Oxygen generator on-demand pulse oxygen supply control method based on respiratory rhythm perception

Abstract

The present application relates to the field of auxiliary medical technology, and particularly relates to a method for pulse oxygen supply control of an oxygen generator based on respiratory rhythm sensing, which comprises the following steps: obtaining a plurality of respiratory pressure sequences, and dividing the respiratory pressure sequences into expiration phases and end-expiratory transition segments; determining signal confidence according to the performance of the expiration phases and the end-expiratory transition segments in the respiratory pressure sequences; introducing a regularization term of physiological rationality constraint into a regression loss function, and imposing a timing smoothing constraint on a model cell state update process to obtain an optimized respiratory rhythm prediction model; taking a three-dimensional feature vector as input data of the respiratory rhythm prediction model to obtain a length proportion of an end-expiratory stable segment in a next respiratory cycle; determining a pre-judgment window period from the start of inspiration according to the length proportion of the end-expiratory stable segment in the next respiratory cycle, and taking the pre-judgment window period as a pulse trigger countdown parameter of the oxygen generator to control pulse oxygen supply of the oxygen generator, so as to realize precise synchronization of oxygen supply action and inspiration action.

Description

Technical Field

[0001] This invention relates to the field of assistive medical technology, specifically to a method for controlling on-demand pulsed oxygen supply in an oxygen concentrator based on respiratory rhythm sensing. Background Technology

[0002] Human respiration has a stable rhythm, characterized by periodic alternation between inhalation and exhalation. Based on this physiological characteristic, pulsed oxygen delivery technology is widely used in portable oxygen concentrators. Its core idea is to release high-concentration oxygen only during the user's inhalation phase and stop oxygen delivery during exhalation or non-breathing periods, thereby achieving on-demand oxygen delivery, saving energy, extending device battery life, and improving wearing comfort.

[0003] Currently, mainstream pulse oxygen supply control methods typically employ a respiratory event triggering mechanism: pressure sensors, flow sensors, or accelerometers monitor the user's respiratory airflow or chest and abdominal movement signals in real time. When the system detects an inhalation action, it immediately activates a solenoid valve to release an oxygen pulse. This method achieves on-demand oxygen supply to a certain extent. However, this type of passive response control strategy has the following inherent drawbacks:

[0004] Because of time delays in respiratory signal detection, algorithm processing, and actuator response, the actual output of the oxygen pulse occurs later than the inhalation start point. Since the human body has a limited effective inhalation window, this delayed oxygen supply prevents oxygen from entering the alveoli during the initial stages of inspiration, significantly reducing oxygenation efficiency.

[0005] Existing methods treat respiration as a discrete event and fail to fully utilize the inherent periodicity and predictability of the respiratory process. Even if the user's breathing is stable, the system still needs to wait for each inhalation before deciding to supply oxygen, which cannot achieve proactive prediction and advance oxygen supply, thus limiting the physiological compatibility of human-machine collaboration.

[0006] When a user’s breathing rate or depth changes, the system needs to relearn the new rhythm. During the transition period, oxygen may be supplied too early or too late, affecting the continuity of oxygen therapy and the user experience.

[0007] In summary, the existing pulse control method based on detecting inhalation and then supplying oxygen is essentially a delayed response mechanism, which makes it difficult to achieve precise spatiotemporal synchronization between the oxygen supply action and the human body's inhalation process, thus restricting further improvements in the efficiency, comfort, and intelligence of pulse oxygen supply. Summary of the Invention

[0008] To address the technical problem of existing pulse oxygen supply control methods, which suffer from response delays and difficulty in achieving precise synchronization between oxygen supply actions and the human inspiratory phase, the present invention aims to provide an on-demand pulse oxygen supply control method for oxygen concentrators based on respiratory rhythm sensing. The specific technical solution adopted is as follows:

[0009] One embodiment of the present invention provides a method for controlling on-demand pulse oxygen supply in an oxygen concentrator based on respiratory rhythm sensing, the method comprising the following steps:

[0010] Obtain a respiratory pressure sequence of several historical respiratory cycles containing the current respiratory cycle, and divide the respiratory pressure sequence into the expiratory phase and the end-expiratory transition phase;

[0011] For each historical respiratory cycle, the signal confidence level of each historical respiratory cycle is determined based on the stability and regularity of the expiratory phase and end-expiratory transition phase in the respiratory pressure sequence.

[0012] For the prediction model, a regularization term for physiological rationality constraint is introduced into the regression loss function, and a temporal smoothing constraint is applied to the cell state update process of the model to obtain the optimized respiratory rhythm prediction model.

[0013] The three-dimensional feature vector of each historical respiratory cycle is used as the input data of the optimized respiratory rhythm prediction model to obtain the proportion of end-expiratory steady segment duration in the next respiratory cycle of the current respiratory cycle. The three-dimensional feature vector includes the proportion of end-expiratory steady segment duration, the total duration of the expiratory phase, and the signal confidence.

[0014] The predicted window period from the start of inspiration is determined based on the proportion of the end-expiratory steady period in the next respiratory cycle, and the predicted window period is used as the pulse trigger countdown parameter of the oxygen concentrator to control the pulse oxygen supply of the oxygen concentrator.

[0015] Further, obtaining the respiratory pressure sequence of several historical respiratory cycles including the current respiratory cycle includes:

[0016] Obtain the user's respiratory time-series waveform during the current preset time period;

[0017] The respiratory time-series waveform is converted into a respiratory pressure time series, and the first-order difference of the respiratory pressure time series is performed to determine the inflection point where the slope of the pressure change changes from positive to negative.

[0018] The moment corresponding to the inflection point is taken as the peak inspiratory position, and the time interval between adjacent inspiratory peaks is used to divide the respiratory cycles into independent ones, so as to obtain a respiratory pressure sequence of several historical respiratory cycles including the current respiratory cycle.

[0019] Furthermore, for each historical respiratory cycle, the signal confidence level for each historical respiratory cycle is determined based on the stability and regularity exhibited by the expiratory phase and end-expiratory transition segment in the respiratory pressure sequence, including:

[0020] For each historical respiratory cycle, the smoothness of expiratory phase pressure decay is calculated based on the performance data of the expiratory phase in the respiratory pressure sequence; the end-expiratory transition pressure variability is calculated based on the performance data of the end-expiratory transition in the respiratory pressure sequence.

[0021] The signal confidence level for each historical respiratory cycle is determined based on the synergistic effect of the changes in the smoothness of the expiratory phase pressure decay and the variability of the end-expiratory transition pressure in each historical respiratory cycle.

[0022] Further, determining the signal confidence level for each historical respiratory cycle based on the synergistic effect of the changes in the smoothness of the expiratory phase pressure decay and the variability of the end-expiratory transition pressure in each historical respiratory cycle includes:

[0023] Any historical respiratory cycle is determined as the target respiratory cycle, and the normalized relative changes of the expiratory phase pressure decay smoothness and the end-expiratory transition pressure fluctuation of the target respiratory cycle are calculated.

[0024] The amplitude matching degree is determined based on the two normalized relative changes, wherein the amplitude matching degree characterizes the relative balance between the two feature changes.

[0025] The degree of change deviation is determined based on the two normalized relative changes, and the degree of change deviation characterizes the consistency of the two characteristic change trends;

[0026] Based on the amplitude matching degree and the variation deviation degree, a synergy coefficient is calculated and used as the signal confidence of the target respiratory cycle.

[0027] Furthermore, the optimized respiratory rhythm prediction model includes:

[0028] An improved loss function is obtained by introducing a regularization term with physiological rationality constraints into the regression loss function. The regularization term is used to suppress abnormal fluctuations in the predicted output.

[0029] A temporal smoothing constraint is applied to the cell state update process of the model to obtain a constrained cell state update function. The temporal smoothing constraint is used to limit the instantaneous change amplitude of the cell state so that the dynamic response of the model matches the gradual change characteristics of the respiratory signal.

[0030] Based on the improved loss function and the constrained cell state update function, an optimized respiratory rhythm prediction model is constructed.

[0031] Furthermore, the expression for the improved loss function is:

[0032] In the formula, This represents the total loss after regularization for a single-target physiological trend. This represents the percentage of the measured end-tidal transition period in the (x+1)th cycle. This represents the percentage of the predicted end-tidal transition period in the (x+1)th cycle. This represents the percentage of the measured end-tidal transition period in the x-th cycle. This represents the preset physiological constraint weight coefficients, and max represents the function for finding the maximum value. This represents the preset upper limit constant for the rate of change in the proportion of people at the end of the expiration period.

[0033] Furthermore, the expression for the constrained cell state update function is:

[0034] In the formula, This represents the feature-compliant cell state value at time t. This represents the output of the forget gate at time t. This represents the cell state value at time t-1. This represents the input gate output value at time t. This represents the candidate cell state value at time t. This represents the preset upper limit threshold for cell state amplitude. This represents the clipping function. This indicates element-wise multiplication.

[0035] Further, obtaining the percentage of the end-tidal steady-state duration includes:

[0036] For any given historical respiratory cycle, the duration of the end-expiratory steady segment is obtained, and the first ratio of the end-expiratory steady segment duration to the total duration of the complete expiratory phase of the corresponding historical respiratory cycle is calculated as the percentage of the end-expiratory steady segment duration.

[0037] Furthermore, determining the predicted window period from the onset of inspiration based on the proportion of the end-expiratory steady period in the next respiratory cycle includes:

[0038] Based on the historical average proportion, the end-expiratory steady segment duration of the next respiratory cycle is corrected using the signal confidence of the current respiratory cycle to obtain the corrected predicted duration proportion; the historical average proportion is the average of the end-expiratory steady segment duration proportions of the N historical respiratory cycles with the highest signal confidence, where N is a positive integer.

[0039] Based on the temporal composition of the expiratory phase in the respiratory cycle, the predicted window period from the start of inspiration is calculated by using the proportional relationship between the expiratory attenuation duration of the current respiratory cycle and the proportion of the corrected predicted duration.

[0040] Furthermore, the calculation of the prediction window period from the onset of inhalation includes:

[0041] Use the expiratory attenuation duration of the current respiratory cycle as the baseline time length;

[0042] The modified prediction duration ratio is used as the numerator of the second ratio, and the difference between 1 and the modified prediction duration ratio is used as the denominator of the second ratio.

[0043] The product of the reference time length and the second ratio is used as the prediction window period for the start of inhalation. The prediction window period is used to characterize the prediction waiting time from the end of the current expiratory decay to the start of the next inhalation.

[0044] The present invention has the following beneficial effects:

[0045] To overcome the problems of asynchronous oxygen supply, low oxygenation efficiency, and poor user experience caused by slow response, lack of foresight, and insufficient dynamic adaptability in existing pulse oxygenation technologies, this invention achieves intelligent oxygenation control driven by respiratory rhythm through the following synergistic technical features: First, it acquires respiratory pressure sequences from several historical respiratory cycles, including the current respiratory cycle, and precisely divides them into the expiratory phase and the end-expiratory transition segment, thus providing a structured temporal basis for subsequent refined feature extraction; based on this, for each historical respiratory cycle, according to the stability of its expiratory phase pressure decay and the regularity of its end-expiratory transition segment fluctuations, it quantifies and generates a single-cycle signal confidence level, effectively identifying and suppressing abnormal cycles affected by motion, noise, and other interferences, ensuring the data reliability of rhythm modeling; furthermore, it constructs an optimal... The optimized respiratory rhythm prediction model incorporates a regularization term with physiological rationality constraints into the regression loss function to prevent the output from exceeding the human respiratory limit. Simultaneously, a temporal smoothing constraint is applied to the model's cell state update process to match the gradual physiological changes in respiratory signals, thus ensuring both numerical accuracy and physiological reliability in the prediction results. Subsequently, the three-dimensional feature vector of each historical respiratory cycle is used as input to this optimized model to accurately predict the proportion of the end-expiratory steady segment in the next respiratory cycle. Signal confidence is also used to dynamically correct the prediction results, enhancing the system's robustness under interference. Finally, based on this corrected prediction proportion, the prediction window period from the start of the next inhalation is determined and directly used as the countdown parameter for the oxygen concentrator's pulse trigger, achieving precise oxygen release at the initial stage of inspiration. This invention not only eliminates the inherent delay of traditional methods but also achieves, to a certain extent, spatiotemporal synchronization between oxygen delivery and the physiological inhalation process. It can adaptively adjust the timing of oxygen delivery when respiratory rate or depth changes abruptly, significantly improving the efficiency, continuity, comfort, and intelligence of oxygen therapy. Attached Figure Description

[0046] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating the steps of an on-demand pulse oxygen supply control method for an oxygen concentrator based on respiratory rhythm sensing, as proposed in one embodiment of the present invention;

[0048] Figure 2 This is a flowchart illustrating the steps for obtaining the optimized respiratory rhythm prediction model in an embodiment of the present invention.

[0049] Figure 3 This is a flowchart illustrating the steps of determining the prediction window period from the start of inspiration based on the proportion of the end-expiratory steady period in the next respiratory cycle, as described in an embodiment of the present invention. Detailed Implementation

[0050] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0052] The application scenarios targeted by this invention can be:

[0053] When using an oxygen concentrator for oxygen therapy, the respiratory signal characteristics collected by sensors are extracted and quantified to predict the time point from the end of respiration to the beginning of human inspiration, and oxygen supply is triggered accordingly, achieving on-demand pulsed oxygen supply control that is dynamically regulated according to the natural breathing rhythm. However, existing pulsed oxygen supply control methods have difficulty achieving precise synchronization between oxygen supply action and the human inspiratory phase due to response delays.

[0054] To address the need for precise synchronization between oxygen delivery and natural breathing in oxygen concentrators, and to eliminate time delays during oxygen delivery, this invention provides a method for controlling on-demand pulsed oxygen delivery in an oxygen concentrator based on respiratory rhythm sensing. This method aims to achieve on-demand pulsed oxygen delivery that aligns with the body's natural breathing rhythm and improves the adaptability of the oxygen supply. Figure 1 As shown, it includes the following steps:

[0055] S1, obtain the respiratory pressure sequence of several historical respiratory cycles including the current respiratory cycle, and divide the respiratory pressure sequence into the expiratory phase and the end-expiratory transition phase.

[0056] In rhythm-based pulsed oxygen supply control, accurate identification of the respiratory phase is a prerequisite for precise timing of oxygen delivery. However, the signal of a single respiratory cycle is susceptible to transient interference (such as body movement, coughing, and sensor noise), leading to incorrect phase judgment. To improve robustness, this embodiment employs a multi-cycle joint analysis strategy: acquiring a continuous respiratory pressure sequence including the current respiratory cycle and several historical respiratory cycles.

[0057] As an exemplary implementation, obtaining a respiratory pressure sequence including several historical respiratory cycles of the current respiratory cycle includes:

[0058] The first step is to obtain the user's respiratory time-series waveform for the current preset time period.

[0059] Here, the respiratory timing waveform covers multiple consecutive respiratory cycles, including the complete inspiratory phase, expiratory phase, and end-expiratory transition, providing a high-quality data foundation for subsequent cycle segmentation, feature extraction, and rhythm prediction.

[0060] In this embodiment, the respiratory timing waveform is obtained in real time by a sleep apnea monitor. Specifically, within the current preset time period (such as a continuous 6-hour window during the user's nighttime sleep), the raw respiratory signals during the user's complete respiratory cycle are continuously collected, i.e., the respiratory timing waveform; the sleep apnea monitor can use at least one of the following: a nasal airflow thermal sensor, a chest and abdominal impedance sensor, or a pressure breathing belt.

[0061] To ensure the accuracy of subsequent rhythm analysis, the following signal acquisition parameters were recorded and configured simultaneously:

[0062] Time resolution: not less than 10ms, i.e., sampling frequency To accurately capture the transient state at the beginning of inhalation and the minute fluctuations at the end of exhalation;

[0063] Data acquisition accuracy: Pressure measurement error not exceeding Or the airflow rate error does not exceed This ensures the reliability of respiratory phase recognition and feature extraction.

[0064] The second step is to convert the respiratory time-series waveform into a respiratory pressure time series, perform first-order difference on the respiratory pressure time series, and determine the inflection point where the slope of the pressure change changes from positive to negative.

[0065] In this embodiment, the respiratory time-series waveform is converted into a respiratory pressure time series. The series consists of pressure sample values ​​arranged in chronological order, with each sample value corresponding to a timestamp. Equivalently, it can be represented as a discrete time series with timestamps as the horizontal axis and respiratory pressure values ​​as the vertical axis, where the timestamp interval is determined by the time resolution of the acquisition system.

[0066] Performing a first-order difference operation on the respiratory pressure-time series yields a pressure change slope series. Each term in the pressure change slope series... , This represents the slope of the pressure change at the k-th sampling point. This represents the respiratory pressure value at the (k+1)th sampling point. This represents the respiratory pressure value at the k-th sampling point. Indicates the sampling interval.

[0067] Subsequently, the pressure change slope sequence is traversed to identify discrete points where the slope sign changes from positive to non-positive (negative). For example, if the pressure change slope of the kth sampling point is greater than zero, it indicates that the pressure is rising and the gas is in the intake phase; if the pressure change slope of the (k+1)th sampling point is less than or equal to zero, it indicates that the pressure has stopped rising or has started to fall. At this time, the kth sampling point can be regarded as the inflection point.

[0068] The third step is to take the moment corresponding to the inflection point as the peak position of inspiratory function, and divide the respiratory cycles into independent ones by the time interval between adjacent inspiratory peaks, so as to obtain a respiratory pressure sequence of several historical respiratory cycles including the current respiratory cycle.

[0069] In this embodiment, the moment corresponding to the inflection point where the identified pressure slope changes from positive to negative is marked as the inspiratory peak position; the time interval between two adjacent inspiratory peaks is used as the range of a single respiratory cycle, thereby dividing the continuous respiratory pressure sequence into a series of independent respiratory cycles.

[0070] After segmenting the current respiratory cycle, select the M most recent consecutive effective respiratory cycles, including the current respiratory cycle, to form a set of historical respiratory cycles for subsequent analysis.

[0071] It should be noted that for an effective respiratory cycle, firstly, the cycle duration is within a preset physiologically reasonable range, such as 2 to 10 seconds; secondly, the cycle includes a complete inspiratory and expiratory phase without signal loss or saturation cutoff; and finally, the peak inspiratory pressure amplitude exceeds the noise baseline threshold, such as 0.3.

[0072] Furthermore, to achieve refined modeling of respiratory dynamics, the pressure sequence of each respiratory cycle needs to be decomposed into sub-phases with clear physiological significance. Specifically: the expiratory phase characterizes the active expulsion of gas from the lungs, and its pressure decay characteristics (such as decay rate and duration) are closely related to airway resistance and lung compliance; the end-expiratory transition phase refers to the brief window between the end of expiration and the start of the next inspiration, reflecting the phase switching readiness state of the respiratory center, and its stability is directly related to the predictability of the respiratory rhythm. Because the mechanisms of change in the expiratory phase and the end-expiratory transition phase are different, their features must be extracted separately to accurately capture the changing patterns of the respiratory rhythm.

[0073] The division between the expiratory phase and the end-expiratory transition phase is based on the exponential decay characteristic of expiratory airflow and the smoothness of the end-expiratory pause. This embodiment utilizes the physical geometric characteristics of the respiratory waveform to clearly define the two states: "exhaling" and "end of exhalation awaiting inhalation." During the expiratory phase, at the beginning of exhalation, the intrapulmonary pressure is greater than atmospheric pressure, airflow is rapid, and the pressure curve shows a steep downward trend; at this time, the absolute value of the pressure change rate is very large. During the end-expiratory transition phase, the emptying of gas from the lungs is nearing completion, alveolar pressure tends to balance with atmospheric pressure, and airflow stops or becomes extremely weak.

[0074] As an exemplary implementation, the respiratory pressure sequence is divided into the expiratory phase and the end-expiratory transition period, including:

[0075] First, a pressure stabilization threshold is set. This threshold can be the product of the respiratory pressure sensor's acquisition accuracy and a preset safety factor. The preset safety factor ranges from 1.5 to 3.0, used to tolerate sensor background noise and ensure criterion robustness. It's worth noting that introducing a dynamic threshold related to sensor accuracy effectively avoids premature or late segmentation caused by a fixed threshold, improving the adaptability and accuracy of stage division. Of course, implementers can also set the pressure stabilization threshold according to specific circumstances; no specific limitations are imposed here.

[0076] Secondly, starting from the inspiratory peak position of the first historical respiratory cycle, the pressure first-order difference slope sequence is traversed along the time axis to identify the first moment where the absolute value of the slope is less than or equal to the pressure change plateau threshold, denoted as . The inhalation peak position corresponds to the time and time. The time interval is used as the expiratory phase to characterize the active expulsion of gas from the lungs; from time... The time interval from the start of the respiratory cycle until the peak inspiratory pressure of the next historical respiratory cycle is defined as the end-expiratory transition period, used to characterize the preparation phase for respiratory phase switching. The first-order differential slope sequence of pressure is the result of performing a first-order differential operation on the respiratory pressure sequence of each historical respiratory cycle.

[0077] It should be noted that explicitly dividing the two key stages mentioned above facilitates the subsequent extraction of highly discriminative features such as expiratory decay duration and end-expiratory stability, providing structured input for subsequent rhythm prediction, confidence assessment, and prediction window calculation. This method of dividing the expiratory phase and end-expiratory transition segment not only conforms to basic physiological principles but also provides a temporal basis for distinguishing between real respiratory signals and random interference. That is, physiological signals exhibit phased and coordinated changes across multiple cycles, while interference typically only affects local sampling points sporadically.

[0078] S2, for each historical respiratory cycle, determines the signal confidence level of each historical respiratory cycle based on the stability and regularity of the expiratory phase and end-expiratory transition phase in the respiratory pressure sequence.

[0079] Here, signal confidence is used at least to represent the usability of the signal. Subsequent processing, such as expiratory initiation prediction, can be based on signal confidence, emphasizing the main role of high-confidence data and ensuring the physiological rationality of signal processing.

[0080] To achieve on-demand pulsed oxygen supply from an oxygen concentrator based on the body's natural respiratory rhythm, high-precision phase recognition and rhythm prediction are required for the acquired respiratory pressure waveform. However, the actual acquired respiratory signals are often affected by various interference factors, including but not limited to: motion artifacts, poor sensor contact, environmental airflow disturbances, or equipment background noise. These interferences often manifest as local pressure abrupt changes, non-physiological oscillations, or signal distortion. If directly used for rhythm modeling, they will lead to misjudgment of the inspiratory initiation point, deviation in the prediction window calculation, and consequently, errors in oxygen supply timing. Based on this, this embodiment introduces a single-cycle signal confidence quantification mechanism before rhythm prediction: for each historical respiratory cycle, the stability and regularity of the expiratory phase and end-expiratory transition segment in the respiratory pressure sequence are analyzed, and the signal confidence of that cycle is calculated accordingly.

[0081] As an exemplary implementation, determining the signal confidence level for each historical respiratory cycle includes:

[0082] The first step is to calculate the smoothness of expiratory pressure decay for each historical respiratory cycle based on the performance data of the expiratory phase in the respiratory pressure sequence; and to calculate the end-expiratory transition pressure variability based on the performance data of the end-expiratory transition in the respiratory pressure sequence.

[0083] Interfering noises exhibit unevenness or large fluctuations in expiratory pressure decay and end-expiratory transition pressure; conversely, physiologically generated fluctuations tend to be coordinated and stable. Since the expiratory phase is easily disturbed by noises, quantifying the uniformity of expiratory pressure decay and the fluctuations in end-expiratory transition pressure helps distinguish between localized abruptness caused by noise and overall effective changes resulting from physiological or pathological state transitions.

[0084] For each historical respiratory cycle, based on the performance data of the expiratory phase in the respiratory pressure sequence, the smoothness of expiratory phase pressure decay is calculated, including:

[0085] In this embodiment, a subsequence of respiratory pressure corresponding to the expiration of a single historical respiratory cycle is obtained. A first-order difference operation is performed on this subsequence to obtain a pressure decay slope sequence. The absolute value of each slope in the pressure decay slope sequence is taken, and the average of all absolute values ​​is calculated. A negative correlation processing is applied to the average of all absolute values, and the resulting negative correlation value is used as the smoothness of the expiratory phase pressure decay. The negative correlation processing can be achieved by taking the reciprocal of the slope.

[0086] As an example, the formula for calculating the smoothness of expiratory phase pressure decay can be:

[0087] In the formula, This indicates the smoothness of expiratory pressure decay, where I represents the number of slopes in the pressure decay slope sequence. This represents the absolute value of the i-th slope.

[0088] It should be noted that, in order to avoid the calculation failure due to the denominator being zero, a very small positive number, such as 0.001, can be introduced into the denominator when calculating the smoothness of expiratory phase pressure decay.

[0089] Based on the performance data of the end-expiratory transition segment in the respiratory pressure sequence, the end-expiratory transition segment pressure variability is calculated, including:

[0090] In this embodiment, the respiratory pressure subsequence corresponding to the end-expiratory transition segment of a single historical respiratory cycle is obtained. The variance or standard deviation of the respiratory pressure subsequence corresponding to the end-expiratory transition segment is calculated. The variance or standard deviation of the respiratory pressure subsequence is used as the end-expiratory transition segment pressure fluctuation, denoted as . .

[0091] The second step is to determine the signal confidence level for each historical respiratory cycle based on the synergistic effect of the changes in the smoothness of the expiratory phase pressure decay and the pressure fluctuation of the end-expiratory transition phase.

[0092] Expiratory pressure decay and end-expiratory pressure fluctuation are two interconnected phases within the same respiratory cycle. When physiological or pathological changes occur, they act synchronously throughout the entire expiratory cycle, causing synergistic changes in both features. Sensor interference, on the other hand, consists of random, short-lived signals that only sporadically affect individual sampling points or single signal segments within the expiratory cycle, failing to impact the overall expiratory process and thus causing only localized fluctuations in a single feature. Therefore, quantifying the synergy between these two features is crucial for distinguishing between genuine respiratory signals and invalid interference signals; that is, quantifying the signal confidence level for each historical respiratory cycle.

[0093] Determine the signal confidence level for each historical respiratory cycle, including:

[0094] The first sub-step involves determining any historical respiratory cycle as the target respiratory cycle and calculating the normalized relative changes in the smoothness of expiratory phase pressure decay and the pressure fluctuation at the end of expiratory phase of the target respiratory cycle.

[0095] First, for ease of understanding and description, any one of the historical respiratory cycles is designated as the target respiratory cycle. Second, in order to effectively distinguish between the real physiological respiratory signal and the abnormal cycle affected by interference, this embodiment introduces the normalized relative change as a quantitative indicator to characterize the degree of deviation of the current respiratory cycle from the historical benchmark in key features.

[0096] As an example, the formula for calculating the normalized relative change in the smoothness of expiratory phase pressure decay can be:

[0097] In the formula, The normalized relative change representing the smoothness of expiratory phase pressure decay. Indicates the smoothness of expiratory phase pressure decay. This represents the historical decay smoothness reference value. This represents the function that takes the absolute value.

[0098] In the formula for calculating the normalized relative change, the historical decay smoothness reference value can be the average of the expiratory phase pressure decay smoothness over several historical respiratory cycles. The number of historical respiratory cycles can be set by the implementer according to the specific circumstances, such as a value of 5. Alternatively, the historical decay smoothness reference value can be the expiratory phase pressure decay smoothness of the previous cycle of the target respiratory cycle, with its initial value during the cold start phase being a preset empirical constant. It is worth noting that a custom historical decay smoothness reference value cannot be zero.

[0099] Similarly, the normalized relative change in pressure fluctuation at the end-expiratory transition can be obtained.

[0100] It should be noted that characteristic synergy analysis based on normalized relative changes can, to some extent, eliminate the influence of differences in breathing amplitude among individuals, focus on the relative trend of characteristic changes, and make the confidence assessment more universal and robust.

[0101] The second sub-step determines the magnitude matching degree based on the two normalized relative changes.

[0102] Here, the magnitude matching degree is used to characterize at least the relative balance between the two feature changes.

[0103] In this embodiment, the maximum and minimum values ​​of the two normalized relative changes are determined, and the ratio of the minimum and maximum values ​​is used as the amplitude matching degree.

[0104] The third sub-step is to determine the degree of deviation based on the two normalized relative changes.

[0105] Here, the degree of variation deviation is used at least to characterize the consistency of the changing trends of two features.

[0106] In this embodiment, the absolute value of the difference between the two normalized relative changes is calculated as the degree of change deviation.

[0107] The fourth sub-step involves calculating the synergy coefficient based on the amplitude matching degree and the degree of variation deviation, which serves as the signal confidence level for the target respiratory cycle.

[0108] As an example, the formula for calculating the signal confidence level of the target respiratory cycle can be:

[0109] In the formula, K represents the signal confidence level of the target respiratory cycle, max represents the maximization function, b represents the minimum of the two normalized relative changes, and B represents the maximum of the two normalized relative changes. This indicates the magnitude matching degree, where D represents the absolute value of the difference between two normalized relative changes, i.e., the degree of change deviation. This represents a non-zero constant, such as 0.01, used to avoid the case where the denominator of a fraction is zero.

[0110] In the formula for calculating signal confidence, The degree of matching is used to measure the magnitude of change of two features. The closer the degree of matching is to 1, the more balanced the magnitude of the relative changes of the two features are. If one feature changes significantly while the other feature remains basically unchanged, the degree of matching may be less than 1 and tend to 0. Normalization is achieved by dividing the absolute difference of the characteristic changes by the magnitude of the dominant change, in order to measure the degree of synchronization between the two characteristic changes; the smaller the deviation, the better. The closer it is to 1; the more coordinated the changes in actual breathing can improve amplitude matching. and All values ​​approach 1, while local fluctuations in individual features caused by random interference may lead to a decrease in amplitude matching, which in turn leads to a decrease in the overall coordination coefficient, resulting in a lower signal confidence of the target respiratory cycle, thus achieving effective differentiation; max is used to limit the range of signal confidence to avoid negative values ​​that could affect subsequent respiratory rhythm prediction.

[0111] It should be noted that the synergy coefficient quantifies the degree of synergy between the smoothness of expiratory phase decay and the variation of end-expiratory pressure fluctuation variance. This synergy is the core basis for distinguishing real respiratory signals from external interference. The level of synergy is directly equivalent to the reliability of the signal. Therefore, the synergy coefficient can be used as the confidence level of the signal.

[0112] S3. For the prediction model, a regularization term for physiological rationality constraint is introduced into the regression loss function, and a temporal smoothing constraint is applied to the cell state update process of the model to obtain the optimized respiratory rhythm prediction model.

[0113] The original Long Short-Term Memory (LSTM) network, with its memory units' ability to model historical temporal information, can effectively capture the changing trends of respiratory rhythms across multiple cycles, thereby predicting the respiratory temporal characteristics of the next cycle. However, standard LSTM models, during training, only aim to minimize the numerical error between predicted and measured values, without incorporating any physiological domain knowledge as constraints. This can lead to predictions that deviate from the physiologically reasonable range of human respiration. Furthermore, the cell states of LSTMs can fluctuate significantly during temporal updates, which is inconsistent with the gradual and smooth characteristics of respiratory signals themselves. This can easily cause instability in the model's fit to respiratory trends, affecting the continuity and smoothness of predictions.

[0114] To address the aforementioned issues, this embodiment introduces a regularization term with physiological rationality constraints into the regression loss function. By using the patterns of respiratory physiological changes as prior knowledge for model learning, the prediction results are not only numerically accurate but also conform to the physiological boundaries of human respiration. Simultaneously, a temporal smoothing constraint is applied to the model's cell state update process, limiting its update amplitude and rhythm to maintain consistency with the continuous and gradual changes in respiratory signals. This enhances the model's learning ability and fitting accuracy for target temporal features. Through this optimization, the resulting respiratory rhythm prediction model not only outputs prediction results that better reflect physiological reality but also effectively suppresses prediction fluctuations caused by sudden changes in cell state, significantly improving the stability and reliability of predictions.

[0115] As an exemplary implementation, an optimized respiratory rhythm prediction model is obtained, such as... Figure 2 As shown, it includes:

[0116] S31, an improved loss function is obtained by introducing a regularization term with physiological rationality constraints into the regression loss function.

[0117] Traditional LSTM models do not incorporate constraints based on human physiological patterns, potentially leading to outliers that exceed physiological limits. Therefore, a physiological trend regularization term is added to force the model to learn and conform to the physiological patterns of human respiration. Here, the regularization term is used to suppress abnormal fluctuations in the predicted output.

[0118] As an example, the expression for the improved loss function can be:

[0119] In the formula, This represents the total loss after regularization for a single-target physiological trend. This represents the percentage of the measured end-tidal transition period in the (x+1)th cycle. This represents the percentage of the predicted end-tidal transition period in the (x+1)th cycle. This represents the percentage of the measured end-tidal transition period in the x-th cycle; This represents the preset physiological constraint weight coefficient, which ranges from 0.1 to 1 and can be optimized using a grid search method; max represents the function for finding the maximum value. This represents the upper limit constant of the pre-set end-tidal percentage change rate, with a value ranging from 0.05 to 0.15, which can be determined by the 95% confidence interval of the statistical distribution of healthy samples.

[0120] In the improved loss function expression, the regularized total loss can characterize the training loss of LSTM in the oxygen concentrator inhalation initiation prediction scenario. It can provide a real breathing reference standard for model training. It is obtained by using LSTM prediction based on historical data. It can be obtained based on offline large-sample statistics, rather than in real time; The default regression loss function of the native LSTM is mean squared error, which can only represent the numerical deviation between the predicted value and the true value. This is a newly added physiological regularization term that only penalizes the model when the predicted change exceeds the physiological limit; otherwise, the penalty is 0.

[0121] In addition, for any historical respiratory cycle, the duration of the end-expiratory steady segment is obtained, and the first ratio of the end-expiratory steady segment duration to the total duration of the complete expiratory phase of the corresponding historical respiratory cycle is calculated as the proportion of the end-expiratory steady segment duration.

[0122] S32, apply temporal smoothing constraints to the cell state update process of the model to obtain the constrained cell state update function.

[0123] While LSTM has a large number of neurons and a relatively milder gradient vanishing problem, cell states can still be updated freely. However, the steady and slow changes in human respiration make it impossible to reliably capture these free fluctuations in cell states. Therefore, it is necessary to impose temporal smoothing constraints on the model's cell state update process. Here, temporal smoothing constraints are used to limit the instantaneous changes in cell states, so that the model's dynamic response matches the gradual and smooth characteristics of the respiratory signal.

[0124] As an example, the expression for the constrained cell state update function can be:

[0125] In the formula, This represents the feature-compliant cell state value at time t. This represents the output of the forget gate at time t. This represents the cell state value at time t-1. This represents the input gate output value at time t. This represents the candidate cell state value at time t. This represents the preset upper limit threshold for cell state amplitude. This represents the clipping function. This indicates element-wise multiplication.

[0126] In the expression of the cell state update function, Used to characterize the cell state values ​​calculated by the optimized LSTM. Used to characterize cell state values ​​from the previous time step, to store historical respiratory time-series features learned by the model. The output of the LSTM input gate is used to characterize the respiratory features of the current respiratory cycle that need to be included in the cell state analysis. The larger the output value of the input gate, the more important the corresponding respiratory feature is. It can represent the potential update amount that respiratory features can bring to the cell state at the current moment, and it is an inherent value of the native LSTM. It can be set based on the range of activation functions, and in general, an empirical value of 0.1 is taken.

[0127] It should be noted that, because human breathing is regular and continuous in the breathing prediction scenario, and the model needs to retain most of the historical breathing memory, a forgetting gate is used to prevent infinite accumulation. Set it to a constant close to 1, such as 0.98; for the update increment Some parts undergo physiological amplitude cropping. It limits the update increment of the current respiratory characteristics to a certain value. Within the range, if it exceeds the range, it is pruned to the boundary value; the above update function expression as a whole combines the retained historical respiratory memory and the current update increment after pruning compliance to obtain the final feature-compliant cell state.

[0128] S33. Based on the improved loss function and the constrained cell state update function, an optimized respiratory rhythm prediction model is constructed.

[0129] After improving the loss function and cell state update mechanism, this embodiment integrates them into the training and inference process of the LSTM network to construct an optimized respiratory rhythm prediction model, specifically including the following steps:

[0130] First, a standard LSTM unit is used as the basic architecture, retaining the computational logic of the input gate, forget gate, output gate, and candidate cell states. Second, the original cell state update formula is replaced with a constrained cell state update function. Then, an improved loss function is used for end-to-end training. Next, using a labeled respiratory pressure sequence dataset, the total loss is minimized through backpropagation to update the network weights. Finally, after training, the model parameters are fixed, and only the constrained cell state update function is used to generate the prediction output during the inference phase, ensuring that the real-time prediction results have both numerical accuracy and physiological rationality, ultimately resulting in an optimized respiratory rhythm prediction model. The training data for the model consists of the three-dimensional feature vectors shown below, and the model training process is based on existing technology and will not be described in detail here.

[0131] S4 uses the three-dimensional feature vector of each historical respiratory cycle as input data for the optimized respiratory rhythm prediction model to obtain the proportion of end-expiratory steady segment duration in the next respiratory cycle of the current respiratory cycle.

[0132] In this embodiment, each historical respiratory cycle is represented as a three-dimensional feature vector, which serves as the input data for the optimized respiratory rhythm prediction model. The three-dimensional feature vector includes the proportion of end-expiratory steady-state time, the total duration of the expiratory phase, and the signal confidence level. Specifically:

[0133] The percentage of end-expiratory steady phase is the ratio of the duration of the end-expiratory steady phase to the total duration of the expiratory phase in the current respiratory cycle. That is, the ratio of the time from the end of the expiratory phase pressure decay to the start of the next inspiration to the total duration of the expiratory phase. It can reflect the relative degree of end-expiratory stability, with a typical value range of 0.2–0.6.

[0134] The total duration of the expiratory phase is the time interval from the end of the current peak inspiratory phase (i.e., the start of expiration) to the start of the next peak inspiratory phase. This parameter directly characterizes the expiratory dynamics of the user at the current respiratory rate.

[0135] For signal confidence, a dimensionless index is calculated based on the synergy of stability and regularity between the expiratory phase and the end-expiratory transition phase within the current respiratory cycle. This index is used to quantify the degree to which the cycle is affected by interference. High confidence indicates that the cycle is a reliable physiological signal, while low confidence suggests possible interference from body movement, coughing, or sensor noise.

[0136] For M consecutive valid historical respiratory cycles, including the current respiratory cycle, the system arranges their three-dimensional feature vectors in reverse chronological order to form an input sequence. After normalization, the sequence is input into the optimized LSTM prediction model, and the output is the predicted value of the end-expiratory steady segment duration for the next respiratory cycle.

[0137] S5 determines the prediction window period from the start of inspiration based on the proportion of the end-expiratory steady period in the next respiratory cycle, and uses the prediction window period as the pulse trigger countdown parameter of the oxygen concentrator to control the pulse oxygen supply of the oxygen concentrator.

[0138] Because human respiration is rhythmic and predictable, the duration of the end-expiratory phase of the next respiratory cycle is closely related to the current and historical respiratory patterns. The proportion of end-expiratory steady-state duration is a key indicator reflecting the readiness for respiratory phase switching: the larger this ratio, the more stable the end-expiratory phase, and the more likely the respiratory center is to initiate the next inhalation in a shorter time; conversely, it suggests possible prolonged expiration or rhythm instability. Based on this, this embodiment dynamically calculates the predicted window period from the start of the next inhalation based on the predicted proportion of end-expiratory steady-state duration of the next respiratory cycle. Here, the predicted window period is used to characterize the predicted waiting time from the end of the current expiratory decay to the start of the next inhalation, i.e., the time required to trigger oxygen supply in advance.

[0139] As an exemplary implementation, the predicted window period from the onset of inspiration is determined based on the proportion of the end-expiratory steady period in the next respiratory cycle. Figure 3 As shown, it includes:

[0140] S51, based on the historical average proportion, uses the signal confidence of the current respiratory cycle to correct the duration of the end-expiratory steady segment of the next respiratory cycle, and obtains the corrected predicted duration proportion.

[0141] Although the proportion of end-expiratory steady-state duration in the next respiratory cycle can be predicted based on historical respiratory cycles, this prediction is essentially a statistical estimate of the population's rhythm trend and does not fully consider the instantaneous reliability of the current respiratory state. If the current respiratory cycle is affected by disturbances and its signal confidence is low, directly using this cycle for prediction or as a control basis will lead to inaccurate prediction windows, thereby causing deviations in oxygen delivery timing. Therefore, this embodiment introduces a dynamic correction mechanism based on the signal confidence of the current cycle: the initially predicted proportion of end-expiratory steady-state duration is weighted and adjusted using the signal confidence of the current respiratory cycle to obtain a corrected predicted proportion.

[0142] Specifically, if the signal confidence level of the current respiratory cycle is close to 1, it indicates that the current respiratory state is stable and reliable, the correction magnitude is small, and the original predicted value is retained. If the signal confidence level of the current respiratory cycle is significantly less than 1, the regression weight to the historical average proportion is increased, making the correction result more dependent on long-term stable rhythms and suppressing the influence of transient interference. Here, the historical average proportion is the average of the end-expiratory steady-state duration proportions of the N historical respiratory cycles with the highest signal confidence, representing the user's typical breathing pattern in a stable state, where N is a positive integer, such as 5.

[0143] As an example, the calculation process for correcting the percentage of predicted duration may include:

[0144] In the formula, This indicates the percentage of the initial revised prediction duration for the next respiratory cycle. This indicates the confidence level of the signal in the current respiratory cycle. This indicates the predicted end-expiratory steady-state duration of the next respiratory cycle. This indicates the historical average percentage.

[0145] While signal confidence-weighted fusion can effectively improve prediction stability, in extreme respiratory modes (such as severe obstructive ventilatory impairment or deep sedation), the model may still output an excessively high proportion of end-expiratory steady-state duration. Directly using such high proportions for subsequent prediction window calculations will lead to an excessively high denominator. Approaching zero, the predictive window period Y... Y A rapid increase, even approaching infinity, can lead to the following problems: oxygen supply triggering is too early, releasing oxygen in the middle or even early stages of exhalation, resulting in gas waste and user discomfort; the system response delay is too long, and the oxygen supply action is not completed even after actual inhalation has begun, thus reducing synchronization; there is a risk of numerical calculation overflow, affecting the stability of the embedded controller.

[0146] Therefore, this embodiment introduces a safety upper limit threshold to trim the initial corrected prediction duration proportion, resulting in the final corrected prediction duration proportion. In the formula, This represents the percentage of the corrected predicted duration for the next respiratory cycle, and 'min' represents the function for finding the minimum value. This indicates the percentage of the initial revised prediction duration for the next respiratory cycle; This indicates the preset safety upper limit threshold, which can be dynamically adjusted according to user type. For example, it can be set to 0.85 for adults and 0.80 for children.

[0147] It should be noted that by using the minimum value processing described above, not only is the mathematical stability of the prediction window period calculation formula ensured, but the prediction results are also constrained within a physiologically reasonable range, avoiding the failure of oxygen supply logic due to extreme output. Furthermore, by determining the proportion of the modified prediction duration, the robustness and physiological reasonableness of the prediction results can be significantly improved while maintaining the ability to respond to rhythm changes, providing a more reliable input for subsequent prediction window period calculations.

[0148] S52, based on the temporal composition of the expiratory phase in the respiratory cycle, calculates the prediction window period from the start of inspiration by utilizing the proportional relationship between the expiratory attenuation duration and the proportion of the corrected prediction duration in the current respiratory cycle.

[0149] In rhythm-based pulsed oxygen supply control, if oxygen supply is triggered only after the actual inspiratory action is detected, the inherent delay in oxygen delivery from the oxygen concentrator outlet to the user's airway will prevent oxygen from entering the alveoli during the initial stage of inspiration, significantly reducing oxygenation efficiency. Therefore, it is essential to predict the start of the next inspiratory phase and initiate oxygen supply before it arrives, introducing a predictive window as a countdown benchmark for pulse triggering. However, a fixed lead time cannot accommodate the differences in breathing patterns among different users. To address this, this embodiment proposes a dynamically adaptive predictive window calculation method: based on respiratory physiology, the expiratory attenuation duration directly reflects the expiratory dynamics of the current respiratory cycle; while the proportion of the predicted duration reflects the stability of the end-expiratory phase and the urgency of phase switching. Both together determine the effective waiting time from the current expiratory state to the start of the next inspiratory phase.

[0150] In this embodiment, the expiratory attenuation duration, which characterizes the time from the peak of inspiration to the stabilization of expiratory pressure in the current respiratory cycle, is used as the baseline time length; the proportion of the corrected predicted duration is used as the numerator of the second ratio, and the difference between 1 and the proportion of the corrected predicted duration is used as the denominator of the second ratio; the product of the baseline time length and the second ratio is used as the prediction window period from the start of inspiration.

[0151] As an example, the formula for calculating the prediction window period from the onset of inhalation can be:

[0152] In the formula, Y represents the predicted window period from the onset of inspiration. This indicates the duration of expiratory attenuation in the current respiratory cycle. This indicates the percentage of the revised predicted duration for the next respiratory cycle.

[0153] In the formula for calculating the predicted window period, The closer the value is to 1, the more stable the breathing process of the next respiratory cycle is for most of the time, with a prolonged end-expiratory phase, commonly seen in deep breathing. In this case, the respiratory center needs more time to complete the expiratory-inspiratory transition, and the actual initiation of inspiration will be significantly delayed. When the value approaches 1, the prediction window increases dramatically with the nonlinearity of the proportion, thus matching the long waiting time required under pathological or deep and slow breathing conditions. The predictive window is used to characterize the speed of a user's expiratory dynamics. Using this as a baseline time length allows the predictive window to automatically adapt to the breathing speed of different users. That is, the expiratory decay time is shorter for fast-breathing users and longer for slow-breathing users. Under normal breathing conditions, the window changes gradually, while under abnormal breathing conditions, the window changes rapidly, forming a safety redundancy and avoiding missed triggers due to abrupt changes in rhythm. This effectively balances efficiency and safety.

[0154] Finally, the predicted window period is used as the pulse trigger countdown parameter for the oxygen concentrator to control the pulse oxygen supply, including:

[0155] To ensure the safety and reliability of the system under complex respiratory conditions, this embodiment can set physical safety upper and lower thresholds for the prediction window period. For example, the physical safety lower threshold is 100ms, and the physical safety upper threshold is 1500ms, corresponding to the physiological lower limit of the effective inspiratory window and the upper limit of the device's response capability, respectively. When the calculated prediction window period exceeds this safe range, it indicates that the current rhythm prediction may be abnormal. Furthermore, if continuous... One respiratory cycle (e.g.) If the confidence level of the signals is lower than the preset confidence level (e.g., 0.4), it is determined that the user's breathing mode has changed significantly (e.g., from resting to talking, coughing, or changing body position), causing the historical rhythm model to fail.

[0156] In any of the above scenarios, the system will automatically execute a safety degradation strategy: clear the historical cache data of the current respiratory rhythm prediction model and switch to a traditional oxygen supply mode based on real-time sensor threshold detection—that is, when the inspiratory airflow is detected to exceed a preset trigger threshold, pulse oxygen supply will be immediately initiated. Although this mode has an inherent response delay, it is highly robust and can ensure basic oxygen therapy functions. The system continuously monitors the confidence level and rhythm stability of subsequent respiratory signals; when continuous One cycle (e.g.) When the stability condition is met, the historical cache is automatically rebuilt and the intelligent predictive oxygen supply mode is restored, achieving a smooth switchback from safety downgrade to advanced control.

[0157] In this embodiment, the predicted window period is directly used as the trigger countdown parameter for pulse oxygen supply: the starting point of the expiratory steady segment of the current respiratory cycle is taken as the timing starting point of the predicted window period. When the system detects that the remaining time between the starting point of the expiratory steady segment of the current respiratory cycle and the start of the next inhalation is equal to the predicted window period, the solenoid valve of the oxygen generator is immediately activated to release a high-concentration oxygen pulse.

[0158] It should be noted that since oxygen takes a certain amount of time to travel from release to arrival at the alveoli, this early triggering mechanism ensures that the oxygen flow enters the airway precisely at the beginning of inhalation, achieving precise synchronization between oxygen supply and the physiological inhalation process, significantly improving oxygenation efficiency and user comfort.

[0159] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for controlling on-demand pulse oxygen supply in an oxygen concentrator based on respiratory rhythm sensing, characterized in that, Includes the following steps: Obtain a respiratory pressure sequence of several historical respiratory cycles containing the current respiratory cycle, and divide the respiratory pressure sequence into the expiratory phase and the end-expiratory transition phase; For each historical respiratory cycle, the signal confidence level of each historical respiratory cycle is determined based on the stability and regularity of the expiratory phase and end-expiratory transition phase in the respiratory pressure sequence. For the prediction model, a regularization term for physiological rationality constraint is introduced into the regression loss function, and a temporal smoothing constraint is applied to the cell state update process of the model to obtain the optimized respiratory rhythm prediction model. The three-dimensional feature vector of each historical respiratory cycle is used as the input data of the optimized respiratory rhythm prediction model to obtain the proportion of end-expiratory steady segment duration in the next respiratory cycle of the current respiratory cycle. The three-dimensional feature vector includes the proportion of end-expiratory steady segment duration, the total duration of the expiratory phase, and the signal confidence. The predicted window period from the start of inspiration is determined based on the proportion of the end-expiratory steady period in the next respiratory cycle, and the predicted window period is used as the pulse trigger countdown parameter of the oxygen concentrator to control the pulse oxygen supply of the oxygen concentrator. For each historical respiratory cycle, the signal confidence level of each cycle is determined based on the stability and regularity of the expiratory phase and end-expiratory transition phase in the respiratory pressure sequence, including: For each historical respiratory cycle, the smoothness of expiratory phase pressure decay is calculated based on the performance data of the expiratory phase in the respiratory pressure sequence; the end-expiratory transition pressure variability is calculated based on the performance data of the end-expiratory transition in the respiratory pressure sequence. The signal confidence level for each historical respiratory cycle is determined based on the synergistic effect of the changes in the smoothness of the expiratory phase pressure decay and the pressure fluctuation of the end-expiratory transition phase in each historical respiratory cycle. The regularization term that introduces physiological rationality constraints into the regression loss function includes: obtaining an improved loss function by introducing a regularization term that introduces physiological rationality constraints into the regression loss function, wherein the regularization term is used to suppress the prediction output of abnormal data fluctuations; The expression for the improved loss function is as follows: In the formula, This represents the total loss after regularization for a single-target physiological trend. This represents the percentage of the measured end-tidal transition period in the (x+1)th cycle. This represents the percentage of the predicted end-tidal transition period in the (x+1)th cycle. This represents the percentage of the measured end-tidal transition period in the x-th cycle. This represents the preset physiological constraint weight coefficients, and max represents the function for finding the maximum value. This represents the preset upper limit constant for the rate of change in the proportion of people at the end of the expiration period; In the improved loss function expression, the regularized total loss characterizes the training loss of the LSTM in the oxygen concentrator inhalation initiation prediction scenario. Provide a realistic breathing reference standard for model training. It is obtained by using LSTM prediction based on historical data. Based on offline large-sample statistics, rather than real-time; The default regression loss function of the native LSTM is the mean squared error, which represents the numerical deviation between the predicted value and the true value. This is a newly added physiological regularization term that only penalizes the model when the predicted change exceeds the physiological limit; otherwise, the penalty is 0.

2. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 1, characterized in that, The step of obtaining a respiratory pressure sequence containing several historical respiratory cycles of the current respiratory cycle includes: Obtain the user's respiratory time-series waveform during the current preset time period; The respiratory time-series waveform is converted into a respiratory pressure time series, and the first-order difference of the respiratory pressure time series is performed to determine the inflection point where the slope of the pressure change changes from positive to negative. The moment corresponding to the inflection point is taken as the peak inspiratory position, and the time interval between adjacent inspiratory peaks is used to divide the respiratory cycles into independent ones, so as to obtain a respiratory pressure sequence of several historical respiratory cycles including the current respiratory cycle.

3. The method for on-demand pulse oxygen supply control of an oxygen concentrator based on respiratory rhythm sensing according to claim 1, characterized in that, The determination of signal confidence for each historical respiratory cycle based on the synergistic effect of changes in the smoothness of expiratory phase pressure decay and the variability of end-expiratory transition pressure in each historical respiratory cycle includes: Any historical respiratory cycle is determined as the target respiratory cycle, and the normalized relative changes of the expiratory phase pressure decay smoothness and the end-expiratory transition pressure fluctuation of the target respiratory cycle are calculated. The amplitude matching degree is determined based on the two normalized relative changes, wherein the amplitude matching degree characterizes the relative balance between the two feature changes. The degree of change deviation is determined based on the two normalized relative changes, and the degree of change deviation characterizes the consistency of the two characteristic change trends; Based on the amplitude matching degree and the variation deviation degree, a synergy coefficient is calculated and used as the signal confidence of the target respiratory cycle.

4. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 1, characterized in that, The optimized respiratory rhythm prediction model includes: A temporal smoothing constraint is applied to the cell state update process of the model to obtain a constrained cell state update function. The temporal smoothing constraint is used to limit the instantaneous change amplitude of the cell state so that the dynamic response of the model matches the gradual change characteristics of the respiratory signal. Based on the improved loss function and the constrained cell state update function, an optimized respiratory rhythm prediction model is constructed.

5. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 4, characterized in that, The expression for the constrained cell state update function is: In the formula, This represents the feature-compliant cell state value at time t. This represents the output of the forget gate at time t. This represents the cell state value at time t-1. This represents the input gate output value at time t. This represents the candidate cell state value at time t. This represents the preset upper limit threshold for cell state amplitude. This represents the clipping function. This indicates element-wise multiplication.

6. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 1, characterized in that, Obtaining the percentage of end-tidal steady-state duration includes: For any given historical respiratory cycle, the duration of the end-expiratory steady segment is obtained, and the first ratio of the end-expiratory steady segment duration to the total duration of the complete expiratory phase of the corresponding historical respiratory cycle is calculated as the percentage of the end-expiratory steady segment duration.

7. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 1, characterized in that, The determination of the predicted window period from the onset of inspiration based on the proportion of the end-expiratory steady period in the next respiratory cycle includes: Based on the historical average proportion, the end-expiratory steady segment duration of the next respiratory cycle is corrected using the signal confidence of the current respiratory cycle to obtain the corrected predicted duration proportion; the historical average proportion is the average of the end-expiratory steady segment duration proportions of the N historical respiratory cycles with the highest signal confidence, where N is a positive integer. Based on the temporal composition of the expiratory phase in the respiratory cycle, the predicted window period from the start of inspiration is calculated by using the proportional relationship between the expiratory attenuation duration of the current respiratory cycle and the proportion of the corrected predicted duration.

8. The method for controlling on-demand pulse oxygen supply to an oxygen concentrator based on respiratory rhythm sensing according to claim 7, characterized in that, The calculation of the predicted window period from the onset of inhalation includes: Use the expiratory attenuation duration of the current respiratory cycle as the baseline time length; The modified prediction duration ratio is used as the numerator of the second ratio, and the difference between 1 and the modified prediction duration ratio is used as the denominator of the second ratio. The product of the reference time length and the second ratio is used as the prediction window period for the start of inhalation. The prediction window period is used to characterize the prediction waiting time from the end of the current expiratory decay to the start of the next inhalation.

Images