Breathing quality evaluation method based on multiple parameters, big data and data mining

Abstract

The invention relates to a breathing quality evaluation method based on multiple parameters, big data and data mining. The method comprises the following steps: acquiring parameter data of an artificial respirator to be predicted; parameter data of the artificial respirator to be predicted are input into an MSAGDBO-LSTM model, parameter indexes influencing the breathing quality are output, the MSAGDBO-LSTM model is obtained after training of a training set and testing of a testing set, and the MSAGDBO-LSTM model is obtained by optimizing an LSTM network through an improved dung beetle optimization algorithm; the parameter indexes are input into a random forest model, a breathing quality evaluation grade result is obtained, and the random forest model is obtained through historical parameter indexes and label training of the corresponding breathing quality. According to the method, the state of the breathing quality can be effectively evaluated in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical and health care, and particularly relates to a respiratory quality evaluation method based on multi-parameters, big data and data mining. BACKGROUND

[0002] Artificial respiration is an important medical measure to replace the life-sustaining function of respiration, and an artificial respirator is a main device for realizing artificial respiration and is widely used in respiratory failure caused by various reasons, respiratory support treatment, emergency cardiopulmonary resuscitation and respiratory management during general anesthesia. When artificial respiration is implemented, each quantity such as a ventilation mode, a ventilation frequency, an expiratory-inspiratory time ratio and an upper limit value of an airway pressure must be properly set to obtain appropriate ventilation conditions. If appropriate ventilation conditions cannot be obtained, the life of a patient may be in danger. For example, if the air pressure in the airway is too high, there is a risk of alveolar rupture, and conversely, if the pressure in the airway is too low, oxygen supply is insufficient. The characteristics of the human respiratory system differ from person to person, and the respiratory rhythm and the elasticity of the lungs differ due to differences in age / sex / physical condition and change as the disease progresses or regresses. When artificial respiration is implemented, each quantity such as a ventilation mode, a ventilation frequency, an expiratory-inspiratory time ratio must be properly set to obtain appropriate ventilation conditions and obtain effective respiration, so the effectiveness and rationality of the parameter setting of the artificial respirator are of the utmost importance and are key issues. In view of the difficulty that there is no set of personalized parameter setting standards for the artificial respirator at present, a respiratory quality evaluation method based on multi-parameters, big data and data mining is proposed. SUMMARY

[0003] The purpose of the present application is to provide a respiratory quality evaluation method based on multi-parameters, big data and data mining, which can effectively evaluate the state in which the respiratory quality is located in real time.

[0004] To achieve the above-mentioned purpose, the present application provides the following scheme:

[0005] A respiratory quality evaluation method based on multi-parameters, big data and data mining, comprising:

[0006] obtaining artificial respirator parameter data to be predicted;

[0007] inputting the artificial respirator parameter data to be predicted into an MSAGDBO-LSTM model to output parameter indexes affecting respiratory quality, wherein the MSAGDBO-LSTM model is obtained based on training of a training set and testing of a test set, the training set and the test set include historical parameter data and corresponding parameter indexes, the MSAGDBO-LSTM model is obtained by optimizing an LSTM network through an improved catharsius optimization algorithm, and the parameter indexes include blood oxygen saturation, dynamic lung compliance and rapid shallow breathing index;

[0008] Input the parameter index into a random forest model to obtain a respiratory quality evaluation grade result, wherein the random forest model is obtained by training historical parameter indexes and corresponding respiratory quality labels.

[0009] Optionally, the artificial respirator parameter data to be predicted comprises airway peak pressure, mean airway pressure, positive end-expiratory pressure, tidal volume, minute ventilation, respiratory rate, minute air leakage, exhalation time, spontaneous breathing rate, and oxygen concentration fraction in inhaled air.

[0010] Optionally, obtaining the training set and the test set comprises:

[0011] Obtaining original historical parameter data and corresponding parameter indexes and performing format conversion processing and preprocessing to obtain the training set and the test set.

[0012] Optionally, the MSAGDBO-LSTM model is obtained by optimizing an LSTM network through an improved Melolontha optimization algorithm, and the method comprises the following steps:

[0013] S1, defining a target function and setting parameters;

[0014] S2, initializing a Melolontha population by using Bernoulli chaotic mapping and calculating the fitness value of each Melolontha, wherein each individual of the Melolontha represents an LSTM network hyperparameter;

[0015] S3, when δ < ST, updating the position of a rolling Melolontha according to a traditional Melolontha optimization algorithm, otherwise, updating the position of the rolling Melolontha according to an improved sine algorithm, δ is a parameter, and ST ∈ (0.5, 1];

[0016] S4, performing boundary judgment on each Melolontha to update the positions of a breeding ball, a small Melolontha, and a stealing Melolontha;

[0017] S5, calculating the fitness value of each Melolontha after updating, obtaining a current optimal solution, perturbing the current optimal solution by using an adaptive Gaussian-Cauchy hybrid mutation disturbance strategy, and obtaining a new optimal solution;

[0018] S6, comparing the fitness of the current optimal solution and the new optimal solution according to a greedy mechanism to determine whether to update the position;

[0019] S7, if the number of iterations has not reached a preset maximum number of iterations, then returning to S3 after the number of iterations is increased by one, and if the number of iterations reaches the preset maximum number of iterations, then outputting the result of the best network hyperparameter.

[0020] Optionally, initializing the Melolontha population by using Bernoulli chaotic mapping comprises:

[0021]

[0022] wherein beta is a mapping parameter, z n is a current position, z n+1 is an updated position.

[0023] Optionally, the step of perturbing the current optimal solution by using an adaptive Gaussian-Cauchy hybrid mutation perturbation strategy comprises:

[0024] H b (t) = X b (t) * (1 + mu1 * Gauss(sigma) + mu μ cauchy(sigma))

[0025] wherein t is a current iteration number, X b (t) is an optimal position of an individual X in the tth iteration, H b (t) is a position of the optimal position X b (t) in the tth iteration after Gaussian-Cauchy hybrid perturbation, Gauss(sigma) is a Gaussian mutation operator, cauchy(sigma) is a Cauchy mutation operator, T max a maximum iteration number.

[0026] Optionally, the step of comparing the fitness of the current optimal solution and the new optimal solution according to a greedy mechanism to determine whether to update the position comprises:

[0027]

[0028] wherein f(x) represents a fitness value of a position x.

[0029] The application has the beneficial effects that the MSAGDBO-LSTM model is used to predict the blood oxygen saturation, Cdyn and RSB reflecting the effectiveness of using an artificial respirator, and the effectiveness of a prediction scheme is obtained based on the predicted values and actual values of the blood oxygen saturation, Cdyn and RSB; the RF random forest algorithm is used to evaluate the grade of respiratory quality based on the blood oxygen saturation, Cdyn and RSB reflecting the effectiveness of using an artificial respirator, so that the state of the respiratory quality can be effectively evaluated, and the ventilation parameters of the artificial respirator are adjusted to improve the quality of respiration, which has practical value and guiding significance. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0031] Figure 1A flow chart of a respiratory quality evaluation method based on multi-parameters, big data and data mining for an embodiment of the present application;

[0032] Figure 2 A workflow diagram of the MSAGDBO-LSTM model for an embodiment of the present application;

[0033] Figure 3 A training fitting diagram of the RSB shallow and rapid breathing index for an embodiment of the present application;

[0034] Figure 4 A prediction fitting diagram of the RSB shallow and rapid breathing index for an embodiment of the present application;

[0035] Figure 5 A training fitting diagram of the Cdyn compliance for an embodiment of the present application;

[0036] Figure 6 A prediction fitting diagram of the Cdyn compliance for an embodiment of the present application;

[0037] Figure 7 A comparison diagram of the predicted grade and the actual grade for an embodiment of the present application;

[0038] Figure 8 A confusion matrix diagram for an embodiment of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0040] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0041] Medical treatment generally includes artificial respiration, and some factors such as air pressure, flow, breathing frequency, inspiration expiration time ratio and the like are factors that obviously affect the state of respiration. How the state of respiration is directly related to the next measures to be taken, and it is of great significance to evaluate the state of respiration. The oxygen saturation in the blood directly represents whether the respiration is sufficient, and the blood oxygen saturation refers to the capacity of oxygenated hemoglobin combined in the blood, which accounts for the percentage of the total combined hemoglobin capacity, that is, the concentration of blood oxygen in the blood, which is a very important parameter in the respiratory cycle; Cdyn(dynamic lung compliance) is an important physiological parameter in medicine, which helps to evaluate the lung function; RSB(shallow and rapid breathing index) is an index for evaluating the success of mechanical ventilation weaning, mechanical ventilation refers to providing artificial ventilation support for patients through a ventilator in an intensive care unit or the like, when the respiratory function recovers, the support of the ventilator needs to be gradually reduced, and finally the weaning is realized. The shallow and rapid breathing index can help doctors improve the prediction ability of weaning success, thereby guiding the management and treatment of mechanical ventilation.

[0042] As shown in Figure 1 The embodiment provides a respiration quality evaluation method based on multi-parameter, big data and data mining, which comprises the following steps:

[0043] Obtaining artificial respiration parameter data to be predicted;

[0044] Inputting the artificial respiration parameter data to be predicted into an MSAGDBO-LSTM model, and outputting parameter indexes affecting respiration quality, wherein the MSAGDBO-LSTM model is obtained based on training and testing of a training set and a test set, the training set and the test set comprise historical parameter data and corresponding parameter indexes, the MSAGDBO-LSTM model is obtained by optimizing an LSTM network through an improved ball rolling optimization algorithm, and the parameter indexes comprise blood oxygen saturation, dynamic lung compliance and shallow and rapid breathing index.

[0045] Inputting the parameter indexes into a random forest model, and obtaining a respiration quality evaluation grade result, wherein the random forest model is obtained by training historical parameter indexes and corresponding respiration quality labels.

[0046] Further, the artificial respiration parameter data to be predicted comprises airway peak pressure, mean airway pressure, positive end-expiratory pressure, tidal volume, minute ventilation, breathing frequency, minute air leakage, expiration time, spontaneous breathing frequency and oxygen concentration fraction in inhaled air.

[0047] Further, obtaining the training set and the test set comprises:

[0048] Obtaining original historical parameter data and corresponding parameter indexes, and performing format conversion processing and preprocessing, to obtain the training set and the test set.

[0049] Further, the MSAGDBO-LSTM model is obtained by improving the optimization of the LSTM network through the improved M. mutans optimization algorithm, which comprises the following steps:

[0050] S1, defining a target function and setting parameters;

[0051] S2, initializing the M. mutans individual position population by using Bernoulli chaotic mapping, and calculating the fitness value of each M. mutans, wherein each M. mutans individual represents an LSTM network hyperparameter;

[0052] S3, when δ < ST, updating the position of the rolling M. mutans according to the traditional M. mutans optimization algorithm, otherwise, updating the position of the rolling M. mutans according to the improved sine algorithm; δ is a parameter, δ = rand(1), and ST ∈ (0.5, 1].

[0053] S4, updating the positions of the breeding ball, the small M. mutans and the stealing M. mutans for each M. mutans through boundary judgment;

[0054] S5, calculating the fitness value of each M. mutans after updating, obtaining the current optimal solution, and using the adaptive Gaussian-Cauchy hybrid mutation disturbance strategy to disturb the current optimal solution to obtain a new optimal solution;

[0055] S6, comparing the fitness of the current optimal solution and the new optimal solution according to the greedy mechanism to determine whether to update the position;

[0056] S7, if the iteration number has not reached the preset maximum iteration number, then the iteration number is increased by one and the process returns to S3, and if the iteration number has reached the preset maximum iteration number, then the result of the best network hyperparameter is output.

[0057] Specifically, the unique feature of the LSTM unit is that it contains three different types of gates: a forget gate, an input gate, and an output gate. LSTM can effectively retain and utilize long-term information when processing long sequence data, thereby significantly improving the performance of traditional RNN, especially in tasks such as respiratory-related time series prediction, thereby improving the prediction ability and accuracy of the model. Although the DBO algorithm is better than other algorithms in optimization, it has strong optimization ability and fast convergence speed, but when solving some complex problems, it still has an imbalance between global exploration and local development, and is prone to fall into local optimum, and the global exploration ability is weak, so in order to improve the exploration ability of DBO, the original optimization algorithm is improved, and the Bernoulli mapping strategy, the embedded improved sine algorithm strategy and the adaptive Gaussian-Cauchy mutation disturbance are used to improve its shortcomings.

[0058] Further, the M. mutans population is initialized by using Bernoulli chaotic mapping, which comprises the following steps:

[0059]

[0060] wherein β is a mapping parameter, z n is a current position, z n+1 is an updated position.

[0061] Further, the disturbing the current optimal solution by using the adaptive Gauss-Cauchy hybrid mutation disturbance strategy comprises:

[0062] H b (t) = X b (t) x (1 + μ1 x Gauss (σ) + μ2 cauchy (σ))

[0063] wherein t is a current iteration number, X b (t) is an optimal position of the individual X in the tth iteration, H b (t) is the optimal position X b (t) in the tth iteration after Gauss-Cauchy hybrid disturbance, Gauss (σ) is a Gauss mutation operator, cauchy (σ) is a Cauchy mutation operator, T max is a maximum iteration number.

[0064] Further, the determining whether to update the position according to the comparison of the fitness of the current optimal solution and the new optimal solution according to the greedy mechanism comprises:

[0065]

[0066] wherein f represents a fitness value.

[0067] In order to obtain effective and high-quality breathing supply, according to the breathing quality evaluation grade result, that is, the states of excellent, good and poor, the parameters of the artificial breathing machine setting are optimized in reverse by using SSA, the adjustment trend of the breathing machine setting parameters is given, and data information for adjusting the parameters of the artificial breathing machine is provided for medical staff.

[0068] The method of the embodiment will be further described below in combination with the drawings:

[0069] The embodiment provides a kind of based on artificial breathing machine parameters such as airway peak pressure pip, average airway pressure pmean, positive end-expiratory pressure Peep, tidal volume VT, minute ventilation Mve, respiratory rate RR, minute air leakage Mvleak, expiratory time Te, spontaneous breathing rate Rrspon, oxygen concentration fraction Fio2 in inhaled air etc.;And MSAGDBO-LSTM model predicts the parameter index such as blood oxygen saturation, dynamic lung compliance Cdyn, shallow and fast breathing index RSB that influence breathing quality;And based on blood oxygen saturation, Cdyn, RSB etc. Parameter and RF random forest model are carried out grade evaluation to breathing quality.

[0070] The MSAGDBO-LSTM model prediction and RF random forest evaluation can predict and evaluate the respiratory state of the patient in real time, so as to effectively adjust the ventilation parameters of the artificial respirator, improve the quality of respiration, and has practical value and guiding significance.

[0071] A respiratory quality evaluation method based on multi-parameter, big data and data mining, specifically comprising:

[0072] Step 1: Obtain effective samples:

[0073] The samples take 1041 groups of data recorded every 30 minutes during the ICU period of 21 people in the ICU of a city-level large three-level first-class hospital as the benchmark, including the data of peak airway pressure pip, mean airway pressure pmean, positive end-expiratory pressure Peep, tidal volume VT, minute ventilation Mve, respiratory rate RR, minute air leakage Mvleak, exhalation time Te, spontaneous breathing rate Rrspon, oxygen concentration fraction in inhaled air Fio2, blood oxygen saturation, Cdyn, RSB and other parameters.

[0074] Step 2: Preprocessing of sample data:

[0075] Convert the original data into a format that can be processed by the program.

[0076] Step 3: MSAGDBO-LSTM model predicts blood oxygen saturation, dynamic lung compliance Cdyn, shallow and rapid breathing index RSB and other parameter indicators affecting respiratory quality:

[0077] 3.1 dung beetle optimization algorithm:

[0078] The dung beetle optimization algorithm (DBO) is a swarm intelligence optimization algorithm based on the behavioral characteristics of dung beetles. The algorithm simulates the behaviors of rolling balls, reproduction, foraging and stealing of dung beetles, and designs a series of update rules and strategies. Each group of dung beetles is composed of four different agent dung beetles, namely rolling dung beetles, breeding dung beetles (breeding balls), small dung beetles and stealing dung beetles.

[0079] (1) Rolling dung beetle:

[0080] Dung beetles roll feces into balls and roll them to suitable locations. When rolling, they use clues such as the sun or wind direction to maintain a straight line. In order to simulate this behavior in the algorithm, the dung beetles need to move in a given direction in the search space. The positions of these dung beetles are updated during the rolling process, and their position changes are:

[0081] x i (t+1)=x i (t)+αkx i (t-1)+bΔx

[0082] Δx = |x i (t) - X w | (1)

[0083] where t represents the iteration number; x i (t) represents the position information of the ith dung beetle at the tth iteration; a represents a natural coefficient, which is assigned a value of 1 or -1, where a = 1 represents no deviation in direction and a = -1 represents deviation in direction; k e (0, 0.2] represents a deflection coefficient; b represents a constant belonging to (0, 1); X w represents the global worst position; and Δx is used to simulate the change in light intensity.

[0084] When these dung beetles encounter an obstacle in front of them that blocks their way forward, they can re-plan their route of advance by means of the dance behavior, in which case the position update formula for these dung beetles is as follows:

[0085] x i (t+1) = x i (t) + tan(0)|x i (t) - x i (t-1)| (2)

[0086] where 0 e [0, π] is the deflection angle, and when 0 = 0, and π, the position of the dung beetle will not be updated.

[0087] (2) Breeding dung beetles (breeding balls):

[0088] In order to safely breed offspring, the dung beetles will roll the fecal ball to a safe place and hide it, and lay eggs inside the fecal ball, so the boundary selection strategy for the dung beetles is as follows:

[0089]

[0090] where and respectively represent the lower and upper bounds of the egg-laying region; X * represents the current optimal position; R = 1 - t / T max , T max is the maximum number of iterations; b L and b U respectively represent the lower and upper bounds of the optimization problem.

[0091] After the position of the egg-laying region is determined, the female dung beetle will select a breeding ball in this region to lay eggs, and each female dung beetle lays only one egg in each iteration. It can be seen from the above formula (9) that the boundary of the egg-laying region is dynamic and depends on the value of R. Therefore, the position of the breeding ball is also dynamically changed during the iteration process, which is represented as:

[0092]

[0093] where B i (t) is the position of the i-th fecal pellet ball at the t-th iteration; b1 is a D-dimensional random vector following a normal distribution, and b2 represents a D-dimensional random vector in [0, 1].

[0094] (3) Small dung beetle:

[0095] When some larvae grow into adult dung beetles and crawl out of the ground to forage, they are called small dung beetles. The boundary of the optimal foraging area of small dung beetles is defined as follows:

[0096]

[0097] where and represent the lower and upper bounds of the optimal foraging area of small dung beetles, respectively; X b represents the global optimal position. Therefore, the foraging position of small dung beetles is updated as follows:

[0098]

[0099] where x i (t) is the position information of the i-th small dung beetle at the t-th iteration; C1 represents a random number following a normal distribution; C2 represents a random vector belonging to (0, 1).

[0100] (4) Stealing dung beetle:

[0101] Some dung beetles will steal their fecal balls from other dung beetles, and these fecal ball stealing dung beetles are called stealing dung beetles. In equation (7), X b is the best food source, so it can be considered that the position near X b is the best position for food competition. During the iteration process, the position of the stealing dung beetle is constantly updated, which can be described as:

[0102] x i (t+1) = x b +S×g(|x i (t)-X * |+|x i (t)-X b |) (7)

[0103] where x i (t) represents the position information of the i-th thief at the t-th iteration; g is a 1 × D random vector following a normal distribution; S represents a constant.

[0104] 3.2 Improved dung beetle optimization algorithm, as shown in Figure 2 :

[0105] Although the DBO algorithm is better than other algorithms in optimization, it has strong optimization ability and fast convergence speed, but when solving some complex problems, it is easy to fall into local optimum and has weak global exploration ability. In order to improve the exploration ability of DBO, the original optimization algorithm is improved by Bernoulli mapping strategy, embedded improved sine algorithm strategy and adaptive Gaussian-Kossi variation disturbance to improve its shortcomings.

[0106] (1) Define the objective function and set the parameters: objective function: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), mean absolute percentage error (MAPE), etc.

[0107] (2) Bernoulli mapping initializes the population:

[0108] The population initialization of the improved DBO algorithm uses random population generation. The shortcomings of this method are uneven distribution of dung beetle positions, weak global exploration ability, low population diversity and easy to fall into local optimum. Chaos mapping is a combination of determinism and randomness. Chaos has randomness, non-periodicity and other characteristics. Chaos initialization can improve the search breadth of the optimization algorithm and can be used to solve global optimization problems. There are many kinds of chaos mapping, and Bernoulli mapping belongs to one of them. It can replace random number initialization population in the optimization field and improve the quality of dung beetle population distribution and global search ability. Therefore, Bernoulli mapping is used to initialize the position of dung beetle individuals. First, the pseudo-random number, i.e. the randomly generated model hyperparameter, is projected into the chaotic variable space using the Bernoulli mapping relationship, and then the generated chaotic value is mapped to the initial space of the algorithm through linear transformation. The specific expression of Bernoulli mapping is:

[0109]

[0110] Where β is the mapping parameter, β∈(0,1), and β=0.518, z0=0.326 to achieve the best value effect.

[0111] This method almost uniformly (equidistantly) distributes the initial points in the unit interval and ensures that the second half of the generated sequence will not overlap. Statistical tests also show that the generated sequence has good randomness. This makes the Bernoulli mapping initialization population more evenly distributed, improves the quality and diversity of the population, and avoids falling into local optimum.

[0112] (3) Improve using improved sine algorithm:

[0113] The improved sine algorithm (MSA) strategy is inspired by the Sine Cosine Algorithm (SCA)

[25] The Sine Algorithm (SA) and Exponential Sine Cosine Algorithm (ESCA) functions and improved sine cosine algorithm (ISCA) and other SCA-related algorithms use the sine function in mathematics for iterative optimization, and have strong global exploration ability. In order to achieve a good level of global exploration and local development ability, an adaptive variable inertia weight coefficient ω t is added in the position update process. The improved sine algorithm position update formula is shown in equation (9):

[0114] x i (t+1)=ω t x i (t)+r1×sin(r2)×[r3p i (t)-x i (t)] (9)

[0115] where t is the current iteration number, ω t is the inertia weight, x i (t) is the i-th position component of individual X in the t-th iteration, p i (t) is the i-th component of the best individual position variable in the t-th iteration, r1 is a nonlinear decreasing function, r2 is a random number in the interval [0, 2π], and r3 is a random number in the interval [-2, 2].

[0116] r1 represents the search distance and direction of the dung beetle, which optimizes the optimization method of the DBO algorithm, and its value is:

[0117]

[0118] where ω max and ω min represent the maximum and minimum values of ω t , t represents the current iteration number, and T max represents the maximum iteration number.

[0119] The adaptive coefficient ω t gradually reduces the search space, and the inertia weight decreases with the increase of the iteration number. The relatively large inertia weight in the early stage of algorithm iteration enables the algorithm to have strong global exploration ability, while the relatively small inertia weight in the later stage of iteration helps to improve its local development ability. The adaptive coefficient ωt The formula is:

[0120]

[0121] To further improve the global exploration and local exploitation capabilities of the DBO algorithm, a sinusoidal guidance mechanism was introduced. This mechanism guides the dung beetle's position update by performing a sinusoidal operation on the entire individual dung beetle during the rolling phase. The improved formula is as follows:

[0122]

[0123] In the improved position update formula, where δ = rand(1) and ST ∈ (0.5, 1), when δ < ST, it indicates that the dung beetle is rolling with a goal in mind and is in the normal global exploration phase. When δ > ST, it means that the dung beetle does not have a clear rolling goal, but will search and move in a sinusoidal manner. By introducing this improved sinusoidal guidance mechanism, on the one hand, it can greatly improve the defect of the DBO algorithm's overly random position update strategy, and on the other hand, it can improve the problem that the original algorithm is prone to getting trapped in local optima. This MSA sinusoidal guidance mechanism allows the dung beetle to conduct global exploration and local optimization within the range given by the algorithm, which expands the search space to a certain extent and gradually converges to the same optimal solution, i.e., the objective function value, thereby improving the algorithm's global optimization ability.

[0124] (4) Adaptive Gaussian-Cauchy hybrid variation perturbation:

[0125] Mutation operators are added to the dung beetle optimization algorithm. Gaussian mutation and Cauchy mutation are two common mutation operators. Considering the advantages and disadvantages of the two, an adaptive Gaussian-Cauchy hybrid perturbation mutation perturbation strategy that combines the advantages of Cauchy mutation and Gaussian mutation is proposed. Its specific formula is shown in (19):

[0126] H b (t)=X b (t)×(1+μ1×Gauss(σ)+μ2cauchy(σ)) (13)

[0127] Among them, X b H(t) represents the optimal position of individual X in the t-th iteration. b (t) represents the optimal position X in the t-th iteration. b (t) represents the position after the Gaussian-Cauchy mixed perturbation, where Gauss(σ) is the Gaussian mutation operator and Cauchy(σ) is the Cauchy mutation operator.

[0128] (5) Greedy mechanism:

[0129] In order to ensure that the new position obtained after subsequent mutation disturbance is better than the original position, the greedy mechanism is added after the mutation disturbance update, and the fitness of the new and old positions is compared to determine whether the position needs to be updated.

[0130] f(x) represents the fitness value of x position, and the formula of the greedy mechanism is shown in (20):

[0131]

[0132] 3.3. The specific steps of MSAGDBO-LSTM are as follows:

[0133] S1, define the mean square error (MSE) of the LSTM model output as the objective function, set the initial parameters: the feature dimension of the input data is 10, the number of neurons in the hidden layer is 16, the iteration number is 10, the maximum training number is 100, the initial learning rate is 0.01, the learning rate is adjusted after 50 times of training, the running environment is cpu, and the regularization parameter is 0.005.

[0134] Optimizing the LSTM model hyperparameters includes: number of neurons, learning rate, and regularization parameter.

[0135] S2, initialize the population according to Bernoulli chaotic mapping, generate by pseudo-random number, calculate the fitness value of individual under the objective function.

[0136] S3, δ<ST select formula (1), otherwise according to formula (9), update the individual position according to the rolling ball behavior, and calculate the updated individual fitness value.

[0137] S4, update the position of the breeding ball according to boundary formula (3) formula (4); update the position of the small scarab according to boundary formula (5) formula (6); update the position of the stealing scarab formula (7).

[0138] S5, calculate the updated individual fitness value, and obtain the optimal solution.

[0139] S6, use the adaptive Gaussian-Cauchy mutation disturbance strategy to disturb the current optimal solution (scarab position), if the new individual fitness value is better than the original individual, replace the original individual, update the fitness value, and obtain the new solution (that is, the hyperparameters of LSTM), otherwise keep the individual position unchanged.

[0140] S7, according to the greedy mechanism formula (14) to judge whether to update the position, and get the new optimal solution.

[0141] S8, judge whether to meet the termination condition, if yes, get the optimal solution, and input the obtained optimal parameters into LSTM; if not, return to step S3.

[0142] The accuracy of the LSTM neural network model prediction is mainly related to the L2 regularization parameter, the learning rate and the neurons of the hidden layer. Although the DBO beetle algorithm can optimize these parameters of the LSTM, the final optimized result may have problems such as uneven initial population distribution, weak local development ability and easy to fall into local optimum. On this basis, the DBO algorithm is improved in the application, and the MSAGDBO sine optimization algorithm is proposed. The population is initialized by Bernoulli mapping, and the improved sine algorithm is used to optimize the position of the rolling beetle, so as to enhance the exploration ability of the algorithm in the local and global, and the adaptive variable inertia weight coefficient is introduced to improve the position of the stealing cockroach, so as to balance the exploration ability of the algorithm in the global and local, and finally the adaptive Gaussian-Cauchy hybrid disturbance mutation disturbance strategy is used to enhance the global search and jump out of the local optimum ability of the algorithm.

[0143] 3.4. Data sources and evaluation indicators:

[0144] (1) Data sources:

[0145] The data set comes from the data set provided by "a municipal third-grade hospital", in order to verify that the model proposed in the application can adapt to different individuals, the relevant data of different individuals using the respirator are selected, which contains 10 characteristic variables as input, including inspiratory peak pressure pip, mean airway pressure pmean, end-expiratory positive pressure peep, tidal volume VT, minute ventilation Mve, respiratory rate RR, air leakage Mvleak, expiratory time Te, spontaneous breathing rate Rrspon and oxygen concentration fraction FiO2. Respiratory data totals 7464 data. RSB and Cdyn both select the first 70% of the data as training samples, which are used to predict the RSB and Cdyn respiratory data of the last 30%, and the ratio of training samples to test samples is 7:3.

[0146] (2) Evaluation indicators:

[0147] The prediction accuracy of the prediction model is evaluated and compared by evaluation indicators, and the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), determination coefficient (R 2 ), and the four evaluation indicators are selected to measure the gap between the predicted value and the actual value from different angles. Among them, MSE is used as the objective function of the evaluation model performance, the smaller the fitness value, the higher the prediction accuracy. RMSE is usually used to represent the degree of dispersion of the result. MAE and MAPE can indicate the deviation of the prediction. In addition, the determination coefficient R 2 is used to measure the linear correlation between the actual value and the predicted value.

[0148] 3.5 Simulation analysis:

[0149] In order to verify the prediction effect of the model proposed in the application, 7464 sets of sample data of RSB and Cdyn are respectively input into the models of LSTM, PSO-LSTM, DBO-LSTM and MSAGDBO-LSTM for comparison, and the comparison fitting of the training of the four models is as shown in Figure 3 , 5 The comparison fitting of the test of the four models is as shown in Figure 4 , 6

[0150] The MSAGDBO-LSTM prediction model is used for prediction, the first 70% of the data is selected as the training set, and the last 30% is selected as the test set for prediction, the iteration number is 10, the maximum training number is 100, the initial learning rate is 0.01, the learning rate is adjusted after 50 times of training, the running environment is CPU, and the regularization parameter is 0.005. As shown in Table 1, the different model prediction evaluation indexes of RSB are compared, and as shown in Table 2, the different model prediction evaluation indexes of Cdyn are compared. According to Tables 1-2, the MAE of the MSAGDBO-LSTM RSB model is the smallest, and the prediction accuracy is the highest.

[0151] Table 1

[0152]

[0153]

[0154] Table 2

[0155]

[0156] Step 4: Respiratory quality evaluation based on RF random forest:

[0157] Based on the parameters of blood oxygen saturation, Cdyn, RSB and the like, the RF random forest algorithm is used to evaluate the level of respiratory quality as excellent, good and poor.

[0158] Based on the three indexes of blood oxygen saturation, Cdyn, RSB and the like, the classification evaluation is carried out, and according to the range of the three indexes in the sample data, the respiratory quality is set to belong to three levels of excellent, good and poor according to the medical rules.

[0159] The input is the three indexes of blood oxygen saturation, Cdyn, RSB and the like, and the output is the label of respiratory quality: excellent, good and poor. The RF classification model is established, and the basic parameter of the RF model is only one, that is, the number of decision trees. When classifying, with the increase of the number of decision trees, the classification error becomes smaller and smaller and finally converges to 0. The number of decision trees is set to 60.

[0160] ​Random forest classifier is a classification model based on ensemble learning that improves classification performance by combining multiple decision trees. In a random forest, each decision tree is built independently, trained using randomly selected features and samples, and finally, the classification results of each decision tree are voted on to obtain the final classification result.

[0161] 4.1 Data Sampling (Bootstrap Sampling):

[0162] The first step in random forest is to randomly sample multiple subsets of samples with replacement from the original dataset. This process is called bootstrap sampling. Each subset of samples is the same size as the original dataset, but because it is sampling with replacement, some samples may be sampled repeatedly while others may not be sampled.

[0163] Let the original dataset be D = {(x1,y1),(x2,y2),...,(x N ,y N )}, where N is the number of samples. Bootstrap sampling can generate B sample subsets D1, D2, ..., D B .

[0164] 4.2 Construction of Decision Tree:

[0165] For each subset of Bootstrap samples, a decision tree is constructed. When constructing the decision tree, the splitting of each node is not performed on all features, but rather on a randomly selected subset of features. Let the original feature set be F = {f1, f2, ..., f...} M When splitting each node, m features are randomly selected for splitting. Typically... Or m = log2(M). The splitting criteria for decision trees can be Gini impurity or information gain.

[0166] 4.3 Integration Results:

[0167] Random forests make predictions by integrating the results of multiple decision trees. For classification tasks, the final classification result is determined by majority voting.

[0168] Majority vote for classification tasks:

[0169] Specifically:

[0170] 1. Input data: Experimental data from 7464 groups, which were normalized and used for model training and testing.

[0171] 2. Confusion Matrix: A confusion matrix generated from the predicted results and the actual results, used to calculate recall, precision, and F1 score.

[0172] 3. Model prediction results: The prediction results of the test set data are compared with the actual results, and the evaluation indicators are calculated.

[0173] 1) Evaluation accuracy: 98.6148%; 2) Average recall rate: 0.98766; 3) Average precision: 0.98054; 4) F1 score: 0.98409.

[0174] These results show that the model has very good classification performance on experimental data, with high accuracy, recall rate and precision, and very high F1 score, indicating that the comprehensive performance of the model is excellent.

[0175] Figure 7 The comparison chart of predicted grade and actual grade. Figure 7 In the figure, the matching situation of predicted value and actual value can be seen intuitively, and the place where the lines coincide indicates correct prediction, and the place where the lines do not coincide indicates error prediction. Figure 7 It can be seen that more than 98% of the data can be accurately predicted. Figure 8 The rows in the figure represent the actual categories, and the columns represent the predicted categories. The values on the diagonal (272, 870, 1065) represent the number of correctly classified samples. The values on the non-diagonal lines represent the number of incorrectly classified samples. The percentages on the right and bottom represent the classification accuracy and classification error of each category. Among them, the accuracy of category 1 is 96.5%, and the error is 3.5%. The accuracy of category 2 is 98.1%, and the error is 1.9%. The accuracy of category 3 is 99.6%, and the error is 0.4%.

[0176] Step 5: Give the adjustment trend of the artificial respirator parameter based on SSA optimization:

[0177] In order to obtain effective and high-quality breathing supply, according to the breathing quality evaluation grade results, i.e. excellent, good, and poor states, when the breathing quality evaluation grade result is poor, the parameters of the artificial respirator are optimized in reverse using SSA, and the adjustment trend of the respirator setting parameters is given, providing data information for medical personnel to adjust the parameters of the artificial respirator.

[0178] SSA (Sparrow Search Algorithm) is a new type of swarm intelligence optimization algorithm. Based on the biological group characteristics of the foraging behavior and anti-predation behavior of sparrow population, the algorithm simulates the behavior pattern of sparrow population in nature to perform optimization search. SSA has the characteristics of high solution accuracy and high efficiency.

[0179] The nonlinear mapping relationship between the ten independent variables and the three dependent variables of blood oxygen saturation, Cdyn, and RSB is established by the method of MSAGDBO-LSTM prediction model in step 3; through step 4, the blood oxygen saturation, Cdyn, RSB and other parameter indexes are used, and the respiratory quality is evaluated based on RF random forest to be in the excellent, good and poor grades. According to the evaluation result of respiratory quality, the input features of the parameters of the artificial respirator, i.e. airway peak pressure pip, mean airway pressure pmean, positive end-expiratory pressure Peep, respiratory rate RR, and exhalation time Te, are optimized by using sparrow search algorithm SSA, and the trend of the above parameter adjustment is given, so that the blood oxygen saturation, Cdyn, and RSB corresponding to the final respiratory quality reach the ideal range.

[0180] In the SSA optimization process, the number of optimization variables is 10, the value range of each variable is determined according to the medical index, the population number of sparrow search algorithm is set to 60, the maximum iteration number is 100, the optimization objective function is the error between the blood oxygen saturation, Cdyn, and RSB of the output result and the mean value of the blood oxygen saturation, Cdyn, and RSB of the corresponding optimal respiratory quality, and the target function is established:

[0181] obj = min ( | RSB output -RSB opt | + | Cdyn output -Cdyn opt | + | Saturation output -Saturation opt |)

[0182] Wherein, obj is the target function, RSB output is the predicted value of RSB output by the MSAGDBO-LSTM prediction model, RSB opt represents the RSB value corresponding to the optimal respiratory quality, Cdyn output is the predicted value of Cdyn output by the MSAGDBO-LSTM prediction model, Cdyn opt represents the Cdyn value corresponding to the optimal respiratory quality, Saturation output is the predicted value of blood oxygen saturation output by the MSAGDBO-LSTM prediction model, Saturation optThe blood oxygen saturation value corresponding to the optimal respiratory quality is represented. When the difference between the corresponding neural network output value in the optimization process and the optimal value of the three indicators corresponding to the optimal respiratory quality is the smallest, the optimization result of the corresponding input variables, i.e. the characteristics 1 airway peak pressure pip, the characteristics 2 average airway pressure pmean, the characteristics 3 positive end-expiratory pressure Peep, the characteristics 6 respiratory rate RR, the characteristics 8 expiratory time Te, etc. is obtained (the characteristics 4, the characteristics 5, the characteristics 7, the characteristics 9, the characteristics 10 are monitoring values, and there is no need to optimize here).

[0183] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for respiratory quality assessment based on multi-parameter, big data and data mining, characterized in that, The method comprises the following steps: obtaining artificial respirator parameter data to be predicted; inputting the artificial respirator parameter data to be predicted into an MSAGDBO-LSTM model to output a parameter index affecting respiratory quality, wherein the MSAGDBO-LSTM model is obtained based on a training set and a test set, the training set and the test set comprising historical parameter data and corresponding parameter indexes, the MSAGDBO-LSTM model is obtained by optimizing an LSTM network through an improved scarab beetle optimization algorithm, and the parameter index comprises blood oxygen saturation, dynamic lung compliance and rapid shallow breathing index; inputting the parameter index into a random forest model to obtain a respiratory quality evaluation grade result, wherein the random forest model is obtained by training historical parameter indexes and corresponding respiratory quality labels.

2. The multi-parameter, big data and data mining based respiratory quality assessment method according to claim 1, characterized in that, The artificial respirator parameter data to be predicted comprises airway peak pressure, mean airway pressure, positive end-expiratory pressure, tidal volume, minute ventilation, respiratory rate, minute air leakage, exhalation time, spontaneous breathing rate and oxygen concentration fraction in inhaled air.

3. The multi-parameter, big data and data mining based respiratory quality assessment method, as claimed in claim 1, wherein, The method for obtaining the training set and the test set comprises the following steps: obtaining original historical parameter data and corresponding parameter indexes, and performing format conversion processing and preprocessing to obtain the training set and the test set.

4. The multi-parameter, big data and data mining based respiratory quality assessment method, as claimed in claim 1, wherein, The method for obtaining the MSAGDBO-LSTM model by optimizing an LSTM network through an improved scarab beetle optimization algorithm comprises the following steps: S1, defining a target function and setting parameters; S2, initializing a scarab beetle population by using a Bernoulli chaotic mapping, and calculating the fitness value of each scarab beetle, wherein each individual of each scarab beetle represents an LSTM network hyperparameter; S3, when δ < ST, updating the position of a rolling scarab beetle according to a traditional scarab beetle optimization algorithm, otherwise, updating the position of the rolling scarab beetle according to an improved sine algorithm, δ is a parameter, and ST ∈ (0.5, 1]; S4, performing boundary judgment on each scarab beetle to update the positions of a breeding ball, a small scarab beetle and a stealing scarab beetle; S5, calculating the fitness value of each scarab beetle after updating, obtaining a current optimal solution, disturbing the current optimal solution by using an adaptive Gaussian-Cauchy hybrid mutation disturbance strategy, and obtaining a new optimal solution; S6, comparing the fitness of the current optimal solution and the new optimal solution according to a greedy mechanism to determine whether to update the position; S7, if the number of iterations has not reached a preset maximum number of iterations, returning to S3 after the number of iterations is increased by one, and if the number of iterations reaches the preset maximum number of iterations, outputting a result of the best network hyperparameter.

5. The multi-parameter, big data and data mining based respiratory quality assessment method, as claimed in claim 4, wherein, The method for initializing a scarab beetle population by using a Bernoulli chaotic mapping comprises the following steps: where β is a mapping parameter, z n is the current position, z n+1 is the updated position.

6. The multi-parameter, big data and data mining based respiratory quality assessment method, as claimed in claim 4, wherein, The method for disturbing the current optimal solution by using an adaptive Gaussian-Cauchy hybrid mutation disturbance strategy comprises the following steps: H b (t) = X b (t) x (1 + μ1 x Gauss(σ) + μ2 cauchy(σ)) where t is the current iteration number, X b (t) is the optimal position of individual X in the tth iteration, H b (t) is the optimal position X b (t) after the Gaussian-Cauchy hybrid disturbance, Gauss(σ) is the Gaussian variation operator, cauchy(σ) is the Cauchy variation operator, T max is the maximum iteration number.

7. The multi-parameter, big data and data mining based respiratory quality assessment method, as claimed in claim 5, wherein, The method for comparing the fitness of the current optimal solution and the new optimal solution according to a greedy mechanism to determine whether to update the position comprises the following steps: wherein f represents the fitness value.

Images