Neurological critical patient aspiration risk assessment system and method
Through the multivariate Logistic regression prediction model and aspiration state judgment model, the signal energy sequence is extracted using physiological data to achieve accurate and timely prediction of the risk of aspiration in critically ill patients, solving the problem of inability to prevent aspiration in the existing technology, and improving the Treatment effect and quality of life.
Patent Information
- Application Number
- CN202411017865.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-10-25
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-22
AI Technical Summary
The prior art cannot accurately predict the risk of aspiration in patients with severe neuropathy, and lacks real-time and adaptability, and cannot calculate the probability of aspiration without aspiration, resulting in the inability to prevent the occurrence of aspiration events.
A multivariate Logistic regression prediction model is used to combine the aspiration state judgment model, and aspiration state judgment model is constructed through physiological data to extract signal energy sequences, a real-time aspiration state judgment model is constructed, and parameters are dynamically adjusted to adapt to patient status changes, providing quantitative aspiration risk prediction probability.
It improves the accuracy and real-time prediction of aspiration status, provides reliable decision-making support for medical staff, reduces the incidence of aspiration events, and improves the treatment effect and quality of life of patients with severe neurological conditions.
Smart Images

Figure CN119007996A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical assessment technology, and in particular to a system and method for assessing the risk of aspiration in critically ill neurological patients. Background Art
[0002] Aspiration refers to the process by which foreign bodies such as oropharyngeal secretions, blood, bacteria, food particles, and gastric contents enter the respiratory tract and lungs. It is one of the main causes of aspiration pneumonia and leads to an increased mortality rate in patients with neurological critical illness. Neurological critical patients often have impaired consciousness and receive various interventions that may impair their swallowing function, such as endotracheal intubation, enteral nutrition via gastrointestinal tube, sedatives, etc., and are at high risk of gastric content aspiration. Methylene blue staining, fiberoptic endoscopic evaluation of swallowing (FEES), video fluoroscopic swallowing study (VFSS), and pepsin assay are commonly used techniques for diagnosing aspiration. However, the current methods used to diagnose aspiration are difficult to implement in patients with neurological critical illness. Methylene blue staining has limited sensitivity and serious side effects, and has been eliminated in clinical practice. In addition, FEES and VFSS examinations cannot be performed for neurological critical patients who cannot swallow orally. In addition, pepsin assay is costly, complicated, and not immediate, and has not been promoted in clinical practice. There is no reliable tool for the diagnosis of aspiration in patients with neurological critical illness. Therefore, it is imperative to establish a non-invasive and easy-to-promote method to screen for aspiration in neurocritical patients and take preventive measures in a timely manner. In recent years, clinical prediction models have been widely used to assist clinical decision-making and predict the incidence of diseases. The development of accurate and reliable prediction models can help healthcare professionals identify medium- or high-risk groups for specific diseases, thereby achieving early prevention. Despite the existence of many clear risk factors, there is currently no prediction model for gastric aspiration in neurocritical patients.
[0003] For example, the patent document with application publication number CN116035564A discloses a method, device and electronic device for intelligent detection of choking and aspiration, the method comprising: obtaining video data of the target object; determining whether the current scene meets the first scene based on the video data, and if so, obtaining facial expression data, body movement data and physiological information data of the target object; determining whether any of the facial expression data, body movement data and physiological information data meets the preset conditions for judging whether choking or aspiration occurs; in response to any of the facial expression data, body movement data and physiological information data meeting the preset conditions, determining that the target object has choked or aspirated, and issuing an early warning reminder. The present application is based on rapid recognition of facial expressions, body movements and physiological information, not only relying on a single data, but also realizing the rapid fusion of facial expressions, body movements and physiological information, and can efficiently analyze the health status of the target object, improving the accuracy of choking and aspiration detection. This patent document determines choking or aspiration by monitoring any conditions in the target object's facial expression data, body movement data and physiological information data. It is a warning method and cannot determine the probability of aspiration of the target object before choking or aspiration occurs.
[0004] For another example, the patent document with application publication number CN116631437A discloses a method for monitoring aspiration, including event recognition model generation and monitoring recognition process; obtaining the vocal cord vibration signal of the subject collected by the contact microphone, the contact microphone is installed above the cricoid cartilage in the throat of the subject; using the vocal cord vibration signal converted into the Mel spectrum to perform deep residual neural network training to generate an event recognition model; when the contact microphone senses the vocal cord vibration signal, the data processing module is activated, and the event recognition model is used to determine the throat movement of the tester, and the results are output, including swallowing, coughing, speaking and other events; after the swallowing behavior is recognized, within the next preset time window, if a coughing event is recognized, the model will determine that aspiration has occurred. This patent document is also judged when aspiration occurs, and it is also impossible to calculate the probability of aspiration of the target object before aspiration occurs.
[0005] The above two technical solutions only focus on a few indicators and lack comprehensive analysis of multiple physiological indicators. Especially when long-term continuous monitoring is required, the above two technical solutions have limitations in real-time and continuity, and lack adaptability and generalization capabilities for different patients and different situations.
[0006] As described above, the current system for monitoring aspiration behavior of a target object can only identify and actively warn when aspiration occurs, but cannot calculate the probability of aspiration of the target object when no aspiration occurs, and therefore cannot prevent aspiration from occurring.
[0007] In order to solve the above problems, the present invention provides a risk assessment system for aspiration in critically ill neurological patients. By building an artificial intelligence model to accurately predict the probability of aspiration in critically ill neurological patients, a simple, intuitive and rapid method for diagnosing aspiration is provided for clinicians.
[0008] In addition, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making the present invention, but due to space limitations, not all details and contents are listed in detail. However, this does not mean that the present invention does not have the characteristics of these prior arts. On the contrary, the present invention already has all the characteristics of the prior art, and the applicant reserves the right to add relevant prior art to the background technology. Summary of the invention
[0009] The current system for monitoring the aspiration behavior of the target object can only identify and actively warn when aspiration occurs. It cannot calculate the probability of aspiration of the target object when aspiration does not occur, and therefore cannot prevent aspiration from occurring. For critically ill neurological patients who cannot swallow orally, FEES and VFSS examinations cannot be performed. In addition, pepsin determination is costly, complicated, and not immediate, and has not been promoted in clinical use. There is no reliable tool for diagnosing aspiration in critically ill neurological patients.
[0010] The prior art has already presented a technical solution for obtaining a prediction model with high discrimination accuracy through optimization of univariate analysis and multivariate analysis. For example, the patent document with publication number CN115317012A discloses a method for STAS discrimination of lung adenocarcinoma based on CT imaging, which includes: obtaining a CT image to be discriminated, reconstructing the CT image through thin-layer images, and obtaining a reconstructed image; extracting multiple preset feature values from the reconstructed image; obtaining a discrimination result of the STAS possibility of lung adenocarcinoma based on multiple preset feature values and a pre-obtained Logistic regression prediction model; the Logistic regression prediction model is obtained by optimizing univariate analysis and multivariate analysis in turn. However, this technical solution is mainly applied to the field of medical imaging, focusing on diagnosing lung adenocarcinoma through CT images. The core of this method lies in image processing technology. First, a high-resolution reconstructed image is obtained through thin-layer image reconstruction; then, multiple preset feature values are extracted from these images, such as the shape, size, and edge features of the tumor. These feature values are the key to training and optimizing the Logistic regression prediction model, and are ultimately used to predict the STAS (tumor invasiveness) of lung adenocarcinoma. The optimization process of this method includes univariate analysis and multivariate analysis to improve the prediction accuracy of the model. On the contrary, the present invention is mainly used in the field of nursing care for critically ill neurological patients, especially in assessing the risk of aspiration. The data acquisition module of the system is responsible for collecting clinical and physiological data of patients, such as heart rate, blood pressure, respiratory rate, etc. The processing module of the present invention screens out effective predictive variables based on these data through a screening model, and these variables are used to construct a multivariate logistic regression prediction model to calculate the predicted probability of aspiration risk. In other words, the two are significantly different in specific analysis objects and data types. The above-mentioned prior art focuses on the analysis of image data, and trains the logistic regression model to predict the aggressiveness of tumors through image reconstruction and feature extraction, while the present invention focuses on the analysis of clinical and physiological data, and extracts key variables that can predict the risk of aspiration through data acquisition, feature screening and signal processing. In addition, the two methods also differ in real-time requirements. The above-mentioned prior art discrimination method usually does not involve real-time requirements, and the analysis process may be offline, and the main purpose is to assist diagnosis. On the contrary, the present invention requires real-time or near-real-time analysis in order to adjust the nursing strategy in time to prevent the occurrence of aspiration events.
[0011] In view of the deficiencies of the prior art, the present invention provides a system for assessing the risk of aspiration in critically ill neurological patients from a first aspect, comprising a data acquisition module and a processing module. The data acquisition module collects clinical data of the target object; the processing module screens effective predictive variables based on the screening model, and calculates the predicted probability of aspiration risk based on the effective predictive variables and the pre-built multivariate logistic regression prediction model, wherein the processing module also includes an aspiration state judgment model, which extracts a signal energy sequence based on the physiological data of the target object, and determines the real-time aspiration state of the target object based on the signal energy sequence; the multivariate logistic regression prediction model adjusts the predicted probability of aspiration risk based on the real-time judgment data input by the aspiration state judgment model.
[0012] Different from the prior art, the aspiration state judgment model of the present invention extracts the corresponding signal energy sequence for the physiological data of the target object, and determines the real-time aspiration state of the target object according to the signal energy sequence. Based on the above-mentioned distinguishing technical features, the problems to be solved by the present invention may include: how to accurately determine the real-time aspiration state of the target object. Specifically, the above-mentioned prior art mainly relies on CT image information, but this information is often non-real-time and can only reflect the physiological state at a specific moment, which is difficult to adapt to the real-time prediction requirements of the target object's respiratory state. Based on the above-mentioned prior art, it is easy to think of adjusting the CT image information to a real-time image data monitoring method to make up for the time difference in the actual physiological state monitoring of the target object. However, although this setting method has a certain positive effect in avoiding the delay of the actual monitoring state of the target object, the analysis deviation caused by factors such as angle error and clarity error in the image, in addition, the expression changes of neurological critical patients are usually not as obvious as those of normal patients, which makes the method of evaluating their aspiration situation through image monitoring limited in effect. Therefore, the use of image monitoring of the target object to monitor the aspiration of the neurological critical patients still cannot achieve a more accurate target object real-time state monitoring effect, which is also the fundamental reason why the prior art cannot be directly used or avoided for monitoring aspiration of neurological critical patients. On the contrary, the present invention directly selects the physiological data of the target object as the analysis object, extracts the corresponding signal energy sequence through the physiological data of the target object, and these sequences are used as key indicators for judging the real-time aspiration state of the target object, thereby not only reducing the monitoring error, but also providing more accurate and real-time monitoring results. In the monitoring of neurological critical patients, the monitoring of physiological signals is particularly important. These signals can not only evaluate the health status of the patient, but also reflect the patient's responsiveness to external stimuli. The present invention can quantify the patient's physiological load by calculating the average energy of these signals in a specific time window, and then be used to analyze and predict the patient's physiological state changes. The application of this method not only improves the accuracy of aspiration state prediction, but also enhances the real-time nature of monitoring, and provides more reliable decision support for medical professionals. The present invention can accurately collect and analyze the patient's clinical data through the collaborative work of the data acquisition module and the processing module, thereby realizing the early assessment of the risk of aspiration. The present invention also provides a quantitative and objective aspiration risk prediction probability by combining the effective prediction variables screened out by the screening model with the multivariate logistic regression prediction model, which helps medical staff to take preventive measures in time, reduce the incidence of aspiration events, and improve the treatment effect and quality of life of critically ill neurological patients. The aspiration state judgment model inputs the real-time aspiration state data of critically ill neurological patients as sample data into the multivariate logistic regression prediction model to help the multivariate logistic regression prediction model dynamically adjust parameters to adapt to changes in patient status and improve the accuracy of prediction.
[0013] According to a preferred embodiment, the method of calculating the predicted probability of aspiration risk based on the effective predictor variables by the multivariate logistic regression prediction model includes: extracting the regression coefficient of the effective predictor variable; calculating the change of the effective predictor variable based on the product of the regression coefficient and the range of the effective predictor variable, and determining the effective predictor variable corresponding to the largest change as the baseline predictor variable; calculating the score of each non-baseline predictor variable respectively, and adding and converting the scores of each effective predictor variable to obtain the predicted probability of aspiration risk.
[0014] Through the precise calculation of the multivariate logistic regression prediction model, the effective predictive variables that have the greatest impact on the risk of aspiration can be identified and used as a benchmark for risk assessment. This method based on regression coefficients and ranges can more accurately reflect the impact of each effective predictive variable on the risk of aspiration, thereby providing a more refined risk prediction. This helps medical staff to adjust treatment plans in a targeted manner, reduce unnecessary medical interventions, and improve the efficiency of medical resource utilization.
[0015] According to a preferred embodiment, the aspiration state judgment model is formed by training a hidden Markov model, and the aspiration state judgment model judges the real-time aspiration state of the target object in the following manner: decoding the input physiologically related parameters to obtain a hidden state sequence that can derive physiologically related parameters with the maximum probability; classifying the hidden states in the hidden state sequence to obtain the load level corresponding to each point in the physiologically related parameters; calculating the time series characteristic parameters of the physiologically related parameters, comparing them with the threshold of the time series characteristic parameters, and determining the curve type of the physiologically related parameters; judging the aspiration state of the target object according to the combination of the load level and the curve type of the physiologically related parameters. The present invention directly judges the aspiration state of the target object according to the patient's physiological parameters, and the information in the judgment process does not need to be re-analyzed and calculated, which reduces the error caused by data conversion and makes the judgment result more accurate.
[0016] According to a preferred embodiment, the multivariate logistic regression prediction model calculates the scores of each non-baseline prediction variable in the following manner: Where F represents the score of the non-baseline predictor variable, A 1 Represents the first change in the baseline predictor variable; A 2 Represents the second change in the nonbaseline predictor variable.
[0017] By calculating the scores of non-baseline predictive variables, the impact of each effective predictive variable on the risk of aspiration can be comprehensively evaluated. This method not only quantifies the contribution of each effective predictive variable, but also obtains a comprehensive aspiration risk prediction probability by adding and converting the scores. This evaluation method helps medical staff to fully understand the patient's aspiration risk status and provide a scientific basis for formulating personalized care plans.
[0018] According to a preferred embodiment, the screening model selects variables with univariate analysis values less than a threshold as effective predictive variables. By selecting variables with univariate analysis values less than a threshold as effective predictive variables, the variables that do not contribute much to the prediction of aspiration risk can be effectively excluded, thereby improving the prediction accuracy and efficiency of the screening model. This screening mechanism helps to reduce the complexity of the model, speed up the calculation speed, and ensure the reliability of the evaluation results.
[0019] According to a preferred embodiment, the processing module also includes a preprocessing model, which is configured to: identify missing values of clinical data; delete variables whose missing values are greater than a missing ratio threshold; and fill in variables whose missing values are not greater than the missing ratio threshold.
[0020] The preprocessing model can effectively process missing values in the data and improve the quality and availability of the data by preprocessing clinical data. By deleting or filling in variables, the system of the present invention ensures that the data input into the prediction model is complete and accurate, thereby improving the accuracy of aspiration risk prediction. In addition, this preprocessing mechanism also helps to reduce the bias caused by missing data and improve the generalization ability of the model.
[0021] According to a preferred embodiment, the preprocessing model is further configured to: analyze the central tendency, dispersion degree and / or distribution of clinical data: visualize the central tendency, dispersion degree and / or distribution. This can help medical staff understand the data characteristics more intuitively, so as to better interpret and apply the aspiration risk prediction results. This visual analysis not only improves the transparency of data analysis, but also helps medical staff quickly identify abnormal patterns in the data and provide support for clinical decision-making.
[0022] According to a preferred embodiment, the processing module also includes an evaluation model, which is configured to: draw a working characteristic curve, an internal calibration curve of the model and / or a clinical decision curve of the multivariate Logistic regression prediction model to evaluate the discrimination ability, consistency and / or net yield of the multivariate Logistic regression prediction model.
[0023] By drawing the operating characteristic curve, the ability of the multivariate logistic regression prediction model in distinguishing the risk level of aspiration in critically ill neurological patients can be effectively evaluated. This evaluation method is intuitive and informative, which helps to quickly identify the strengths and weaknesses of the model, thereby guiding the optimization and adjustment of the model.
[0024] By drawing the internal calibration curve of the model, the evaluation model can detect the consistency between the predicted probability and the actual probability of occurrence, thereby verifying the calibration of the multivariate logistic regression prediction model.
[0025] Through the analysis of clinical decision curves, the evaluation model can quantify the net benefits of the multivariate logistic regression prediction model at different thresholds, helping medical staff weigh the actual benefits and potential risks when adopting a specific prediction threshold. This evaluation helps to refine clinical decision-making, improve the pertinence and effectiveness of medical interventions, and reduce the risk of ineffective or over-treatment.
[0026] Through the comprehensive performance evaluation provided by the evaluation model, medical staff and researchers can have a deeper understanding of the performance of the multivariate logistic regression prediction model in practical applications, including its strengths and potential shortcomings. This transparency not only enhances users' confidence in the model, but also promotes the continuous improvement and customized application of the model.
[0027] The present invention provides a method for assessing the risk of aspiration in critically ill neurological patients from a second aspect, the method comprising: collecting clinical data of a target object: screening effective predictive variables based on a screening model, and calculating the predicted probability of aspiration risk based on the effective predictive variables and a pre-constructed multivariate Logistic regression prediction model, wherein the aspiration state judgment model extracts a signal energy sequence based on the physiological data of the target object, and determines the real-time aspiration state of the target object based on the signal energy sequence; the multivariate Logistic regression prediction model adjusts the predicted probability of aspiration risk based on the real-time judgment data input by the aspiration state judgment model.
[0028] By inputting the real-time aspiration status data of neurocritical patients as sample data into the multivariate Logistic regression prediction model, the multivariate Logistic regression prediction model can be helped to dynamically adjust parameters to adapt to changes in patient status and improve the accuracy of prediction. This method can ensure that there is a significant statistical association between the selected variables and the risk of aspiration in neurocritical patients, thereby improving the accuracy and reliability of the prediction model.
[0029] The predicted probability of aspiration risk is calculated based on the screened effective predictive variables and the pre-constructed multivariate logistic regression prediction model. This method can provide medical staff with a quantitative tool to assess the risk of aspiration in patients and help formulate more accurate care and treatment plans.
[0030] According to a preferred embodiment, the method of calculating the predicted probability of aspiration risk based on effective predictive variables includes: extracting the regression coefficient of the effective predictive variable; calculating the change of the effective predictive variable based on the product of the regression coefficient and the range of the effective predictive variable, and determining the effective predictive variable corresponding to the largest change as the baseline predictive variable; calculating the score of each non-baseline predictive variable respectively, and adding and converting the scores of each effective predictive variable to obtain the predicted probability of aspiration risk.
[0031] By extracting the regression coefficient of the effective predictor variable and calculating its product with the range, this method can identify the baseline predictor variable with the greatest impact on the risk of aspiration, thereby ensuring the sensitivity of the model to key factors. The scores of each non-baseline predictor variable are calculated separately, and these scores are added and converted to obtain the predicted probability of aspiration risk. This method not only improves the prediction accuracy of the multivariate logistic regression prediction model, but also enhances the interpretability of the model, allowing medical staff to better understand the logic behind the prediction results.
[0032] According to a preferred embodiment, the method of calculating the scores of each non-baseline predictor variable includes: Where F represents the score of the non-baseline predictor variable, A 1 Represents the first change in the baseline predictor variable; A 2 Represents the second change in the nonbaseline predictor variable.
[0033] This calculation method simplifies the calculation process of non-baseline predictor variable scores, making it easier for medical staff to understand and apply the model, reducing the complexity of model use and improving its acceptance and application efficiency in clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 It is a flow chart of the method for constructing and evaluating the aspiration probability web calculator for critically ill neurological patients provided by the present invention;
[0035] Figure 2 It is a schematic diagram of a display interface for constructing risk probability using a multivariate logistic regression model provided by the present invention;
[0036] Figure 3 It is a schematic diagram of the ROC curve of the internal validation provided by the present invention;
[0037] Figure 4 It is a schematic diagram of the ROC curve of external validation provided by the present invention;
[0038] Figure 5 It is a schematic diagram of the self-service sampling internal calibration curve provided by the present invention;
[0039] Figure 6 is a schematic diagram of a clinical decision curve provided by the present invention;
[0040] Figure 7 It is a simplified schematic diagram of the module connection relationship of the aspiration risk assessment system provided by the present invention.
[0041] Reference numerals list
[0042] 100: data acquisition module; 200: processing module; 210: screening model; 220: multivariate logistic regression prediction model; 230: pretreatment model; 240: evaluation model; 250: aspiration status judgment model. DETAILED DESCRIPTION
[0043] The following is a detailed description with reference to the accompanying drawings.
[0044] The present invention explains some of the noun terms.
[0045] Clinical data include: demographic characteristics, medical history, physical examination, family history, blood and biochemical tests, vital signs, ventilator use, treatment during hospitalization, enteral nutrition plan, etc. Preferably, the data acquisition module 100 establishes a communication connection with the hospital's HIS system to obtain the clinical data of the target object (patient). The target object includes patients with diseases such as stroke, brain trauma, and brain tumor.
[0046] Existing monitoring systems can only detect and sound an alarm when aspiration has already occurred. They lack the ability to calculate and prevent aspiration risk by predicting the target subject before aspiration actually occurs. For critically ill neurological patients who cannot swallow orally, examinations such as FEES and VFSS cannot be performed. At the same time, the determination of pepsin is not only costly, cumbersome, and the results are not immediate, but it has not been widely used in clinical practice, resulting in a lack of effective tools for the diagnosis of aspiration in critically ill neurological patients.
[0047] Example 1
[0048] In view of the shortcomings of the prior art, the present invention provides a system for assessing the risk of aspiration of critically ill neurological patients, including a data acquisition module 100 and a processing module 200. The data acquisition module 100 and the processing module 200 are connected to each other in a wired or wireless manner. The wired manner is, for example, to transmit information through an optical fiber or a signal line. The wireless manner is, for example, to transmit information through wireless signals such as WiFi signals and Bluetooth signals.
[0049] The data acquisition module 100 collects clinical data of the target object. The processing module 200 screens effective predictive variables based on the screening model 210, and calculates the predicted probability of aspiration risk based on the effective predictive variables and the pre-built multivariate logistic regression prediction model 220. The screening model 210 and the multivariate logistic regression prediction model 220 are arranged inside the processing module 200. The processing module 200 is preferably a server, processor, cloud server or server group capable of executing the coding program of the screening model 210 and the multivariate logistic regression prediction model 220.
[0050] Within the processing module 200, the clinical data received by its communication port is input into the screening model 210. The screening model 210 analyzes and screens out single predictive variables with predictive value for aspiration risk. This step is performed based on univariate analysis techniques, meaning that the impact of each variable is evaluated individually to determine whether it is important enough to be included in the final multivariate logistic regression prediction model 220.
[0051] Figure 2 The figure shows a schematic diagram of the display interface for constructing risk probability using a multivariate logistic regression model. Figure 2 As shown, the processing module 200 is integrated in the smart terminal device, or is connected to the smart terminal device in communication, so that the smart terminal device displays a display interface on the display component. The content of the display interface includes clinical data and prediction variables, and the calculation of the predicted probability of aspiration risk is implemented. In the display interface, GRV (mL) represents gastric residual volume and its unit; GCS score represents Glasgow score; velocity of enteral nutrition (mL / h) represents enteral nutrition speed and its unit; tracheal cuff pressure (cm H 2 O) represents the cuff pressure and its unit; HOB (°) represents the head of bed angle and its unit; Aspiration rate represents the probability of aspiration. The horizontal line and its circle corresponding to each predictor variable change adaptively based on the value of the input predictor variable.
[0052] The valid predictor variables processed by the screening model 210 will be sent to the multivariate logistic regression prediction model 220. In the multivariate logistic regression prediction model 220, all selected valid predictor variables will be used to calculate the aspiration risk probability of the target object. This calculation is performed through a pre-constructed multivariate logistic regression equation, which integrates the values of multiple predictor variables and outputs a probability prediction of whether aspiration will occur.
[0053] The present invention can accurately collect and analyze the patient's clinical data through the collaborative work of the data acquisition module 100 and the processing module 200, thereby realizing early assessment of the risk of aspiration. The present invention also provides a quantitative and objective prediction probability of aspiration risk by combining the effective prediction variables screened out by the screening model 210 with the multivariate logistic regression prediction model 220, which helps medical staff to take preventive measures in time, reduce the incidence of aspiration events, and improve the treatment effect and quality of life of critically ill neurological patients.
[0054] According to a preferred embodiment, the processing module 200 further includes a preprocessing model 230. The preprocessing model 230 is connected to the communication port of the processing module 200, and is used to preprocess the clinical data and send the processed clinical data to the screening model 210.
[0055] The preprocessing model 230 is configured to:
[0056] S11: Identify missing values in clinical data.
[0057] S12: Delete variables with missing values greater than the missing proportion threshold.
[0058] S13: Fill in the variables whose missing values are not greater than the missing proportion threshold.
[0059] S14: Analyze the central tendency, dispersion and / or distribution of clinical data. Visualize the central tendency, dispersion and / or distribution.
[0060] The preprocessing model 230 preprocesses the clinical data to effectively process missing values in the data and improve the quality and availability of the data. By deleting or filling in variables, the system of the present invention ensures that the data input into the prediction model is complete and accurate, thereby improving the accuracy of aspiration risk prediction. In addition, this preprocessing mechanism also helps to reduce the bias caused by missing data and improve the generalization ability of the model.
[0061] This can help medical staff understand the data characteristics more intuitively, so as to better interpret and apply the aspiration risk prediction results. This visual analysis not only improves the transparency of data analysis, but also helps medical staff quickly identify abnormal patterns in the data and provide support for clinical decision-making.
[0062] The screening model 210 screens effective predictive variables by analyzing a single factor. According to a preferred embodiment, the screening model 210 selects variables whose single factor analysis values are less than a threshold as effective predictive variables. The threshold is, for example, 0.1.
[0063] The screening model 210 selects variables with univariate analysis values less than a threshold as valid predictive variables, which can effectively exclude variables that do not contribute much to the prediction of aspiration risk, thereby improving the prediction accuracy and efficiency of the multivariate logistic regression prediction model 220. This screening mechanism helps to reduce the complexity of the model, speed up the calculation, and ensure the reliability of the evaluation results.
[0064] According to a preferred embodiment, the multivariate logistic regression prediction model 220 calculates the predicted probability of aspiration risk based on the effective predictor variables in the following manner:
[0065] S21: Extract the regression coefficients of valid predictor variables.
[0066] S22: Calculate the change in the effective predictor variable based on the product of the regression coefficient and the range of the effective predictor variable.
[0067] S23: Determine the effective predictor variable corresponding to the largest change amount as the benchmark predictor variable. The score of the benchmark predictor variable is 1 00.
[0068] S24: Calculate the scores of each non-baseline predictor variable separately.
[0069] Specifically, the multivariate logistic regression prediction model 220 calculates the scores of each non-baseline prediction variable in the following manner:
[0070]
[0071] F represents the score of the non-baseline predictor variable. 1 Represents the first change in the baseline predictor variable. 2 Represents the second change in the nonbaseline predictor variable.
[0072] By calculating the score F of the non-baseline predictor variable, the impact of each effective predictor variable on the risk of aspiration can be comprehensively evaluated. This method not only quantifies the contribution of each effective predictor variable, but also obtains a comprehensive aspiration risk prediction probability by adding and converting the scores. This evaluation method helps medical staff to fully understand the patient's aspiration risk status and provide a scientific basis for formulating personalized care plans.
[0073] S25: The scores of each valid predictor variable are added and transformed to obtain the predicted probability of aspiration risk.
[0074] Through the precise calculation of the multivariate logistic regression prediction model 220, the effective predictive variables that have the greatest impact on the risk of aspiration can be identified and used as a benchmark for risk assessment. This method based on regression coefficients and ranges can more accurately reflect the degree of impact of each effective predictive variable on the risk of aspiration, thereby providing a more refined risk prediction. This helps medical staff to adjust treatment plans in a targeted manner, reduce unnecessary medical interventions, and improve the efficiency of medical resource utilization.
[0075] According to a preferred embodiment, the processing module 200 further includes an evaluation model 240. The evaluation model 240 establishes a communication connection relationship with the multivariate logistic regression prediction model 220.
[0076] The evaluation model 240 is configured to:
[0077] S31: Draw a receiver operating characteristic (ROC) curve of the multivariate logistic regression prediction model 220 to evaluate the discrimination ability of the multivariate logistic regression prediction model 220.
[0078] Figure 3 Schematic diagram showing the internal validation ROC curve. Figure 4 A schematic diagram of the external validation ROC curve is shown. Figure 3 and Figure 4 In the figure, the horizontal axis represents specificity and the vertical axis represents sensitivity. The evaluation model 240 draws the operating characteristic curve (ROC) of the multivariate logistic regression prediction model 220 and calculates the area under the curve AUC.
[0079] like Figure 3 As shown in Figure 2, the area under the ROC curve (AUC) was 0.788 when internally validated. Figure 4 As shown, when externally validated, the area under the ROC curve (AUC) was 0.784. The 95% confidence interval was 0.686-0.883, indicating that the multivariate logistic regression prediction model 220 had good discrimination ability.
[0080] S32 : Drawing an internal calibration curve of the multivariate Logistic regression prediction model 220 to evaluate the consistency of the multivariate Logistic regression prediction model 220 .
[0081] Figure 5 A schematic diagram of the bootstrap internal calibration curve is shown. Figure 5The horizontal axis represents the nomogram predicted probability (Nomogram predicted probaility); the vertical axis represents the actual probability (actual probability). The inclined dotted straight line (Ideal) represents the ideal straight line, and its slope is l. The curved solid line (Bias-corredted) represents the bias-corrected line; the curved dotted line (Apparent) represents the apparent line. Figure 5 In, B = 1000 repeated starts. Absolute mean error = 0.021: n = 81.
[0082] Evaluation model 240 uses the Bootstrap method to perform repeated sampling with replacement for a preset number of times, such as 1000 times, to draw an internal calibration curve and perform a goodness of fit test. Figure 5 As shown, the internal calibration curve of the multivariate Logistic regression prediction model 220 fits well with the ideal straight line, indicating that there is good consistency between the predicted value and the actual value of the multivariate Logistic regression prediction model 220.
[0083] S33: Draw a clinical decision curve of the multivariate logistic regression prediction model 220 to evaluate the net yield of the multivariate logistic regression prediction model 220.
[0084] Figure 6 The clinical decision curve (DCA) is shown as a color image. The decision curve analysis quantifies the net benefit at different threshold probabilities to evaluate the clinical utility of the model. Figure 6 In the figure, the horizontal axis represents the threshold probability and the vertical axis represents the net benefit. The red curve represents the net benefit of treating all patients at different threshold probabilities. The blue curve represents the net benefit of treating patients based on this model at different threshold probabilities. The green straight line represents the data of not treating all patients.
[0085] like Figure 6 As shown, when the clinical decision curve (DCA) is within the probability range of 0.1 to 0.9, the multivariate logistic regression prediction model 220 has a higher net yield rate and good clinical application performance.
[0086] As described above, the present invention can effectively evaluate the ability of the multivariate logistic regression prediction model 220 in distinguishing the risk level of aspiration in critically ill neurological patients by drawing operating characteristic curves such as ROC curves. This evaluation method is intuitive and informative, and helps to quickly identify the strengths and weaknesses of the model, thereby guiding the optimization and adjustment of the model.
[0087] The present invention draws a calibration curve inside the model, and the evaluation model 240 can detect the consistency between the predicted probability and the actual probability of occurrence, thereby verifying the calibration of the multivariate logistic regression prediction model 220.
[0088] Through the analysis of the clinical decision curve, the evaluation model 240 of the present invention can quantify the net benefits of the multivariate logistic regression prediction model 220 at different thresholds, helping medical staff to weigh the actual benefits and potential risks when adopting a specific prediction threshold. This evaluation helps to refine clinical decision-making, improve the pertinence and effectiveness of medical intervention, and reduce the risk of ineffective or over-treatment.
[0089] The present invention also provides an example of aspiration risk probability obtained based on the patient's clinical data. Table 1 is related to Figure 6 Corresponding patient baseline data sheet.
[0090] The effective predictor variables for the patient were as follows: gastric residual volume of 165 mL, Glasgow score of 10, enteral nutrition rate of 40 mL / h, and balloon pressure of 26 cmH 2 O, the head of bed angle is 25°.
[0091] The five valid predictive variables of the patient mentioned above are input into the processing module 200 .
[0092] The calculation of the multivariate logistic regression prediction model 220 is:
[0093] P=exp(-4.93+0.01*a+0.16*b+0.01*c+0.59*d+1.35*e) / [1+
[0094] exp(-4.93+0.01*a+0.16*b+0.01*c+0.59*d+1.35*e)].
[0095] a represents gastric residual volume, unit: mL, value range: [0, 800].
[0096] -4.93 is the formula coefficient.
[0097] The regression coefficient of gastric residual volume was 0.01.
[0098] b represents the Glasgow score, unit: points, value range [3, 15]. The regression coefficient of the Glasgow score is 0.16.
[0099] c represents the enteral nutrition rate, unit: mL / h, value range [0, 80]. The regression coefficient of enteral nutrition rate is 0.01.
[0100] d represents the airbag pressure, unit: cmH 2O, when <20cmH 2 The value is 1 when the pressure is 0, otherwise it is 0. The regression coefficient of the balloon pressure is 0.59.
[0101] e: represents the angle of the head of the bed, unit: °. When it is less than 30°, the value is 1, otherwise it is 0. The regression coefficient of the head of the bed angle is 1.35.
[0102] The multivariate logistic regression prediction model 220 calculated the aspiration risk probability to be 0.654. The present invention can intuitively and conveniently obtain accurate aspiration risk probability, which is convenient for individualized evaluation.
[0103] The beneficial effects of the system of the present invention on doctors include: (1) Accurate diagnosis and personalized treatment: By accurately screening out the key predictive variables (effective predictive variables) that affect the risk of aspiration, doctors can more accurately diagnose high-risk patients based on the quantified risk prediction probability. This provides doctors with a scientific basis to formulate more personalized and targeted prevention and treatment plans, thereby improving treatment outcomes and the quality of life of patients. (2) Improving clinical decision-making efficiency: Through an automated and systematized risk assessment process, doctors can quickly obtain risk assessment results for patients, reducing the time required for traditional manual analysis.
[0104] The beneficial effects of the system of the present invention on nurses include: (1) Improving nursing management efficiency: Nurses can use the system to quickly identify high-risk patients and prioritize nursing resources and attention, thereby more efficiently performing nursing management and risk control. (2) Strengthening nursing preventive measures: The system's assessment results can help nurses better understand the patient's risk level and take more targeted nursing preventive measures, such as strengthening monitoring and taking special nursing procedures to minimize the occurrence of aspiration events.
[0105] The beneficial effects of the system of the present invention on patients include: (1) Reducing the risk of aspiration: By accurately assessing the risk of patients, medical staff can identify high-risk patients and take special preventive measures for these patients, thereby effectively reducing the risk of aspiration. This is crucial to improving patient safety and reducing the risk of potential complications. (2) Enhancing patient confidence and satisfaction: The assessment system can provide patients with more accurate and personalized medical services, enhance patient confidence in treatment plans, and improve patient satisfaction and treatment compliance.
[0106] Example 2
[0107] This embodiment is a further improvement of embodiment 1, and the repeated contents will not be repeated here.
[0108] At present, although the present invention improves the predicted probability of aspiration risk for critically ill neurological patients, helps medical staff identify patients at high risk of aspiration and provides preventive measures, thereby improving the level of monitoring. However, the predicted probability of aspiration risk for critically ill neurological patients cannot be equated with timely determination of the occurrence of aspiration events in critically ill neurological patients, and it is even impossible to determine whether aspiration has occurred. When a critically ill neurological patient has mild aspiration, there is still a possibility that it will not be discovered in time. Although the prior art uses images for scene monitoring, which has certain positive effects, the analysis deviations caused by the angle errors and clarity errors of the images, and the expressions of critically ill neurological patients cannot change significantly like normal patients, therefore, the use of images to monitor the aspiration of critically ill neurological patients is very poor. This is also the fundamental reason why the prior art cannot be directly used for aspiration monitoring of critically ill neurological patients.
[0109] Therefore, how to further timely judge the aspiration situation of critically ill neurological patients and whether an aspiration event occurs on the basis of improving the predicted probability of aspiration risk of critically ill neurological patients without using image monitoring is the technical problem that this embodiment hopes to solve.
[0110] This embodiment provides a model for determining the aspiration state of a neurologically critically ill patient, and provides a system and method for monitoring aspiration of a neurologically critically ill patient.
[0111] Hidden Markov Model (HMM) is a statistical model that is used to describe an observation sequence generated by an implicit, not directly visible sequence of states. Hidden Markov Model can model the implicit states of the system that cannot be directly observed, and infer the changes of these implicit states through observed data. This makes it very useful in many real-world problems, such as when the internal state of the system or the human body cannot be directly observed. Hidden Markov Model is particularly suitable for processing time series data. It can capture the dependencies and sequence patterns of the states in the time series, so as to effectively analyze and predict the changing trends of time series data.
[0112] The principle of the present invention is:
[0113] Hidden Markov models are particularly good at processing time series data and can capture and analyze patterns and dependencies in time series. The physiological indicators of neurocritical patients (such as heart rate, respiratory rate, blood oxygen saturation, etc.) are usually time series data that change over time. Hidden Markov models can judge the aspiration state by analyzing the changing trends of these data. Hidden Markov models can model implicit states, which represent states that cannot be directly observed within the system. In the aspiration detection of neurocritical patients, the implicit state can represent the patient's physiological state (such as normal, mild aspiration, moderate aspiration, severe aspiration), while the observed data is the time series changes of physiological indicators. By training the hidden Markov model, these implicit states and their transition probabilities can be learned. The hidden Markov model provides the ability of probabilistic inference, which can infer the most likely sequence of implicit states from the observed data. In aspiration detection, this means that the hidden Markov model can infer the most likely sequence of aspiration states based on the patient's physiological indicator data, thereby realizing the judgment of the aspiration state. The hidden Markov model can perform model training and parameter estimation through training data. In the aspiration detection of critically ill neurological patients, historical physiological indicator data can be used as training data, and the initial state probability, state transition probability and observation probability of the hidden Markov model can be estimated through methods such as the Baum-Welch algorithm or maximum likelihood estimation. The hidden Markov model can comprehensively analyze the time series changes of multiple physiological indicators and learn the complex relationship between the indicators through model training. This enables the hidden Markov model to more comprehensively analyze the patient's physiological state and improve the accuracy of aspiration detection. Therefore, the present invention uses the hidden Markov model as the basic model to train and construct an aspiration state judgment model 250 for critically ill neurological patients. Figure 7 As shown, the aspiration state determination model 250 is disposed in the processing module 200 and is controlled and operated by the processing module 200 .
[0114] The method of constructing the aspiration state determination model 250 is as follows.
[0115] S100: Perform data cleaning and preprocessing.
[0116] Data cleaning and preprocessing were performed on the historical physiological monitoring data of critically ill neurological patients. The energy characteristics of time series signals, such as heart rate, respiratory rate, blood oxygen saturation, etc., were introduced. By calculating the energy value of each sample point, the input physiological indicator sequence was converted into a signal energy sequence, which was used as the training data of the hidden Markov model.
[0117] Methods for data cleaning and preprocessing include:
[0118] Delete outliers: Due to uncontrollable factors such as acquisition equipment failure and transmission errors, there may be outliers such as NaN (non-numeric value), Inf (infinity), and null (empty value) in the data. These outliers will affect the training and prediction accuracy of the model, so they need to be deleted or replaced with reasonable values (such as mean, median, etc.).
[0119] Removing zero values: In some physiological monitoring data, there may be a large number of zero values, which may be caused by equipment calibration, data recording errors, etc. The presence of a large number of zero values will make the distribution of observed data very sparse, which may lead to inaccurate estimates of state transition probabilities and observation probabilities during the training of the hidden Markov model, thus affecting the performance of the model. Therefore, these zero values need to be removed or replaced.
[0120] Adjust data types: According to the requirements of the hidden Markov model for input data, adjust the data type to ensure that the data format meets the requirements of the model input. For example, convert the data to appropriate types such as floating point numbers and integers.
[0121] For the physiological monitoring data of neurological critical patients, such as electrocardiogram, respiratory rate, etc., these data are usually time series signals. In order to better reflect the physiological state of the patient, the energy characteristics of the time series signal can be introduced. The energy characteristics of the time series signal are a set of characteristics that describe the energy distribution of the signal in the time domain, and are often used to analyze the signal's intensity, spectral components, and overall energy.
[0122] Calculate the energy sequence of the signal: By performing a square operation on the signal, we can get the energy sequence of the signal.
[0123] E(n)=x(n)2.
[0124] x(n) represents a discrete time series signal, n represents the time step, and x(n) represents the signal value at time step n. N represents the length of the signal.
[0125] Calculate the average energy of the energy sequence: The average energy of the energy sequence can be calculated by summing the energy sequence and dividing by the signal length N:
[0126] The average energy formula of the energy series represents the square of the value of each time step of the signal, then summing and dividing it by the total length of the signal to obtain the average energy of the signal. This average energy can be used to describe the overall strength and energy distribution of the signal. In the physiological monitoring data of neurocritical patients, this average energy can be used as an important feature of the time series signal for further analysis and modeling.
[0127] In the analysis of physiological monitoring data of neurocritical patients, the average energy in the time window can be used to describe the load level. This method converts the input load sequence into a signal energy sequence by calculating the average energy value in each sample window, which serves as the training data of the hidden Markov model (HMM). In this way, the hidden Markov model (HMM) can learn the distribution of load levels in the sequence and the transition probability.
[0128] A load sequence, also called an observation sequence, refers to a data sequence that can be directly observed in the model. For the neurocritical patient aspiration state judgment model 250, the load sequence generally includes a series of measurable parameters related to the patient's physiological state, which can reflect the state of the patient's breathing, swallowing or other related functions. The load sequence of neurocritical patients may include the following types:
[0129] Respiration-related parameters: such as respiratory rate, tidal volume, end-tidal carbon dioxide concentration, respiratory waveform, etc.
[0130] Physiological signals: such as heart rate, blood oxygen saturation, blood pressure, electrocardiogram signals, etc.
[0131] Swallowing related parameters: such as laryngeal movement, esophageal pressure, oral secretion volume, etc.
[0132] Clinical scores: such as Glasgow Coma Scale (GCS), swallowing function score, etc.
[0133] Other monitoring data: such as EEG, EMG, body temperature, etc.
[0134] In the monitoring of neurologically critically ill patients, physiological signals (such as heart rate, blood pressure, blood oxygen saturation, etc.) are usually used to assess the patient's health status and physiological load. These signals can reflect the patient's physiological state and ability to respond to external stimuli. The load level refers to the signal strength or change in the physiological monitoring data of neurologically critical patients. These signals may include electroencephalogram (EEG), electrocardiogram (ECG), respiratory rate, etc. By calculating the average energy of these signals within a specific time window, the patient's physiological load can be quantified, which can then be used to analyze and predict changes in the patient's condition.
[0135] The time window size is set to ten sample points. A time window that is too small will cause the average energy to degenerate into the load value of a single sample point, while a time window that is too large will cause a time lag in the calculated energy value, affecting the real-time performance of the model.
[0136] For each time window, calculate the average energy value of all sample points in it. Assuming that the signal in the time window is x(n), x(n+1), ... x(n+9), the calculation formula for the average energy in the window is:
[0137]
[0138] By sliding the time window, the average energy value in each window is calculated for the entire input load sequence to generate the corresponding signal energy sequence. This energy sequence reflects the change of load level over time.
[0139] Data segmentation: Split the cleaned and preprocessed data into training sets and test sets to ensure the reliability of model training and verification.
[0140] S200: Training a Hidden Markov Model (HMM).
[0141] The generated signal energy sequence is used as the training data of the hidden Markov model. The hidden Markov model can learn the distribution of load levels and the transition probability between different levels based on these energy data.
[0142] In the monitoring of neurological critical patients, hidden Markov models (HMMs) are used to describe and analyze the relationship between the patient's physiological state and observed data. The hidden Markov model contains a set of parameters, including the initial state probability vector π, the state transition probability matrix A, and the observation probability matrix B, which together define the behavior of the model. HMM = (π, A, B).
[0143] π is a vector; it represents the initial probability distribution of the N hidden states in the model. Formally, π = [π 1 , π 2 ,…,π N ], where π i Indicates that at time t = 1, it is in state S i The probability of i =P(q 1 =S i ).
[0144] For example, if the HMM has three hidden states (e.g., stable, mildly stressed, severely stressed), π might be expressed as [0.7, 0.2, 0.1], meaning that the HMM starts out with a 70% probability of being in the stable state, a 20% probability of being in the mildly stressed state, and a 10% probability of being in the severely stressed state.
[0145] A is a matrix that represents the probability of transitioning from one hidden state to another. Formally, A = [a ij ], where a ij Indicates that it is in state S at time step t i In the case of transfer to state S j The probability of ij =P(q t+1 =S j |q t=S i ).
[0146] For example, if the hidden Markov model has three hidden states, A might be represented as a 3x3 matrix where each element represents the probability of transitioning from one state to another.
[0147] B is a matrix that represents the probability distribution of observations in each hidden state.
[0148] In form, B = [b i (ok)], where b i (ok) means it is in state S at time step t i The probability of observing ok under the condition of b i (ok) = P(ok|q t =S i ).
[0149] For example, if the observations are discrete (e.g., normal, abnormal), B might be represented as a 2x2 matrix where each element represents the probability of observing a certain value in a particular state.
[0150] As mentioned above, the hidden Markov model of a neurological critical patient can be expressed as a triple HMM = (π, A, B), where π, A, and B represent the initial state probability vector, the state transition probability matrix, and the observation probability matrix, respectively. These parameters are learned through training data and are used to describe and predict the relationship between the patient's physiological state and the observed data.
[0151] In the analysis of aspiration state, the energy sequence (i.e., the observation sequence) is usually a discrete count time series, such as the number of aspiration events. Since the observation data is a discrete count type, it can be assumed that the distribution of the observation data is a Poisson distribution. At this time, the observation probability can be expressed by the probability mass function of the Poisson distribution:
[0152]
[0153] Among them, b i (ot) indicates that it is in state S at time step t i The probability of observing data ot under the condition λ i represents the mean of the Poisson distribution, also called the average incidence rate, for each hidden state S i , there is a corresponding λ i The value represents the average number of occurrences of events (count data) observed in this state. i It is estimated by the hidden Markov model during the parameter learning process.
[0154] During the training process, the hidden Markov model adjusts λ based on the observed data through methods such as maximum likelihood estimation. i The value of λ is used to make the model fit the observed data better. i The different values of reflect the average occurrence intensity of events under different hidden states.
[0155] The present invention adopts the Baum-Welch iterative estimation algorithm to learn the parameters of the hidden Markov model. The steps of the Baum-Welch iterative estimation algorithm are as follows:
[0156] (1) Initialize state probability π 0 , state transition probability A O and the observation probability B O .
[0157] (2) Use the forward-backward algorithm to calculate the maximum likelihood estimation function Q of the observed data.
[0158] (3) Using the expectation maximization algorithm, the obtained function Q is maximized to obtain the model parameters A, B and π for the next iteration.
[0159] (4) Iteration: Repeat the above steps until the model parameters converge or the preset number of iterations is reached.
[0160] The hyperparameters during training include the number of hidden states n and π of the model. 0 , A 0 , B 0 and the maximum number of iterations I. The choice of each hyperparameter will have an important impact on the performance of the final model. In the actual training of the present invention, in order to determine the optimal number of hidden states, a plurality of different numbers of hidden states are pre-set and each setting is trained separately. In this way, the impact of each number of hidden states on the performance of the model can be evaluated, and the model with the best effect can be selected. This ensures that the complexity of the model matches the characteristics of the data, avoiding overly simple or overly complex models.
[0161] At the beginning of each training, a random initialization method is used to set π 0 , A 0 , B 0. Generally, the more groups that are randomly initialized, the stronger the model's ability to explore the parameter space and the greater the possibility of finding the global optimal solution. However, this also means that more computing resources and time are required. Specifically, for the judgment of the aspiration state of neurological critical patients, considering that the patient's condition may change at any time, it is necessary to make judgments quickly and accurately. Therefore, the number of randomly initialized groups should not be too many. In this case, selecting 3 to 5 groups of random initial values for experiments can reduce unnecessary computational burdens while ensuring the model's exploration capabilities, and speed up the model's training and evaluation process. The random initialization method can increase the model's ability to explore the parameter space and help find better parameter settings.
[0162] In addition, in order to save computing resources and prevent model overfitting, the present invention limits the maximum number of iterations to 10 to 20 times. This setting can ensure that the model has sufficient training time without wasting computing resources or risking model overfitting due to too many iterations. Overfitting can cause the model to perform well on training data but poorly on new data, so controlling the number of iterations is an effective way to prevent this problem.
[0163] After all training is completed, each model is scored using the likelihood assessment method. The likelihood assessment method can quantify the degree of fit of the model to the observed data. The higher the score, the stronger the model's ability to explain the data. Finally, the model with the highest score is selected as the optimal model for subsequent applications and analysis. This ensures that the constructed hidden Markov model is in the best state in terms of both fitting the data and predictive ability.
[0164] S300: Classifying the hidden state of the trained hidden Markov model into an aspiration state.
[0165] S310: Analyze the similarity of hidden states.
[0166] By looking at the observed probability distribution of hidden states in the training data, we can identify which hidden states behave similarly on the observed data. We cluster the hidden states using a clustering algorithm (such as K-means or hierarchical clustering) to identify similar groups of hidden states.
[0167] S320: Merge similar hidden states.
[0168] For similar hidden state groups identified in cluster analysis, merge them into one hidden state. Recalculate the observation probability distribution and state transition probability of the merged hidden state.
[0169] S330: Retrain the model.
[0170] The hidden Markov model is retrained using the merged hidden states to adjust the initial probability distribution, state transition probability matrix, and observation probability matrix.
[0171] S340: Calculate the distribution probability and transfer direction of each hidden state.
[0172] In the hidden Markov model, the distribution probability and transition direction of the hidden state are the core parameters of the model, which determine how the model generates observation data from the hidden state and how the hidden states transform into each other.
[0173] S341: Calculate the initial probability distribution.
[0174] The initial probability distribution (π) represents the probability of the model being in each hidden state at the initial moment. Use the training data to calculate the initial probability of each hidden state:
[0175] π i =P(S i , at time t=1).
[0176] S342: Calculate the state transition probability matrix.
[0177] The state transition probability matrix (A) represents the probability of transitioning from one hidden state to another. The transition probability between each hidden state is calculated using the training data.
[0178] a ij =P(S j , at time t+1|S i , at time t)
[0179] S343: Calculate the observation probability matrix.
[0180] The observation probability matrix (B) represents the probability of generating a specific observation state under a certain hidden state. Use the training data to calculate the probability of each hidden state generating each observation state:
[0181] b ij =P(Q i |S i ).
[0182] S350: Classify the hidden state into different aspiration states according to the distribution probability and transfer direction of the hidden state. By analyzing the distribution probability and transfer direction of the hidden state, the hidden state can be mapped to the actual aspiration state, thereby achieving accurate judgment of the aspiration state of the neurological critically ill patient.
[0183] Specifically, we observe the observed probability distribution of each latent state and identify which latent states are consistent with a specific aspiration state in the observed data. For example, if a latent state shows frequent coughing and low blood oxygen saturation in the observed data, it may correspond to a severe aspiration state.
[0184] Observe the direction of transitions between latent states and identify which transitions between latent states correspond to changes in aspiration status. For example, if there are frequent transitions from a mild aspiration state latent state to a severe aspiration state latent state, this may indicate that the patient's aspiration state is worsening.
[0185] S360: Map each hidden state to the actual aspiration state according to the distribution probability and transfer direction of the hidden state. 0 Mapped to the normal state, the hidden state S 1 Mapped to the mild aspiration state, the hidden state S 2 Mapped to the moderate aspiration state, the hidden state S 3 Mapped to severe aspiration state.
[0186] An independent validation dataset was used to evaluate the accuracy of the latent state to aspiration state mapping. Based on the validation results, the mapping was adjusted and optimized as necessary.
[0187] S400: Calculating a threshold of a time series characteristic parameter related to curve type detection.
[0188] For example, the time series characteristic parameters include: the average value of respiratory rate, the standard deviation of blood oxygen saturation, and the peak value of coughing times.
[0189] The characteristic mean and standard deviation under normal and aspiration conditions were obtained through statistical analysis:
[0190] Average respiratory rate (normal): μ 1 =15 times / min, σ 1 =2 times / minute.
[0191] Standard deviation of blood oxygen saturation (normal): μ 2 =1%,σ 2 =0.5%.
[0192] Peak cough frequency (normal): μ 3 =2 times, σ 3 =1 time.
[0193] Set the initial threshold:
[0194] The average value threshold of respiratory rate: [μ 1 -2σ 1 , μ 1 +2σ 1 ] = [11, 19] times / minute.
[0195] Standard deviation threshold of blood oxygen saturation: [μ 2 -2σ 2 , μ 2 +2σ 2 ]=[0,2]%.
[0196] Peak threshold of cough frequency: [μ 3 -2σ 3 , μ 3 +2σ 3 ]=[0,4] times.
[0197] After completing the above steps, the hidden Markov model is trained to form the aspiration state judgment model 250 .
[0198] S500: Inputting physiological data of a neurological critical care patient to be tested and detecting the aspiration state.
[0199] The aspiration state determination model 250 determines the real-time aspiration state of the target object in the following manners:
[0200] Decoding the input physiologically relevant parameters to obtain a hidden state sequence that can derive the physiologically relevant parameters with maximum probability;
[0201] Classify the hidden states in the hidden state sequence to obtain the load level corresponding to each point in the physiologically related parameters;
[0202] Calculating a time series characteristic parameter of a parameter related to physiology, comparing the time series characteristic parameter with a threshold value of the time series characteristic parameter, and determining a curve type of the parameter related to physiology;
[0203] The aspiration state of the target subject is determined based on a combination of the load level and the curve type of the physiologically relevant parameters.
[0204] Specifically, the aspiration state judgment model 250 uses the Viterbi algorithm to decode the observation sequence to determine the most likely hidden state sequence. The Viterbi algorithm finds the optimal path by maximizing the joint probability of the observation sequence and the hidden state sequence. An appropriate time window size is selected, such as 30 seconds or 1 minute, to ensure that sufficient physiological changes are captured while maintaining sufficient real-time performance. The size of the time window can also be determined by specific physiological parameters and clinical needs.
[0205] The aspiration state judgment model 250 classifies the decoded latent state sequence to obtain the load level corresponding to each point in the physiologically related parameters. 0 Represents the normal state, S 1 Represents mild aspiration state, S 2 Represents moderate aspiration state, S 3 Represents a severe aspiration state.
[0206] According to the duration and switching frequency of the latent state sequence, the patient's aspiration state is judged. For example, if S 1If it continues to occur, the patient is judged to be in a state of mild aspiration.
[0207] In each time window, the time series characteristics of the physiological parameters, such as the average value, standard deviation, peak value, trend, etc., are calculated. The calculated time series characteristic parameters are compared with the preset threshold value. For example, if the average value of the respiratory frequency exceeds the preset threshold range, it indicates that the respiratory frequency is abnormal. According to the comparison result between the time series characteristic parameters and the threshold value, the change curve type of the physiological parameters is determined. For example, if the average value of the respiratory frequency is continuously higher than the threshold value, it indicates that the respiratory frequency is showing an upward trend.
[0208] Combined with the latent state classification results and the change curve type of the time series characteristic parameters, the patient's aspiration state is comprehensively judged. For example, if the latent state sequence shows that the patient is in a mild aspiration state and the respiratory rate shows an upward trend, it can be judged that the patient's current aspiration state is mild aspiration and the respiratory rate is abnormal.
[0209] Monitor the patient's physiological parameters in real time and provide timely feedback and intervention measures based on the judgment results of the aspiration judgment model. For example, if the patient is judged to be in a state of severe aspiration, the medical staff should be notified immediately for intervention.
[0210] Assume that the patient's physiological parameters within a certain time window are as follows:
[0211] Respiratory rate: 18 times / minute; blood oxygen saturation: 95%; number of coughs: 3 times.
[0212] The calculated time series characteristic parameters are:
[0213] The average respiratory rate is 18 times / minute; the standard deviation of blood oxygen saturation is 1%; the peak number of coughs is 3 times.
[0214] The preset thresholds of the time series characteristic parameters are:
[0215] The average threshold of respiratory rate is [11, 19] times / minute; the standard deviation threshold of blood oxygen saturation is [0, 2]%; the peak threshold of cough frequency is [0, 4] times.
[0216] The time series characteristic parameters are compared with the threshold, and according to the latent state classification and the change curve type of the time series characteristic parameters, the aspiration state judgment model 250 judges that the patient is currently in a mild aspiration state, and the respiratory rate and coughing times are both within the normal range.
[0217] Example 3
[0218] This embodiment is a further improvement of Embodiments 1 and 2, and the repeated contents will not be repeated here.
[0219] like Figure 7As shown, in the processing module 200, the aspiration state judgment model 250 is connected to the multivariate logistic regression prediction model 220 for communication to transmit information to each other. The aspiration state judgment model 250 uses the real-time aspiration state data of the neurological critically ill patient as sample data and inputs it into the multivariate logistic regression prediction model 220 to help the multivariate logistic regression prediction model 220 dynamically adjust parameters to adapt to changes in the patient's state and improve the accuracy of the prediction. For example, if the aspiration state judgment model 250 determines that the patient is in a mild aspiration state, the probability weight of aspiration occurring in the future can be increased, thereby improving the accuracy of the prediction.
[0220] Specifically, the characteristic parameters that have the greatest impact on aspiration prediction in the multivariate logistic regression prediction model 220 are determined, such as respiratory rate, blood oxygen saturation, latent state sequence, etc. An initial weight is set for each characteristic parameter based on historical data and expert experience.
[0221] The real-time judgment data (absorption state and its physiological data) of the aspiration state judgment model 250 is input into the multivariate logistic regression prediction model 220 to monitor the changes of each characteristic parameter. The weight of the characteristic parameter is adjusted according to the changes of the real-time data. For example, if the hidden Markov model judges that the respiratory rate shows an upward trend, the weight of the respiratory rate parameter is increased to better reflect the real-time state of the patient.
[0222] Real-time retraining: The multivariate logistic regression prediction model 220 is retrained periodically or in real time, and the model parameters are updated using new real-time data.
[0223] Verification and testing: During the retraining process, the performance of the multivariate logistic regression prediction model 220 is evaluated using cross-validation and test data sets to ensure the prediction accuracy of the multivariate logistic regression prediction model 220 after adjusting the parameters.
[0224] The multivariate logistic regression prediction model 220 and the aspiration status judgment model 250 can promote each other, improve prediction accuracy, dynamically adjust parameters, and provide timely and effective decision support for medical staff, so as to better manage and prevent the risk of aspiration in critically ill neurological patients.
[0225] The present invention may also provide an electronic device, comprising an intelligent terminal and a display terminal. The intelligent terminal comprises at least one processor. The processor is configured to execute the method of the present invention. The display terminal is used to display process data, result data and curve images processed by the intelligent terminal. The electronic device may also include an interactive component for implementing input or output of data and control instructions. The interactive component is, for example, a display screen, a touch control screen, a mouse, a keyboard, etc.
[0226] The present invention may also provide a computer device, comprising: a processor and a memory, which is communicatively coupled to the processor and has instructions stored thereon. The processor executes the method of the present invention.
[0227] The present invention also provides a computer-implemented method for aspiration risk assessment performed by one or more processors, comprising: preprocessing the collected clinical data of the patient by one or more processors; screening predictive variables and selecting potential predictive variables; constructing a predictive model and constructing a risk probability web page calculator based on the predictive model; and outputting the predicted probability.
[0228] Table 1: Baseline data table
[0229]
[0230]
[0231] It should be noted that the above-mentioned specific embodiments are exemplary, and those skilled in the art can come up with various solutions inspired by the disclosure of the present invention, and these solutions also belong to the disclosure scope of the present invention and fall within the protection scope of the present invention. Those skilled in the art should understand that the present invention specification and its drawings are illustrative and do not constitute a limitation on the claims. The scope of protection of the present invention is defined by the claims and their equivalents. The present invention specification contains multiple inventive concepts, such as "preferably", "according to a preferred embodiment" or "optionally", which means that the corresponding paragraph discloses an independent concept, and the applicant reserves the right to file a divisional application based on each inventive concept.
Claims
1. A system for assessing the risk of aspiration in critically ill neurological patients, characterized in that: include: Data collection module (100): collects clinical data of target subjects; The processing module (200) is configured to screen effective predictive variables based on the screening model (210), and calculate the predicted probability of aspiration risk based on the effective predictive variables and a pre-constructed multivariate logistic regression prediction model (220), wherein: The processing module (200) further comprises an aspiration state judgment model (250), wherein the aspiration state judgment model (250) extracts a signal energy sequence based on physiological data of the target object, and determines a real-time aspiration state of the target object based on the signal energy sequence; The multivariate logistic regression prediction model (220) adjusts the aspiration risk prediction probability based on the real-time judgment data input by the aspiration state judgment model (250).
2. The aspiration risk assessment system for critically ill neurological patients according to claim 1, characterized in that: The multivariate logistic regression prediction model (220) calculates the predicted probability of aspiration risk based on the effective prediction variables in the following manner: Extracting the regression coefficients of the effective predictor variables; Calculating the change amount of the effective prediction variable based on the product of the regression coefficient and the range of the effective prediction variable, and determining the effective prediction variable corresponding to the largest change amount as the reference prediction variable; The scores of each non-baseline predictor variable are calculated separately, and the scores of each of the effective predictor variables are added and transformed to obtain the predicted probability of aspiration risk.
3. The aspiration risk assessment system for critically ill neurological patients according to claim 1 or 2, characterized in that: The mis-absorption state judgment model (250) is formed by training a hidden Markov model, wherein: The aspiration state judgment model (250) judges the real-time aspiration state of the target object in the following manner: Decoding the input physiologically relevant parameters to obtain a hidden state sequence that can derive the physiologically relevant parameters with maximum probability; Classifying the hidden states in the hidden state sequence to obtain the load level corresponding to each point in the physiologically related parameters; Calculating the time series characteristic parameter of the parameter related to physiology, comparing it with the threshold value of the time series characteristic parameter, and determining the curve type of the parameter related to physiology; The aspiration state of the target object is determined according to a combination of the load level of the physiologically related parameter and the curve type.
4. The aspiration risk assessment system for critically ill neurological patients according to any one of claims 1 to 3, characterized in that: The screening model (210) selects variables whose univariate analysis values are less than a threshold as effective predictor variables.
5. The aspiration risk assessment system for critically ill neurological patients according to any one of claims 1 to 4, characterized in that: The processing module (200) further comprises a pre-processing model (230), wherein the pre-processing model is configured to: identifying missing values for the clinical data; Variables with missing values greater than the missing proportion threshold were deleted; Variables with missing values no greater than the missing proportion threshold were filled.
6. The aspiration risk assessment system for critically ill neurological patients according to claim 5, characterized in that: The preprocessing model (230) is also configured to: Analyze the central tendency, dispersion and / or distribution of the clinical data; The central tendency, dispersion and / or distribution are visualized.
7. The aspiration risk assessment system for critically ill neurological patients according to any one of claims 1 to 6, characterized in that: The processing module (200) further comprises an evaluation model (240), wherein the evaluation model (240) is configured to: The operating characteristic curve, internal calibration curve and / or clinical decision curve of the multivariate logistic regression prediction model (220) are plotted to evaluate the discrimination ability, consistency and / or net yield of the multivariate logistic regression prediction model (220).
8. A method for assessing the risk of aspiration in critically ill neurological patients, characterized in that: The method comprises: Collect clinical data of target subjects; Based on the screening model (210), effective predictive variables are screened, and based on the effective predictive variables and a pre-constructed multivariate logistic regression prediction model (220), the predicted probability of aspiration risk is calculated, wherein: The aspiration state judgment model (250) extracts a signal energy sequence based on the physiological data of the target object, and determines the real-time aspiration state of the target object based on the signal energy sequence; The multivariate logistic regression prediction model (220) adjusts the aspiration risk prediction probability based on the real-time judgment data input by the aspiration state judgment model (250).
9. The method for assessing the risk of aspiration in critically ill neurological patients according to claim 8, characterized in that: The method of calculating the predicted probability of aspiration risk based on the effective predictive variables includes: Extracting the regression coefficients of the effective predictor variables; Calculating the change amount of the effective prediction variable based on the product of the regression coefficient and the range of the effective prediction variable, and determining the effective prediction variable corresponding to the largest change amount as the reference prediction variable; The scores of each non-baseline predictor variable are calculated separately, and the scores of each of the effective predictor variables are added and transformed to obtain the predicted probability of aspiration risk.
10. The method for assessing the risk of aspiration in critically ill neurological patients according to claim 8 or 9, characterized in that: Scores for each non-baseline predictor variable are calculated as follows: Where F represents the score of the non-baseline predictor variable, A1 represents the first change of the baseline predictor variable, and A2 represents the second change of the non-baseline predictor variable.
Citation Information
Cited By
Failure patient tracking and analyzing system based on intelligent chip
CN119541859A
Swallowing aspiration risk prediction method
CN119786052A
Integrated management system for ankylosing spondylitis
CN119920487A
Ankylosing spondylitis comprehensive management system
CN119920487B
Enterprise audit risk assessment method and system based on hidden Markov model
CN120410753A
