state vector;

[0139] Specifically, in this embodiment, the similarity is preferably cosine similarity.

[0140] A2-2. Arrange the similarities in ascending order according to the associated sampling times to form a similarity sequence, resulting in D similarity sequences;

[0141] A2-3. Identify the mutation similarity of each of the D similar sequences;

[0142] Specifically, the mutation similarity represents the critical point at which the similarity between adjacent states undergoes a mutation during the evolution of physiological states. It can be achieved by directly calculating the first-order difference between adjacent similarities and comparing the absolute value of the first-order difference with a set threshold. When the difference is greater than the set threshold, the corresponding similarity is marked as mutation similarity.

[0143] It should be noted that while mutation similarity and critical cohesion are obtained in similar forms, both based on first-order differential mutation detection of sequences, their levels of meaning differ:

[0144] Mutation similarity reflects the critical moment when the stability of an individual trajectory of a single historical dialysis object significantly decreases during its state evolution, and is used to locate the object's pre-mutation vector. Critical cohesion, on the other hand, reflects the common boundary moment when the group trajectories of multiple objects within a physiological state evolution cluster shift from high consistency to significant divergence during backtracking, and is used to define the length of the individual evolution window for that cluster. In other words, mutation similarity is used for individual pattern anchoring, while critical cohesion is used for group commonality anchoring.

[0145] A2-4. Extract the adjacent physiological state vectors corresponding to the mutation similarity, and label the physiological state vector of the previous moment as the vector before mutation.

[0146] In this embodiment, a similarity sequence is generated by calculating the cosine similarity of adjacent physiological state vectors in each historical state evolution sequence, and mutation similarity is identified based on first-order difference mutation detection, thereby marking the previous physiological state vector at the corresponding time as the vector before mutation.

[0147] In this embodiment, step A4 specifically includes:

[0148] A4-1. Vectorize each pre-mutation local sequence to generate D pre-mutation evolution vectors;

[0149] Specifically, in this embodiment, the vectorization can be obtained by mean pooling the H physiological state vectors of the pre-mutation local sequence in the time dimension. Specifically, in N dimensions, the H physiological features in the same dimension are summed and averaged to obtain the pooled physiological features in that dimension; then, the N pooled physiological features are reassembled in their original order to generate a pre-mutation evolution vector of length N.

[0150] Of course, other equivalent vectorization strategies can also be adopted. For example, the physiological state vector at the last moment of the local sequence before mutation can be used directly as its representative vector; or a pre-trained LSTM encoder can be used to encode the local sequence before mutation and extract its hidden state as the evolution vector before mutation; the specific choice can be determined according to the actual data characteristics and model requirements.

[0151] A4-2. From the D pre-mutation evolution vectors, randomly select K pre-mutation evolution vectors as the first round cluster centers;

[0152] A4-3. For any pre-mutation evolution vector, calculate its vector distance with the K first-round cluster centers, and assign it to the first-round cluster center with the smallest vector distance; the vector distance is preferably Euclidean distance.

[0153] A4-4. Traverse all pre-mutation evolution vectors until each pre-mutation evolution vector is assigned to the first-round cluster center with the minimum vector distance, forming K initial physiological state evolution clusters;

[0154] A4-5. Based on all the pre-mutation evolution vectors within each initial physiological state evolution cluster, calculate its cluster center vector and anchor the cluster center vector as the cluster center for the next round.

[0155] A4-6. Calculate the vector distance between the cluster centers of the next round and the cluster centers of the first round;

[0156] A4-7. If the vector distance is greater than the preset threshold, then all pre-mutation evolution vectors are iteratively allocated based on the cluster center in the next round.

[0157] A4-8. Perform the iterative allocation on the D pre-mutation evolution vectors respectively until the distance between all vectors is less than the preset threshold, and obtain K physiological state evolution clusters.

[0158] In this embodiment, time-mean pooling is performed on D pre-mutation local sequences to generate pre-mutation evolution vectors, and an iterative K-means clustering algorithm based on Euclidean distance is used to finally obtain K physiological state evolution clusters.

[0159] In this embodiment, step A5 specifically includes:

[0160] A5-1. Within the K physiological state evolution clusters, select any one target evolution cluster;

[0161] A5-2. Starting from the latest sampling time of the G pre-mutation local sequences contained in the target evolutionary cluster, backtrack towards the historical direction at each sampling time, and extract G local evolutionary sub-sequences with each backtracking time as the endpoint.

[0162] A5-3. Calculate the F subsequence distances between each pair of the G local evolutionary subsequences; where F = G(G-1) / 2;

[0163] The formula for calculating the distance between the subsequences is:

[0164] ;

[0165] in, Represents the distance between subsequences. and Let L represent two local evolutionary subsequences of length L. and Let represent the l-th physiological state vector in each of the two local evolutionary subsequences. L represents the Euclidean distance between two physiological state vectors; L represents the subsequence length (i.e., the number of sampling points within the backtracking window).

[0166] The subsequence distance is calculated by taking the average Euclidean distance (normalized root mean square) between two locally evolved subsequences at all sampling times. It is used to quantify the overall difference between the two trajectories on the state trajectory. The smaller the value, the closer the two trajectories are; the larger the value, the more obvious the trajectory divergence.

[0167] A5-4. Obtain the mean and standard deviation of the distances between F subsequences;

[0168] A5-5. Based on the mean and standard deviation of the distances between the F subsequences, calculate the cohesion that characterizes the population consistency of the G locally evolved subsequences.

[0169] Specifically, in this embodiment, the formula for calculating cohesion is:

[0170] ;

[0171] in, Indicates the degree of cohesion. and σ and σ represent the mean and standard deviation of the subsequence distances, respectively. In other words, cohesion comprehensively reflects the average level and dispersion of distances within the population: when all locally evolved subsequences are highly similar and concentrated, both μ and σ approach 0, and cohesion approaches 1; conversely, if the distances between subsequences are large or their distribution is dispersed, μ or σ increases, and cohesion significantly decreases. Therefore, this cohesion can effectively characterize the population cohesion of physiological state evolution clusters at a specific retrospective moment.

[0172] A5-6. Arrange the cohesion corresponding to each retrospective moment in ascending time order to construct the cohesion sequence of the target evolution cluster;

[0173] A5-7. Traverse the K physiological state evolution clusters and repeat the cohesion sequence construction to form K cohesion sequences.

[0174] In this embodiment, by backtracking step by step from the latest sampling time of each physiological state evolution cluster, the subsequence distance between each pair of corresponding local evolution subsequences is calculated at each backtracking position, and cohesion is generated based on the mean and standard deviation of the distance. Finally, the cohesion of each backtracking time is arranged in chronological order to form K cohesion sequences.

[0175] Furthermore, this embodiment of the invention also discloses a method for constructing a complete evolutionary template sequence, the steps of which include:

[0176] B1. Local sequence of the cluster center of the anchored physiological state evolution cluster;

[0177] It should be noted that the historical dialysis subjects within each physiological state evolution cluster exhibit highly similar state evolution patterns in the local stages before the mutation. Therefore, their total duration from the start of dialysis to the latest sampling time usually shows strong consistency. Based on this, the time range of the retrospective start and end points corresponding to the local sequence at the cluster center can reasonably represent the overall time span of the cluster.

[0178] B2. Calculate the distances between the cluster center local sequence and the G pre-mutation local sequences within the cluster;

[0179] B3. Locate the optimal local sequence based on the minimum distance among the G local sequence distances;

[0180] It should be noted that the cluster center local sequence may correspond to the pre-mutation local sequence of a real historical dialysis object, or it may only be an abstract representation in the feature space; while the whole-process state evolution sequence must be constructed based on the real recorded complete physiological state evolution process. Therefore, it is necessary to select the local sequence with the smallest distance from the cluster center local sequence from the G real pre-mutation local sequences in the cluster as the optimal local sequence, so as to ensure that the constructed whole-process evolution template sequence has real physiological record support and reflects the common evolutionary trajectory of the cluster.

[0181] B4. Anchor the historical state evolution sequence to which the optimal local sequence belongs, and use it as the whole evolution template sequence associated with the physiological state evolution cluster;

[0182] B5. Traverse the K physiological state evolution clusters and repeat the anchoring of the whole evolution template sequence to obtain K whole evolution template sequences.

[0183] In this embodiment, by anchoring the local sequence of the cluster center for each physiological state evolution cluster, and selecting the optimal local sequence with the smallest distance from the local sequence before the actual mutation within the cluster, and using the historical state evolution sequence to which it belongs as the full-process evolution template sequence corresponding to the cluster, K full-process evolution template sequences are finally obtained.

[0184] This invention also provides a time-series physiological signal early warning system for the risk of hypotension during hemodialysis. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that achieve a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0185] like Figure 6 As shown, Figure 6 This is a structural block diagram of the time-series physiological signal early warning system for the risk of hypotension during hemodialysis according to the present invention. The system includes:

[0186] The real-time sequence construction module is used to construct the real-time state evolution sequence of the target object during the dialysis process;

[0187] The similarity calculation module is used to calculate the K sequence similarities between the real-time state evolution sequence and the K pre-constructed full-process evolution template sequences;

[0188] The K whole-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters. The physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events.

[0189] The similarity selection module is used to select the maximum similarity among the K sequence similarities;

[0190] The evolution length extraction module is used to determine: if the maximum similarity is greater than the set evolution similarity threshold, then based on the full evolution template sequence corresponding to the maximum similarity, the associated personalized evolution window length is extracted;

[0191] The subsequence backtracking module is used to backtrack and extract a state evolution subsequence with a length equal to the personalized evolution window length from the real-time state evolution sequence, with the latest sampling time as the endpoint.

[0192] The risk probability output module is used to input the state evolution subsequence into a pre-trained time series classification model and output the risk probability of the target object experiencing a hypotension event within the prediction interval; wherein, the time length of the prediction interval is equal to the length of the personalized evolution window;

[0193] The risk warning module is used to determine whether a low blood pressure risk warning signal is issued if the risk probability is greater than a set warning risk threshold.

[0194] In the above system, a real-time state evolution sequence is constructed using a real-time sequence construction module; the similarity calculation module calculates the similarity of K sequences; wherein, the K whole-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to each of the K physiological state evolution clusters, and the physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events; the maximum similarity is selected using a similarity selection module; the personalized evolution window length is extracted using an evolution length extraction module; the state evolution sub-sequence is truncated using a sub-sequence backtracking module; the risk probability output module outputs the risk probability of hypotension events; and the hypotension risk warning signal is issued using a risk warning module, thus solving the problem of individual adaptation in state pattern matching.

[0195] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for early warning of hypotension risk during hemodialysis based on temporal physiological signals, characterized in that, include: Construct a real-time state evolution sequence of the target object during the dialysis process; Calculate the K sequence similarities between the real-time state evolution sequence and the K pre-constructed full-process evolution template sequences; The K full-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters, and the physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events. Select the maximum similarity among the K sequence similarities; If the maximum similarity is greater than the set evolutionary similarity threshold, then the personalized evolutionary window length associated with the full evolutionary template sequence corresponding to the maximum similarity is extracted. Using the latest sampling time as the endpoint, backtrack and extract a subsequence of state evolution with a length equal to the personalized evolution window length from the real-time state evolution sequence; The state evolution subsequence is input into a pre-trained temporal classification model, which outputs the risk probability of a hypotension event occurring in the target object within the prediction interval; wherein the time length of the prediction interval is equal to the length of the personalized evolution window. If the probability of the risk is greater than the set warning risk threshold, a low blood pressure risk warning signal will be issued.

2. The method for early warning of hypotension risk during hemodialysis according to claim 1, characterized in that, Construct a real-time state evolution sequence of the target object during dialysis, including: Obtain N raw physiological parameters of the target object at the sampling time; The N original physiological parameters are characterized and concatenated into a physiological state vector representing the instantaneous physiological state during dialysis. Using the start time of dialysis as the zero point, the physiological state vector of the target object at each sampling time is continuously collected and updated; The physiological state vectors are arranged in chronological order according to the sampling time to form a real-time state evolution sequence.

3. The method for early warning of hypotension risk during hemodialysis according to claim 1, characterized in that, The method for constructing the personalized evolution window length includes: A1. Obtain the historical state evolution sequence of D hypotension events; where the latest sampling time of the historical state evolution sequence corresponds to the hypotension event. A2. Anchor the pre-mutation vectors of each of the D historical state evolution sequences; A3. Using the pre-mutation vector as the endpoint, extract D pre-mutation local sequences of fixed length H from its corresponding historical state evolution sequence in reverse order. A4. Cluster the D pre-mutation local sequences to obtain K physiological state evolution clusters; where each physiological state evolution cluster contains G pre-mutation local sequences, and G is a variable that is equal or unequal; A5. Construct the cohesion sequences of K physiological state evolution clusters; A6. Identify the critical cohesion degree in the cohesion degree sequence; A7. The sampling time corresponding to the anchor critical cohesion is taken as the cluster-level critical time; where the cluster-level critical time represents the time point at which the evolution trajectory of the population state within the cluster begins to diverge significantly. A8. Calculate the time interval between the cluster-level critical moment and the latest sampling moment, and use it as the personalized evolution window length of the target evolving cluster; A9. Traverse the K physiological state evolution clusters and repeat the calculation of the personalized evolution window length until the K personalized evolution window lengths are obtained.

4. The method for early warning of hypotension risk during hemodialysis according to claim 3, characterized in that, Obtain the historical state evolution sequence of D hypotension events, including: A1-1. Select D hypotension events from the hypotension event database; A1-2. Anchor the occurrence times of D hypotension events on the time axis and mark them as the latest sampling time; A1-3. Starting from the dialysis initiation time and ending at each latest sampling time, obtain the physiological state vector of the hypotension event at all sampling times. A1-4. Arrange the physiological state vectors of all sampling times in ascending chronological order to form D historical state evolution sequences.

5. The method for early warning of hypotension risk during hemodialysis according to claim 3, characterized in that, Anchoring the pre-mutation vectors of each of the D historical state evolution sequences includes: A2-1. In each historical state evolution sequence, calculate the similarity between adjacent physiological state vectors; where each similarity is associated with the sampling time of the previous physiological state vector; A2-2. Arrange the similarities in ascending order according to the associated sampling times to form a similarity sequence, resulting in D similarity sequences; A2-3. Identify the mutation similarity of each of the D similar sequences; A2-4. Extract the adjacent physiological state vectors corresponding to the mutation similarity, and label the physiological state vector of the previous moment as the vector before mutation.

6. The method for early warning of hypotension risk during hemodialysis according to claim 3, characterized in that, Clustering the D pre-mutation local sequences yields K physiological state evolution clusters, including: A4-1. Vectorize each pre-mutation local sequence to generate D pre-mutation evolution vectors; A4-2. From the D pre-mutation evolution vectors, randomly select K pre-mutation evolution vectors as the first round cluster centers; A4-3. For any pre-mutation evolution vector, calculate its vector distance to the K first-round cluster centers, and assign it to the first-round cluster center with the smallest vector distance. A4-4. Traverse all pre-mutation evolution vectors until each pre-mutation evolution vector is assigned to the first-round cluster center with the minimum vector distance, forming K initial physiological state evolution clusters; A4-5. Based on all the pre-mutation evolution vectors within each initial physiological state evolution cluster, calculate its cluster center vector and anchor the cluster center vector as the cluster center for the next round. A4-6. Calculate the vector distance between the cluster centers of the next round and the cluster centers of the first round; A4-7. If the vector distance is greater than the preset threshold, then all pre-mutation evolution vectors are iteratively allocated based on the cluster center in the next round. A4-8. Perform the iterative allocation on the D pre-mutation evolution vectors respectively until the distance between all vectors is less than the preset threshold, and obtain K physiological state evolution clusters.

7. The method for early warning of hypotension risk during hemodialysis according to claim 3, characterized in that, Construct the cohesion sequences for each of the K physiological state evolution clusters, including: A5-1. Within the K physiological state evolution clusters, select any one target evolution cluster; A5-2. Starting from the latest sampling time of the G pre-mutation local sequences contained in the target evolutionary cluster, backtrack towards the historical direction at each sampling time, and extract G local evolutionary sub-sequences with each backtracking time as the endpoint. A5-3. Calculate the F subsequence distances between each pair of the G local evolutionary subsequences; where F = G(G-1) / 2; A5-4. Obtain the mean and standard deviation of the distances between F subsequences; A5-5. Based on the mean and standard deviation of the distances between the F subsequences, calculate the cohesion that characterizes the population consistency of the G locally evolved subsequences. A5-6. Arrange the cohesion corresponding to each retrospective moment in ascending time order to construct the cohesion sequence of the target evolution cluster; A5-7. Traverse the K physiological state evolution clusters and repeat the cohesion sequence construction to form K cohesion sequences.

8. The method for early warning of hypotension risk during hemodialysis according to claim 7, characterized in that, The steps for constructing the full-process evolution template sequence include: B1. Local sequence of the cluster center of the anchored physiological state evolution cluster; B2. Calculate the distances between the cluster center local sequence and the G pre-mutation local sequences within the cluster; B3. Locate the optimal local sequence based on the minimum distance among the G local sequence distances; B4. Anchor the historical state evolution sequence to which the optimal local sequence belongs, and use it as the whole evolution template sequence associated with the physiological state evolution cluster; B5. Traverse the K physiological state evolution clusters and repeat the anchoring of the whole evolution template sequence to obtain K whole evolution template sequences.

9. A time-series physiological signal early warning system for the risk of hypotension during hemodialysis, characterized in that, include: The real-time sequence construction module is used to construct the real-time state evolution sequence of the target object during the dialysis process; The similarity calculation module is used to calculate the K sequence similarities between the real-time state evolution sequence and the K pre-constructed full-process evolution template sequences; The K whole-process evolution template sequences are obtained by backtracking and completing the pre-mutation local sequences corresponding to the K physiological state evolution clusters. The physiological state evolution clusters are obtained by clustering the pre-mutation local sequences of historical hypotension events. The similarity selection module is used to select the maximum similarity among the K sequence similarities; The evolution length extraction module is used to determine: if the maximum similarity is greater than the set evolution similarity threshold, then based on the full evolution template sequence corresponding to the maximum similarity, the associated personalized evolution window length is extracted; The subsequence backtracking module is used to backtrack and extract a state evolution subsequence with a length equal to the personalized evolution window length from the real-time state evolution sequence, with the latest sampling time as the endpoint. The risk probability output module is used to input the state evolution subsequence into a pre-trained time series classification model and output the risk probability of the target object experiencing a hypotension event within the prediction interval; wherein, the time length of the prediction interval is equal to the length of the personalized evolution window; The risk warning module is used to determine whether a low blood pressure risk warning signal is issued if the risk probability is greater than a set warning risk threshold.

Images