Construction method and prediction system of extubation outcome prediction model based on respiratory variability and machine learning

Abstract

The present invention provides a method for constructing an extubation outcome prediction model based on respiratory variability and machine learning. By introducing comprehensive respiratory indicators and complex, comprehensive statistical methods for physiological variability, a number of respiratory variability indicators are developed. Using feature engineering, multiple respiratory variability indicators are selected to train a machine learning model for extubation outcome prediction. The prediction system of the present invention is not limited to extubation prediction but has general applicability to other types of clinical decision-making support analysis. Compared with existing technologies, the prediction system proposed in the present invention utilizes respiratory variability indicators in a machine learning model, achieving a higher accuracy rate for extubation outcome prediction, and the method is highly interpretable.

Description

Technical Field

[0001] The present invention relates to a method for predicting the extubation outcome of patients undergoing invasive mechanical ventilation based on respiratory variability and a machine learning model, belonging to the technical field. Background Art

[0002] Invasive mechanical ventilation (IMV) is an important means of life support for critically ill patients, and its ultimate goal is to wean the ventilator. However, the important step before weaning, that is, determining the timing of extubation, has always been a difficult decision for clinicians. Delayed extubation will result in a higher incidence of pneumonia, longer hospital stays, and more hospital costs, while premature extubation will increase the risk of death. Today, there are three main methods to solve the timing of extubation decisions. Clinically, there are spontaneous breathing tests (SBTs), respiratory indicators include the shallow rapid breathing index (RSBI), and the field of artificial intelligence has a variety of machine learning models using electronic medical records. However, these methods are still unsatisfactory in the accuracy of extubation prediction due to the lack of the complexity of obtaining respiratory system control. For example, patients who pass SBT still have a mortality rate of 10-15%. Therefore, the use of respiratory variability in continuous respiratory waveforms to predict extubation failure has gradually been accepted. Current studies have found that respiratory rate RR, tidal volume V T Variability in peak inspiratory pressure (PIP), and minute ventilation (MV) has been associated with extubation outcomes. However, these studies have only considered these basic respiratory parameters obtained from the ventilator using simple variability analysis methods, lacking studies on comprehensive respiratory parameters and more complex physiological variability methods. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method and device for predicting the extubation outcome of patients undergoing invasive mechanical ventilation based on respiratory variability and a machine learning model.

[0004] The object of the present invention is achieved like this:

[0005] A method for constructing an extubation outcome prediction model based on respiratory variability and machine learning includes the following steps:

[0006] S1. Collect the extubation outcome, respiratory waveform data before extubation and under SBT, and corresponding ventilator parameters of each selected patient;

[0007] S2. Calculate the corresponding respiratory parameter sequence based on the respiratory waveform data before extubation and under SBT and the corresponding ventilator parameters; wherein the respiratory parameter sequence corresponding to each patient includes peak inspiratory pressure PIP, respiratory rate RR, tidal volume V T , one or more of minute ventilation MV, rapid shallow breathing index RSBI, mechanical work MP;

[0008] S3. Calculate a respiratory variability index for each respiratory parameter sequence based on a variability statistical analysis method, wherein the respiratory variability index includes a basic statistical method standard deviation SD, a coefficient of variation CV, a comprehensive statistical method standard deviation 1SD1, a standard deviation 2SD2, a Porta index PI, a Guzik index GI, a Slope index SI, an approximate entropy ApEn, a sample entropy SampEn, a fuzzy entropy FuzzEn, an acceleration performance AC, a deceleration performance DC, and one or more of a mean AVE, a median MED, a quartile QUA, and a three-quarter digit TQUA reflecting the distribution of the respiratory parameter itself;

[0009] S4. Select one or more component sets from the calculated respiratory variability indicators as the respiratory characteristics of the corresponding patients, and construct sample data consisting of the respiratory characteristics of each patient and the corresponding extubation outcome to obtain a training data set; based on the obtained training data set, use the patient's respiratory characteristics as input and the corresponding extubation outcome as output to train and obtain a patient extubation outcome prediction model.

[0010] Furthermore, in step S1, the patient information mainly comes from the electronic medical record and admission information, and the selection of respiratory waveform data depends on the extubation time and extubation information determined by the clinician, that is, the data around the extubation time is selected, and the ventilation mode is used to determine whether the obtained respiratory data is under SBT.

[0011] Furthermore, the respiratory signal contains three data sources: pressure, flow rate, and tidal volume. Different ventilators use different sampling frequencies. To ensure consistent respiratory data length across patients, all three data sources need to be resampled to the set frequency. The same applies to ventilator parameter settings. For example, if the sampling frequency is set to 50Hz, waveform data with a sampling frequency of 62.5Hz needs to be downsampled to 50Hz.

[0012] Furthermore, in S2, each breath of the respiratory data is processed, and according to the parameter settings of the ventilator at that time, it is determined whether it is pressure support ventilation PSV, and the respiratory parameters are calculated based on the pressure, flow rate and tidal volume data of this breath, which are the basic respiratory parameters PIP, RR, V T , MV and RSBI, and the comprehensive respiratory parameter MP. Since the PEEP value of the selected samples is set relatively consistent, it will affect the variability analysis effect. Here we only consider the dynamic part of MP MP d , MP d According to the different units, it can be divided into MP d [J / breath],MP d [J / L], MP d [J / min].

[0013] Furthermore, the step S2 further includes a screening step:

[0014] Invalid breaths are screened based on the thresholds corresponding to the breathing parameters and / or based on outliers. If one of the breathing parameters exceeds the threshold and is an outlier, the breath corresponding to the breathing parameter is considered to be invalid breath, and the invalid breathing parameters in all breathing parameter sequences are deleted.

[0015] Furthermore, the total length of the breathing sequence is greater than or equal to 1 hour.

[0016] Furthermore, step S4 is specifically as follows:

[0017] S4-1. Statistical analysis was performed on the distribution of each respiratory variability indicator in the extubation success and extubation failure groups. A respiratory variability indicator with a p < 0.05 was considered statistically significant. Receiver-operating characteristic (ROC) analysis was also performed to estimate the area under the curve (AUC) of each respiratory variability indicator.

[0018] S4-2, combining statistically significant respiratory variability indicators to form a feature set H;

[0019] S4-3, initialize or update the length g of the feature subset G, where the initial value of g is 0; select a respiratory variability index from the feature set H and merge it with the feature subset to form a new set G * , length is g+1; the feature set H consists of significant respiratory variability indicators;

[0020] S4-4. Calculate G * Average AUC of five-fold cross validation under the logistic regression LR model;

[0021] S4-5. Repeat steps S4-3 to S4-4 until all respiratory variability indices in feature set H have participated in the AUC calculation; add the respiratory variability indices with the highest average AUC to feature subset G and delete them from feature set H;

[0022] S4-6. Based on the feature subset G, select a set of calculated respiratory variability indices as the respiratory features of the corresponding patient, and construct a training dataset by combining the respiratory features of each patient and the corresponding extubation outcome as sample data. Based on the obtained training dataset, use the patient's respiratory features as input and the corresponding extubation outcome as output to train a patient extubation outcome prediction model and evaluate the performance of the patient extubation outcome prediction model.

[0023] S4-7. Repeat S4-3 to S4-6 until the average AUC of the five-fold cross validation of the feature subset G is greater than the threshold; the feature subset G with the maximum AUC under the five-fold cross validation and the patient extubation outcome prediction model obtained through training are used to predict the extubation outcome.

[0024] Furthermore, when statistically analyzing the distribution of each respiratory variability index in the extubation success and extubation failure groups, if the respiratory variability index followed a normal distribution, a t-test was used, otherwise a Wilcoxon test was used.

[0025] Furthermore, the step S4-4 also includes using the synthetic minority class oversampling technology SMOTE to balance the patient samples of failed extubation and successful extubation.

[0026] Furthermore, the performance indicators of the prediction model for patient extubation outcomes included: AUC, accuracy ACC, sensitivity SEN, and specificity SPE.

[0027] Furthermore, the structure of the patient extubation outcome prediction model is a random forest, support vector machine, decision tree or neural network model.

[0028] A prediction system for an extubation outcome prediction model based on respiratory variability and machine learning, the system comprising a prediction module, which inputs the patient's respiratory characteristics into the patient extubation outcome prediction model constructed by the construction method to obtain the corresponding extubation outcome.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention is a method for constructing an extubation outcome prediction model based on respiratory variability and machine learning. By introducing comprehensive respiratory indicators and complex and comprehensive statistical methods of physiological variability, many respiratory variability indicators are developed, and multiple respiratory variability indicators are selected under feature engineering to train a machine learning model for extubation outcome prediction. The introduction of comprehensive respiratory indicators such as MP can better reflect respiratory control, a complex system involving the respiratory system, central nervous system, chemoreceptors, motor receptors, and limbic system, and avoid the limitations caused by focusing only on respiratory movement. Secondly, the outstanding contribution of complex variability methods such as PRSA also shows that there is asymmetry in respiratory parameters in acceleration and deceleration, as well as other variability characteristics that need to be discovered before extubation.

[0031] The prediction system proposed in this paper is not limited to extubation prediction; it has general applicability to other types of clinical decision-making support. Compared with existing technologies, the proposed prediction system, which uses respiratory variability indicators in a machine learning model, can achieve a higher accuracy in predicting extubation outcomes, and the method is highly interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The figure is a flow chart of a method for constructing an extubation outcome prediction model based on respiratory variability and machine learning according to the present invention.

[0033] Figure 2This is a flow chart of the method for extracting respiratory parameters proposed in the present invention. DETAILED DESCRIPTION

[0034] The following embodiments of the present invention are further described in conjunction with the accompanying drawings and examples. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the invention.

[0035] In order to solve the low accuracy problem caused by the lack of consideration of the complexity of the respiratory control system in the current extubation timing decision-making, the present invention proposes a method for constructing an extubation outcome prediction model based on respiratory variability and machine learning and uses it for extubation outcome prediction.

[0036] See also Figure 1 The present invention provides a method for constructing an extubation outcome prediction model based on respiratory variability and machine learning. By introducing comprehensive respiratory indicators and a comprehensive statistical method for physiological variability, multiple respiratory variability indicators are developed. Through feature engineering, multiple respiratory variability indicators are selected to train a machine learning model for extubation outcome prediction. The details are as follows:

[0037] S1. Respiratory waveform data acquisition

[0038] Obtain respiratory waveform data of selected patients under SBT before extubation, as well as corresponding ventilator parameter settings (including ventilation mode, positive end-expiratory pressure (PEEP), support pressure (PS), etc.). The sampling frequency should be 50 Hz or above.

[0039] Extubation label information, i.e., the patient's extubation outcome, was annotated by a professional physician. Based on the physician's annotated extubation time, waveform data was collected for the three hours prior to extubation. If continuous data remained after the extubation time, data was collected up to six hours after extubation. Based on the ventilation pattern obtained by sampling every half hour, respiratory waveform data obtained under conditions other than SBT was removed. In this example, a total of 163 patients were enrolled, of whom 12 failed extubation, resulting in an extubation failure rate of 7.4%.

[0040] Respiratory waveform data includes pressure, flow rate, and tidal volume data. Different ventilators use different sampling frequencies. To ensure consistent respiratory data length across patients, all three data channels need to be resampled to the set frequency. The same applies to ventilator parameter settings. For example, if the sampling frequency is set to 50Hz, waveform data with a sampling frequency of 62.5Hz needs to be downsampled to 50Hz.

[0041] S2. Calculate and obtain the corresponding respiratory parameter sequence;

[0042] The process for this step is shown in Figure 2. According to the three-way data of the respiratory waveform and the ventilator parameters, start traversing each breath from the starting position of one of the data. First, check the mode setting of the ventilator parameters corresponding to this breath, and judge whether it is pressure support ventilation PSV through the ventilation mode of different ventilator models. Secondly, calculate the respiratory parameters, including the basic respiratory parameters PIP (peak inspiratory pressure), RR, V T , MV and RSBI, and comprehensive respiratory parameters MP.

[0043] RR is calculated as the number of breaths per minute, in breaths per minute (breath / min). The calculation formula is as follows:

[0044]

[0045] Where T(n) is the period of the nth breath. T The tidal volume of one breath is calculated in milliliters (mL). The calculation formula is as follows:

[0046]

[0047] Where F(n,t) is the flow rate of the nth breath, T i (n) is the inspiratory cycle of the nth breath. MV is the minute ventilation of one breath, expressed in liters per minute (L / min), using the following formula:

[0048]

[0049] RSBI is calculated as RR and V T The ratio is a dimensionless value and is calculated as follows:

[0050]

[0051] MP d [J / breath] calculates the dynamic part of the mechanical work of a breath, in joules per breath (J / breath), using the following formula:

[0052] MP d [J / breath]=0.098×∫0 VT (P(V)-PEEP)dV

[0053] Where P(V) is the pressure-volume loop function and PEEP is the positive end-expiratory pressure. d [J / L] The calculation is based on the mechanical work of one breath, taking V into account. T The unit is joules per liter (J / L), and the calculation formula is as follows:

[0054]

[0055] MP d [J / min] calculates the portion of the mechanical work of one breath that takes into account RR, expressed in joules per minute (J / min). The calculation formula is as follows:

[0056] MP d [J / min]=MP d [J / breath]×RR(n).

[0057] In order to avoid the influence of invalid data on the constructed patient extubation outcome prediction model, this step also includes an invalid breathing screening step, including method 1:

[0058] Each respiratory parameter is screened by a threshold. Breaths corresponding to respiratory parameters that are not within the threshold range are considered invalid breaths. Only breaths with all respiratory parameters valid are included in the respiratory sequence. Table 1 shows the valid threshold ranges for all respiratory parameters in this embodiment.

[0059] Table 1 Effective threshold ranges for all respiratory parameters in the embodiment

[0060]

[0061]

[0062] Breaths whose ventilator parameter settings are not PSV or whose respiratory parameters are invalid will be omitted in the traversal until the total duration of valid breaths reaches 1 hour.

[0063] Method 2: Outlier screening is performed on each respiratory parameter sequence in the respiratory sequence, and the breath with outlier respiratory indicators is judged as invalid breath; similarly, invalid breath parameters in all respiratory parameter sequences are deleted.

[0064] Specifically, the outlier screening method is as follows: the respiratory sequences are merged into respiratory parameter sequences of the same length according to the categories of respiratory parameters, and outlier detection is performed on each respiratory parameter sequence. N The distance d between each respiratory parameter and the mean in the respiratory parameter sequence of}, N represents the number of respiratory parameters, and the sequence D={d1,d2,...,d N}, if d appears i Greater than the threshold, such as d i If the value is greater than 4× standard deviation (X), the respiratory parameter is considered as an outlier and is deleted from the corresponding respiratory parameter sequence.

[0065] S3, calculating the values ​​of the respiratory parameter sequence under different variability analysis methods to obtain a respiratory variability index;

[0066] Variability analysis is performed on each respiratory parameter sequence. The variability methods include basic SD, CV, and comprehensive SD1, SD2, PI, GI, SI, ApEn, SampEn, FuzzEn, AC, DC, etc.

[0067] SD calculates the standard deviation of the respiratory parameter sequence, and the calculation formula is as follows:

[0068]

[0069] where x i is the respiratory parameter value, is the mean of the respiratory parameter sequence. CV calculates the coefficient of variation of the respiratory parameter sequence, and the calculation formula is as follows:

[0070]

[0071] where x i and The meaning is the same as that in SD. SD1 and SD2 are two variability analysis methods that rely on Poincaré analysis. The coordinates of each point in the Poincaré scatter plot drawn based on the respiratory parameter sequence are (x i ,x i+1 ), this embodiment uses V T For example.

[0072] SD1 and SD2 are the minor and major axes of the fitted ellipse in the Poincare scatter plot, respectively. SD1 is defined as the degree of dispersion of points perpendicular to the characteristic line at the center of the plot, while SD2 is defined as the degree of dispersion of points along the characteristic line. The calculation formulas for both are as follows:

[0073]

[0074]

[0075] where X N and X N+1 are the vectors of the X-axis and Y-axis of the scatter points, Var is the calculation formula for the variance, x i and x are X N The elements and means of .

[0076] PI, GI, and SI are three asymmetry analysis methods for respiratory parameter sequences using the Poincare scatter plot. In the Poincare scatter plot, LI is the diagonal line starting from the origin, P is the point in the plot, and the set of points above LI is defined as A with a number of a, the set of points below LI is defined as B with a number of b, and the set of points not on LI is defined as TO with a number of tot. PI calculates the percentage of b and tot using the following formula:

[0077]

[0078] GI is calculated as the percentage of the distance from A to LI and the distance from TO to LI. The calculation formula is as follows:

[0079]

[0080]

[0081] in is the distance from a point in A to LI, is the distance from the point in B to LI. SI is calculated as the percentage of the angle between A and LI and the angle between TO and LI. The calculation formula is as follows:

[0082]

[0083]

[0084] where θ i is the angle between the line connecting point P and the origin and the X-axis, θ LI is the angle between LI and the X axis.

[0085] ApEn calculates the complexity of the sequence. First, a respiratory parameter sequence of length N is divided into N-m+1 vectors using a window of width m. The calculation formula is as follows:

[0086] X=x(1),x(2),...,x(i),...,x(Nm),x(N-m+1)

[0087] Where each vector is defined as x(i) = [x i ,x i+1 ,...,x i+m-1 Based on the maximum distance from each vector to other vectors, we can get the ratio of the number of vectors with offsets greater than the threshold to the total number. The calculation formula is as follows:

[0088] d[x(i),x(j)]=max(|x i+k-1 -x j+k-1 |,1≤k≤m)

[0089]

[0090] Among them, the maximum distance d[x(i),x(j)] is the calculation formula of the maximum distance between vectors. Finally, calculate C i m The mean of (r) is obtained to obtain ApEn:

[0091]

[0092] ApEn(m,r,N)=Φ m (r)-Φ m+1 (r)

[0093] Where m is the size of the window, r is the tolerance value, and N is the length of the sequence. SampEn is based on ApEn and removes x(i) The impact of In the solution, i≠j, then calculate The mean C m (r), we can get SampEn:

[0094]

[0095]

[0096] FuzzEn adds fuzzy sets based on the previous two. First, normalize all vectors:

[0097]

[0098] Calculate the maximum distance between vectors and their similarity:

[0099]

[0100]

[0101] Where μ represents the fuzzy function, n represents the setting value in the fuzzy function, and finally the mean of all similarities except the vector itself is calculated to obtain FuzzEn:

[0102]

[0103]

[0104] The definitions of m and r are the same as SampEn, and n is set to 2.

[0105] AC and DC are two indicators in the Phase Rectified Signal Average (PRSA). First, the anchor point is determined by comparing each value in the sequence with the mean of the T values ​​before and after it:

[0106]

[0107]

[0108] Take 2L points around each anchor point to form a new sequence, which can be expressed as:

[0109]

[0110] in is the vth anchor point. Align each new sequence at the anchor point position and calculate the average:

[0111]

[0112] Where M represents the number of anchor points. Finally, a simple wavelet analysis is performed on the obtained sequence for evaluating AC and DC:

[0113]

[0114]

[0115] Where s is the scale of the wavelet, p is the phase, is the Haar wavelet, This can be equivalent to the AC / DC coefficient of variation of the respiratory parameter sequence. In this example, L is set to 120, and T and s are tested from 0 to 10. Finally, the AC / DC coefficient of variation of the specific respiratory parameter is calculated based on the T and s combination that maximizes the AUC.

[0116] Finally, the mean (AVE), median (MED), quartile (QUA), and third-quartile (TQUA) were calculated to reflect the distribution of the respiratory parameters themselves. The resulting 128 respiratory variability indices were recorded as "variability method - respiratory parameter." For example, the coefficient of variation for respiratory rate can be expressed as "CV-RR." Table 2 categorizes all respiratory variability indices into four groups based on respiratory metric type and variability method.

[0117] Table 2 Summary and classification of respiratory variability indicators

[0118]

[0119] S4. Select one or more component sets from the calculated respiratory variability indicators as the respiratory characteristics of the corresponding patients, and construct sample data consisting of the respiratory characteristics of each patient and the corresponding extubation outcome to obtain a training data set; based on the obtained training data set, use the patient's respiratory characteristics as input and the corresponding extubation outcome as output to train and obtain a patient extubation outcome prediction model.

[0120] Specifically, in this step, selecting one or more components of the calculated respiratory variability index as the respiratory characteristics of the corresponding patient can be based on random selection, and then the final set is determined based on the performance evaluation of the patient extubation outcome prediction model; or other reasonable screening methods can be used. As an optional solution, step S4 is specifically as follows:

[0121] S4-1. Evaluate the distribution of each respiratory variability indicator in the included samples, the significance level (p), and the area under the receiver operating characteristic curve (ROC) (AUC);

[0122] The distribution of respiratory variability indices in the successful and failed extubation groups was statistically analyzed. If the respiratory variability indices followed a normal distribution, a t-test was used; otherwise, a Wilcoxon test was used. Respiratory variability indices with a p < 0.05 were considered statistically significant. Receiver operating characteristic (ROC) analysis was also performed to estimate the area under the curve (AUC) of individual respiratory variability indices.

[0123] S4-2. Based on feature engineering, a feature set H is formed from significant (p < 0.05) respiratory variability indicators.

[0124] S4-3. Initialize or update the length g of the feature subset G, where the initial value of g is 0; select a feature from the feature set H and merge it with the feature subset to form a new set G. * , the length is g+1.

[0125] S4-4. Calculate G * Average AUC for five-fold cross-validation in the logistic regression (LR) model. Due to the significant imbalance between extubation failure and extubation success samples, the samples needed to be balanced before model training. This was achieved using the synthetic minority class oversampling technique (SMOTE).

[0126] S4-5. Repeat steps S4-3 to S4-4 until all respiratory variability indicators in the feature set H have participated in the AUC calculation. The respiratory variability indicator with the highest average AUC is added to the feature subset G and deleted from the feature set H.

[0127] S4-6. Based on the feature subset G, a set of calculated respiratory variability indicators is selected as the respiratory characteristics of the corresponding patients, and the respiratory characteristics of each patient and the corresponding extubation outcome are composed of sample data to obtain a training data set; based on the obtained training data set, the patient's respiratory characteristics are used as input and the corresponding extubation outcome is used as output to train a patient extubation outcome prediction model and evaluate the performance of the patient extubation outcome prediction model; the patient extubation outcome prediction model can be a general machine learning model, such as random forest, decision tree, support vector machine, neural network, etc. Taking the gradient boosting decision tree XGBoost as an example, the performance of the training data set under the gradient boosting decision tree XGBoost under the feature subset G is calculated, including AUC, accuracy ACC, sensitivity SEN and specificity SPE.

[0128] S4-7. Repeat S4-3 to S4-6 until the average AUC of the five-fold cross validation of the feature subset G is greater than the threshold; the feature subset G with the maximum AUC under the five-fold cross validation and the patient extubation outcome prediction model obtained through training are used to predict the extubation outcome.

[0129] Table 3 shows the comparison of the ablation experiment results on the XGBoost model for the best feature combination obtained after selection by the above method in different indicator groups.

[0130] Table 3 Comparison of ablation experiments on multiple respiratory variability indicators

[0131]

[0132] The results show that the respiratory variability index proposed in the present invention can achieve a higher accuracy in predicting extubation outcomes in the machine learning model.

[0133] This invention describes a method for constructing a model to predict extubation outcomes in patients undergoing invasive mechanical ventilation based on respiratory variability and machine learning. In this example, respiratory variability is analyzed before extubation in patients with non-brain injuries and neurological disorders, and a machine learning model is trained for extubation prediction. However, this invention is not limited to predicting extubation outcomes and has general applicability to other clinical decision-making support analyses.

[0134] An embodiment of the present invention also provides a prediction system for an extubation outcome prediction model based on respiratory variability and machine learning. The system includes a prediction module, which inputs the patient's respiratory characteristics into the patient extubation outcome prediction model constructed by the above-mentioned construction method to obtain the corresponding extubation outcome.

[0135] The respiratory variability index training machine learning model proposed in the present invention is used as an extubation outcome prediction model, which can obtain a high accuracy rate for extubation outcome prediction, and the method is highly interpretable. By introducing comprehensive respiratory indicators and complex and comprehensive statistical methods of physiological variability, many respiratory variability indicators have been developed, and multiple respiratory variability indicators have been selected under feature engineering to train machine learning models for extubation outcome prediction. The introduction of comprehensive respiratory indicators such as MP can better reflect respiratory control, a complex system involving the respiratory system, central nervous system, chemoreceptors, motor receptors and limbic system, and avoid the limitations caused by focusing only on respiratory movement. Secondly, the outstanding contribution of complex variability methods such as PRSA also shows that there is asymmetry in the acceleration and deceleration of respiratory parameters, as well as other variability characteristics that exist before extubation to be discovered.

[0136] In the above embodiments, the present invention is only described for exemplary purposes. However, after reading this patent application, those skilled in the art may make various modifications to the present invention without departing from the spirit and scope of the present invention.

Claims

1. A method for constructing an extubation outcome prediction model based on respiratory variability and machine learning, characterized in that: The following steps are involved: S1. Collect the extubation outcome, respiratory waveform data before extubation and under SBT, and corresponding ventilator parameters of each selected patient; S2, based on the respiratory waveform data before extubation and under SBT and the corresponding ventilator parameters, the corresponding respiratory parameter sequence is calculated; wherein the respiratory parameter sequence corresponding to each patient includes PIP, RR, V T , one or more of MV, RSBI, and MP; S3. Calculate a respiratory variability index for each respiratory parameter sequence based on a variability statistical analysis method, wherein the respiratory variability index includes a basic statistical method standard deviation SD, a coefficient of variation CV, a comprehensive statistical method standard deviation 1SD1, a standard deviation 2SD2, a Porta index PI, a Guzik index GI, a Slope index SI, an approximate entropy ApEn, a sample entropy SampEn, a fuzzy entropy FuzzEn, an acceleration performance AC, a deceleration performance DC, and one or more of a mean AVE, a median MED, a quartile QUA, and a three-quarter digit TQUA reflecting the distribution of the respiratory parameter itself; S4. Select one or more components of the calculated respiratory variability index as the respiratory characteristics of the corresponding patient, and construct a training dataset by combining the respiratory characteristics of each patient and the corresponding extubation outcome sample data; based on the obtained training dataset, train a patient extubation outcome prediction model using the patient's respiratory characteristics as input and the corresponding extubation outcome as output; The step S2 further includes a screening step: Invalid breaths are screened based on thresholds corresponding to the respiratory parameters and / or based on outliers. If one of the respiratory parameters exceeds the threshold and is an outlier, the breath corresponding to the respiratory parameter is considered invalid, and the invalid breath parameters in all respiratory parameter sequences are deleted. RR is the number of breaths per minute calculated in breaths per minute. The calculation formula is as follows: Where T(n) is the period of the nth breath, V T The tidal volume of one breath is calculated in milliliters. The calculation formula is as follows: Where F(n,t) is the flow rate of the nth breath, T i (n) is the inspiratory cycle of the nth breath; MV is calculated as the minute ventilation of one breath in liters per minute. The calculation formula is as follows: RSBI is calculated as RR and V T The ratio is a dimensionless value and is calculated as follows: MP includes MP d [J / breath]、MP d [J / L] and MP d [J / min], where MP d [J / breath] calculates the dynamic part of the mechanical work of a breath, in joules per breath, using the following formula: Where P(V) is the pressure-volume loop function, and PEEP is the positive end-expiratory pressure; MP d [J / L] The calculation is based on the mechanical work of one breath, taking V into account. T The unit is joules per liter, and the calculation formula is as follows: MP d [J / min] calculates the portion of the mechanical work of one breath that takes into account RR, expressed in joules per minute, using the following formula: MP d [J / min]=MP d [J / breath]×RR(n)。 2. The method according to claim 1, characterized in that Step S4 is specifically as follows: S4-1. Statistical analysis was performed on the distribution of each respiratory variability indicator in the extubation success and extubation failure groups. Respiratory variability indicators with p < 0.05 were considered statistically significant. Receiver-operating characteristic (ROC) analysis was also performed to estimate the area under the curve (AUC) of each respiratory variability indicator. S4-2, combining statistically significant respiratory variability indicators to form a feature set H; S4-3, initialize or update the length g of the feature subset G, where the initial value of g is 0; select a respiratory variability index from the feature set H and merge it with the feature subset to form a new set G * , length is g+1; S4-4. Calculate G * Average AUC of five-fold cross validation under the logistic regression LR model; S4-5. Repeat steps S4-3 to S4-4 until all respiratory variability indices in feature set H have participated in the AUC calculation; add the respiratory variability indices with the highest average AUC to feature subset G and delete them from feature set H; S4-6. Based on the feature subset G, select a set of calculated respiratory variability indices as the respiratory features of the corresponding patient, and construct a training dataset by combining the respiratory features of each patient and the corresponding extubation outcome as sample data. Based on the obtained training dataset, use the patient's respiratory features as input and the corresponding extubation outcome as output to train a patient extubation outcome prediction model and evaluate the performance of the patient extubation outcome prediction model. S4-7. Repeat S4-3 to S4-6 until the average AUC of the five-fold cross validation of the feature subset G is greater than the threshold; the feature subset G with the maximum AUC under the five-fold cross validation and the patient extubation outcome prediction model obtained through training are used to predict the extubation outcome.

3. The method according to claim 2, characterized in that When statistically analyzing the distribution of each respiratory variability index in the extubation success and extubation failure groups, if the respiratory variability index obeyed the normal distribution, the t test was used, otherwise the Wilcoxon test was used.

4. The method according to claim 2, characterized in that The step S4-4 further includes using the synthetic minority class oversampling technique SMOTE to balance the patient samples of failed extubation and successful extubation.

5. The method according to claim 2, characterized in that The performance indicators of the prediction model for patient extubation outcomes included: AUC, accuracy ACC, sensitivity SEN, and specificity SPE.

6. The method according to claim 1, characterized in that The structure of the patient extubation outcome prediction model is a random forest, support vector machine, decision tree or neural network model.

7. A prediction system for extubation outcome prediction model based on respiratory variability and machine learning, characterized by: The system includes a prediction module, which inputs the patient's respiratory characteristics into the patient extubation outcome prediction model constructed by the construction method according to any one of claims 1 to 6 to obtain the corresponding extubation outcome.

Images