Domestic sleep breathing machine ventilation pressure prediction method based on deep learning

Abstract

The invention discloses a deep learning-based ventilation pressure prediction method for a household sleep breathing machine. The method comprises the following steps of 1, selecting obstructive sleep apnea and hypopnea symptom influence factors; 2, determining the influence weight of each influence factor; 3, generating an original data set; 4, data standardization; 5, building a long-short term memory neural network for training; 6, verifying the model on the verification set to output an optimal model for predicting the optimal ventilation pressure; 7, real-time sleep data, including snore, drinking amount, head angle and body angle, of the patient are collected, and the data serve as input to be used for predicting the optimal ventilation pressure model; and 8, outputting a predicted optimal ventilation pressure value by using the trained model. The method has the beneficial effects that the embodiment of the invention can realize accurate dynamic pressure prediction for most people by adopting a mode of performing pressure prediction by combining the multi-sample training set with the long and short-term time network.

Description

Technical Field

[0001] The present invention relates to the technical field of home sleep apnea machines, and in particular to a method for predicting ventilation pressure of home sleep apnea machines based on deep learning. Background Art

[0002] People with obstructive sleep apnea-hypopnea syndrome will repeatedly experience complete or partial obstruction of the upper airway during sleep. Due to the airway obstruction, patients cannot breathe smoothly, which leads to frequent sleep interruptions, daytime sleepiness, and may induce cerebrovascular diseases. Home sleep apnea machines can effectively prevent upper airway obstruction during sleep by continuously delivering a certain positive pressure of gas into the airway, thereby keeping the airway open, significantly improving the patient's sleep quality, relieving the symptoms of daytime sleepiness, and reducing the risk of cerebrovascular diseases. However, it is worth noting that the setting of ventilation pressure needs to be precisely controlled: if it is too low, the ideal treatment effect cannot be achieved; if it is too high, it may reduce the patient's comfort and affect their compliance with treatment, which may worsen the patient's quality of life and may induce cerebrovascular diseases.

[0003] At present, home sleep apnea-hypopnea syndrome treatment sleep apnea-hypopnea syndrome uses a sleep apnea machine to adjust ventilation pressure according to the patient's airway patency and the patient's responsiveness to previous pressure changes. When the patient's airway is detected to be open, the ventilator does not increase the pressure; once the airway is blocked, the pressure is increased by a preset amount (such as 1 cm H2O). If the patient does not respond to the previous pressure increase, the ventilator will automatically reduce the pressure by a predetermined amount (such as 1 cm H2O) until the optimal ventilation pressure level is found and maintained. However, the existing technology has the problem of being unable to dynamically predict the optimal ventilation pressure.

[0004] In view of the above situation, it is necessary to improve the existing method for predicting the ventilation pressure of a home sleep apnea machine so that it can meet the current needs for predicting the optimal ventilation pressure of a home sleep apnea machine. Summary of the invention

[0006] To achieve the above-mentioned purpose, the technical solution of the present invention is a method for predicting ventilation pressure of a home sleep apnea machine based on deep learning, comprising the following steps:

[0007] Step 1: Select the factors affecting obstructive sleep apnea and hypopnea symptoms;

[0008] Step 2: Determine the weight of each influencing factor;

[0009] Step 3: Generate the original data set;

[0010] Step 4: Data standardization;

[0011] Step 5: Build a long short-term memory neural network for training;

[0012] Step 6: Validate the model on the validation set to output the optimal model for predicting the optimal ventilation pressure;

[0013] Step 7: Collect the patient's real-time sleep data, including snoring, alcohol consumption, head angle and body angle. These data will be used as input to predict the optimal ventilation pressure model;

[0014] Step 8: Use the trained model to output the predicted optimal ventilation pressure value.

[0015] As a further supplement to the technical solution, in step 1, snoring, alcohol concentration, head angle and body angle are selected as monitoring variables.

[0016] As a further supplement to the present technical solution, the step 2 study deeply explored the effects of snoring, alcohol concentration, head angle and body angle on the upper respiratory tract. Through detailed analysis, it was found that alcohol concentration and snoring volume were positively correlated with ventilation pressure; in the standard supine position, the head angle and body angle were set to 0°, and the changes in head angle and body angle were negatively correlated with ventilation pressure. The specific impact weights were set as follows: when the alcohol concentration reached or exceeded 100 mg / 100 ml, the weight was 0.35; when the head angle was 0°, the weight was 0.05; when the body angle was 0°, the weight was 0.3; when the snoring volume was 80 dB, the weight was 0.3. At the same time, considering the potential effects of individual differences, mood, and environmental factors, impact deviations were introduced into each impact weight, specifically: alcohol concentration impact deviation ±0.1, head angle impact deviation ±0.02, body angle impact deviation ±0.1, and snoring volume deviation ±0.1.

[0017] As a further supplement to the present technical solution, in step three, in view of the limited data in the public data set, the data randomly generated by the influencing factors are combined for joint training, and finally a complete data set containing 445 groups of data is formed; among them, 400 groups of data are used as training sets, and the remaining 45 groups of data are used as validation sets.

[0018] As a further supplement to the technical solution, in step 4, in view of the fact that the range and distribution differences between different feature data may affect the performance of the prediction model, a standardized and normalized data processing method is adopted to ensure that each feature is compared and calculated at the same scale. In particular, for the feature of ventilation pressure, which has a larger value than other parameters, a min-max standardization method is adopted to constrain the data of it and all influencing factors to the range of [0, 1]. This method maps the original data to the specified interval through linear transformation, and its formula is:

[0019]

[0020] Where: x ＊ is the standardized value; x max and x min are the maximum and minimum values ​​in the sample data respectively.

[0021] To further supplement the technical solution, in step 5, in view of the need to process the continuous snoring, alcohol consumption, head angle and body angle data of the patient's sleep stage, an LSTM neural network that can efficiently process sequence data is selected. The specific architecture includes a single LSTM layer, trained using the Adam optimizer, and using the ReLU activation function and dropout mechanism (set to 0.3) to prevent the model from overfitting. The ReLU activation function is:

[0022] f(x)=max(0,x)

[0023] When the input value is greater than zero, the output of ReLU is non-zero, and when z is less than or equal to zero, the output is zero; this means that only inputs greater than zero will have non-zero outputs;

[0024] After many debugging and testing, it was determined that the LSTM layer contains 32 hidden units, and the fully connected layer is used to extract 64 eigenvalues. The root mean square error is used as the loss function to evaluate the performance of the model. The formula is:

[0025]

[0026] Where N is the total number of forecast model moments; X pred.t is the predicted value at time t output by the prediction model; X act.t is the actual data at time t.

[0027] As a further supplement to the technical solution, in step five, the maximum number of iterations in the training process is set to 100 times, the learning rate is set to 0.001, and the neuron retention rates of the input layer and the hidden layer are 0.8 and 0.5 respectively.

[0028] The beneficial effect lies in that the example of the present invention can achieve accurate dynamic pressure prediction for most people by adopting a multi-sample training set combined with a long short-term time network (LSTM) for pressure prediction. Compared with the existing technology, the latter needs to set ventilator-related parameters including the inhalation pressure range and rely on the guidance of professional technicians. The example of the present invention can automatically predict and adjust to the optimal ventilation pressure based on real-time data such as snoring, alcohol consumption, head angle and body angle, without manually setting parameters such as the inhalation pressure range, so it is more convenient to use; the example of the present invention collects the patient's sleep-related data in real time, including snoring, alcohol consumption records, head and body angles. When the patient's sleeping posture changes, the upper airway critical closing pressure will change accordingly. With the help of the ventilation pressure prediction model, the system can adjust the ventilation pressure to the optimal value in the current sleep state in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a specific implementation step diagram of the present invention;

[0030] Figure 2 It is a long short-term memory neural network model diagram of the present invention. DETAILED DESCRIPTION

[0031] In order to make the technical solution more clear to those skilled in the art, Figure 1-2 The technical solution of the present invention is described in detail:

[0032] Step 1: Select the factors affecting obstructive sleep apnea and hypopnea symptoms:

[0033] The symptoms of obstructive sleep apnea and hypopnea are characterized by patients repeatedly experiencing complete or partial upper airway obstruction during sleep; when the upper airway is obstructed, the airflow generates vortices through the narrow part, which in turn causes vibrations in the soft palate and other parts, producing abnormal sounds, namely snoring. In addition, the occurrence of upper airway obstruction is closely related to the critical closing pressure of the upper airway; therefore, in order to understand the situation of upper airway obstruction, it is necessary to pay attention to the size of snoring and the size of the critical closing pressure of the upper airway. Studies have shown that the critical closing pressure of the upper airway is affected by multiple factors such as drinking, head angle and body angle. Based on this, snoring, alcohol concentration, head angle and body angle were selected as monitoring variables.

[0034] Step 2: Determine the weight of each influencing factor:

[0035] Existing studies have explored in depth the effects of snoring, alcohol concentration, head angle, and body angle on the upper respiratory tract. Through detailed analysis, it is known that alcohol concentration and snoring volume are positively correlated with ventilation pressure; in the standard supine position (head angle and body angle are set to 0°), changes in head angle and body angle are negatively correlated with ventilation pressure. The specific impact weights are set as follows: when the alcohol concentration reaches or exceeds 100mg / 100ml, the weight is 0.35; when the head angle is 0°, the weight is 0.05; when the body angle is 0°, the weight is 0.3; when the snoring volume is 80dB, the weight is 0.3. At the same time, considering the potential impact of individual differences, mood, environment and other factors, impact deviations are introduced in each impact weight, specifically: alcohol concentration impact deviation ±0.1, head angle impact deviation ±0.02, body angle impact deviation ±0.1, snoring volume deviation ±0.1.

[0036] Step 3: Generate the original data set:

[0037] Given the limited data in the public dataset, we combined the data randomly generated by the influencing factors for joint training, and finally formed a complete dataset containing 445 sets of data, of which 400 sets were used as training sets and the remaining 45 sets were used as validation sets.

[0038] Step 4: Data standardization:

[0039] Given that the range and distribution differences between different feature data may affect the performance of the prediction model, data processing methods such as standardization and normalization were adopted to ensure that each feature is compared and calculated at the same scale. In particular, for the feature of ventilation pressure, which has a larger value than other parameters, the min-max standardization method was used to constrain the data of it and all influencing factors to the range of [0, 1]. This method maps the original data to the specified interval through linear transformation, and its formula is:

[0040]

[0041] Where: x ＊ is the standardized value; x max and x min are the maximum and minimum values ​​in the sample data respectively.

[0042] Step 5: Build a long short-term memory neural network (LSTM) for training:

[0043] Considering the need to process the patient's continuous snoring, alcohol consumption, head angle and body angle data during sleep, we chose an LSTM neural network that can efficiently process sequence data. Figure 2The example includes a single LSTM layer, trained using the Adam optimizer, and using the ReLU activation function and dropout mechanism (set to 0.3) to prevent the model from overfitting. The ReLU activation function is:

[0044] f(x)=max(0,x)

[0045] ReLU is a nonlinear activation function that can introduce nonlinear factors into neural networks, enabling them to learn and represent more complex functions. When the input value is greater than zero, the output of ReLU is a non-zero output, and when z is less than or equal to zero, the output is zero. This means that only inputs greater than zero will have non-zero outputs, so it has a certain sparsity when activating neurons, which helps to reduce the common activation of neurons. The ReLU activation function is very simple and computationally efficient because it only involves a simple threshold comparison and maximum value operation. Alleviate the gradient vanishing problem: Relative to some traditional activation functions (such as Sigmoid or Tanh), ReLU helps alleviate the gradient vanishing problem because its gradient is 1 in the positive interval, which helps to better propagate the gradient.

[0046] After multiple debugging and testing, it was determined that the LSTM layer contains 32 hidden units, and the fully connected layer (Dense layer) is used to extract 64 eigenvalues. The root mean square error is used as the loss function to evaluate the model performance, and its formula is:

[0047]

[0048] Where N is the total number of forecast model moments; X pred.t is the predicted value at time t output by the prediction model; X act.t is the actual data at time t; in addition, the maximum number of iterations during training is set to 100, the learning rate is set to 0.001, and the neuron retention rates of the input layer and hidden layer are 0.8 and 0.5, respectively.

[0049] Step 6: Validate the model on the validation set to output the optimal model for predicting the optimal ventilation pressure.

[0050] Step 7: Collect the patient's real-time sleep data, including snoring, alcohol consumption, head angle, and body angle. These data will be used as input to predict the optimal ventilation pressure model.

[0051] Step 8: Use the trained model to output the predicted optimal ventilation pressure value.

[0052] The above technical solutions only reflect the preferred technical solutions of the technical solutions of the present invention. Some changes that may be made to certain parts thereof by technicians in this technical field all reflect the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A method for predicting ventilation pressure of a home sleep apnea machine based on deep learning, characterized in that: The following steps are involved: Step 1: Select the factors affecting obstructive sleep apnea and hypopnea symptoms; Step 2: Determine the weight of each influencing factor; Step 3: Generate the original data set; Step 4: Data standardization; Step 5: Build a long short-term memory neural network for training; Step 6: Validate the model on the validation set to output the optimal model for predicting the optimal ventilation pressure; Step 7: Collect the patient's real-time sleep data, including snoring, alcohol consumption, head angle and body angle. These data will be used as input to predict the optimal ventilation pressure model; Step 8: Use the trained model to output the predicted optimal ventilation pressure value.

2. A method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 1, characterized in that: In the step 1, snoring, alcohol concentration, head angle and body angle are selected as monitoring variables.

3. A method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 2, characterized in that: The step 2 study deeply explored the effects of snoring, alcohol concentration, head angle and body angle on the upper respiratory tract. Through detailed analysis, it was found that alcohol concentration and snoring volume were positively correlated with ventilation pressure; in the standard supine position, the head angle and body angle were set to 0°, and the changes in head angle and body angle were negatively correlated with ventilation pressure. The specific impact weights were set as follows: when the alcohol concentration reached or exceeded 100mg / 100ml, the weight was 0.35; when the head angle was 0°, the weight was 0.05; when the body angle was 0°, the weight was 0.3; when the snoring volume was 80dB, the weight was 0.

3. At the same time, considering the potential effects of individual differences, mood, and environmental factors, impact deviations were introduced into each impact weight, specifically: alcohol concentration impact deviation ±0.1, head angle impact deviation ±0.02, body angle impact deviation ±0.1, snoring volume deviation ±0.

1.

4. The method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 1, characterized in that: In step three, in view of the limited data in the public data set, the data randomly generated by the influencing factors were combined for joint training, and finally a complete data set containing 445 groups of data was formed; among them, 400 groups of data were used as training sets, and the remaining 45 groups of data were used as validation sets.

5. The method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 1, characterized in that: In step 4, considering that the range and distribution differences between different feature data may affect the performance of the prediction model, a standardized and normalized data processing method is adopted to ensure that each feature is compared and calculated at the same scale. In particular, for the feature of ventilation pressure, which has a larger value than other parameters, a min-max standardization method is adopted to constrain the data of it and all influencing factors to the range of [0, 1]. This method maps the original data to the specified interval through linear transformation. The formula is: Where: x ＊ is the standardized value; x max and x min are the maximum and minimum values ​​in the sample data respectively.

6. A method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 5, characterized in that: In step 5, in view of the need to process the continuous snoring, alcohol consumption, head angle and body angle data of the patient's sleep stage, an LSTM neural network that can efficiently process sequence data is selected. The specific architecture includes a single LSTM layer, trained using the Adam optimizer, and using the ReLU activation function and dropout mechanism (set to 0.3) to prevent the model from overfitting. The ReLU activation function is: f(x)=max(0,x) When the input value is greater than zero, the output of ReLU is non-zero, and when z is less than or equal to zero, the output is zero; this means that only inputs greater than zero will have non-zero outputs; After many debugging and testing, it was determined that the LSTM layer contains 32 hidden units, and the fully connected layer is used to extract 64 eigenvalues. The root mean square error is used as the loss function to evaluate the performance of the model. The formula is: Where N is the total number of forecast model moments; X pred.t is the predicted value at time t output by the prediction model; X act.t is the actual data at time t.

7. A method for predicting ventilation pressure of a home sleep apnea machine based on deep learning according to claim 6, characterized in that: In step 5, the maximum number of iterations in the training process is set to 100, the learning rate is set to 0.001, and the neuron retention rates of the input layer and the hidden layer are 0.8 and 0.5 respectively.

Images