A method for evaluating acute kidney injury induced by rhabdomyolysis
By optimizing the distance measurement and clustering process, the hidden Markov model was trained, and the problem of poor prediction ability caused by the uneven distribution of K-means clustering algorithm in the multi-dimensional detection data space was solved, and the risk of acute renal injury induced by rhabdomyolysis was achieved Accuracy and stability of the assessment.
Patent Information
- Application Number
- CN202411328105.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-09-24
AI Technical Summary
The hidden Markov model trained based on the K-means clustering algorithm in the prior art has poor prediction ability under the uneven distribution of multidimensional detection data space of rhabdomyolysis patients, and cannot accurately evaluate the risk of acute renal injury inducing rhabdomyolysis.
By obtaining N-dimensional detection index data of rhabdomyolysis patients, building N-dimensional data space, optimizing distance measurement, obtaining clustering distance, performing clustering division, and training hidden Markov models to improve prediction capabilities.
The prediction ability of the Hidden Markov model is improved, ensuring the accuracy of the risk assessment of acute renal injury induced by rhabdomyolysis, and able to capture status transitions in a timely manner to provide stable and reliable risk assessment.
Smart Images

Figure CN118866388B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical data analysis, and in particular to a method for evaluating acute kidney injury induced by rhabdomyolysis. Background Art
[0002] For patients with rhabdomyolysis, there is a risk of acute kidney injury induced by rhabdomyolysis. Therefore, it is necessary to predict the risk of rhabdomyolysis-induced acute kidney injury through various detection indicators, and to use the risk situation of acute kidney injury as a reference to optimize the treatment process of patients with rhabdomyolysis.
[0003] In the prior art, when performing an acute kidney injury risk assessment on a patient with rhabdomyolysis, a hidden Markov model is trained using the various test data of the patient with rhabdomyolysis, and the trained hidden Markov model is used to evaluate the hidden state transition probability of the real-time test data of the patient with rhabdomyolysis, and the probability of transitioning to an acute kidney injury state is used as the risk prediction result for the patient with rhabdomyolysis. In the training process of the hidden Markov model, the various test data of the patient with rhabdomyolysis are first divided into hidden states using a K-means clustering algorithm, and the initial state probability, state transition probability, and observation probability in the hidden Markov model are obtained based on the hidden state division results, and then the model parameters in the hidden Markov model are set based on the obtained results to obtain a trained hidden Markov model.
[0004] It is worth noting that the initial state probability refers to the number of samples in each cluster divided by the total number of samples in the data space; the state transition probability refers to the transition frequency of each state in the test data corresponding to the i-th rhabdomyolysis patient, and for the state Transfer to state The transition probability from state Transfer to state The number of times divided by the state The total number of times; the observation probability matrix refers to discretizing the test data of patients with rhabdomyolysis, determining the set of observation symbols, for example, dividing indicators such as CK and SCr into several intervals, counting the frequency of different observation symbols in each state, and dividing the frequency of each observation symbol by the total number of observations in that state to obtain the observation probability matrix.
[0005] However, in the process of dividing the hidden states of various test data of patients with rhabdomyolysis through the K-means clustering algorithm, K-means clustering is performed under the assumption that the data distribution is uniform and the clusters are spherical. However, in the multidimensional test data space of patients with rhabdomyolysis, the distribution is usually uneven, which will lead to deviations in the K-means clustering results, resulting in poor predictive ability of the trained hidden Markov model, and thus unable to accurately predict the risk of rhabdomyolysis-induced acute kidney injury.
[0006] Therefore, how to improve the predictive ability of the trained hidden Markov model obtained based on the K-means clustering algorithm to ensure the accuracy of risk assessment of rhabdomyolysis-induced acute kidney injury has become an urgent problem to be solved. Summary of the invention
[0007] In view of this, an embodiment of the present invention provides a method for assessing acute kidney injury induced by rhabdomyolysis to solve the problem of how to improve the predictive ability of a trained hidden Markov model obtained based on a K-means clustering algorithm to ensure the accuracy of risk assessment of acute kidney injury induced by rhabdomyolysis.
[0008] An embodiment of the present invention provides a method for evaluating acute kidney injury induced by rhabdomyolysis, the method comprising the following steps:
[0009] Obtaining N-dimensional detection index data of each rhabdomyolysis patient, and correspondingly obtaining N-dimensional detection index time series data of each rhabdomyolysis patient;
[0010] After standardizing all N-dimensional detection indicator time series data, construct an N-dimensional data space of all N-dimensional detection indicator time series data, wherein one N-dimensional detection indicator data corresponds to one data point in the N-dimensional data space, and for any two data points in the N-dimensional data space, obtain an initial distance measurement between the two data points according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space;
[0011] According to the volatility and continuity of the two data points in the N-dimensional data space, a distance optimization factor between the two data points is obtained, and the initial distance metric is optimized using the distance optimization factor to obtain a cluster distance between the two data points;
[0012] According to the clustering distance between every two data points in the N-dimensional data space, all data points are clustered to obtain a preset number of clusters. According to the acute kidney injury risk status corresponding to each cluster, the hidden Markov model is trained to obtain a trained hidden Markov model. The trained hidden Markov model is used to perform real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis.
[0013] Preferably, obtaining the initial distance metric between the two data points according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space includes:
[0014] Obtaining the Euclidean distance between every two data points in the N-dimensional data space, obtaining a maximum Euclidean distance and a minimum Euclidean distance, obtaining M scales according to the maximum Euclidean distance and the minimum Euclidean distance, and setting a weight for each scale;
[0015] For any data point of the two data points, respectively obtain the local density of the data point at each scale, use a logarithmic function with a preset value as the base to map the sum between a constant 1 and each local density to obtain a corresponding mapping value, and perform weighted summation on all mapping values according to the weight of each scale to obtain the local data distribution complexity of the data point;
[0016] According to the local data distribution complexity corresponding to the two data points, the mean of the local data distribution complexity is calculated as the local data distribution difference index between the two data points; according to the feature importance of the two data points in different dimensions, the overall feature difference index is obtained;
[0017] The product of the local data distribution difference index and the overall feature difference index between the two data points is squared twice, and the corresponding result is used as the initial distance metric between the two data points.
[0018] Preferably, acquiring M scales according to the maximum Euclidean distance and the minimum Euclidean distance includes:
[0019]
[0020] in, represents the i-th scale, represents the minimum Euclidean distance, represents the maximum Euclidean distance, i<6.
[0021] Preferably, obtaining the overall feature difference index according to the feature importance of the two data points in different dimensions includes:
[0022] For any dimension of the two data points, according to the N-dimensional detection index data corresponding to the two data points, respectively obtain the information entropy of each dimension, use the exponential function with a natural constant as the base to normalize the information entropy of the dimension to obtain the corresponding normalized result, and obtain the feature importance index of the dimension according to the ratio between the normalized result and the accumulated value of the information entropy of all dimensions;
[0023] Obtain the square result of the numerical difference between the two data points in the dimension, and obtain the multiplication result of the feature importance index and the square result as the feature difference index of the two data points in the dimension; accumulate the feature difference index of the two data points in each dimension to obtain the overall feature difference index between the two data points.
[0024] Preferably, obtaining the distance optimization factor between the two data points according to the volatility and continuity of the two data points in the N-dimensional data space includes:
[0025] Obtaining a timestamp difference between the two data points, calculating an absolute value of a ratio between the timestamp difference and a preset time scale parameter, and performing negative mapping on the absolute value of the ratio using an exponential function with a natural constant as a base to obtain a corresponding mapping result;
[0026] Obtaining a JS divergence between the two data points, obtaining a difference between a constant 1 and the JS divergence, and taking a product of the difference and the mapping result as a correlation index between the two data points;
[0027] According to the correlation index between the two data points and the numerical difference between the two data points in each dimension, a distance optimization factor between the two data points is obtained.
[0028] Preferably, obtaining the distance optimization factor between the two data points according to the correlation index between the two data points and the numerical difference between the two data points in each dimension includes:
[0029] Obtaining the dynamic weight of each dimension, using the dynamic weight of each dimension, performing weighted summation on the absolute value of the numerical difference between the two data points in each dimension to obtain the characteristic difference degree between the two data points, normalizing the product of the correlation index and the characteristic difference degree between the two data points to obtain the dynamic transfer characteristic index between the two data points;
[0030] The difference between the constant 1 and the dynamic transfer characteristic index between the two data points is obtained as a distance optimization factor between the two data points.
[0031] Preferably, the optimizing the initial distance metric by using the distance optimization factor to obtain the cluster distance between the two data points includes:
[0032] The product of the distance optimization factor and the initial distance metric is taken as the cluster distance between the two data points.
[0033] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:
[0034] The present invention obtains N-dimensional detection index data of each rhabdomyolysis patient, and correspondingly obtains N-dimensional detection index time series data of each rhabdomyolysis patient; constructs an N-dimensional data space of the N-dimensional detection index time series data of all rhabdomyolysis patients, one N-dimensional detection index data corresponds to one data point in the N-dimensional data space, and for any two data points in the N-dimensional data space, obtains an initial distance metric between the two data points according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space; obtains a distance optimization factor between the two data points according to the volatility and continuity of the two data points in the N-dimensional data space, optimizes the initial distance metric using the distance optimization factor, and obtains a clustering distance between the two data points; clusters all data points according to the clustering distance between every two data points in the N-dimensional data space to obtain a preset number of clusters, trains a hidden Markov model according to the acute kidney injury risk state corresponding to each cluster, obtains a trained hidden Markov model, and uses the trained hidden Markov model to perform real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis. Among them, the distance metric in the clustering process is initially optimized through the data point distribution information of each rhabdomyolysis patient in the N-dimensional data space to obtain the initial distance metric. On the basis of the initial distance metric, the distance metric in the clustering process is further optimized through the fluctuation and continuous changes of the detection index data presented by the rhabdomyolysis patients to filter out the impact of short-term detection fluctuations, and the corresponding clustering distance in the clustering process is obtained. The clustering distance can accurately track the asymptotic changes in the state of rhabdomyolysis patients and can capture important state transitions in time, so that the acute kidney injury risk state represented by the clustering clusters divided according to the clustering distance is more stable and reliable, thereby improving the effect of training the hidden Markov model according to the acute kidney injury risk state corresponding to each cluster, making the prediction ability of the trained hidden Markov model more rigorous, and ensuring the accuracy of risk assessment of rhabdomyolysis-induced acute kidney injury. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0036] Figure 1 This is a method flow chart of a method for evaluating acute kidney injury induced by rhabdomyolysis provided in Example 1 of the present invention. DETAILED DESCRIPTION
[0037] Embodiments of the present disclosure are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present disclosure, but should not be understood as limiting the present disclosure.
[0038] It should be noted that the terms "first", "second", etc. in the specification of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure.
[0039] In order to illustrate the technical solution of the present invention, specific embodiments are provided below for illustration.
[0040] See also Figure 1 , is a method flow chart of a method for evaluating acute kidney injury induced by rhabdomyolysis provided in Example 1 of the present invention, such as Figure 1 As shown, the method may include:
[0041] Step S101, obtaining N-dimensional detection index data of each rhabdomyolysis patient, and correspondingly obtaining N-dimensional detection index time series data of each rhabdomyolysis patient.
[0042] The N-dimensional detection index data of each rhabdomyolysis patient is obtained in the patient medical data system, where N is greater than or equal to 2, and the N-dimensional detection index data of patients with rhabdomyolysis-induced acute kidney injury are additionally annotated for subsequent model training. For any rhabdomyolysis patient, the N-dimensional detection index data obtained by each test can be timestamped according to the patient's test time, so that all the patient's test data can be formed into an N-dimensional detection index time series data, and the N-dimensional detection index time series data is obtained by arranging multiple N-dimensional detection index data in chronological order. In this way, the N-dimensional detection index time series data of each rhabdomyolysis patient can be obtained.
[0043] It should be noted that in the process of acquiring the N-dimensional detection index time series data of each rhabdomyolysis patient, the relevant information of each rhabdomyolysis patient needs to be desensitized. The desensitization process is to encrypt or blur the patient's personal information through encryption means to ensure the patient's information security.
[0044] At this point, the N-dimensional detection index time series data of each rhabdomyolysis patient is obtained.
[0045] Step S102, after standardizing all N-dimensional detection indicator time series data, construct an N-dimensional data space of all N-dimensional detection indicator time series data, one N-dimensional detection indicator data corresponds to one data point in the N-dimensional data space, and for any two data points in the N-dimensional data space, obtain the initial distance measurement between the two data points based on the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space.
[0046] In the actual scenario of rhabdomyolysis-induced acute kidney injury, the data distribution of N-dimensional detection indicators of patients with rhabdomyolysis showed obvious uneven characteristics, and this uneven characteristic was mainly reflected in the fact that patients with rhabdomyolysis at low risk of inducing acute kidney injury had more N-dimensional detection indicator data, followed by patients with medium risk of inducing acute kidney injury, and patients with high risk of inducing acute kidney injury had relatively scarce N-dimensional detection indicator data. The above data distribution characteristics are derived from the natural progression of the disease: most patients with rhabdomyolysis may only show mild symptoms and will not develop acute kidney injury; some patients may be at medium risk; only a few patients will progress to a high risk state or actually develop acute kidney injury.
[0047] In the prior art, it is necessary to divide all N-dimensional detection index data of all rhabdomyolysis patients into three corresponding categories according to the above three risk states (high, medium and low) through K-means clustering, which is used as the training set of the hidden Markov model, and use the trained hidden Markov model to perform real-time assessment of acute kidney injury in rhabdomyolysis patients. However, since there are more N-dimensional detection index data of rhabdomyolysis patients with low risk of inducing acute kidney injury, it is easy to mistakenly classify a large number of N-dimensional detection index data of rhabdomyolysis patients with low risk of inducing acute kidney injury into stroke during the clustering process. Or, the rare N-dimensional detection indicator data corresponding to high-risk induced acute kidney injury are completely ignored, and they are mistakenly classified into the N-dimensional detection indicator data corresponding to medium-risk induced acute kidney injury. This erroneous clustering result will directly affect the subsequent training results of the hidden Markov model based on the clustering results, that is, the prediction accuracy of the trained hidden Markov model, which will lead to the underestimation of patients with rhabdomyolysis corresponding to high-risk induced acute kidney injury, thereby delaying treatment.
[0048] Therefore, in an embodiment of the present invention, after obtaining the N-dimensional detection index time series data of each rhabdomyolysis patient, the detection index data of all rhabdomyolysis patients are standardized so that the numerical ranges of different detection index data are unified to avoid dimensional interference, thereby ensuring that the distance metric in the subsequently constructed multidimensional data space is meaningful. Among them, the standardization process belongs to the prior art and will not be repeated here. Afterwards, an N-dimensional data space of all N-dimensional detection index time series data is constructed, and the N-dimensional data space does not include the timestamp information of each N-dimensional detection index data in the N-dimensional detection index time series data, but each detection index data in the N-dimensional detection index data corresponds to a dimension in the N-dimensional data space.
[0049] After constructing the N-dimensional data space, each N-dimensional detection indicator data in the N-dimensional detection indicator time series data corresponds to a data point in the N-dimensional data space, and then for any two data points in the N-dimensional data space, according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space, the initial distance measurement between the two data points is obtained, which is used to cluster all data points in the N-dimensional data space to divide the N-dimensional detection indicator time series data of all patients with rhabdomyolysis into different risk states.
[0050] It is worth noting that the risk status includes low-risk status, medium-risk status, and high-risk status, and the low-risk status is for the test data corresponding to low-risk induced acute kidney injury, the medium-risk status is for the test data corresponding to medium-risk induced acute kidney injury, and the high-risk status is for the test data corresponding to high-risk induced acute kidney injury.
[0051] Among them, according to the spatial distribution characteristics of two data points at different scales in the N-dimensional data space, the specific process of obtaining the initial distance measurement between two data points is as follows:
[0052] (1) Obtaining the Euclidean distance between every two data points in the N-dimensional data space, obtaining a maximum Euclidean distance and a minimum Euclidean distance, obtaining M scales according to the maximum Euclidean distance and the minimum Euclidean distance, and setting a weight for each scale.
[0053] Wherein, according to the maximum Euclidean distance and the minimum Euclidean distance, M scales are obtained, including:
[0054]
[0055] in, represents the i-th scale, represents the minimum Euclidean distance, represents the maximum Euclidean distance, i<6.
[0056] In one embodiment, the maximum value of the Euclidean distance is obtained according to the Euclidean distance between every two data points in the N-dimensional data space. and minimum value Will As the basis for selecting multiple scales (that is, radius), set M=5, that is, select 5 scales. Each scale corresponds to different data distribution characteristics. The 5 scales are as follows:
[0057] Minimum scale; Small scale; medium scale; Large scale; Maximum scale.
[0058] It should be noted that the scale selection method is based on geometric series, which can be evenly distributed on a logarithmic scale, covering different ranges from local to global. (Minimum scale) can capture the most local data distribution characteristics, which helps to identify high-risk but rare data points; (Small scale) can reflect the data distribution within a small range, which helps to distinguish patients with similar but different risk levels; (Medium scale) can balance local and global features and reflect the general data distribution; (Large scale): Capturing a larger range of data distribution, which helps understand the transition between different risk states; (Maximum scale): reflects data distribution close to the global level, which helps to understand the overall risk distribution.
[0059] For each scale, a weight is assigned according to its importance. The weight assignment method is:
[0060]
[0061] in, Represents the weight of the qth scale. The value range of q is [1, 5], and 1 represents a constant.
[0062] It is worth noting that this weight distribution method gives higher weights to smaller scales, and local features are generally more important when identifying patients with rhabdomyolysis corresponding to high-risk states.
[0063] (2) For any data point among the two data points, obtain the local density of the data point at each scale respectively, use a logarithmic function with a preset value as the base to map the sum between a constant 1 and each local density respectively to obtain a corresponding mapping value, and perform weighted summation on all mapping values according to the weight of each scale to obtain the local data distribution complexity of the data point.
[0064] In one implementation, taking the j-th data point and the i-th data point as examples, for the i-th data point, the calculation expression of the local data distribution complexity of the i-th data point is:
[0065]
[0066] in, represents the local data distribution complexity of the i-th data point, M represents the number of scales, represents the weight of the qth scale, log() represents the logarithmic function with a constant of 2 as the base, 1 represents a constant, represents the local density of the i-th data point calculated at the q-th scale.
[0067] It is worth noting that local density belongs to the prior art and will not be described in detail here. By considering local density at multiple scales, the complexity of data distribution can be better captured. In the rhabdomyolysis scenario, the N-dimensional detection index data of rhabdomyolysis patients with low risk of inducing acute kidney injury are more, while the N-dimensional detection index data of rhabdomyolysis patients with high risk of inducing acute kidney injury are sparse. Being able to identify such density differences at different scales, for sparse high-risk data points, even if there are fewer surrounding data points, They can also be given higher weights to prevent them from being overlooked. For example, when a patient with high-risk rhabdomyolysis has significantly elevated creatine kinase (CK) and serum creatinine (SCr) values, even if the surrounding data points are sparse, The value of will also be relatively large, which makes it easier for this data point to be correctly identified as high risk during the clustering process.
[0068] Similarly, obtain the local data distribution complexity of the j-th data point.
[0069] (3) According to the local data distribution complexity corresponding to the two data points, the mean of the local data distribution complexity is calculated as the local data distribution difference index between the two data points; according to the feature importance of the two data points in different dimensions, the overall feature difference index is obtained.
[0070] Among them, according to the feature importance of the two data points in different dimensions, the overall feature difference index is obtained, including:
[0071] For any dimension of the two data points, according to the N-dimensional detection index data corresponding to the two data points, respectively obtain the information entropy of each dimension, use the exponential function with a natural constant as the base to normalize the information entropy of the dimension to obtain the corresponding normalized result, and obtain the feature importance index of the dimension according to the ratio between the normalized result and the accumulated value of the information entropy of all dimensions;
[0072] Obtain the square result of the numerical difference between the two data points in the dimension, and obtain the multiplication result of the feature importance index and the square result as the feature difference index of the two data points in the dimension; accumulate the feature difference index of the two data points in each dimension to obtain the overall feature difference index between the two data points.
[0073] In one embodiment, after obtaining the local data distribution complexity of the jth data point and the local data distribution complexity of the ith data point, the mean between the two is calculated. As an indicator of the local data distribution difference between the j-th data point and the i-th data point.
[0074] Furthermore, in the assessment of rhabdomyolysis, different detection index data contribute differently to risk assessment. Therefore, according to the N-dimensional detection index data corresponding to the j-th data point and the i-th data point, the information entropy of each dimension is obtained respectively, and then the feature importance index of the m-th dimension is obtained according to the information entropy. The calculation expression of the feature importance index of the m-th dimension is:
[0075]
[0076] in, represents the feature importance index of the mth dimension, represents the information entropy of the mth dimension, represents a set of dimensions, and e represents a natural constant.
[0077] It should be noted that By calculating the information entropy of the features corresponding to each dimension, important features can be automatically identified and assigned higher weights. For example, CK and SCr may reflect the risk status of patients with rhabdomyolysis better than other detection indicators. These features will be automatically given higher weights, and the higher the feature importance index of the corresponding dimension is.
[0078] Similarly, the feature importance index of each dimension between the jth data point and the ith data point is obtained, and then the overall feature difference index between the jth data point and the ith data point is obtained according to the difference between the jth data point and the ith data point in each dimension. The calculation formula of the overall feature difference index is:
[0079]
[0080] in, represents the overall feature difference index between the j-th data point and the i-th data point, represents the feature importance index of the mth dimension, Represents the data difference between the j-th data point and the i-th data point in the m-th dimension.
[0081] (4) Taking the square root of the product of the local data distribution difference index and the overall feature difference index between the two data points, and using the corresponding result as the initial distance measure between the two data points.
[0082] In one embodiment, the calculation formula of the initial distance metric between the j-th data point and the i-th data point is:
[0083]
[0084] in, represents the initial distance metric between the j-th data point and the i-th data point, represents the overall feature difference index between the j-th data point and the i-th data point, Represents the mean between the local data distribution complexity of the j-th data point and the local data distribution complexity of the ith data point.
[0085] It should be noted that the initial distance metric Combining MLDF and FIW, when calculating the distance between data points, both the unevenness of data distribution and the importance of features are taken into account. Suppose there are two patients A and B, most of their indicators are similar, but A's CK value is significantly higher than B. The traditional Euclidean distance may think that A and B are close, but through the initial distance metric: due to the large FIW of CK (because it is an important indicator), the difference in CK values will be magnified; if A is in a data sparse area (high-risk area), its MLDF value will be large, further increasing the distance from B. As a result, although A and B are similar in most indicators, the initial distance metric can still correctly identify them as patients with different risk states.
[0086] At this point, the initial distance metric between every two data points in the N-dimensional data space can be obtained.
[0087] Step S103, obtaining a distance optimization factor between the two data points according to the volatility and continuity of the two data points in the N-dimensional data space, optimizing the initial distance metric using the distance optimization factor, and obtaining a clustering distance between the two data points.
[0088] After obtaining the initial distance measurement between every two data points in the N-dimensional data space, since there are also N-dimensional detection index time series data of any rhabdomyolysis patient in the N-dimensional data space, and the state change of each rhabdomyolysis patient must be stable, and the initial distance measurement only solves the problem of uneven data distribution, in actual medical scenarios, the health status of rhabdomyolysis patients usually does not fluctuate violently in a short period of time, but shows a certain continuity and asymptotic changes. If the distribution optimization is based only on the N-dimensional data space, this important feature may be ignored, resulting in the continuous detection data indicators of the same rhabdomyolysis patient being incorrectly classified into different risk states, resulting in unreasonable risk state jumps. For example, a patient's CK) and SCr levels may gradually rise within a few days, reflecting the asymptotic process from a low-risk state to a medium-risk state or a high-risk state.
[0089] However, if the clustering distance only considers the data distribution at a single time point during the clustering process, this asymptotic process may be misjudged as a discrete state jump, resulting in unstable and inaccurate risk assessment. Therefore, it is necessary to avoid continuous state changes in the N-dimensional test data of a single rhabdomyolysis patient, and use this as the distance limit in the clustering process to optimize the initial distance measurement of the clustering process, thereby avoiding inaccurate risk status assessment after clustering division.
[0090] Based on the above feature analysis, in an embodiment of the present invention, for any two data points in the N-dimensional data space, first, a distance optimization factor between the two data points is obtained according to the volatility and continuity of the two data points in the N-dimensional data space, and then, the distance optimization factor is used to optimize the initial distance metric to obtain the clustering distance between the two data points, so as to reduce the erroneous clustering division caused by data fluctuations, and at the same time, it can timely capture important risk status changes.
[0091] Wherein, obtaining the distance optimization factor between the two data points according to the volatility and continuity of the two data points in the N-dimensional data space includes:
[0092] (1) Obtain the timestamp difference between the two data points, calculate the absolute value of the ratio between the timestamp difference and a preset time scale parameter, and negatively map the absolute value of the ratio using an exponential function with a natural constant as the base to obtain a corresponding mapping result; obtain the JS divergence between the two data points, obtain the difference between a constant 1 and the JS divergence, and use the product of the difference and the mapping result as a correlation indicator between the two data points.
[0093] In one embodiment, taking the jth data point and the ith data point as an example, the timestamp difference is obtained according to the timestamps of the jth data point and the ith data point. At the same time, obtain the JS divergence between the jth data point and the i-th data point, recorded as JS divergence belongs to the prior art and will not be described here. Then, according to the timestamp difference and JS divergence between the jth data point and the ith data point, the correlation index between the jth data point and the ith data point is obtained, and the calculation expression of the correlation index is:
[0094]
[0095] in, represents the correlation index between the j-th data point and the i-th data point, represents the timestamp difference between the jth data point and the ith data point, Represents a time scale parameter, which is used to control the speed of time decay. , e represents a natural constant, | | represents an absolute value symbol, 1 represents a constant, represents the JS divergence between the j-th data point and the i-th data point.
[0096] It should be noted that Taking time continuity into consideration, the smaller the timestamp difference, the closer the value of this part is to 1, and the stronger the correlation; The similarity of feature distribution is taken into account. The more similar the distribution is, the smaller the JS divergence is, the closer the value of this part is to 1, and the stronger the correlation between the jth data point and the ith data point is. In the scenario of rhabdomyolysis, this means that not only the proximity of the detection time is taken into account, but also the similarity of the indicator changes. For example, if a patient's CK and SCr show a similar upward trend in two similar tests, even if the absolute values are different, TFCF will give a higher value, indicating that there is a stronger association between the two states, and the corresponding correlation index is larger.
[0097] TFCF takes into account the temporal continuity and similarity of feature distribution. In practical applications, this means that the model not only pays attention to the time interval of the test data, but also compares whether the feature distributions at different time points are similar. For example, if a patient has a small increase in CK and SCr levels at two adjacent test time points, TFCF will believe that there is a high correlation between the two data points and tend to classify them into the same or similar risk levels.
[0098] (2) Obtaining a distance optimization factor between the two data points according to a correlation index between the two data points and a numerical difference between the two data points in each dimension.
[0099] Specifically, the dynamic weight of each dimension is obtained, and the absolute value of the numerical difference between the two data points in each dimension is weighted and summed using the dynamic weight of each dimension to obtain the characteristic difference degree between the two data points, and the product of the correlation index and the characteristic difference degree between the two data points is normalized to obtain the dynamic transfer characteristic index between the two data points;
[0100] The difference between the constant 1 and the dynamic transfer characteristic index between the two data points is obtained as a distance optimization factor between the two data points.
[0101] In one embodiment, the calculation expression of the dynamic transfer characteristic index between the j-th data point and the i-th data point is:
[0102]
[0103] in, Represents the dynamic transfer characteristic index between the j-th data point and the i-th data point, represents the correlation index between the jth data point and the ith data point, softmax() represents the normalized exponential function, represents the data difference between the j-th data point and the i-th data point in the m-th dimension, || represents the absolute value symbol, and N represents the number of dimensions. Represents the dynamic weight of the mth dimension.
[0104] It should be noted that DSTP further introduces the concept of dynamic weights based on TFCF, so that the subsequent hidden Markov model can adapt to the speed of change of risk status. In the process of rhabdomyolysis, there may be different key indicators at different stages. For example, in the early stage, the change of CK may be more important. At this time, the dynamic weight In the later stage, the change of SCr may better reflect the deterioration of renal function. DSTP dynamically adjusts the weight of features so that the importance of such changes in data indicators can be captured during clustering, thereby more accurately assessing the patient's risk status. DSTP combines TFCF with feature differences and The function converts the results into probabilities, so that we can get a measure that reflects the likelihood of state transitions.
[0105] Among them, the dynamic weight of each dimension is updated through a dynamic learning algorithm. The dynamic learning algorithm is an existing technology and will not be described here.
[0106] According to the dynamic transfer characteristic index between the jth data point and the ith data point, the calculation expression for obtaining the distance optimization factor between the jth data point and the ith data point is:
[0107]
[0108] in, represents the distance optimization factor between the jth data point and the ith data point, 1 represents a constant, Represents the dynamic transfer characteristic index between the j-th data point and the i-th data point.
[0109] It should be noted that the larger the dynamic transfer characteristic index between the j-th data point and the i-th data point, the greater the possibility of the corresponding risk state transition between the j-th data point and the i-th data point, indicating that the risk state change between the two data points belongs to a smooth state transition, and the corresponding distance optimization factor should be smaller.
[0110] The initial distance metric is optimized by using the distance optimization factor to obtain the cluster distance between the two data points, including:
[0111] The product of the distance optimization factor and the initial distance metric is taken as the cluster distance between the two data points.
[0112] In one embodiment, the calculation formula for the cluster distance between the j-th data point and the i-th data point is:
[0113]
[0114] in, represents the clustering distance between the jth data point and the ith data point, represents the distance optimization factor between the jth data point and the ith data point, represents the initial distance metric between the j-th data point and the i-th data point.
[0115] It should be noted that the clustering distance Combining the spatial distribution optimization of step S102 and the time-feature dynamic optimization of step 103, in actual scenarios, the role of this clustering distance can be explained by the following example:
[0116] Assume that patient A has three consecutive days of test data A1, A2, and A3. If both CK and SCr increase slightly from A1 to A2, TFCF will give a higher value and DSTP will also be larger, resulting in a clustering distance between A1 and A2. If the CK and SCr of A2 to A3 suddenly increase significantly, even if the time is close, the TFCF value will decrease due to the large change in feature distribution, and the DSTP will also decrease accordingly, resulting in a clustering distance of Relatively large, reflecting a significant change in risk status.
[0117] Therefore, the optimized cluster distance can effectively capture the asymptotic changes in the patient's risk status and avoid drastic fluctuations in risk assessment caused by short-term fluctuations. And the effect of this optimization in practical applications can be illustrated by the following example: Assume that a patient's CK level begins to rise slowly, but other indicators remain stable for the time being. In the traditional clustering process, the patient may be classified as high-risk when CK begins to rise, resulting in unstable risk assessment. The cluster distance obtained by the distance optimization factor will take into account the gradualness of this change. The patient may be classified as medium-risk first, and as CK continues to rise and other indicators (such as SCr) begin to change, it will gradually adjust to a high-risk state. This smooth state transition is more in line with the actual progression of the disease and avoids drastic fluctuations in risk assessment.
[0118] It is worth mentioning that another important application scenario of cluster distance is to deal with noise and fluctuations in test data. In actual medical environments, test results may be affected by a variety of factors and experience short-term fluctuations. The optimized cluster distance can effectively filter out these short-term fluctuations by considering the time continuity and the comprehensive changes of multiple features, providing a more stable and reliable risk assessment.
[0119] At this point, the clustering distance between every two data points in the N-dimensional data space can be obtained.
[0120] Step S104, clustering all data points according to the clustering distance between every two data points in the N-dimensional data space to obtain a preset number of clusters, training the hidden Markov model according to the acute kidney injury risk status corresponding to each cluster to obtain a trained hidden Markov model, and using the trained hidden Markov model to perform real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis.
[0121] The clustering distance effectively solves the problem of temporal continuity and dynamic changes of patient status by introducing temporal continuity, feature similarity and dynamic feature importance. This method enables the clustering process to more accurately track the gradual changes of patient status, reduces misclassification caused by data fluctuations, and can capture important state transitions in a timely manner. In the assessment of rhabdomyolysis-induced acute kidney injury, this optimization makes risk assessment smoother and more reliable, better reflects the actual progression of the disease, and provides more reliable support for clinical decision-making. Therefore, after obtaining the clustering distance between every two data points in the N-dimensional data space, the K-means clustering algorithm is used to cluster all data points according to the clustering distance between every two data points in the N-dimensional data space to obtain three clusters.
[0122] After the clustering process is completed, the three clusters obtained are used to determine the high, medium and low risk states according to the N-dimensional detection indicator data in each cluster. One cluster corresponds to one risk state, and each N-dimensional detection indicator data in the corresponding cluster corresponds to one risk state. It should be noted that the risk state refers to the risk level of acute kidney injury induced by rhabdomyolysis patients, and the risk level includes high risk, low risk and medium risk.
[0123] After classifying the data points in the N-dimensional data space, the hidden Markov model can be trained according to the acute kidney injury risk status corresponding to each cluster (the risk level of acute kidney injury induced by rhabdomyolysis patients) to obtain a trained hidden Markov model, wherein the training of the hidden Markov model is an existing well-known technology and will not be repeated here. The trained hidden Markov model is then used to conduct a real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis. Specifically: through the current state of the rhabdomyolysis patient, the transition probability of the current state to the high-risk state of acute kidney injury is obtained according to the hidden Markov model, and the transition probability is used as the risk assessment result for the auxiliary assessment process of the renal injury assessment of rhabdomyolysis patients.
[0124] It should be noted that the focus of the present invention is on training the hidden Markov model. Using the trained hidden Markov model to conduct risk assessment of acute kidney injury induced by rhabdomyolysis belongs to the prior art and will not be described in detail here.
[0125] In summary, the embodiment of the present invention obtains N-dimensional detection index data of each rhabdomyolysis patient, and correspondingly obtains N-dimensional detection index time series data of each rhabdomyolysis patient; constructs an N-dimensional data space of the N-dimensional detection index time series data of all rhabdomyolysis patients, one N-dimensional detection index data corresponds to one data point in the N-dimensional data space, and for any two data points in the N-dimensional data space, according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space, obtains the initial distance metric between the two data points; according to the spatial distribution characteristics of the two data points in the N-dimensional space, obtains the initial distance metric between the two data points; According to the volatility and continuity in the space, a distance optimization factor between the two data points is obtained, and the initial distance metric is optimized using the distance optimization factor to obtain a clustering distance between the two data points; according to the clustering distance between every two data points in the N-dimensional data space, all data points are clustered to obtain a preset number of clusters, and a hidden Markov model is trained according to the acute kidney injury risk state corresponding to each cluster to obtain a trained hidden Markov model, and the trained hidden Markov model is used to perform real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis. Among them, the distance metric in the clustering process is initially optimized through the data point distribution information of each rhabdomyolysis patient in the N-dimensional data space to obtain the initial distance metric. On the basis of the initial distance metric, the distance metric in the clustering process is further optimized through the fluctuation and continuous changes of the detection index data presented by the rhabdomyolysis patients to filter out the impact of short-term detection fluctuations, and the corresponding clustering distance in the clustering process is obtained. The clustering distance can accurately track the asymptotic changes in the state of rhabdomyolysis patients and can capture important state transitions in time, so that the acute kidney injury risk state represented by the clustering clusters divided according to the clustering distance is more stable and reliable, thereby improving the effect of training the hidden Markov model according to the acute kidney injury risk state corresponding to each cluster, making the prediction ability of the trained hidden Markov model more rigorous, and ensuring the accuracy of risk assessment of rhabdomyolysis-induced acute kidney injury.
[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.
Claims
1. A method for evaluating acute kidney injury induced by rhabdomyolysis. It is characterized in that The method comprises: Obtaining N-dimensional detection index data of each rhabdomyolysis patient, and correspondingly obtaining N-dimensional detection index time series data of each rhabdomyolysis patient; After standardizing all N-dimensional detection indicator time series data, construct an N-dimensional data space of all N-dimensional detection indicator time series data, wherein one N-dimensional detection indicator data corresponds to one data point in the N-dimensional data space, and for any two data points in the N-dimensional data space, obtain an initial distance measurement between the two data points according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space; According to the volatility and continuity of the two data points in the N-dimensional data space, a distance optimization factor between the two data points is obtained, and the initial distance metric is optimized using the distance optimization factor to obtain a cluster distance between the two data points; Clustering all data points according to the cluster distance between every two data points in the N-dimensional data space to obtain a preset number of clusters, training a hidden Markov model according to the acute kidney injury risk state corresponding to each cluster to obtain a trained hidden Markov model, and using the trained hidden Markov model to perform real-time assessment of the risk of acute kidney injury induced by rhabdomyolysis; The obtaining the initial distance metric between the two data points according to the spatial distribution characteristics of the two data points at different scales in the N-dimensional data space includes: Obtaining the Euclidean distance between every two data points in the N-dimensional data space, obtaining a maximum Euclidean distance and a minimum Euclidean distance, obtaining M scales according to the maximum Euclidean distance and the minimum Euclidean distance, and setting a weight for each scale; For any data point of the two data points, respectively obtain the local density of the data point at each scale, use a logarithmic function with a preset value as the base to map the sum between a constant 1 and each local density to obtain a corresponding mapping value, and perform weighted summation on all mapping values according to the weight of each scale to obtain the local data distribution complexity of the data point; According to the local data distribution complexity corresponding to the two data points, the mean of the local data distribution complexity is calculated as the local data distribution difference index between the two data points; according to the feature importance of the two data points in different dimensions, the overall feature difference index is obtained; Taking a quadratic square root of the product of the local data distribution difference index and the overall feature difference index between the two data points, and using the corresponding result as the initial distance metric between the two data points; The obtaining M scales according to the maximum Euclidean distance and the minimum Euclidean distance includes: in, represents the i-th scale, represents the minimum Euclidean distance, represents the maximum Euclidean distance, i<6; The obtaining of the overall feature difference index according to the feature importance of the two data points in different dimensions includes: For any dimension of the two data points, according to the N-dimensional detection index data corresponding to the two data points, respectively obtain the information entropy of each dimension, use the exponential function with a natural constant as the base to normalize the information entropy of the dimension to obtain the corresponding normalized result, and obtain the feature importance index of the dimension according to the ratio between the normalized result and the accumulated value of the information entropy of all dimensions; Obtaining the square result of the numerical difference between the two data points in the dimension, obtaining the multiplication result of the feature importance index and the square result as the feature difference index of the two data points in the dimension; accumulating the feature difference index of the two data points in each dimension to obtain the overall feature difference index between the two data points; The step of obtaining a distance optimization factor between the two data points according to the volatility and continuity of the two data points in the N-dimensional data space includes: Obtaining a timestamp difference between the two data points, calculating an absolute value of a ratio between the timestamp difference and a preset time scale parameter, and performing negative mapping on the absolute value of the ratio using an exponential function with a natural constant as a base to obtain a corresponding mapping result; Obtaining a JS divergence between the two data points, obtaining a difference between a constant 1 and the JS divergence, and taking a product of the difference and the mapping result as a correlation index between the two data points; Obtaining the dynamic weight of each dimension, using the dynamic weight of each dimension, performing weighted summation on the absolute value of the numerical difference between the two data points in each dimension to obtain the characteristic difference degree between the two data points, normalizing the product of the correlation index and the characteristic difference degree between the two data points to obtain the dynamic transfer characteristic index between the two data points; Obtaining a difference between a constant 1 and a dynamic transfer characteristic index between the two data points as a distance optimization factor between the two data points; The step of optimizing the initial distance metric by using the distance optimization factor to obtain the cluster distance between the two data points includes: The product of the distance optimization factor and the initial distance metric is taken as the cluster distance between the two data points.
Citation Information
Patent Citations
Hidden Markov model food quality evaluation method based on grey correlation analysis
CN110428121A
Acute kidney injury early prediction method and device
CN113903458A
Cited By
Rhabdomyolysis early warning system based on flexible wearable device
CN120895220A
A rhabdomyolysis early warning system based on a flexible wearable device
CN120895220B
