Intelligent analysis and management method for nursing and physical examination data

Abstract

The invention relates to the technical field of machine learning, in particular to a nursing physical examination data intelligent analysis and management method which comprises the following steps: extracting body temperature, blood pressure and heart rate from a physical examination data set, performing Gaussian kernel mapping processing, and mapping physical examination data into a high-dimensional feature space to obtain high-dimensional feature data; and training a support vector machine model by using the high-dimensional feature data. Through the Gaussian kernel mapping of the high-dimensional feature space and the application of the support vector machine model, the processing capability of the nursing physical examination data is improved, and abnormal and normal data points can be distinguished. And the nonlinear representation of the data is enhanced through mapping, so that the model can process a complex mode more effectively, and outlier and error data are effectively removed. Real-time classification ensures fast processing of data streams, allowing instant updates and responses. And by dynamically marking the severity of the data points and adjusting the processing strategy, the flexibility and individuation of data management are further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for intelligent analysis and management of nursing physical examination data. Background Art

[0002] Machine learning is a core area of ​​artificial intelligence that enables computer systems to improve their performance using experience. Machine learning techniques are primarily based on creating and using algorithms that can learn and recognize patterns from large amounts of data.

[0003] In existing technologies, feature selection and adjustment processes often lack automation and adaptability, which limits the performance of models on new or unseen data. In addition, many existing algorithms have difficulty updating and adjusting model parameters in real time, which may cause models to become outdated in a rapidly changing data environment. For example, in the analysis and management of nursing physical examination data, processing delays may lead to obvious anomalies and erroneous data being entered. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a method for intelligent analysis and management of nursing physical examination data.

[0005] In order to achieve the above object, the present invention adopts the following technical scheme, a method for intelligent analysis and management of nursing physical examination data, comprising the following steps:

[0006] Extracting body temperature, blood pressure and heart rate from the physical examination data set, performing Gaussian kernel mapping processing, mapping the physical examination data into a high-dimensional feature space, and obtaining high-dimensional feature data; using the high-dimensional feature data to train a support vector machine model, adjusting model parameters, and generating a classification model;

[0007] Using the classification model, newly added physical examination data is classified in real time to distinguish normal data from abnormal data, and obtain real-time classification results; abnormal data identified in the real-time classification results are analyzed, and the severity level of each abnormal data point is marked to form an abnormal data marking record;

[0008] According to the abnormal data annotation record, the attribute of each annotated abnormal value is analyzed through decision support, and the processing strategy of each data point is determined, including deletion, correction or retention, to obtain a processing decision result; according to the processing decision result, the abnormal value in the real-time classification result is deleted or modified to generate an updated data set;

[0009] Applying an incremental gradient descent method to the updated data set, optimizing the weights and biases of the classification model, adjusting the learning rate according to the weights and biases, and then adjusting the learning rate according to error feedback in consecutive batches to generate a recognition model.

[0010] Preferably, the steps of acquiring the high-dimensional feature data are:

[0011] Extract body temperature, blood pressure and heart rate information from the physical examination data set, remove noise and outliers, and obtain cleaned physical examination data;

[0012] Based on the cleaned physical examination data, applying Gaussian kernel mapping to perform spatial transformation on the body temperature, blood pressure and heart rate data, and forming data mapped to a high-dimensional feature space by calculating the similarity between each data point and the kernel center, thereby obtaining the mapped physical examination data;

[0013] Based on the mapped physical examination data, the features of each dimension are integrated, and through statistical analysis, key high-dimensional features are extracted to obtain high-dimensional feature data.

[0014] Preferably, the steps of obtaining the classification model are:

[0015] Based on the high-dimensional feature data, start the training of the support vector machine model, the initial configuration includes selecting the kernel function type and setting the value of the regularization parameter, and obtain the initially trained support vector machine model;

[0016] Adjusting model parameters from the initially trained support vector machine model, including optimizing kernel function parameters and regularization parameters by grid search, and evaluating the performance of the model under various parameter configurations by cross validation, to form an adjusted support vector machine model;

[0017] Based on the adjusted support vector machine model, final training is performed to obtain a classification model.

[0018] Preferably, the steps of obtaining the real-time classification results are:

[0019] Deploy the classification model, receive newly added physical examination data in real time, and apply the classification model to classify the data instantly, including calculating each physical examination data point, determining the classification label, and obtaining a preliminary classification result;

[0020] Based on the preliminary classification result, by comparing the predicted label with the known label, adjusting the classification threshold to obtain adjusted classification decision data;

[0021] According to the adjusted classification decision data, each data point is marked as normal or abnormal to obtain a real-time classification result.

[0022] Preferably, the steps of obtaining the abnormal data annotation record are:

[0023] Extract all data points marked as abnormal from the real-time classification results, analyze the position and distribution of the abnormal data points in the multidimensional feature space, and obtain preliminary feature analysis results of the abnormal data points based on the relationship between each data point and its neighboring points;

[0024] Based on the preliminary feature analysis results of the abnormal data points, the severity level of each abnormal data point is calculated using the following formula:

[0025]

[0026] Among them, S i represents the severity level of the ith data point, represents the value of the i-th data point on the k-th feature, μ k is the average value of the kth feature, w k is the weight of feature k, p is the power index, and K represents the total number of features;

[0027] Based on the severity level, each data point is labeled, and the degree of abnormality of each point is recorded to obtain an abnormal data annotation record.

[0028] Preferably, the steps of obtaining the processing decision result are:

[0029] Extracting the abnormal level and attribute information of each data point from the abnormal data annotation record, analyzing the abnormal characteristics and context of the data point, and obtaining a characteristic analysis result of each abnormal data point;

[0030] Based on the characteristic analysis results, the processing priority of each abnormal data point is evaluated, and the calculation formula is:

[0031]

[0032] Among them, P i represents the processing priority of the i-th data point, r i Indicates the severity level, d i It indicates the degree of deviation from normal data, and t is the decisive threshold;

[0033] Based on the processing priority, a processing strategy is formulated for each abnormal data point, including deletion, correction or retention, to obtain a processing decision result.

[0034] Preferably, the steps of acquiring the updated data set are:

[0035] Based on the processing decision results, access the real-time classification results and perform target processing on the specified operations for each marked data point, including deleting data that does not meet the standards or correcting data errors to form a preliminary corrected data set;

[0036] Based on the initially revised data set, the integrity and consistency of the data set are checked to verify that all data points have been correctly processed, thereby obtaining an updated data set.

[0037] Preferably, the steps of obtaining the recognition model are:

[0038] Applying incremental gradient descent to the updated data set includes processing points in the data set one by one, calculating error gradients for each data point, updating model parameters, gradually reducing the error, and obtaining an optimized model parameter set;

[0039] Based on the optimized model parameter set, the learning rate is recalculated using the following formula:

[0040]

[0041] Among them, η represents the new learning rate, η0 is the basic learning rate, γ is the adjustment factor, and e i is the error of the ith data point, and n represents the total number of data points;

[0042] The new learning rate is used to optimize the classification model to obtain the recognition model.

[0043] Compared with the prior art, the advantages and positive effects of the present invention are:

[0044] The present invention improves the processing capability of nursing physical examination data through the application of Gaussian kernel mapping and support vector machine model in high-dimensional feature space, and can distinguish abnormal and normal data points. The nonlinear representation of data is enhanced by mapping, so that the model can process complex patterns more effectively, thereby effectively removing outliers and erroneous data. Real-time classification ensures rapid processing of data streams, allowing instant updates and responses. By dynamically marking the severity of data points and adjusting processing strategies, the flexibility and personalization of data management are further enhanced. The model adaptively adjusts weights and learning rates by applying incremental gradient descent method, and continuously optimizes to adapt to new data features and changes, improves overall efficiency and accuracy, and effectively avoids the entry of erroneous body temperature, blood pressure and heart rate data. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0047] See also Figure 1 The present invention provides a technical solution, a method for intelligent analysis and management of nursing physical examination data, comprising the following steps:

[0048] Body temperature, blood pressure and heart rate are extracted from the physical examination data set, and Gaussian kernel mapping processing is performed to map the physical examination data into a high-dimensional feature space to obtain high-dimensional feature data; the high-dimensional feature data is used to train a support vector machine model, adjust the model parameters, and generate a classification model.

[0049] Use the classification model to classify the newly added physical examination data in real time, distinguish normal data from abnormal data, and obtain real-time classification results; analyze the abnormal data identified in the real-time classification results, mark the severity level of each abnormal data point, and form an abnormal data annotation record.

[0050] According to the abnormal data annotation records, the attributes of each annotated outlier are analyzed through decision support to determine the processing strategy for each data point, including deletion, correction or retention, to obtain the processing decision result; according to the processing decision result, the outliers in the real-time classification results are deleted or modified to generate an updated data set.

[0051] Apply the incremental gradient descent method to the updated dataset to optimize the weights and biases of the classification model, adjust the learning rate based on the weights and biases, and then adjust the learning rate based on the error feedback in consecutive batches to generate the recognition model.

[0052] The steps to obtain high-dimensional feature data are:

[0053] Extract body temperature, blood pressure and heart rate information from the physical examination data set, remove noise and outliers, and obtain cleaned physical examination data;

[0054] Based on the cleaned physical examination data, Gaussian kernel mapping is used to perform spatial transformation on the temperature, blood pressure and heart rate data. By calculating the similarity between each data point and the kernel center, data mapped to the high-dimensional feature space is formed to obtain the mapped physical examination data.

[0055] Based on the mapped physical examination data, the features of each dimension are integrated, and through statistical analysis, the key high-dimensional features are extracted to obtain high-dimensional feature data.

[0056] Specifically, remove noise and outliers. When processing body temperature, compare each measured value with the pre-established effective range of body temperature. For example, 36°C to 39.5°C can be regarded as a reference range. If the body temperature value of a record is lower than 36°C or higher than 39.5°C, the data will be marked as abnormal and combined with the collection environment to check for possible deviations. For blood pressure data, the systolic pressure can be set to 60mmHg to 200mmHg, and the diastolic pressure can be set to 40mmHg to 120mmHg as the numerical range. If the systolic pressure or diastolic pressure is not within this range, the corresponding record will be identified as noise or abnormal. The numerical range can be derived based on past observations and clinical practice. At the same time, a comparison range of 40bpm to 200bpm is set for the heart rate with reference to historical statistics. If it is detected that the heart rate data deviates from this range, the record is extracted separately. For each suspected abnormal record, the temperature, blood pressure and heart rate information of the adjacent time periods will be compared item by item. If it is indeed an extreme value and exceeds the pre-established normal physiological parameters, the data will be deleted or corrected. The method of deletion or correction is determined by established empirical rules or based on further judgment of the contextual data. After completing this process, a more consistent processing result is obtained.

[0057] Gaussian kernel mapping is applied to spatially transform the temperature, blood pressure and heart rate data. First, the sample vector is determined for each cleaned record, and the temperature, blood pressure and heart rate are combined in a fixed order. Then a kernel function form is selected, such as Among them, σ can usually be determined based on the standard deviation of the current sample distribution or the empirical value obtained from previous statistics. When calculating the similarity between each record and the specified kernel center, the temperature, blood pressure and heart rate values ​​of the record are read one by one, and the exponential operation is completed according to the selected kernel function. The similarity results are then summarized to form a new feature representation. At this time, if the similarity result of any record deviates significantly from the overall level, the background attributes of the record can be combined to analyze whether repeated confirmation is needed. If there are no special circumstances, the mapping results are integrated into a high-dimensional matrix, and the sequence index is maintained so that it can correspond one-to-one with the original record in subsequent links. In this way, a mapping set in a high-dimensional feature space can be obtained.

[0058] Integrate the features of each dimension, extract key high-dimensional features through statistical analysis, first perform descriptive statistics on the high-dimensional mapping matrix obtained above, such as calculating the data distribution, extreme value and dispersion of each dimension, and then use common correlation measures to compare the multidimensional features in pairs to determine whether each feature has a significant impact on the subsequent model training. If some features are extremely concentrated and have a very low correlation with other features, they may be considered not to be selected or combined with other indicators for auxiliary judgment. In this process, typical dimensionality reduction methods can also be used, such as retaining several principal components with higher contribution in the inter-column covariance matrix. These principal components can be selected in turn after sorting the feature vectors. At the same time, the remaining features are screened. If it is found that the feature is abnormally distributed in different recording scenarios, the feature will be included in the special review scope and continue to be paid attention to in subsequent steps. In the statistical analysis process, the accumulated mean and variance of continuous observation data can be used, and the reliability of each dimension can be evaluated in combination with the normal physiological range listed above. Finally, the features that meet the selection criteria are numbered and output as a key feature set.

[0059] The steps to obtain the classification model are:

[0060] Based on the high-dimensional feature data, the training of the support vector machine model is started. The initial configuration includes selecting the kernel function type and setting the value of the regularization parameter to obtain the initially trained support vector machine model.

[0061] From the initially trained support vector machine model, perform model parameter adjustments, including optimizing kernel function parameters and regularization parameters through grid search, and use cross-validation to evaluate the performance of the model under various parameter configurations to form an adjusted support vector machine model;

[0062] Based on the adjusted support vector machine model, final training is performed to obtain a classification model.

[0063] Specifically, based on high-dimensional feature data, first clarify the feature dimensions contained in each record and number them accordingly, then load all records in a given order to form a training sample set, and on this basis, combine past experience to select the kernel function type as the radial basis function. If the kernel function needs to be changed, it needs to be evaluated separately during subsequent parameter fine-tuning. Then, a numerical range from 1 to 100 is preset for the regularization parameter, which can be determined based on the sample size and statistical analysis results collected previously. If in some cases the sample distribution is extremely dispersed or concentrated, consider expanding or shrinking the range. Then, before starting training, input all training samples one by one into the initialized support vector machine calculation process. During calculation, the initial form of the discrimination boundary needs to be obtained through vector inner product and kernel function transformation. , and record the classification scores corresponding to each record in the memory. If any record has an abnormally high or low classification score during this process, it is necessary to back-check the characteristics of the record and refer to the known dimension distribution interval to evaluate whether it belongs to the normal range. When it is found that it is indeed out of the range, the record can be temporarily marked for subsequent processing. The majority of other records maintain the normal classification path. After all samples are calculated, a complete training iteration information will be formed, including the current kernel function type and the selected regularization parameter in this round of classification results. The iteration information may include sample misclassification rate, number of support vectors, and classification result labels corresponding to each record. Finally, the iteration information is fixed in this stage to obtain the initially trained support vector machine model.

[0064] From the initial training of the support vector machine model, it is necessary to comprehensively optimize the parameters related to the internal kernel function, such as the value and the regularization parameter. In the specific implementation, the grid search strategy is usually used to enumerate multiple groups of values ​​in the range of 0.01 to 1, and the regularization parameter C is listed in the range of 1 to 100 in a given step size. Each pair of regularization parameter C combinations is cross-validated to observe its classification performance. For example, k=5 can be set for segmentation and then tested one by one. After the cross-validation of each combination is completed, the number of misclassifications, accuracy and other statistical information will be recorded, and then the results of all combinations will be compared. If the best combination exceeds For the pre-selected interval, the interval can be expanded based on the additional data distribution evaluation and the above search process can be repeated. If multiple combinations are found to be similar in terms of indicators, the support vector distribution and the recorded classification confidence are further checked and compared with other available business reference intervals to determine whether there is an overfitting tendency or underfitting signal. If an overfitting tendency is found, the parameter range can be appropriately tightened or the number of training samples can be expanded to enhance stability. After the best combination of C is finally determined, these parameters will be written back to the model configuration, and all cross-validation results and classification results will be recorded together to form an adjusted support vector machine model.

[0065] Based on the adjusted support vector machine model, all available high-dimensional feature data can be reintegrated into a unified training batch during the training phase and read one by one. When each record is input, the kernel function value is calculated using the previously selected value and C value and the label is associated. The predicted output of each sample is compared with its true label, and the comparison information is recorded in the form of classification error or accurate count. The support vector distribution inside the model is continuously updated in combination with continuous iteration. If the classification results of some records are observed to be too different from those in previous rounds in a certain iteration round, it is necessary to check the feature vector of the record and compare whether there are significant outliers. If the record feature does deviate from the normal data distribution, it can be temporarily checked whether it is an error caused by inaccurate prior data labeling. At the same time, the classification error distribution of the current iteration is merged and stored as evaluation information. When all samples have completed the iteration and the classification error can no longer be significantly reduced within the model, it indicates that the current model has converged to a usable state. At this time, the overall training situation is summarized, and the final classification labels and support vector indexes of all records are centrally organized to obtain the classification model.

[0066] The steps to obtain real-time classification results are:

[0067] Deploy the classification model, receive the newly added physical examination data in real time, and apply the classification model to classify the data immediately, including calculating each physical examination data point, determining the classification label, and obtaining the preliminary classification results;

[0068] Based on the preliminary classification results, the classification threshold is adjusted by comparing the predicted label with the known label to obtain the adjusted classification decision data;

[0069] Based on the adjusted classification decision data, each data point is labeled as normal or abnormal to obtain real-time classification results.

[0070] Specifically, the classification model is deployed, the previously obtained classification model is loaded in the specified device or environment and the necessary parameters are initialized, the newly added physical examination data received in real time is parsed one by one and characteristic values ​​such as body temperature, blood pressure and heart rate are extracted. If it is necessary to ensure the integrity of each data, it can be compared to see whether it contains all the required fields, and then the record is immediately input into the classification model and calculated. During the calculation process, the input vector is processed according to the kernel function form used previously and the internal configuration of the support vector machine, and the vector is converted into a discriminant value in a high-dimensional space. If the discriminant value is within an initial critical interval, a predetermined classification label will be output. The interval range can be determined based on the previous The label distribution and empirical statistical settings during training, such as setting the classification boundary at convergence as the basis, refer to whether the body temperature is in the range of 36℃ to 39.5℃, whether the systolic and diastolic blood pressures are in the range of 60mmHg to 200mmHg and 40mmHg to 120mmHg, and whether the heart rate is in the range of 40bpm to 200bpm. If the discriminant value falls within the acceptable range and the corresponding eigenvalue does not exceed the range mentioned above, it can be directly judged as normal. If it exceeds these ranges or the discriminant value deviates from the normal distribution, it can be judged as abnormal. When all input records are judged, the labels corresponding to each record are output to the memory, thereby obtaining a preliminary classification result.

[0071] Based on the preliminary classification results, the predicted labels of each piece of data are obtained and compared with the known labels one by one. The source of the known labels can be the true value information determined by manual or other processes before the new data arrives. By comparison, it can be found that the current classification model may have positive or negative errors in different feature dimensions. If the number of positive errors, that is, the number of samples predicted to be abnormal but marked as normal, is too high in some intervals, the boundary range of the original abnormal judgment can be appropriately reduced. If the number of negative errors, that is, the number of samples predicted to be normal but actually abnormal, is too large in some intervals, the critical threshold range can be statistically adjusted to make it easier to trigger abnormal marking. If the specific value exceeds the judgment balance point previously determined based on the training set, the balance point can be moved in a more suitable direction. For example, the classification confidence is first based on 0.5, and then the confidence is adjusted up or down according to the statistical distribution. The difference is recalculated after each adjustment. If repeated fluctuations occur, multiple comparisons are required to find a more stable critical value. After determining the optimal classification threshold at this time, the threshold will be solidified and saved in the record, and finally the adjusted classification decision data will be obtained.

[0072] According to the adjusted classification decision data, the final judgment threshold corresponding to each data point is read one by one and the key features stored in the record are checked, including whether the body temperature value is in the range of 36℃ to 39.5℃, whether the blood pressure is in the range of 60mmHg to 200mmHg and 40mmHg to 120mmHg, and whether the heart rate is in the range of 40bpm to 200bpm. Then, the final screening is performed in combination with the previously established correction threshold range. If the record is within the correction threshold range and all feature values ​​also meet the set normal range, it will be marked as normal. If the classification score calculated for the record is still higher than the correction threshold and its corresponding feature value has been confirmed to deviate from the normal range, it will be marked as abnormal. During this process, it will be combined with the previously saved classification decision results for consistency verification to prevent repeated labeling conflicts. When all records have completed the above screening process, their labels can be aggregated to obtain real-time classification results.

[0073] The steps to obtain abnormal data annotation records are as follows:

[0074] Extract all data points marked as abnormal from the real-time classification results, analyze the position and distribution of the abnormal data points in the multi-dimensional feature space, and obtain preliminary feature analysis results of the abnormal data points based on the relationship between each data point and its neighboring points;

[0075] Based on the preliminary feature analysis results of abnormal data points, the severity level of each abnormal data point is calculated using the following formula:

[0076]

[0077] Among them, S i represents the severity level of the ith data point, represents the value of the i-th data point on the k-th feature, μ k is the average value of the kth feature, w k is the weight of feature k, p is the power index, and K represents the total number of features;

[0078] Based on the severity level, label each data point, record the degree of abnormality of each point, and obtain the abnormal data annotation record.

[0079] Specifically, all data points marked as abnormal are extracted from the real-time classification results, and the body temperature, blood pressure and heart rate information of these data points at different times or in the same time period are read one by one. While reading the information, its multi-dimensional coordinate form is mapped to a pre-configured multi-dimensional feature space. The multi-dimensional feature space usually includes multiple reference dimensions such as body temperature distribution, blood pressure category distinction and heart rate fluctuation characteristics. When the position of each data point in the feature space is centrally counted, it is necessary to first verify whether these positions fall within the previously determined area. For example, comparing the body temperature with the range of 36°C to 39.5°C, and the systolic blood pressure with the range of 60mmHg to 200mmHg. The diastolic pressure is compared with the interval of 40mmHg to 120mmHg, and the heart rate is compared with the interval of 40bpm to 200bpm. If the value of any dimension deviates from the specified range, it is regarded as a potential outlier. Then, the similarity of this point with other neighboring data points can be compared. If abnormal deviations occur in multiple dimensions, a higher abnormal density will be presented. In this process, the context records will be combined to view the occurrence sequence of these data points and check the distribution of their distances from the surrounding neighboring points. If the performance of multiple dimensions in the adjacent time period is significantly different from that of the surrounding points, it will be recorded as an abnormal category. Finally, these analysis results will be concentrated to obtain the preliminary feature analysis results of the abnormal data points.

[0080] The benefit of the formula is that it comprehensively considers the difference between multi-dimensional indicators and their average values, and strengthens the capture of extreme deviations through weight factors and power exponents;

[0081] S i Indicates the severity level of the i-th data point. This value is obtained by weighting the deviation of each dimension feature. i k represents the observed value of the i-th data point on the k-th feature. These observations can be obtained through continuous monitoring equipment. For example, in the process of monitoring body temperature, blood pressure and heart rate 24 hours a day, they are collected and recorded every 5 minutes. k is the average value of the kth feature, which can be obtained by summing up all valid data and dividing it by the number of valid records. k is the weight coefficient of feature k, and its value needs to be set according to the importance of each dimension. After statistically analyzing the historical abnormal situations, the contribution of each dimension to health risks when there is a large fluctuation can be analyzed, and then the corresponding weight value can be extracted. p is the power index, indicating that the deviation is amplified by square or higher power. K represents the total number of all features. In the current scenario, for example, K=3 can be set to represent the three dimensions of body temperature, systolic blood pressure, and heart rate.

[0082] Calculation process:

[0083] Get μ kFor example, in the past month of continuous monitoring, 5000 pieces of temperature data were recorded, and the total cumulative value was 185000℃, then μ k =185000 / 5000=37℃, if the second and third dimensions are systolic blood pressure and heart rate respectively, by the same calculation, μ2=120mmHg, μ3=70bpm;

[0084] Get w k In the 6-month abnormal distribution statistics, there were 50 warning events caused by high body temperature, 120 warning events caused by high systolic blood pressure, and 80 warning events caused by high heart rate. The three were combined to calculate the frequency of occurrence and refer to the medical evaluation standard, and the calculation results were w1=0.25, w2=0.45, and w3=0.30;

[0085] Set p = 2, if we observe the eigenvalue of an abnormal point i The calculation process is as follows:

[0086]

[0087] The result shows that the data point deviates greatly from the average value, and the recognition accuracy of obvious anomalies is improved by the participation of the aforementioned weights and power exponents. i Values ​​above 400 are recorded as major abnormalities, values ​​between 100 and 400 are moderate abnormalities, and values ​​below 100 are classified as mild or negligible abnormalities. The current value of 593.5 represents a severe abnormality level, which is directly related to the final judgment of this step.

[0088] Based on the severity level, the S of each data point is i The value is compared with the corresponding abnormal level interval. In actual execution, the severity level of each data point and the corresponding dimension values ​​such as body temperature, blood pressure, and heart rate are usually collected and summarized into memory. Then, multiple data in the same time period or similar time period are compared with the threshold range defined above. If the S value obtained by a certain record in multiple dimensions is i If the value falls in a higher range, it can be labeled as "high abnormality". If it is in a relatively low range but still significantly deviates from the normal range, it can be labeled as "moderate abnormality" or "mild abnormality". When all data points have completed label assignment, the final labels are output one by one. In addition to indicating the abnormality level, these labels also need to record the specific dimension exceeding the limit, such as whether the body temperature is greater than 39.5℃ or the heart rate is greater than 200bpm, etc. For problem points that are indeed abnormal, further processing can be done in subsequent links, so that the degree of abnormality of each point can be summarized and the abnormal data annotation record can be obtained.

[0089] The steps to obtain the decision results are:

[0090] Extract the abnormal level and attribute information of each data point from the abnormal data annotation record, analyze the abnormal characteristics and context of the data point, and obtain the characteristic analysis results of each abnormal data point;

[0091] Based on the characteristic analysis results, the processing priority of each abnormal data point is evaluated, and the calculation formula is:

[0092]

[0093] Among them, P i represents the processing priority of the i-th data point, r i Indicates the severity level, d i It indicates the degree of deviation from normal data, and t is the decisive threshold;

[0094] Based on the processing priority, a processing strategy is formulated for each abnormal data point, including deletion, correction or retention, to obtain the processing decision result.

[0095] Specifically, the abnormal level and attribute information of each data point are extracted from the abnormal data annotation record. When determining these levels and attribute information, the severity score of the previous record is read and combined with the actual monitored environmental parameters. If it is found that the record contains abnormal situations in multiple different time periods, it is necessary to compare the historical reference range of the environment where the data point is located one by one. For example, the temperature may be compared with the 36℃ to 39.5℃ range, the blood pressure may be compared with the 60mmHg to 200mmHg and 40mmHg to 120mmHg range, and the heart rate may be compared with the 40bpm to 200bpm range. If one of the dimensions deviates significantly from the pre-set range, it is necessary to combine the collected values ​​at adjacent times to evaluate whether it is indeed abnormal. After completing After marking or verifying these dimensions, a comprehensive analysis should be conducted in combination with the attribute information of the data point, including judging whether there is temporary external interference based on the living environment or medical scenario in which the data point is located. If external interference does exist, the data point can be marked as a suspected anomaly or a subject of continuous observation. If the external interference is not obvious or other factors have been ruled out, it will be directly included in the formal abnormal record. In this process, the type and source of the abnormal point must be confirmed based on the collected reference materials and contextual data. If there are special marks, they must be recorded synchronously so that they can be distinguished in the subsequent severity calculation. After all data points have completed the above judgment, their attribute information will be archived and summarized to obtain the characteristic analysis results of each abnormal data point.

[0096] The benefit of the formula is that it combines the severity level with the degree of deviation through an exponential mapping function, and uses the decisive threshold as a shift factor, so that more differentiated processing priorities can be obtained in different situations;

[0097] P iIndicates the processing priority of the i-th data point. This value can be directly used to sort abnormal data points. i Indicates the severity level, which can be obtained through the anomaly data annotation records obtained previously. For example, in the six months of monitoring collected previously, a total of 150 significant anomalies were counted based on the temperature and blood pressure exceeding the limit. The outlier amplitudes of these anomalies are combined with the potential health risks to obtain the severity level value and save it in the record. i Indicates the degree of deviation from normal data, which can be accumulated by multi-dimensional deviations. For example, if a record shows a body temperature of 40°C and a blood pressure of 180 mmHg, it can be compared with the normal average value to form a deviation degree. This value is calculated through multiple comparisons and statistics and stored in the database. t is the decisive threshold, and its specific value needs to be fitted based on the central position of the overall abnormal distribution. For example, when the average deviation degree obtained by statistics is about 5, t=5 can be used;

[0098] Calculation process:

[0099] Retrieve the r of a particular outlier from the record i and d i , for example, the severity level of the anomaly point is r i =6.3, deviation degree d i =5.4, decisive threshold t=5;

[0100] First calculate the exponential term (d i -t), that is, 5.4-5=0.4;

[0101] calculate

[0102] According to the index function table or the processing flow, e is obtained. -0.4 ≈0.6703;

[0103] Then calculate the denominator 1+e -0.4 =1+0.6703=1.6703;

[0104] r i =6.3 into the division operation:

[0105] The results show that when the severity level is 6.3 and the deviation degree is 5.4, the final processing priority is about 3.77. When it is greater than 2, it can be judged as a higher processing level. If the value is higher, it means that further attention is needed. This formula can be used to combine the severity level with the deviation degree, and adaptively adjust according to the decisive threshold obtained previously to handle abnormal data points in a layered manner.

[0106] Based on the processing priority, the priority value of each abnormal data point is retrieved from high to low. During the execution process, it is necessary to first unify the processing priorities of each data point and compare them with the previously specified value range. If the priority of a data point obviously exceeds the pre-established safety line, it will be ranked at the top. If the priority is in the middle interval, it can be ranked in the subsequent position. If the priority is in the lower interval, it can be regarded as a relatively slight abnormality. After the data points are sorted in this way, the corresponding dimensional values ​​such as body temperature, blood pressure and heart rate are read one by one. If the priority of the data point is high but the relevant dimensional values ​​only deviate in a local period, it can be judged based on historical information whether there is a one-time excessive measurement situation. It can also be combined with context data to check whether there are special events in the sampling time period. When it is confirmed that the abnormal data point has deviated stably for many times and may cause adverse interference to the subsequent data processing process, it can be decided to delete it. If it is found that the source of the deviation is reasonable and it is only a small fluctuation that exceeds the limited range, it can be corrected. If it is confirmed that the data point is abnormal but does not affect the main process, it can be retained. Finally, an overall decision summary is formed based on the final disposal method of all data points to obtain the processing decision result.

[0107] The steps to obtain the updated dataset are:

[0108] Based on the processing decision results, access the real-time classification results and perform target processing on the specified operations for each marked data point, including deleting data that does not meet the standards or correcting data errors to form a preliminary corrected data set;

[0109] Based on the preliminary revised dataset, the completeness and consistency of the dataset are checked to verify that all data points have been processed correctly to obtain an updated dataset.

[0110] Specifically, based on the processing decision results, first retrieve all marked body temperature, blood pressure or heart rate data points from the corresponding records and obtain their operation categories, and verify these categories with the previous comparisons within the reference ranges of body temperature 36°C to 39.5°C, blood pressure 60mmHg to 200mmHg and 40mmHg to 120mmHg, and heart rate 40bpm to 200bpm. If a data record is marked as needing to be deleted, confirm whether the difference between it and the adjacent time periods is indeed significantly deviated during execution. For example, compare whether the body temperature difference between the adjacent time periods before and after is greater than 3°C or whether the blood pressure difference exceeds If the heart rate is over 40mmHg, if it is indeed too large, the record will be completely removed from the data set. If a data record is marked as needing to be modified, it is necessary to find out the possible cause of the error and revise the value. For example, when the heart rate was originally recorded as 230bpm, the adjacent heart rate records and other physiological parameters at the time can be compared to determine whether it should be corrected to closer to 170bpm or 180bpm. The final correction value can be determined based on the statistical methods of existing scientific literature or clinical experience. After all deletions or modifications are completed, a preliminary revised data set can be formed. All deleted or modified records in the data set will be centrally marked to prevent subsequent confusion.

[0111] Based on the initially revised data set, each data field will be checked one by one to see if it is consistent with the existing data format and numerical specifications in the previous article, including whether the body temperature item is maintained between 36℃ and 39.5℃, whether the blood pressure meets the dual parameter intervals of 60mmHg to 200mmHg and 40mmHg to 120mmHg, and whether the heart rate is maintained in the valid range of 40bpm to 200bpm. In this step of the inspection, if it is found that some correction operations do not put the values ​​in a reasonable range as required, they will be recorded again and reviewed by manual or automatic logic. For example, if it is found that a blood pressure value fails to return to the range that meets the specifications, it may be a misjudgment in the previous link. It is necessary to trace back to the original record to compare whether the reference parameters used are correct. In addition, it is necessary to ensure that there are no duplicate entries and no missing information. If a record is deleted, its index position is guaranteed to no longer participate in statistical operations. If a record is revised, its primary key number is maintained to maintain the overall data order unchanged. When all data items are verified to be correct, they can finally be summarized into a new data set to obtain an updated data set.

[0112] The steps to obtain the recognition model are:

[0113] Apply incremental gradient descent to the updated data set, which includes processing the points in the data set one by one, calculating the error gradient for each data point, updating the model parameters, and gradually reducing the error to obtain the optimized model parameter set;

[0114] Based on the optimized model parameter set, the learning rate is recalculated using the following formula:

[0115]

[0116] Among them, η represents the new learning rate, η0 is the basic learning rate, γ is the adjustment factor, and e i is the error of the ith data point, and n represents the total number of data points;

[0117] Use the new learning rate to optimize the classification model and obtain the recognition model.

[0118] Specifically, incremental gradient descent is applied to the updated data set. In the specific implementation process, each record in the data set is first read according to the existing classification task and decomposed into an input vector and a corresponding label. The parameters of each dimension of the input vector are uniformly numbered to facilitate sequential processing. For each record, the error between the current predicted output and the actual label is first calculated, and then the gradient value of the record is calculated using the current algebraic method. The gradient value can be obtained by taking the partial derivative of the error function, such as taking the derivative of the deviation of the classification result component by component. After obtaining the gradient value, the relevant parameters inside the model are adjusted in time. For example, in a common two-classification scenario, the product of the learning rate multiplied by the gradient term is subtracted from each element of the weight vector, and then the process is repeated from the next record. The operation is repeated in this way until all records are processed. In this process, the errors of different rounds are compared. If it is found that the errors of some records are continuously high, their eigenvalues ​​will be rechecked to confirm whether these eigenvalues ​​are in the reasonable range pre-set for body temperature, blood pressure and even heart rate. For example, body temperature is normal only between 36℃ and 39.5℃, blood pressure is reasonable only between 60mmHg and 200mmHg and 40mmHg to 120mmHg, and heart rate is valid only between 40bpm and 200bpm. Through this step-by-step update process of one record at a time, the model will continuously correct its internal parameters in each iteration, thereby reducing the overall error and finally obtaining a new set of model parameters after convergence.

[0119] The benefit of the formula is that by incorporating the root mean square form of the sum of squared errors into the denominator, the learning rate can be automatically reduced when the error is large, and the learning rate can be automatically maintained or increased when the error is small, thereby enhancing the adaptability of the training process to different error levels;

[0120] η represents the new learning rate, which will dynamically affect the parameter update amplitude. η0 is the basic learning rate, which can be obtained by statistics in a previous training history. For example, the accuracy performance of each round of training in the past month is summarized and the learning rate η0 that balances the error reduction rate and training stability is selected. γ is the adjustment factor, which can be set according to the sensitivity of the training task to the error and the convergence speed requirements. For example, after collecting experimental data for more than two weeks, it is statistically measured that if γ is between 0.01 and 0.1, it can better meet the steady-state convergence requirements. i Refers to the error of the ith data point, which can be obtained by recording the difference between the point in a feedforward prediction and its true label. If the label is 0 or 1, the difference between the two can be regarded as the error. n is the total number of data points, the number of samples from this training batch. For example, if the current batch is 3000 records, n = 3000;

[0121] Calculation process:

[0122] Square the error of each record Accumulation, for example, in a certain calculation, the sum of squares of the errors of each record is about 570, and the total number of records n = 3000;

[0123] calculate First calculate 570 / 3000 = 0.19, then square it to get about 0.4359;

[0124] Set η0 = 0.05, γ = 0.08;

[0125] Calculate γ·0.4359=0.08×0.4359=0.03487;

[0126] Calculate the denominator: 1 + 0.03487 = 1.03487;

[0127] Substitute η0=0.05 into:

[0128] The result shows that the learning rate based on the new error level after the current batch training is about 0.0483. If the value exceeds 0.1, it will easily lead to excessive updates. In practical applications, the changes in this value will be strictly monitored. When it drops below 0.01, it can indicate that the current error is relatively stable. We can also continue to observe whether it remains in a relatively balanced range in subsequent iterations.

[0129] The classification model is optimized using a new learning rate. During the execution process, the model parameters obtained in the previous stage are first loaded into the memory and connected to the corresponding classification scenario. The parameters are updated for each piece of data according to the new learning rate, and the residual value and weight change corresponding to the record are immediately recorded after one update. If a piece of data has a body temperature, blood pressure or even heart rate that deviates from the previously set threshold, an abnormal penalty item will be added to the calculation. For example, separate tracking is performed for situations where blood pressure exceeds 200 mmHg or heart rate exceeds 200 bpm, so as to serve as a statistical control in subsequent updates. Then, after multiple iterations, the updated parameters are stored in structured records and compared with the classification performance presented by the new data one by one. If it is observed that the errors of some records are still high, the characteristic value of the record will be checked to see if it is correct. If the value is confirmed to be out of touch with reality and may be erroneous, it will be corrected or deleted again according to the cleaning rules obtained previously. After the entire process is completed, the latest version of the classification model parameters can be obtained and the training of the overall model is completed, and the final output is the recognition model.

[0130] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for intelligent analysis and management of nursing physical examination data, characterized in that: The following steps are involved: Extracting body temperature, blood pressure and heart rate from the physical examination data set, performing Gaussian kernel mapping processing, mapping the physical examination data into a high-dimensional feature space, and obtaining high-dimensional feature data; using the high-dimensional feature data to train a support vector machine model, adjusting model parameters, and generating a classification model; Using the classification model, the newly added physical examination data is classified in real time to distinguish normal data from abnormal data and obtain real-time classification results; Analyze the abnormal data identified in the real-time classification result, mark the severity level of each abnormal data point, and form an abnormal data marking record; According to the abnormal data annotation record, the attributes of each annotated abnormal value are analyzed through decision support, and the processing strategy of each data point is determined, including deletion, correction or retention, to obtain a processing decision result; Deleting or modifying outliers in the real-time classification results according to the processing decision results to generate an updated data set; Applying an incremental gradient descent method to the updated data set, optimizing the weights and biases of the classification model, adjusting the learning rate according to the weights and biases, and then adjusting the learning rate according to error feedback in consecutive batches to generate a recognition model.

2. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the high-dimensional feature data are: Extract body temperature, blood pressure and heart rate information from the physical examination data set, remove noise and outliers, and obtain cleaned physical examination data; Based on the cleaned physical examination data, applying Gaussian kernel mapping to perform spatial transformation on the body temperature, blood pressure and heart rate data, and forming data mapped to a high-dimensional feature space by calculating the similarity between each data point and the kernel center, thereby obtaining the mapped physical examination data; Based on the mapped physical examination data, the features of each dimension are integrated, and through statistical analysis, key high-dimensional features are extracted to obtain high-dimensional feature data.

3. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the classification model are: Based on the high-dimensional feature data, start the training of the support vector machine model, the initial configuration includes selecting the kernel function type and setting the value of the regularization parameter, and obtain the initially trained support vector machine model; Adjusting model parameters from the initially trained support vector machine model, including optimizing kernel function parameters and regularization parameters by grid search, and evaluating the performance of the model under various parameter configurations by cross validation, to form an adjusted support vector machine model; Based on the adjusted support vector machine model, final training is performed to obtain a classification model.

4. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the real-time classification results are as follows: Deploy the classification model, receive newly added physical examination data in real time, and apply the classification model to classify the data instantly, including calculating each physical examination data point, determining the classification label, and obtaining a preliminary classification result; Based on the preliminary classification result, by comparing the predicted label with the known label, adjusting the classification threshold to obtain adjusted classification decision data; According to the adjusted classification decision data, each data point is marked as normal or abnormal to obtain a real-time classification result.

5. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the abnormal data annotation record are: Extract all data points marked as abnormal from the real-time classification results, analyze the position and distribution of the abnormal data points in the multidimensional feature space, and obtain preliminary feature analysis results of the abnormal data points based on the relationship between each data point and its neighboring points; Based on the preliminary feature analysis results of the abnormal data points, the severity level of each abnormal data point is calculated using the following formula: Among them, S i represents the severity level of the ith data point, represents the value of the i-th data point on the k-th feature, μ k is the average value of the kth feature, w k is the weight of feature k, p is the power index, and K represents the total number of features; Based on the severity level, each data point is labeled, and the degree of abnormality of each point is recorded to obtain an abnormal data annotation record.

6. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the processing decision result are: Extracting the abnormal level and attribute information of each data point from the abnormal data annotation record, analyzing the abnormal characteristics and context of the data point, and obtaining the characteristic analysis result of each abnormal data point; Based on the characteristic analysis results, the processing priority of each abnormal data point is evaluated, and the calculation formula is: Among them, P i represents the processing priority of the i-th data point, r i Indicates the severity level, d i It indicates the degree of deviation from normal data, and t is the decisive threshold; Based on the processing priority, a processing strategy is formulated for each abnormal data point, including deletion, correction or retention, to obtain a processing decision result.

7. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps for obtaining the updated data set are: Based on the processing decision results, access the real-time classification results and perform target processing on the specified operations for each marked data point, including deleting data that does not meet the standards or correcting data errors to form a preliminary corrected data set; Based on the initially revised data set, the integrity and consistency of the data set are checked to verify that all data points have been correctly processed, thereby obtaining an updated data set.

8. The method for intelligent analysis and management of nursing physical examination data according to claim 1, characterized in that: The steps of obtaining the recognition model are: Applying incremental gradient descent to the updated data set includes processing points in the data set one by one, calculating error gradients for each data point, updating model parameters, gradually reducing the error, and obtaining an optimized model parameter set; Based on the optimized model parameter set, the learning rate is recalculated using the following formula: Among them, η represents the new learning rate, η0 is the basic learning rate, γ is the adjustment factor, and e i is the error of the ith data point, and n represents the total number of data points; The new learning rate is used to optimize the classification model to obtain the recognition model.

Images